THE WORLD BANK ECONOMIC REVIEWdocuments.worldbank.org/curated/en/765411468159910121/... ·...
Transcript of THE WORLD BANK ECONOMIC REVIEWdocuments.worldbank.org/curated/en/765411468159910121/... ·...
THE WORLD BANKECONOMIC REVIEW
Is There a Metropolitan Bias? The relationship between povertyand city size in a selection of developing countries
Céline Ferré, Francisco H.G. Ferreira, and Peter Lanjouw
Impact of SMS-Based Agricultural Information on Indian Farmers Marcel Fafchamps and Bart Minten
Crises, Food Prices, and the Income Elasticity ofMicronutrients:Estimates from Indonesia
Emmanuel Skoufias, Sailesh Tiwari, and Hassan Zaman
Economic Geography and Economic Development in Sub-Saharan Africa
Maarten Bosker and Harry Garretsen
The Decision to Import Capital Goods in India: Firms’ FinancialFactors Matter
Maria Bas and Antoine Berthou
Coffee Market Liberalisation and the Implications for Producers in Brazil, Guatemala and India
Bill Russell, Sushil Mohan, and Anindya Banerjee
Implications of COMTRADE Compilation Practices for Trade Barrier Analyses and Negotiations
Alexander J. Yeats
Volume 26 • 2012 • Number 3
www.wber.oxfordjournals.org
THE WORLD BANK1818 H Street, NWWashington, DC 20433, USAWorld Wide Web: http://www.worldbank.org/E-mail: [email protected]
2V
olume 26 • N
umber 3 • 2012
TH
E W
OR
LD
BA
NK
EC
ON
OM
IC R
EV
IEW
Pages 351–555
ISSN 0258-6770 (PRINT) ISSN 1564-698X (ONLINE)
Wber.j_26_3_Cover.qxd 9/25/12 11:53 AM Page 1
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
Pub
lic D
iscl
osur
e A
utho
rized
THE WORLD BANKECONOMIC REVIEW
editorsAlain de Janvry and Elisabeth Sadoulet, University of California at Berkeley
assistant to the editor Marja Kuiper
editorial boardHarold H. Alderman, World Bank (retired)Chong-En Bai, Tsinghua University, ChinaPranab K. Bardhan, University of California,
BerkeleyThorsten Beck, Tilburg University,
Netherlands Johannes van Biesebroeck, K.U. Leuven,
Belgium Maureen Cropper, University of Maryland,
USAAsli Demirgüç-Kunt, World Bank Jean-Jacques Dethier, World BankQuy-Toan Do, World BankFrédéric Docquier, Catholic University of
Louvain, BelgiumEliana La Ferrara, Università Bocconi, ItalyFrancisco H. G. Ferreira, World BankAugustin Kwasi Fosu, United Nations
University, WIDER, Finland
Caroline Freund, World BankPaul Glewwe, University of Minnesota,
USAPhilip E. Keefer, World BankNorman V. Loayza, World BankWilliam F. Maloney, World BankDavid J. McKenzie, World BankJaime de Melo, University of GenevaUgo Panizza, UNCTADNina Pavcnik, Dartmouth College, USAVijayendra Rao, World BankMartin Ravallion, World BankJaime Saavedra-Chanduvi, World BankClaudia Paz Sepúlveda, World BankJonathan Temple, University of Bristol,
UKDominique Van De Walle, World BankChristopher M. Woodruff, University of
California, San Diego
The World Bank Economic Review is a professional journal used for the dissemination of research indevelopment economics broadly relevant to the development profession and to the World Bank inpursuing its development mandate. It is directed to an international readership among economists andsocial scientists in government, business, international agencies, universities, and development researchinstitutions. The Review seeks to provide the most current and best research in the field of quantita-tive development policy analysis, emphasizing policy relevance and operational aspects of economics,rather than primarily theoretical and methodological issues. Consistency with World Bank policy playsno role in the selection of articles.
The Review is managed by one or two independent editors selected for their academic excellence inthe field of development economics and policy.The editors are assisted by an editorial board composedin equal parts of scholars internal and external to the World Bank. World Bank staff and outsideresearchers are equally invited to submit their research papers to the Review.
For more information, please visit the Web sites of the Economic Review at Oxford University Pressat www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals.
Instructions for authors wishing to submit articles are available online at www.wber.oxfordjournals.org.Please direct all editorial correspondence to the Editor at [email protected].
Forthcoming papers in
• The Impact of the Global Food Crisis on Self-Assessed Food SecurityDerek D. Headey
• How Is the Liberalization of Food Markets Progressing? MarketIntegration and Transaction Costs in Subsistence EconomiesWouter Zant
• Decomposing the Labor Market Earnings Inequality:The Public and Private Sectors in Vietnam, 1993–2006Clément Imbert
• Chinese Trade Reforms, Market Access and Foreign Competition:the Patterns of French ExportersMaria Bas and Pamela Bombarda
• Firms Operating under Electricity Constraints in Developing CountriesPhilippe Alby, Jean-Jacques Dethier & Stéphane Straub
• Antidumping, Retaliation Threats, and Export PricesVeysel Avsar
• Information and Participation in Social ProgramsDavid Coady, César Martinelli, and Susan W. Parker
THE WORLD BANKECONOMIC REVIEW
Wber.j_26_3_Cover.qxd 9/25/12 11:53 AM Page 2
THE WORLD BANK ECONOMIC REVIEW
Volume 26 † 2012 † Number 3
Is There a Metropolitan Bias? The relationship between povertyand city size in a selection of developing countries 351
Celine Ferre, Francisco H.G. Ferreira, and Peter Lanjouw
Impact of SMS-Based Agricultural Information on Indian Farmers 383Marcel Fafchamps and Bart Minten
Crises, Food Prices, and the Income Elasticity ofMicronutrients:Estimates from Indonesia 415
Emmanuel Skoufias, Sailesh Tiwari, and Hassan Zaman
Economic Geography and Economic Development in Sub-SaharanAfrica 443
Maarten Bosker and Harry Garretsen
The Decision to Import Capital Goods in India: Firms’ FinancialFactors Matter 486
Maria Bas and Antoine Berthou
Coffee Market Liberalisation and the Implications for Producers inBrazil, Guatemala and India 514
Bill Russell, Sushil Mohan, and Anindya Banerjee
Implications of COMTRADE Compilation Practices for TradeBarrier Analyses and Negotiations 539
Alexander J. Yeats
SUBSCRIPTIONS: A subscription to The World Bank Economic Review (ISSN 0258-6770) comprises 3 issues. Prices
include postage; for subscribers outside the Americas, issues are sent air freight.
Annual Subscription Rate (Volume 26, 3 Issues, 2012): Institutions—Print edition and site-wide online access: £176/
$264/E264, Print edition only: £162/$242/E242, Site-wide online access only: £147/$221/E221; Corporate—Print
edition and site-wide online access: £264/$395/E395, Print edition only: £242/$362/E362, Site-wide online access
only: £219/$330/E330; Personal—Print edition and individual online access: £43/$64/E64. US$ rate applies to US &
Canada, EurosE applies to Europe, UK£ applies to UK and Rest of World. There may be other subscription rates
available; for a complete listing, please visit www.wber.oxfordjournals.org/subscriptions. Readers with mailing
addresses in non-OECD countries and in socialist economies in transition are eligible to receive complimentary
subscriptions on request by writing to the UK address below.
Full prepayment in the correct currency is required for all orders. Orders are regarded as firm, and payments
are not refundable. Subscriptions are accepted and entered on a complete volume basis. Claims cannot be con-
sidered more than four months after publication or date of order, whichever is later. All subscriptions in Canada
are subject to GST. Subscriptions in the EU may be subject to European VAT. If registered, please supply details to
avoid unnecessary charges. For subscriptions that include online versions, a proportion of the subscription price
may be subject to UK VAT. Personal rates are applicable only when a subscription is for individual use and are
not available if delivery is made to a corporate address.
BACK ISSUES: The current year and two previous years’ issues are available from Oxford University Press. Previous
volumes can be obtained from the Periodicals Service Company, 11 Main Street, Germantown, NY 12526,
USA. E-mail: [email protected]. Tel: (518) 537-4700. Fax: (518) 537-5899.
CONTACT INFORMATION: Journals Customer Service Department, Oxford University Press, Great Clarendon Street,
OxfordOX2 6DP, UK. E-mail: [email protected]. Tel: þ44 (0)1865 353907. Fax: þ44 (0)1865 353485. In the
Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary,
NC 27513, USA. E-mail: [email protected]. Tel: (800) 852-7323 (toll-free in USA/Canada) or (919) 677-0977.
Fax: (919) 677-1714. In Japan, please contact: Journals Customer Service Department, Oxford University Press,
Tokyo, 4-5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. E-mail: [email protected]. Tel: þ81 3 5444 5858.
Fax: þ81 3 3454 2929.
POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published three times a year, in
February, June, and October, by Oxford University Press for the International Bank for Reconstruction and
Development/THE WORLD BANK. Send address changes to The World Bank Economic Review, Journals
Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Periodicals
postage paid at Cary, NC and at additional mailing offices. Communications regarding original articles and editorial
management should be addressed to The Editor, The World Bank Economic Review, The World Bank, 3, Chemin
Louis Dunant, CP66 1211 Geneva 20, Switzerland.
ENVIRONMENTAL AND ETHICAL POLICIES: Oxford Journals, a division of Oxford University Press, is committed to
working with the global community to bring the highest quality research to the widest possible audience.
Oxford Journals will protect the environment by implementing environmentally friendly policies and practices
wherever possible. Please see http://www.oxfordjournals.org/ethicalpolicies.html for further information on
environmental and ethical policies.
DIGITAL OBJECT IDENTIFIERS: For information on dois and to resolve them, please visit www.doi.org.
PERMISSIONS: For information on how to request permissions to reproduce articles or information from this
journal, please visit www.oxfordjournals.org/jnls/permissions.
ADVERTISING: Advertising, inserts, and artwork enquiries should be addressed to Advertising and Special Sales,
Oxford Journals, Oxford University Press, Great Clarendon Street, Oxford, OX2 6DP, UK. Tel: þ44 (0)1865
354767; Fax: þ44(0)1865 353774; E-mail: [email protected].
DISCLAIMER: Statements of fact and opinion in the articles in The World Bank Economic Review are those of the
respective authors and contributors and not of the International Bank for Reconstruction and Development/THE
WORLD BANK or Oxford University Press. Neither Oxford University Press nor the International Bank for
Reconstruction and Development/THE WORLD BANK make any representation, express or implied, in respect of the
accuracy of the material in this journal and cannot accept any legal responsibility or liability for any errors or
omissions that may be made. The reader should make her or his own evaluation as to the appropriateness or
otherwise of any experimental technique described.
PAPER USED: The World Bank Economic Review is printed on acid-free paper that meets the minimum require-
ments of ANSI Standard Z39.48-1984 (Permanence of Paper).
INDEXING AND ABSTRACTING: The World Bank Economic Review is indexed and/or abstracted by CAB Abstracts,
Current Contents/Social and Behavioral Sciences, Journal of Economic Literature/EconLit, PAIS International,
RePEc (Research in Economic Papers), and Social Services Citation Index.
COPYRIGHT # 2012 The International Bank for Reconstruction and Development/THE WORLD BANK
All rights reserved; no part of this publication may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means, electronic, mechanical, photocopying, recording, or otherwise without prior written
permission of the publisher or a license permitting restricted copying issued in the UK by the Copyright
Licensing Agency Ltd, 90 Tottenham Court Road, London W1P 9HE, or in the USA by the Copyright
Clearance Center, 222 Rosewood Drive, Danvers, MA 01923.
Typeset by Techset Composition Limited, Chennai, India; Printed by Edwards Brothers Incorporated, USA.
Is There a Metropolitan Bias? The relationshipbetween poverty and city size in a selection of
developing countries
Celine Ferre, Francisco H.G. Ferreira, and Peter Lanjouw1
This paper provides evidence from eight developing countries of an inverse relation-ship between poverty and city size. Poverty is both more widespread and deeper invery small and small towns than in large or very large cities. This basic pattern is gen-erally robust to the choice of poverty line. The paper shows, further, that for all eightcountries, a majority of the urban poor live in medium, small or very small towns.Moreover, it is shown that the greater incidence and severity of consumption povertyin smaller towns is generally compounded by similarly greater deprivation in terms ofaccess to basic infrastructure services, such as electricity, heating gas, sewerage andsolid waste disposal. We illustrate for one country – Morocco – that inequalitywithin large cities is not driven by a severe dichotomy between slum dwellers andothers. Robustness checks are performed to assess whether the findings in the paperhinge on a specific definition of “urban area”; are driven by differences in the cost ofliving across city-size categories; by reliance on an income-based concept of well-being; or by the application of small-area estimation techniques for estimating povertyrates at the town and city level. JEL Codes: I32, O18, R12
In the late 1970s and in the 1980s, there was much discussion of “urban bias”in development circles. Following Lipton (1977), development economistsincreasingly recognized a widespread tendency among (almost always urban-based) governments to pursue policies that – explicitly or implicitly – taxedagriculture and transferred resources to industry and other urban activities.
1. Celine Ferre is an independent consultant based in Amsterdam, the Netherlands; her email
address is [email protected]. Francisco Ferreira is a lead economist in the Development Research
Group at the World Bank and a research fellow at IZA; his email address is [email protected].
Peter Lanjouw (corresponding author) is the manager of the Poverty and Inequality Group at the
Development Research Group at the World Bank; his email address is [email protected]. We
are grateful to Johan Mistiaen for setting us off on this project. We are also much indebted to Victoria
Fazio, Philippe George Leite, Ericka Rascon, Martin Ravallion, and Timothy Thomas for advice and
numerous contributions. Marianne Fay and participants at a World Bank Workshop on Urban Poverty
in June 2007, the World Bank/InWent Development Policy Forum meeting in Berlin, 2007, and the
LCSPP Seminar in November 2010, provided useful comments. These are the views of the authors and
they should not be attributed to the World Bank.
THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 351–382 doi:10.1093/wber/lhs007Advance Access Publication February 14, 2012# The Author 2012. Published by Oxford University Press on behalf of the International Bankfor Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions,please e-mail: [email protected]
351
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
The motivation was not exclusively urban self-interest. There was a widespreadbelief, based on the influential early views of Rosenstein-Rodan (1943),Prebisch (1950), and others, that development was to a large extentsynonymous with industrialization – and that industrialization inevitablyimplied urbanization. As markets could not solely be relied upon to allocateresources to that most dynamic sector, government was required to provide a“big push” to help economies along the righteous path of urban growth.
Against that view, Lipton and his followers argued that urban bias implied a“sacrifice of efficient and equitable growth to rapid urban advance” (p.310).By distorting relative prices and the “intersectoral terms of trade”, such policiesinduced an inefficient allocation of resources that could lead to perpetuallyinfant industries, at the expense of farmers, many of whom were the poorestpeople in the land.2 That was a time when an estimated 80-90% of the world’spoor lived in rural areas, and an important part of the argument against urbanbias was that, in addition to distorting the allocation of capital and otherresources, these policies were also anti-poor.3
In 2012, the situation is somewhat different. Urbanization has proceededapace in the last quarter century such that the world’s urban population is nowas large as its rural population. Extreme poverty remains a predominantly ruralphenomenon, with some 75% of those who subsist on expenditures below$1-a-day still residing in rural areas in 2002, even when higher cost of living inurban areas is taken into account.4 But urban poverty has been falling moreslowly than rural poverty – in part because urbanization has been a key driverbehind rural poverty reduction, but some of those who migrate to urban areasremain poor. Urban poverty therefore accounts for a growing share of globalpoverty: Ravallion et al. (2007) estimate that the urban share of total extremepoverty rose from 19% in 1993 to 25% in 2002. In some regions, like LatinAmerica (76.2%); Eastern Europe and Central Asia (63.5%) and Middle-Eastand North Africa (55.8%), urban poverty is already dominant.
Poverty is expected to continue to urbanize, in the sense that the share ofthe total number of poor who live in urban areas is expected to continue togrow (with some exceptions, notably in Eastern Europe). Some expect that theurban share of $1-a-day ($2-a-day) poverty may reach 40% (51%) around2030. Urban poverty is also thought to be accompanied by a different set ofcharacteristics and challenges, including health and sanitation problems inurban slums, unemployment, and a greater incidence of violent crime. Inresponse, strategies to fight urban poverty – and its specific peculiarities – aregrowing in importance, both at the national and at the international level.
2. A classic study by Bates (1981) documented the use of price regulation and marketing boards in
Ghana, Nigeria and Zambia to extract surplus from farmers to the benefit of urban food consumers.
3. Ravallion et al. (2007) produced arguably the first global poverty statistics that cover the majority
of the world’s population and disaggregate between urban and rural areas. They estimate that the urban
share of the world’s extreme poor in 1993 was 19%.
4. See Ravallion et al. (2007).
352 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Yet urban poverty is far from a homogeneous phenomenon, even within asingle country. It is often remarked that poverty is spatially heterogeneous.Usually this is stated with reference to a marked rural-urban dichotomy inmeasured poverty. But there is also considerable spatial heterogeneity amongurban areas, and one important dimension of that heterogeneity is across citysizes. In Brazil, for instance, while most anecdotal discussion of urban povertyfocuses on the sprawling slums of Rio de Janeiro or Sao Paulo, over 50% ofthe country’s urban poor live in towns with fewer than 50,000 inhabitants.Only around 10% live in cities with populations greater than a million. InKazakhstan, the incidence of poverty in smaller towns is six times larger thanin Almaty. And there are large differences in access to local public goods andservices too: in Morocco, average access to sewerage is over 80% in citiesgreater than a million, but less than 50% in the smallest towns.
A greater understanding of how poverty – both in terms of incomes orconsumption expenditures and in terms of access to public services – variesacross different types of cities should help inform the discussion of appropriatepoverty reduction strategies in most countries. Yet, the evidence base neededfor this disaggregated analysis is seldom available, since household surveys –on which most poverty assessments are based – are seldom representative atthe level of any but the largest metropolitan areas in the developing world.They are certainly not representative for smaller towns and cities, and informa-tion is not usually disaggregated along these lines.5
In this paper, we draw upon insights generated by small area poverty estima-tion (based on the combination of welfare estimates from household surveyswith “sample” sizes from National Censuses) to investigate the relationshipbetween poverty and city size in eight developing countries, namely Albania,Brazil, Kazakhstan, Kenya, Mexico, Morocco, Thailand and Sri Lanka. Wefind substantial variation in the incidence and depth of consumption povertyacross city sizes in seven of the eight countries. For all seven countries wherethe data permits some kind of disaggregation of the incidence of public serviceaccess, there is also considerable variation across city sizes. In all cases, povertyis lowest and service availability is greatest in the largest cities – precisely thosewhere governments, the middle-classes, opinion-makers and airports aredisproportionately located.
At a minimum this “poverty gradient” across city sizes needs to be borne inmind whenever considering options and priorities for addressing urbanpoverty. More speculatively, this evidence might leads us to ask whether,
5. It is rare for household surveys to include identifiers for the specific town or city within which the
survey respondent resides – unless the city constitutes a specific stratum. In those settings where such
identifiers are included, one can estimate urban poverty rates for different city size intervals (see for
example, the poverty profile of Brazil by Ferreira, Lanjouw and Neri (2003) and of India by World
Bank (2011)). While city size intervals constructed from survey data are able to reveal differential
poverty outcomes across the urban spectrum, they do not convey the differences that may exist between
individual towns and cities (see further below).
Ferre, Ferreira and Lanjouw 353
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
alongside Lipton’s original urban bias, there exists also a “metropolitan bias”in the allocation of resources (including policy attention) to larger cities, at theexpense of smaller towns, where most of the urban poor are located.
There are a number of caveats which require that our results be treated withcare. First, although our samples within countries are representative, oursample of countries is not. Although these eight countries are located in all sixregions into which the World Bank routinely divides the developing world,they are not random draws.6 They are countries where there was an earlyinterest in (and the data required for) constructing a poverty map. Second, weuse national, rather than international, poverty lines. This has the advantagethat poverty is measured in the terms which each particular country’s residentsfeel is appropriate. But it has the disadvantage that poverty does not mean thesame living standard across the eight countries. Third, we do not apply auniform definition of “urban area” across our eight countries. Instead, we relyon the definition of urban settlements that each respective country reports in itscensus documents. Such administrative definitions are likely to vary acrosscountries, and may not correspond closely to economic definitions of townsand cities (linked to population density).7 This could have a bearing on ourresults. We probe the robustness of our conclusions, in the context of Brazil, byre-defining urban settlements so as take explicit account of population density.
A fourth caveat concerns our inability to systematically adjust forcost-of-living differences between cities. These differences might be expected tobe smaller than those between urban and rural areas, but they may still matter,and we report a second robustness test with respect to cost-of-living differencesin the only country in our sample for which data permits it, namely Brazil.
Fifth, we do not gauge our findings to variation in equivalence scales –which would matter if family sizes and composition varied systematicallyacross city sizes. Sixth, we assess poverty in a fairly restrictive way: focusing onthe share of the population with incomes or consumption levels below thepoverty line. It is widely acknowledged that poverty can be viewed morebroadly, reflecting multiple dimensions of wellbeing. We seek to mitigate thisconcern by reporting the association between city size and access to variouspublicly-provided services and, in one instance, by looking separately at ahealth outcome (child malnutrition). Nevertheless, it should be acknowledgedthat the patterns observed in these spaces need not be repeated when otherdimensions of poverty are considered.
6. The World Bank divides the developing world into sub-Saharan Africa (AFR), East Asia and the
Pacific (EAP), Eastern Europe and Central Asia (ECA), Latin America and the Caribbean (LAC), the
Middle-East and North Africa (MENA); and South Asia (SAR).
7. As noted by Lanjouw and Lanjouw (2001), definitions of rural and urban locality vary widely
across countries. For example, in Malawi and Zimbabwe settlements with populations of 3000 and
2500 inhabitants, respectively, are defined as urban while in Taiwan a settlement with less than
250,000 inhabitants is designated as rural.
354 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Each of the limitations of the analysis presented in this paper points to theneed for additional research. This paper simply documents the existence ofsystematic differences in the breadth and depth of poverty (and access to ser-vices) across city sizes in eight geographically diverse developing countries.
The paper is structured as follows. Section I provides an overview of thepoverty mapping methodology which was used in each of the six countries, inorder to generate reliable poverty estimates for every urban area captured in thepopulation census. Section II describes the data sources to which this methodwas applied, in each country. Section III presents the consumption poverty pro-files by city size in each country. Section IV turns to the evidence on access topublicly provided services across city sizes. Section V looks at differences inpoverty (and inequality) within specific cities in Morocco, focusing in particularon poverty and inequality differences between slums and non-slum areas inlarger towns. Section VI subjects our findings to some robustness checks. Wefirst look to data from Brazil, in order to gauge the sensitivity of our findings tochanges in the definition of urban areas and city sizes. We also use Braziliandata to probe for evidence of spatial cost of living variation across city-sizes. Asubsequent robustness check is applied to the case of Mexico, to examinewhether an alternative dimension of deprivation, namely child malnutrition,exhibits the same gradient across city size categories as income poverty. In a finalrobustness check we show that, in India, a poverty-city size gradient can beobserved both directly from survey data and from small area estimation techni-ques. We conclude that the inverse relationship is thus not driven exclusively byour reliance on a particular estimation method. Section VII offers tentative con-clusions and discusses some of the questions that this descriptive paper raises forfurther research into urban poverty.
I . M E T H O D O L O G Y
The economic analysis of the distribution of living standards in developing coun-tries relies almost entirely on household surveys. If their samples are selectedappropriately, these surveys can collect detailed information from a relativelysmall number of households (perhaps 0.1% of the country’s total population),and yet generate information that is representative of the population as a whole.The law of large numbers ensures that the uncertainty about the populationwhich results from sampling (the ‘sampling error’) becomes very small at samplesizes that are still cost effective. This enables researchers the world over to askdetailed questions from small groups of people, at a fraction of the cost thatwould be required if entire populations needed to be polled.
But there is one drawback: the samples that are designed to be representativeof large populations are not, in general, representative of specific non-randomsub-divisions of that population. Indeed, the typical (nationally representative)household survey is not representative of sub-national units, such as states, pro-vinces or districts. There are exceptions, mostly in large countries, such as
Ferre, Ferreira and Lanjouw 355
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
China, India or Brazil. But even in those countries, the problem is simplyshifted down one level: living standards will vary enormously across differentlocalities of (or towns and cities in) the states of Uttar Pradesh or MinasGerais; but the Indian National Sample Survey and the Brazilian PesquisaNacional por Amostra de Domicılios are not representative at those levels.
A number of small-area estimation techniques have been developed to seekto address this missing data problem. In this paper, we rely on applications ofthe “poverty mapping” approach developed by Elbers, Lanjouw and Lanjouw(2002, 2003). This approach typically involves a household survey and a popu-lation census as data sources. First, the survey data are used to estimate aprediction model for either consumption or incomes. The selection of explana-tory variables is restricted to those variables that can also be found in thecensus (or some other large dataset) or in a tertiary dataset that can be linkedto both the census and survey. The parameter estimates are then applied to thecensus data, expenditures are predicted, and poverty (and other welfare) statis-tics are derived. The key assumption is that the models estimated from thesurvey data apply to census observations.
Let W be a welfare indicator based on the distribution of a household levelvariable of interest, yh. Using a detailed household survey sample, we estimatethe joint distribution of yh and observed correlates xh. By restricting theexplanatory variables to those that also occur at the household level in thepopulation census, parameter estimates from this “first stage” model can beused to generate the distribution of yh for any target population in the censusconditional on its observed characteristics and, in turn, the conditional distri-bution of W. Elbers et al. (2002, 2003) study the precision of the resulting esti-mates of W and demonstrate that prediction errors will fall (or at least not rise)with the number of households in the target population, and will also beaffected by the properties of the first stage models, in particular the precisionof parameter estimates. A general rule of thumb is that welfare estimatesobtained on this basis will be estimated fairly precisely as long as the targetpopulation comprises at least 1,000-5,000 households.
The first-stage estimation is carried out using household survey data.8 Theempirical models of household consumption allow for an intra-cluster correlationin the disturbances (see Elbers, Lanjouw and Lanjouw, 2002, 2003, Elbers,Lanjouw and Leite, 2008, and Demombynes et al., 2007, for more details).Failing to take account of spatial correlation in the disturbances would result inunderestimated standard errors in the final poverty estimates (Tarozzi andDeaton, 2009). Different models are estimated for each region and the specifica-tions include census mean variables and other aggregate level variables in order tocapture latent cluster-level effects. All regressions are estimated with household
8. These surveys are stratified at the region or state level, as well as for rural and urban areas.
Within each region there are further levels of stratification, and also clustering. At the final level, a small
number of households (a cluster) are randomly selected from a census enumeration area.
356 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
weights and with parsimonious specifications to be cautious about overfitting.Heteroskedasticity is also modeled in the household-specific part of the residual.
Parameter estimates from all the first-stage models are then taken, in thesecond stage, to the population census. Since predicted household-level percapita consumption in the census is a function not only of the parameter esti-mates from the first stage consumption models estimated in the survey, but alsoof the precision of these estimates and of those parameters describing thedisturbance terms in the consumption models, we do not produce just one pre-dicted consumption level per household in the census. Rather, a reasonablylarge number of predicted expenditures are simulated for each household (typic-ally around 100 simulations). The full set of simulated household-level percapita expenditures are then used to calculate the welfare estimates for eachtarget population. Demombynes, Elbers, Lanjouw and Lanjouw (2007) describea variety of simulation approaches that are available and document that these allyield closely similar welfare estimates. Validation studies of the poverty mappingmethodology remain rare; in settings where one can rigorously check themethod, it is likely that it was not needed in the first place. However, a few suchstudies have been conducted and have yielded encouraging findings (see forexample Demombynes, et al., 2007, and Elbers et al., 2008). We examine inSection VI whether there are grounds for suspecting that our broad findings con-cerning the relationship between urban poverty and city size are due to ouremployment of small-area estimates of poverty as opposed to direct measures.
I I . D A T A
Poverty-mapping exercises based on the methodology just described have nowbeen conducted in a number of countries. We have selected eight of these coun-tries, on the basis of the availability of the micro-data files and of regionalcoverage, for analysis in this paper. Table 1 lists the total urban population (atthe Census year) in the eight countries, both in absolute numbers and as ashare of the total. The sample includes a wide variety of countries, from therelatively small (e.g. Albania) to the relatively large (e.g. Brazil), and from thepredominantly rural (e.g. Sri Lanka) to the highly urbanized (e.g. Brazil). TheTable also indicates which household survey was used for the estimation of thehousehold expenditure model, including year and sample size. The year of thenearest available population Census, which was used to generate the small-areawelfare estimates, is also included.9
9. Further details about the poverty maps analyzed here can be found in respectively (INSTAT,
2004) for Albania, IBGE (2003) for Brazil, Kenya Central Bureau of Statistics (2003) for Kenya,
Lopez-Calva et al. (2005) for Mexico, Haut Commissariat au Plan (2005) for Morocco, Healy and
Jitsuchon (2007) for Thailand, Department of Census and Statistics (2005) for Sri Lanka. In the case of
Kazakhstan, the poverty map for that country was produced on a pilot basis in collaboration with the
Agency of Statistics of the Republic of Kazakhstan. The results of this exercise have not been placed in
the public domain.
Ferre, Ferreira and Lanjouw 357
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E1
:D
ata
Sourc
es
Alb
ania
Bra
zil
Kaza
kh-s
tan
Ken
yaM
exic
oM
oro
cco
Thailand
Sri
Lanka
Urb
an
Popula
tion
1.3
m125m
8.2
m5.0
m54.5
m12.7
m18.5
m2.2
mU
rban
Popula
tion
share
0.4
20.8
30.5
70.1
90.6
00.5
10.3
10.1
2C
ensu
sY
ear
2001
2000
1999
1999
2000
1994
2000
2001
Surv
eyY
ear
2002
2002-3
2001
1997
2000
1998
2000
2002
Surv
eyN
am
eL
SM
SPO
FH
BS
WM
SII
IE
NIG
HE
NN
VM
SE
SH
IES
Surv
eySam
ple
Siz
e3,6
00
48,4
70
11,8
83
10,8
74
10,1
08
5,1
84
24,7
47
20,1
00
Pove
rty
Lin
e1A
LL
4,8
91
BR
L100
KZ
T3,1
57
KE
S2,6
48
PE
S768
DH
S283
BA
H1370
LK
R1423
Pove
rty
Lin
e(2
005
PPP$)
85
83
63
147
126
57
88
45
Equiv
ale
nce
Sca
leN
oN
oN
oY
esN
oN
oN
oN
o
1A
llpove
rty
lines
are
month
ly,
and
dis
pla
yed
usi
ng
nat
ional
curr
enci
es.
Inte
rnat
ional
acro
nym
sapply
.
Note
:L
SM
S:
Liv
ing
Sta
ndard
Mea
surm
ent
Surv
ey;
PO
F:
Pes
quis
ade
Orc
am
ento
sFam
ilia
res;
HB
S:
House
hold
and
Budget
Surv
ey;
WM
S:
Wel
fare
Monit
ori
ng
Surv
ey;
EN
IGH
:E
ncu
esta
Nac
ional
de
Ingre
sos
yG
ast
os
de
los
Hogare
s;E
NN
VM
:E
nquet
eN
atio
nale
sur
les
Niv
eaux
de
Vie
des
Men
ages
;SE
S:
Soci
o-E
conom
icSurv
ey;
HIE
S:
House
hold
Inco
me
and
Expen
dit
ure
Surv
ey.
Sourc
e:T
he
above
surv
eys
and
Nat
ional
Cen
suse
s.C
onsu
mer
pri
cein
dic
esand
PPP
exch
ange
rate
suse
dfo
rpove
rty
line
conve
rsio
ns
com
efr
om
Povc
alN
et.
358 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Finally, the Table also lists the poverty line used in each country, both innational currency (in survey year prices) and US dollars (in 2006 prices and atPPP exchange rates)10. As noted in the introduction, we have opted to usenational poverty lines, which better capture the meaning of poverty in eachspecific country. This has the drawback that poverty measures are not definedwith reference to comparable standards of living across countries. The alterna-tive of imposing a constant poverty line across countries, however, would havean even greater disadvantage. Had we selected a low internationally compar-able poverty line, such as $1-a-day, we would be comparing traces of povertydriven largely by measurement error and transitory shocks in the richer coun-tries (such as Albania and Brazil) with real poverty in Kenya and Sri Lanka.Had we instead selected a higher line, like those used in Albania, Brazil orMorocco, we would be comparing “reasonable” poverty incidences in thericher countries, with the bulk of the population in the poorer countries. Sincethis paper is largely about the relative extent of poverty in larger and smallertowns, the absolute level of the poverty line is of limited importance. We do,nevertheless, examine the sensitivity of our results to varying the poverty linein some of the countries, in the next section.
Further caveats relate to the fact that we make no attempt to apply auniform definition of urban, or to systematically correct for differences in thecost-of-living across different urban categories. In some settings, these differ-ences may be substantial, and future research should attempt to take them intoaccount.11 In addition, with the exception of Kenya, we have used consump-tion expenditure per capita as the individual welfare indicator throughout. Ifthere are substantive differences in family size or composition across differenturban categories, one might like to investigate the robustness of the resultswith respect to different assumptions regarding equivalence scales. Note thatwith respect to both cost of living differences and equivalence scales, our find-ings will be sensitive to systematic differences between large cities and smallertowns. We are not as vulnerable, here, to differences that might exist betweenurban areas, generically, and rural areas. It is an important empirical questionjust how much variation there is between cities of different sizes in terms ofprices, consumption patterns, and demographic characteristics.
I I I . C O N S U M P T I O N P O V E R T Y B Y C I T Y S I Z E
Table 2 presents our estimates of the three standard FGT poverty measures (aswell as population shares and the share of the poor) for each country as a
10. Each poverty line is per capita per month.
11. Section VI reports on a robustness check indicating that our findings for Brazil are not
overturned after refining the definition of urban we employ and correcting by cost of living differences
across city-size categories.
Ferre, Ferreira and Lanjouw 359
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA B L E 2: Poverty measures and shares for different city sizes in eightcountries
Population share1 FGT0 FGT1 FGT2 Share of the Poor2
Albania 0.25Rural 0.30Urban 0.42 0.18 0.04 0.01 0.31M 0.15 0.18 0.04 0.02 0.11S 0.13 0.18 0.04 0.01 0.09XS 0.14 0.20 0.05 0.02 0.11
Brazil 0.22Rural 0.37Urban 0.83 0.19 0.07 0.04 0.72XL 0.22 0.09 0.03 0.01 0.09L 0.07 0.17 0.06 0.03 0.06M 0.24 0.15 0.05 0.03 0.17S 0.01 0.19 0.07 0.04 0.01XS 0.28 0.30 0.11 0.06 0.39
Kazakhstan 0.18
Rural 0.23Urban 0.57 0.14 0.04 0.01 0.43XL 0.08 0.03 0.01 0.00 0.01M 0.29 0.13 0.04 0.01 0.21S 0.05 0.18 0.05 0.02 0.05XS 0.15 0.19 0.05 0.02 0.15
Kenya 0.51Rural 0.52Urban 0.19 0.47 0.17 - 0.16XL 0.07 0.44 0.14 - 0.06L 0.02 0.44 0.16 - 0.02M 0.03 0.46 0.17 - 0.03S 0.02 0.55 0.22 - 0.02XS 0.04 0.49 0.21 - 0.04
Mexico 0.32Rural 0.52Urban 0.60 0.19 0.06 0.03 0.39XL 0.27 0.18 0.06 0.03 0.16L 0.13 0.14 0.04 0.02 0.06M 0.11 0.19 0.05 0.03 0.07S 0.04 0.25 0.07 0.04 0.03XS 0.06 0.31 0.09 0.05 0.07
Morocco 0.17Rural 0.23Urban 0.51 0.11 0.03 0.01 0.34XL 0.12 0.04 0.01 0.00 0.03L 0.09 0.14 0.04 0.02 0.07M 0.27 0.13 0.03 0.01 0.20S 0.03 0.16 0.04 0.02 0.03
(Continued)
360 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
whole, and then for their urban areas, first as an urban aggregate, and then dis-aggregated into five size categories: towns smaller than 50,000 (“very small” orXS); between 50,000 and 100,000 (“small” or S); between 100,000 and500,000 (“medium” or M); between 500,000 and 1 million (“large” or L) andabove 1 million (“metropolitan areas” or XL). Two countries have no metro-politan areas: Albania and Sri Lanka. Two countries have no large cities:Albania and Kazakhstan.
In all eight countries, both poverty incidence (FGT(0)) and depth (FGT(1))are highest in either the very small (Albania, Brazil, Kazakhstan, Mexico, SriLanka, Thailand) or the small (Kenya and Morocco) categories.12 This patternis particularly pronounced in the larger, more urbanized countries of Brazil,Kazakhstan, Mexico, Morocco, and Thailand where FGT(0) in the very smallcities is up to six times larger than in the metropolitan areas, and often onlyslightly lower than in rural areas. In these countries, the more distribution-sensitive poverty measures paint a similar picture: FGT(2) is six times largerfor very small towns than for metropolitan areas in Brazil; FGT(1) is five timeslarger in Kazakhstan.
TABLE 2: Continued
Population share1 FGT0 FGT1 FGT2 Share of the Poor2
XS 0.01 0.12 0.03 0.01 0.01
Sri Lanka 0.23Rural 0.25Urban 0.12 0.09 0.02 0.01 0.05L 0.03 0.08 0.02 0.01 0.01M 0.03 0.07 0.02 0.01 0.01S 0.02 0.09 0.03 0.01 0.01XS 0.04 0.12 0.03 0.00 0.02
Thailand 0.14
Rural 0.17Urban 0.31 0.08 - - 0.17XL 0.12 0.02 - - 0.01M 0.03 0.04 - - 0.01S 0.02 0.09 - - 0.01XS 0.14 0.14 - - 0.13
1 Proportion of the population living in each category: urban, XL,L,M,S,XS.2 Proportion of the country’s poor living in each category: urban, XL,L,M,S,XS.
XL: .1,000, L: 500-1,000, M: 100-500, S: 50-100, XS: ,50 (‘000 inhabitants).
Source: Authors’ analysis based on data described in the text.
12. FGT(a) denotes a member of the Foster, Greer and Thorbecke (1984) family of poverty indices,
with parameter a.
Ferre, Ferreira and Lanjouw 361
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
In the other three countries – Albania (heavily urbanized, but small in totalpopulation and area), Kenya and Sri Lanka (predominantly rural) – thepattern is less pronounced, but it is still present. In fact, the coefficient onpopulation in a simple OLS regression of poverty on city size is negative in allcases, and significant at the 10% level in five (four) out of eight cases forFGT(0) (FGT(1)). See Table 3.13
The inverse relation between poverty and city size can also be discerned inFigure 1, which presents the distribution of poverty incidence within each sizecategory, by means of box-plots. The box-plots indicate that there is muchgreater variance in poverty rates among smaller towns, as one might expectfrom their sheer number. But the median poverty rate falls markedly andconsistently with city size in Brazil and Kazakhstan. It also falls in Albania,although less markedly. In Mexico, the gradient is clear across all city-sizecategories except for the metropolitan areas for which the median poverty rateis slightly higher than for the large city-size category (but below all othercategories). In Morocco, the negative correlation detected in Table 3 is drivenby much lower poverty in metropolitan areas, with no clear pattern among theother size categories. In Sri Lanka and Kenya, the relationship owes to greaterpoverty incidence in small and very small towns, with no clear pattern amongmedium and larger towns. Similarly in Thailand it is noteworthy that the over-whelming majority of urban centers belong to the extra small category, withmetropolitan Bangkok representing the one very large exception. (Chiang Mai,Thailand’s second largest city, had a population of only just over 280,000 in2008). The median poverty rate in Thailand’s smallest towns is markedlyhigher than in all other city-size categories.
These patterns are also clearly visible in Figure 2, which presents thenon-parametric regressions of FGT(0) on the logarithm of city size for eachcountry. Here again, it is least visible in Kenya and Sri Lanka. In Morocco, aswe have seen, the negative relationship is driven by markedly lower poverty inCasablanca. With the exception of Kenya and Mexico, metropolitan povertyincidence is less than half of the average urban poverty in every country in oursample that has at least one metropolitan area.
The data underpinning Figures 1 and 2 are of interest not only in providingevidence of a poverty gradient with respect to city size in our eight countries,but also in documenting the great heterogeneity in poverty across towns withina given city size category. Thus, while in Brazil or Thailand the median povertyrate amongst towns in the XS town-size category is clearly higher than in theother categories, there evidently are very small towns that also enjoy very lowpoverty rates (as well as small towns with near universal poverty). Indeed, in
13. These regression coefficients are presented as illustrative of correlations only. City size is clearly
endogenous, and there are evidently many omitted variables, so no inference of causality is possible.
Some countries do not display the full set of regressions for lack of data (the Kenyan census for instance
being very short, no information is available on infrastructure access).
362 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E3
:Sim
ple
regre
ssio
ns
of
pove
rty
indic
ators
on
city
size
,O
LS
Dep
enden
tvari
able
:
FG
T0
FG
T1
Wat
erE
lect
rici
tySew
er
Cou
ntr
yco
eff
p-v
alu
eco
eff
p-v
alu
eco
eff
p-v
alu
eco
eff
p-v
alu
eco
eff
p-v
alu
e
Alb
an
ia2
0.1
70.2
12
0.0
40.2
72
22
22
2
Bra
zil
20.1
00.0
32
0.0
40.0
30.0
30.0
20.0
10.0
30.0
30.0
4K
aza
khst
an
20.2
00.0
02
0.0
60.0
02
20.0
00.0
60.9
20.0
2K
enya
20.0
20.7
32
0.0
40.3
82
22
22
2
Mex
ico
20.0
20.0
62
0.0
04
0.2
80.0
07
0.2
40.0
02
0.0
70.0
07
0.1
6M
oro
cco
20.0
30.0
12
0.0
10.0
20.0
60.0
10.0
50.0
20.0
80.2
3T
hail
an
d2
0.0
40.1
52
0.0
10.2
21.0
00.0
40.0
60.0
0Sri
Lan
ka
20.0
80.0
82
0.0
30.0
50.4
00.2
32
0.0
30.5
42
2
Expla
nat
ory
vari
able
:ci
tysi
zein
mil
lion
of
inhabit
ants
(‘000,0
00).
All
pove
rty
indic
ators
take
valu
esbet
wee
n0
and
1.
Sourc
e:A
uth
ors
’analy
sis
base
don
dat
ades
crib
edin
the
text.
Ferre, Ferreira and Lanjouw 363
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Albania the evidence indicates that the lowest estimated urban poverty rates arefound amongst towns in the XS category. It is thus important to bear in mindthat while on average poverty rates in smaller towns tends to be higher than inmedium and large cities, this is far from a general rule.
To investigate the robustness of the inverse poverty-city size relationship tovariations in the poverty line, we plotted the cumulative distribution function
FIGURE 1.
Source: Authors’ calculations based on the household surveys and censuses listed in Table 1.Note: XL: .1,000, L: 500–1,000, M: 100–500, S: 50–100, XS: ,50 (thousands
inhabitants).
364 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
separately by size category for each country.14 Poverty is always higher in thesmallest towns (XS), for any poverty line, in Albania, Brazil, Sri-Lanka andThailand (up to the 90th percentile). It is generally lowest for metropolitanareas in the vicinity of the national poverty lines, but this ranking is not every-where robust to larger changes in the poverty line. Figures 3 and 4 illustratetwo polar cases: Brazil and Morocco. Figure 3 shows that metropolitan areas
FIGURE 2.
Source: Authors’ calculations based on the household surveys and censuses listed in Table 1.Note: fitted with Lowess regression – bandwidth ¼ 1.
14. Again with the exception of Kenya, for which we do not have the disaggregated poverty
mapping data.
Ferre, Ferreira and Lanjouw 365
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
first-order stochastically dominate all other size categories in Brazil: poverty islower in these large cities than in any other type of town, by any poverty line.Conversely, very small towns are first-order stochastically dominated by everyother size grouping: poverty is higher in this size category than in any other, byany poverty line.
FIGURE 3.
Source: Brazilian poverty map, constructed from the POF survey and the 2000 census.Note: XL: .1,000, L: 500–1,000, M: 100–500, S: 50–100, XS: ,50 (thousands
inhabitants).
FIGURE 4.
Source: Moroccan poverty map, constructed from the ENNVM survey and the 1994 census.Note: XL: .1,000, L: 500–1,000, M: 100–500, S: 50–100, XS: ,50 (thousands
inhabitants).
366 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
A very different picture (in terms of dominance relationships) is that ofMorocco, shown in Figure 4. The poverty ranking between metropolitan areasand large towns which is observed at the country’s poverty line of Dhs 3,400reverses at higher poverty lines (above Dhs 8,000). Similarly, there is no domin-ance relationship among very small, small and medium towns in Morocco:their cumulative distribution functions cross many times. Even in Morocco,however, which displays the largest number of cumulative distribution functioncrossings in our sample, there is still one broad regularity: taken as a group,large and very large cities (L, XL) do provide a lower envelope for the smallertowns (XS, S, M). There is no strict stochastic dominance but it is evident that,for almost every poverty line one could think of, poverty is lower in the groupof larger cities than in other urban settings.
It is possible, of course, that poverty is both more widespread and deeper insmaller towns, but that population is so concentrated in large cities that thebulk of the poor live there. If this were the case, greater attention to (andresources for) metropolitan poverty might be justified on the basis that theshare of poverty is greatest there. But Table 2 shows that this is nowhere thecase. In fact, the share of the poor is lower than the population share in everycountry that has a metropolitan area: the difference is relatively small inKenya, but very substantial elsewhere. In Brazil, although 22% of the popula-tion live in cities greater than 1 million, only 9% of the country’s poor do. InKazakhstan, 14% of the population lives in Almaty, but only 3% of the poor.In Mexico, 27% of the population resides in Mexico City and the other verylarge metropolitan areas of the country, but only 16% of the poor live in theseconurbations. In Morocco, 12% of the population lives in Casablanca but only3% of the poor.
At the other end of the size distribution, a majority of the country’s urbanpoor live in small or very small towns in four of our eight countries: Albania,Brazil, Sri Lanka and Thailand. If we add medium towns to the list, this rises toseven of the eight countries, including Kenya. And even in the case of Mexico,where the population weight of metropolitan areas is particularly large, theshare of the urban poor in medium or smaller sized cities exceeds 40%.
I V. AC C E S S T O S E R V I C E S B Y C I T Y S I Z E
Even though people are poorer in smaller towns than in large cities and eventhough a greater number of the poor live in those smaller towns, one mightthink that, due to the higher population densities in metropolitan areas, per-capita availability of publically provided basic services was lower there. Thisdoes not appear to be the case, however. Table 4 presents the proportion ofhouseholds with access to various basic infrastructure services by city size, inseven of our eight countries.15
15. Kenya is once again omitted for data reasons.
Ferre, Ferreira and Lanjouw 367
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA B L E 4: Access to services for different city sizes in seven developingcountries
Water Electricity Sewer Gas Garbage Fridge Electric Heat
AlbaniaUrban 0.88 0.62M 0.91 0.68S 0.87 0.57XS 0.87 0.60
BrazilUrban 0.96 0.99 0.92 0.86XL 0.98 1.00 0.94 0.92L 0.97 1.00 0.91 0.89M 0.97 1.00 0.93 0.91S 0.96 0.99 0.94 0.89XS 0.92 0.98 0.90 0.76
KazakhstanUrban 1.00 0.68 0.55XL 1.00 0.73 0.81M 1.00 0.80 0.62S 1.00 0.67 0.36XS 1.00 0.40 0.31
MexicoUrban 0.93 0.99 0.94 0.83XL 0.95 0.99 0.97 0.84L 0.93 0.99 0.93 0.87M 0.92 0.98 0.93 0.81S 0.91 0.98 0.91 0.78XS 0.89 0.98 0.90 0.76
Morocco
Urban 0.77 0.82 0.87XL 0.84 0.87 0.87L 0.86 0.87 0.80M 0.71 0.79 0.91S 0.73 0.78 0.91XS 0.75 0.78 0.45
ThailandUrban 0.65 0.16 0.86XL 0.87 0.29 0.88M 0.76 0.23 0.90S 0.61 0.14 0.87XS 0.50 0.07 0.84
Sri LankaUrban 0.57 0.89L 0.53 0.86M 0.68 0.90S 0.60 0.92XS 0.51 0.89
XL: .1,000, L: 500-1,000, M: 100-500, S: 50-100, XS: ,50 (‘000 inhabitants).
Source: Authors’ analysis based on data described in the text.
368 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Access to piped water is generally quite high in Brazil, but it declines from98% in metropolitan areas, to 92% in very small towns. In Mexico andThailand the comparable figures are 95% to 89%, and 87% to 50%, respect-ively. In Morocco and Sri Lanka, the picture is less clear. In Morocco, as forincome poverty, access to piped water is higher in the two largest sizecategories (L, XL), than in the other three (M, S, XS). In Sri Lanka, there is aninverted U curve, with access lowest in large and very small towns. Similar pat-terns hold in each of these countries with respect to access to electricity,although overall access rates tend to be higher. Access increases monotonicallywith city size in Brazil and Mexico; it is higher in L and XL cities than in S,XS and M towns in Morocco (but with no clear pattern within these twoblocks), and it follows an inverted U in Sri Lanka.
Access to networked sanitation and sewerage facilities is on average scarcerthan piped water or electricity in most developing countries. And in our sampleof countries, there is also a clear positive association between city size andaccess to networked sewerage services. In all five countries that report data onthis service (Brazil, Kazakhstan, Mexico, Morocco and Thailand), very smalltowns have the lowest access rates – in two cases just barely half the ratesobserved in larger towns. Interestingly, however, in both Kazakhstan andMorocco, medium-sized towns report higher access rates than metropolitanareas.
Access to piped natural gas is an important infrastructure service inKazakhstan (for cooking and heating). Access is clearly and monotonicallyincreasing with city size. The differences are quite sizable, with 81% of con-nected households in Almaty, but only 31% in very small towns. A similarpattern attains for electric heating apparatus in Albania. Access to organizedsolid waste disposal (garbage collection) is only reported for Brazil, where it isonce again highest in metropolitan areas, and lowest in very small towns.16
V. LO O K I N G W I T H I N C I T I E S : T H E C A S E O F M O R O C C O
A further plausible argument for focusing one’s poverty-reduction efforts onmetropolitan areas might be that – even if poverty is less widespread or intensethere; even if a smaller share of the poor live there; and even if they alreadyenjoy superior access to services – these very large urban centers are deeplydivided between rich and poor. If relative incomes matter for well-being, thenthe stark contrast between the crowded and steep hillsides of Rocinha and theneighboring verdant gardens of Gavea in Rio de Janeiro may be so inherentlyobjectionable as to raise the priority that should be accorded to fightingpoverty in large cities.
16. Although the relationship for intermediate size categories is not monotonic, and there is very
little difference between large, medium and small towns in this respect.
Ferre, Ferreira and Lanjouw 369
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
There may well be something to the argument that stark local inequalitiesmay have greater costs than geographically diffuse inequality. There is someevidence that relative incomes in one’s vicinity do affect well-being directly(Luttmer, 2005), and that local inequality may lead to increased propertyviolence (Demombynes and Ozler, 2005). But there is much less evidence thatinequality is indeed so much greater in metropolitan areas than in smallertowns. Although this is a popular notion, it is one for which very limited statis-tical backing exists – in large part for previously mentioned reasons: householdsurveys are not representative at the level of smaller towns, and so we knowvery little about local inequality in them. It may be that the accumulation ofanecdotal evidence of large inequalities in developing country metropolises isitself simply another reflection of metropolitan bias: Journalists and photogra-phers, like most economists and policy analysts, prefer to visit Casablanca thanFiguig17, and Rio de Janeiro than Bertolınia.18
To shed some additional light on this matter, we now turn to some evidencefrom Morocco. Table 5 presents FGT(0) and three inequality measures (theGini coefficient and the two Theil indices) for each of the five largest cities inthe country, as well as the aggregate inequality for three city size categories.Overall intra-city inequality does not appear to be positively correlated withcity size in this small sample, but this is not the main point. Taking advantageof the fine spatial disaggregation made possible by a poverty map, we calcu-lated inequality measures for various individual neighborhoods within each ofthese cities. We further classified these neighborhoods into slums and non-slums19. We then decomposed the two Theil indices of inequality20 for each ofthe five cities, into a component due to inequality within each of the twogroups of neighborhoods, and a component between the two. For the argumentthat within-city inequality is egregiously large in metropolitan areas and largecities to hold (in Morocco), it would be necessary (but not sufficient) that thebetween-group shares reported in the last two columns of Table 5 be substan-tial. In the event, it appears that most inequality in the five largest cities inMorocco is not due to some great divide between slum areas and other parts ofthe town. Inequality appears to be considerably more widely dispersed withinthese two broad groups.
17. Casablanca is the biggest agglomeration of Morocco (2.9 million inhabitants), Figuig is a small
town in L’Oriental (49,000).
18. Bertolinia is a small town in the Brazilian state of Piauı, with fewer than 40,000 inhabitants.
19. A district (smallest level of disaggregation after the census track) was considered as a slum if less
than 10% of the population had access to water and less than 10% of the population had access to
electricity.
20. GE(0), or mean log deviation, is the Theil-L index. GE(1) is the Theil-T index. Both are
perfectly decomposable into within- and between-group components, in the sense that the
decomposition has no residual. GE(0) weighs within-group inequalities by population shares, while
GE(1) weighs them by incomes shares.
370 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
V I . F O U R R O B U S T N E S S C H E C K S
We noted at the outset several important caveats attached to the broad findingsin this paper. While we are unable to subject all of them to exhaustive scrutiny,we attempt, in this section, to probe several with a view to gauging the robust-ness of our findings. We assess, first, our reliance on official, country-specific,definitions of what constitutes an urban setting in our eight countries. In mostcountries, the designation of an area as urban is based on several criteria,combining both administrative functions and population characteristics. As astarting point, many countries designate as urban any area, irrespective ofpopulation size or density, that happens to have been declared a municipalityor corporation by state law, or that has a Cantonment Board or notified townarea committee. In general such administrative criteria are combined withadditional criteria such as a minimum population (e.g. 5000 inhabitants), aminimum density of population (e.g. at least 400 persons per sq. km), and/or aminimum degree of economic diversification out of agriculture (e.g. at least75% of male working population engaged in non-agricultural pursuits). Asmentioned in the introduction, we take in this paper the definition of urban
TA B L E 5: The role of metropolitan slums: Poverty and inequalitydecompositions within five cities in Morocco
Morocco
All Urban Areas2 groups: slums and
non-slums1
Population FGT0 GINI GE02 GE12 W03 W13 B03 B13
Casablanca 2,875,326 0.05 0.39 0.25 0.29 0.24 0.29 0.01 0.01Rabat 604,680 0.04 0.41 0.27 0.30 0.27 0.29 0.01 0.01Sale 575,600 0.12 0.38 0.24 0.26 0.22 0.25 0.02 0.02Marrakech 503,802 0.07 0.37 0.23 0.23 0.23 0.23 0.00 0.00Fes 501,592 0.23 0.47 0.38 0.41 0.38 0.41 0.00 0.00200,000-500,000 4,930,980 0.12 0.36 0.21 0.23 0.21 0.22 0.00 0.00,200,000 2,736,390 0.15 0.22 0.13 0.14 0.13 0.11 0.00 0.00All urban4 12,728,370 0.11 0.37 0.23 0.25 0.23 0.24 0.01 0.01
1 A slum is a district (smallest disaggregation above the census tract) where less than 10% ofthe population has access to water and less than 10% has access to electricity.
2 GE0, or mean log deviation, is the Theil-L index. GE1 is the Theil-T index. Both are perfect-ly decomposable into within- and between-group components, in the sense that the decompos-ition has no residual. GE0 weighs within-group inequalities by population shares, while GE1weighs them by incomes shares.
3 W0 and W1 display within-group inequality associated with the GE0 and GE1 measures re-spectively; B0 and B1 display the corresponding between-group inequality component.
4 Each index presented here was computed at the city level and then aggregated into each cat-egory (all urban, etc).
Source: Authors’ calculations based on the ENNVM: Enquete Nationale sur les Niveaux deVie des Menages.
Ferre, Ferreira and Lanjouw 371
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
employed by each respective country, and do not try to impose a uniformdefinition across the countries in our sample. The question thus arises whetherour finding of a town size–poverty gradient would survive the adoption of analternative definition of urban. In particular, it is of interest to know whetherusing a more data driven, and less administrative, definition would affect ourconclusions.
To explore this concern we take the Brazil census data and derive our ownurban area definitions for that country.21 We start with census tracts (group-ings of roughly 100 geographically contiguous households) that have beendesignated as either rural or urban by IBGE, the Brazilian National StatisticalOffice. We then assemble these census tracts into towns and cities by joiningcontiguous urban census tracts together. In some cases these form huge units.For example, a coastal area near Rio de Janeiro is 10 km wide and over170 km long. Compared to the GRUMP (Global Rural Urban Mapping Projectat the Columbia University Center for International Earth Science InformationNetwork) settlement area database, this approach would appear to jointogether several distinct cities into megacities. Because these seem too large, wecut them up using Brazilian micro-regions (officially recognized agglomerationof municipios, but subunits of states) which divides up these “super agglomera-tions” into areas that closely approximate what GRUMP considers independentcities in Brazil, together with their immediate metropolitan areas. While thedefinition of a town or city, under this approach, is not tied strictly to a popu-lation density criterion (probably the most economically logical way ofdefining a town) it does employ a single, uniform, criterion and is likely tocorrelate reasonably well with a purely density-based definition.
Table 6 recalculates poverty across city size categories when cities have beendefined in this alternative fashion. We also drop from consideration, here, thelarge subset of very small towns with less than 25,000 inhabitants22. One might
TA B L E 6: Poverty and City Size in Urban Brazil: Modified Definition ofTowns and Cities based on Contiguous Census Tracts
City SizeTotal Population
(000’s) FGT0Share of Urban
PopulationShare of Urban
Poor
25,000-100,000 21,261 0.22 0.20 0.29100,000-500,000 22,384 0.16 0.21 0.22.500,000 61,655 0.12 0.59 0.49All Urban 105,300 0.15 1.00 1.00
Source: Thomas (2008).
21. We draw here on work undertaken by Timothy Thomas in support of the project described in
this paper (Thomas, 2008).
22. There are nearly 10,000 such settlements in Brazil.
372 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
argue that such very small towns are hard to distinguish clearly from rural areas,particularly where those rural areas are reasonably densely populated.23 Weexamine whether our city size – poverty gradient for Brazil survives the exclusionof these very smallest towns. We see from Table 6 that amongst all urban centerswith a population of at least 25,000 inhabitants, poverty in small towns (with apopulation of 25,000 to 100,000) is markedly higher than in the medium sizedcities (100,000-500,000) which in turn record greater poverty than the large citiesof 500,000 inhabitants or more. In terms of poverty shares, our earlier findingsalso survive: while the inhabitants of Brazil’s largest cities represent 59% of theurban population (as per our new definition), they represent just 50% of theurban poor. Conversely, small towns of 25,000 to 100,000 inhabitants accountfor just 20% of the urban population but nearly 30% of the urban poor.
Our second robustness check assesses the possibility that there may beimportant cost of living differences between urban settlements of differentsizes. The findings reported above have not attempted to adjust for suchcost-of-living differences, because spatial price indices across city-sizecategories are not generally available. It is well recognized in the literature,however, that observed differences in poverty rates between urban and ruralareas can be significantly attenuated once one corrects for the fact that the costof living in urban areas may be much higher than in rural areas (generallybecause of the higher cost of food, housing and transport services). The possi-bility exists that our broad findings of greater poverty in small towns than inmetropolitan areas might also be driven, at least in part, by our failure to allowfor a higher cost of living in metropolitan areas.
While household survey datasets are not generally large enough in samplesize to permit the construction of a cost-of-living index across different city-sizecategories, our survey data for Brazil constitute an important exception. Weare able to draw on the 2002 POF data (see Table 1) to construct acost-of-living index across the broad city size categories employed in thispaper, and can check whether our findings for Brazil, reported in previous sec-tions, are robust to this correction.
There are many ways in which spatial price indices can be constructed. Wefollow here the approach applied by Ferreira, Lanjouw and Neri (2003) to theconstruction of a regional price index (that distinguished also between urbanand rural areas) using 1996 PPV data for Brazil. This approach was subse-quently applied in World Bank (2007) to produce a regional price index basedon the 2002 POF data, and is based on unit-value information provided in the
23. Nevertheless, the large number of small towns in developing countries, and the high poverty
rates that characterize them (Table 2), raises the interesting question of the role of such towns in the
classic rural-urban debates in development economics (for a recent overview, see Christiaensen, Demery
and Kuhl, 2011). Lanjouw and Murgai (2009) suggest, in the context of India, that growth in such
small towns has promoted rural non-farm diversification which, in turn, contributes to rural poverty
reduction by both providing new employment opportunities for the poor and by raising agricultural
wages via agricultural labor market tightening (see also World Bank, 2011).
Ferre, Ferreira and Lanjouw 373
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
POF survey on food items, as well as a hedonic model of rent. We re-apply themethod here, but focus solely on urban areas and construct a Laspeyres priceindex that captures price differences across city-size categories (and regions).Because of the limited sample size of even the unusually large POF householdsurvey we employ a three-way city size breakdown, distinguishing between citieslarger than 500,000 persons, large towns with a population between 100,000and 500,000, and towns with fewer than 100,000 inhabitants. For reasons ofdata availability our index captures only food and housing price differentials.24
The reference basket employed in our Laspeyres index is based on the consump-tion patterns of the second quintile of the national urban per capita consumptiondistribution.
Table 7 presents our Laspeyres spatial price index based on the cost of foodand housing in urban Brazil. Relative to the reference region of metropolitan SaoPaulo, the cost of living in other regions and city size categories of Brazil is gener-ally lower. In the regions of the North, South and Center West there is evidence oflower cost of living as towns become smaller. This pattern is less clear-cut in theNorth East, where the cost of living appears particularly high in large towns, andthe South East, where those living in small towns appear to face the highest costof living in the region (although still well below metropolitan Sao Paulo).
TA B L E 7: Spatial Price Indices across City Size Categories in Urban BrazilLaspeyres Price Indices Based on the Cost of Food and Housing
Region City Size Category Laspeyres Price Index
North Large (.500,000) 0.94Medium (100,000-500,000) 0.75Small (,100,000) 0.68
North-East Large (.500,000) 0.66Medium (100,000-500,000) 0.72Small (,100,000) 0.60
South-East Large (.500,000) 0.55Medium (100,000-500,000) 0.49Small (,100,000) 0.84
Sao Paulo 1.00South Large (.500,000) 0.76
Medium (100,000-500,000) 0.65Small (,100,000) 0.62
Center-West Large (.500,000) 0.86Medium (100,000-500,000) 0.80Small (,100,000) 0.64
Source: Authors’ calculations based on the POF: Pesquisa de Orcamentos Familiares.
24. As noted above, transport costs are also likely to vary substantially across city size categories.
Among the poor, these costs typically account for a lower share of expenditures than food and housing.
Nevertheless, the sensitivity of our results to spatial variations in the cost of living that include transport
costs remains an issue for future work.
374 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
We next take the price indices reported in Table 7 and apply them to thesmall-area based estimates of per capita consumption for each household in thepopulation census. We then re-calculate poverty rates across region and city-size categories. Does this cost of living adjustment overturn our conclusion thaturban poverty in smaller towns is significantly higher than in the large andmetropolitan areas? Table 8 indicates that it does attenuate the “gradient”between poverty and city size somewhat, but is far from sufficient to negate oroverturn our broad finding. In Brazil, it remains the case that the incidence ofpoverty in the smallest towns is roughly three times higher than inMetropolitan centers.
In a third robustness check we investigate whether the finding of a negativegradient between poverty and city-size is somehow an artifact of the focus, inthis paper, on income poverty as opposed to a broader conceptualization ofdeprivation. While the broad pattern of higher poverty and lower access to ser-vices in small towns was found to be quite robust across our eight countries,an important potential caveat to this assessment concerns health outcomes. Ithas been suggested in the literature that health outcome indicators in largecities in the developing world may lag behind those in smaller towns. Forexample, Chattopadhyay and Roy (2005) demonstrate that a variety of indica-tors of child mortality are more pronounced in the large cities of India than intowns and medium sized cities. This study finds that while infant mortalityamongst the wealthiest classes in large cities are particularly low, infantmortality rates amongst the poorest classes are quite pronounced – and indeedare higher than amongst the poorer segments in small and medium sizedtowns. These are suggestive findings and may be related to the particularlyunhealthy living conditions in over-crowded slum areas of large cities.However, evidence on health outcomes across city sizes categories remainsscarce and there does not appear to be a broad consensus in the literature onthe relatively higher health risks in large cities. For example, Kapadia-Kunduand Kanitkar (2002) argue, also with reference to India, that urban publichealth services generally place greater emphasis on mega-cities and metro-centers, to the relative neglect of smaller cities and towns.
TA B L E 8: Poverty measures for different city sizes in Brazil Checking forRobustness to Cost of Living Differences
Population share1 FGT0 (nominal expenditure) FGT0 (real expenditure)
Urban 0.83 0.19 0.18XL 0.22 0.09 0.06L 0.07 0.17 0.10M 0.24 0.15 0.10S 0.01 0.19 0.11XS 0.28 0.30 0.19
XL: .1,000, L: 500-1,000, M: 100-500, S: 50-100, XS: ,50 (‘000 inhabitants).
Source: Authors’ analysis based on data described in the text.
Ferre, Ferreira and Lanjouw 375
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
We probe this concern by examining the gradient of child anthropometricoutcomes across city size categories in Mexico. We draw on a small area esti-mation effort undertaken by Rascon (2010) that parallels the work reported inpreceding sections, but focuses on anthropometric outcomes rather thanincome poverty. Rascon combines the Mexican National Survey of Health andNutrition 2006 with the Second Count of the Population and Dwellings 2005in order to apply a variant of the Elbers et al. (2003) small area estimationprocedure to the incidence of stunting and underweight amongst children aged5 and below in Mexico.25 Lanjouw and Rascon (2010) examine the correlationof child health outcomes in urban areas with city size. Table 9 summarizestheir results and documents that the incidence of low height for age (stunting)and low weight for age (underweight) amongst children displays a similargradient across city sizes as we have seen for income poverty. In Mexican citiesthat are larger than 500,000 inhabitants, the incidence of stunting and ofunderweight amongst children is 9%. In the case of stunting the incidence risesmonotonically as cities decline in size. Amongst the smallest cities (of less than10,000 inhabitants) the incidence is as high as 16%. In the case of under-weight, the incidence also rises, but less markedly: from 9% in the largest citiesto 11% in the small cities. The higher incidence of child malnutrition in smalltowns also translates into more malnourished children: 27% of stunted chil-dren in urban areas are found in the largest cities, while 29% are in townswith less than 15,000 inhabitants. Similarly, 27% of underweight urban chil-dren reside in the largest cities, but 30% reside in the smallest towns.
In a final robustness check, we ask whether our observed gradient betweenpoverty and city-size is driven by our reliance on small area estimates ofpoverty for each city rather than direct measures of poverty for such local-ities.26 We have already noted that direct measures of poverty for individualtowns and cities are not generally available in developing country settings. Thehousehold surveys that underpin poverty analysis in these countries do notgenerally cover sufficiently large samples to permit poverty measurement at thisdetailed level. As was described in Section 2, the small area estimationprocedure applied in the present paper combines household survey withunit-record population census data in an effort to circumvent this small sampleproblem. The approach takes advantage of the full population coverage of the
25. Fujii (forthcoming) adapts the Elbers et al. procedure for the estimation of anthropometric
outcomes and applies this methodology to Cambodia. Rascon adapts this procedure further to apply it
to Mexican data.
26. Tarozzi and Deaton (2007) have recently expressed a concern that the small area estimation
procedure employed by ELL (2002, 2003) may overstate the precision of local level poverty estimates.
They base their argument on Monte Carlo simulation results. (See also Molina and Rao, 2010) Elbers,
Lanjouw and Leite (2008) examine this issue with data for the state of Minas Gerais in Brazil, and find
little evidence in that specific setting for concern. It remains true, though, that the ELL procedure
estimates poverty, rather than directly measuring it, and as such there is interest in assessing whether the
findings reported in this paper would also hold had poverty been directly measured.
376 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
population census and then applies statistical techniques to insert into thecensus an indicator of per capita expenditure or income for each household.This is necessary because in most developing (and developed) countries, thepopulation census fails to collect detailed income or expenditure data.
India offers an opportunity to probe the contention that our findings aremerely an artifact of the methods we have employed. The Indian NationalSample Survey Organization (NSSO) fields a very large sample survey everyfive years with a sample size that is sufficiently large to permit a breakdown ofurban areas into city-size categories.27 Table 10 draws on a World Bank study(World Bank, 2011) to illustrate that at the national level for the years 1983,1993 and 2004/5, National Sample Survey data show a clear gradient inpoverty by city size. This gradient holds both at the national level, as well as atthe level of individual states.28 A recent study applies the small area estimationmethodology used here to estimate poverty at the local level in three states ofIndia in 2004/5 (Gangopadhyay et al., 2010). The study confirms that in WestBengal, Orissa and Andhra Pradesh the poverty-city size gradient observedfrom NSS data also emerges from estimates derived out of the small-area esti-mation procedure (Table 11). Thus, at least in India, the finding of an inversepoverty-city size gradient is robust to alternative empirical methods. This pro-vides some support to the claim that the findings reported in preceding sectionsare not driven by our reliance on small-area estimation techniques.
TA B L E 9: Child malnutrition estimates for different city sizes in Mexico Smallarea estimates of malnutrition amongst children under 5 in urban areas
Locality size(inhabitants)
Stunting Underweight
Incidence
Share ofUrban
Population
Share ofNational
Population Incidence
Share ofUrban
Population
Share ofNational
Population
L 0.09 0.27 0.15 0.09 0.27 0.18M 0.11 0.44 0.24 0.09 0.43 0.28S 0.16 0.29 0.15 0.11 0.30 0.19
L: .500, M: 15-500, S: 2.5–15 (’000 inhabitants).
Source: Lanjouw and Rascon (2010).
27. Every five years the NSS fields a “thick round” with a sample size of around 120,000
households, The “thin rounds” fielded in the other years have sample sizes of around 30-40,000
households.
28. World Bank (2011) also shows that the pattern of differential per capita access to public
services across city size categories is skewed in India, with small towns faring more poorly than large
cities.
Ferre, Ferreira and Lanjouw 377
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
V I I . C O N C L U D I N G R E M A R K S
Using highly disaggregated poverty map data from eight countries drawn fromall six regions of the developing world, we have shown evidence of a common– although not universal – inverse relationship between poverty and city size.In all countries in our sample, poverty is both more widespread (higherFGT(0)) and deeper (higher FGT(1)) in very small and small towns (those witha population below 100,000) than in large or very large cities (those with apopulation greater than 0.5 million). Metropolitan poverty, in particular, isconsiderably lower than poverty in other urban areas in all countries in oursample, except for Kenya and Mexico. Dominance analysis of cumulativedistribution functions indicates that the basic pattern is generally robust to thechoice of poverty line.
Neither is it true that, because of sheer population size, most poor people inthese countries live in large cities. In fact, in all eight countries, a majority ofthe urban poor lives in medium, small or very small towns. In four of them
TA B L E 10: Poverty in India’s Small Towns Exceeds Poverty in the LargeCities: Direct Evidence from the NSS.
1983 1993-94 2004-05
Rural 0.465 0.368 0.281Urban: 0.423 0.328 0.258
Small towns 0.497 0.434 0.300Medium towns 0.423 0.315Large towns 0.290 0.202 0.147
Notes: Poverty rates based on NSS 1983, 1993 and 2004/5 surveys using Uniform ReferencePeriod consumption and official poverty lines. Small,50K, Medium 50K-1m, Large.¼ 1m.
Source: World Bank (2011).
TA B L E 11: Small area estimates reveal high poverty in small towns in threeIndian states
CitySize
West Bengal Orissa Andhra Pradesh
No.of
towns
Shareof
Pop
Shareof
Poor FGT0
No.of
towns
Shareof
Pop
Shareof
Poor FGT0
No.of
towns
Shareof
Pop
Shareof
Poor FGT0
XL 1 0.20 0.08 0.05 - - - - 1 0.18 17 0.23L 1 0.05 0.04 0.12 2 0.21 0.20 0.34 3 0.13 7 0.14M 54 0.48 0.46 0.13 6 0.22 0.19 0.31 37 0.39 37 0.24S 28 0.09 0.12 0.17 15 0.19 0.19 0.36 40 0.15 20 0.33XS 298 0.18 0.31 0.23 121 0.38 0.42 0.39 104 0.15 18 0.31
Note: XL . 1m; L: 500K-1m; M: 100K-500K; S: 50K-100K; XS , 50K.
Source: Gangophadyay et al. (2010) and World Bank (2010).
378 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
(Albania, Brazil, Sri Lanka and Thailand), a majority of the urban poor lives intowns smaller than 100,000 people.
The greater incidence and severity of consumption poverty in smaller towns iscompounded by similarly greater deprivation in terms of access to basic infra-structure services, such as electricity, heating gas, sewerage and solid wastedisposal. This pattern is not absolute. It does vary by type of service and acrosscountries. Access rates seldom increase strictly monotonically with city size, butthey do generally increase, so that for most services and in most countries, largecities and metropolitan areas have higher coverage rates than smaller towns.
Finally, we have also shown for one particular country – Morocco – thatinequality within large cities is not driven by a severe dichotomy between slumdwellers and others. The notion of a single cleavage between slum residents andwell-to-do burghers as the driver of urban inequality in the developing worldappears to be unsubstantiated – at least in this case. Perhaps more important thanthe highly visible inequalities within our large cities are the less obvious differ-ences between them and smaller urban settlements. In countries like those studiedhere (with the possible exception of Kenya), poverty is greater and deeper insmaller towns, in both income and (at least some) non-income dimensions.
While the findings reported in this paper are suggestive, we have alsopointed to a number of caveats that merit further attention in order to gaugethe robustness of the poverty - city size gradient, as well as its applicability inother developing countries. To begin with, while our household samples arerepresentative of the populations within countries, our sample of countries isnot statistically representative of the developing world as a whole.
In a number of cases, we have been able to perform tentative robustnesschecks for a subset of these caveats. First, we have indicated that the criteriaused to define urban areas vary across countries. It is not uncommon toobserve large numbers of small towns designated as urban on the basis of someadministrative criterion rather than one linked to population characteristics(e.g. population size, density or economic diversification). We have found, inthe case of Brazil, that the inverse relationship between city size and poverty isrobust to an alternative, non-administrative definition of urban areas. But it isimportant to probe more generally whether the gradient we observe in othercountries might be a construct of the way cities are defined.
Second, differences in cost of living between cities of different sizes can besignificant. Failure to correct for such differences could generate a misleadingsense of higher living standards in larger cities. Again in Brazil, we have foundthat adjusting for cost of living variation attenuates, but does not overturn, ourobserved city size-poverty gradient.
Third, we have noted a contrasting literature suggesting that health outcomesmight be worse in large cities than small towns. We have explored this conjecture byexamining child nutritional outcomes across Mexican towns and cities, and havefound that in this context a gradient for city size-child health mirrors, rather thanoffsets, the city size-poverty gradient uncovered in this paper. The general
Ferre, Ferreira and Lanjouw 379
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
applicability of this finding also merits further attention.29 Finally, we acknowledgethat our findings in this paper are based on small area estimates of poverty thatcombine survey-based prediction models with population census data. It isimportant to probe the findings of a city size-poverty gradient also with directly mea-sured poverty data for different cities. We have illustrated that, in the case of India,direct measures and small area estimates yield the same conclusion. It is extremelydifficult to find survey data that is representative at sufficient levels of disaggregationto measure poverty directly in small towns – which is why we have used povertymapping data. Nevertheless, to the extent that additional comparisons with directsurvey-based measures are possible elsewhere, they are definitely of interest. A fullresearch agenda also presents itself with respect to uncovering possible reasons forthe city size-poverty gradient observed in our sample of countries. We have, some-what provocatively, postulated that higher poverty rates in small towns might be theconsequence of a “metropolitan bias” among policy makers, resulting in a dispro-portionate allocation of resources to large cities. Such a bias would be consistentwith the evidence we have presented of differential per-capita availability of a varietyof infrastructure and public services across the city-size spectrum. However, there aremany other plausible explanations, including the possibility that the cost of infra-structure provision may be lower in larger towns and cities.
More generally, we have noted in Figures 1 and 2, the considerable heterogen-eity in poverty outcomes amongst towns and cities within all of the size-categories we have considered. Thus, even amongst small towns there are somewith low poverty and others with high poverty rates. Important questions ariseas to what explains this variation. The “new” economic geography literatureand also the recent World Development Report, entitled “Reshaping EconomicGeography” devote considerable attention to the mechanisms through whichconcentration of population and economic activities can generate various kindsof externalities (for example, Krugman 1999, Henderson, Shalizi and Venables,2001, World Bank, 2009). It is possible that the inverse association betweenurban poverty and city size reflects primarily urban agglomeration effects.
An additional possibility is that the location of towns matters: small townslocated near major urban centers may experience low poverty rates while thosein remote areas are poorer. There may also be a tendency for towns locatednear major metropolitan areas to tend to be larger in size.30 Although beyond
29. A potentially interesting exercise is to directly estimate infant mortality rates from census data
and to examine differentials of this welfare outcome across the urban city-size spectrum. We leave this
exercise to future work.
30. An examination of these questions in the Indian state of West Bengal reveals that when towns
are split into three groups – within a 100-kilometer radius of Kolkata, in a radius of 100-200
kilometers, and more than 200 kilometers – poverty sharply rises with distance from Kolkata when the
towns are within a 100 kilometer radius (World Bank 2011). The relationship is weaker in the second
group, and completely absent in the third, most distant, group. Thus in the first group the
agglomeration effect that really matters is the one generated by Kolkata. However, in the second group,
and even more strongly so in the third group, a separate agglomeration effect (proxied by the town’s
own size) is discernible.
380 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
the scope of this paper, it is clear that a great deal of additional work is neededto better understand how the empirical patterns we have uncovered comeabout, and how they might best be addressed. What is also clear, though, isthat wherever a clear city size-poverty gradient holds, it should be acknowl-edged and confronted in any strategy for urban poverty reduction.
RE F E R E N C E S
Bates, Robert H. 1981. Markets and States in Tropical Africa. Berkeley, C.A.: University of California
Press.
Chattopadhyay, A., and T.K. Roy. 2005. “Are urban poor doing better than their rural counterpart in
India? A study of fertility, family planning and health.” Demography India 34(2):299–312.
Christiaensen, Luc, Lionel Demery, and Jesper Kuhl. 2011. “The (evolving) role of agriculture in
poverty reduction – An empirical perspective”, Journal of Development Economics 96: 239–254.
Demombynes, Gabriel, and Berk Ozler. 2005. “Crime and Local Inequality in South Africa.” Journal of
Development Economics 76 (2):265–92.
Demombynes, Gabriel, Chris Elbers, Jean O. Lanjouw, and Peter Lanjouw (2007) “How Good a Map?
Putting Small Area Estimation to the Test” Policy Research Working Paper No. 4155, The World
Bank.
Department of Census and Statistics, Sri Lanka. 2005. A Poverty Map for Sri Lanka: Findings and
Lessons (Colombo: Department of Census and Statistics).
Elbers, Chris, Jean O. Lanjouw, and Peter Lanjouw. 2002. “Micro-Level Estimation of Welfare”, Policy
Research Working Paper No. 2911, The World Bank.
———. 2003. “Micro-level Estimation of Poverty and Inequality”, Econometrica, 71 (1), pp. 355–364.
Elbers, Chris, Peter Lanjouw, Johan Mistiaen, Berk Ozler, and Kenneth Simler. 2004. “On the Unequal
Inequality of Poor Communities.” World Bank Economic Review 18(3):401–21.
Elbers, Chris, Peter Lanjouw, and Phillippe Leite. 2008. ‘Brazil within Brazil: Testing the Poverty
Mapping Methodology in Minas Gerais’ World Bank Policy Research Working Paper No. 4513,
The World Bank.
Ferreira, Francisco, H.G., Peter Lanjouw, and Marcelo Neri. 2003. “A Robust Poverty Profile for Brazil
Using Multiple Data Sources”, Revista Brasileira de Economia, 57 (1), pp. 59–92.
Foster, James, Joel Greer, and Erik Thorbecke. 1984. “A Class of Decomposable Poverty Measures”,
Econometrica 52 (3): 761–766.
Fujii, Tomoko. Forthcoming. “Micro-level Estimation of Child Undernutrition Indicators in
Cambodia”, World Bank Economic Review.
Gangopadhyay, S., P. Lanjouw, T. Vishwanath, and N. Yoshida 2010. Identifying Pockets of Poverty:
Insights from Poverty Mapping Experiments in Andhra Pradesh, Orissa and West Bengal, forthcom-
ing Indian Journal of Human Development.
Henderseon, J.V., Z Shalii, and A.J. Venables. 2001. “Geography and Development” Journal of
Economic Geography 1:81–105
Haut Commissariat au Plan, Maroc. 2005. Pauvrete, Developpement Humain et Developpement Social
au Maroc: Donnees Cartographiques et Statistiques d’apres le Recensement General de la Population
et de l’Habitat (Rabat: Haut Commissariat au Plan).
Healy, A., and S. Jitsuchon 2007. ‘Finding the Poor in Thailand’ Journal of Asian Economies 18(5),
739–759.
Instituto Brasileiro de Geografia e Estatıstica, IBGE. 2003. ‘Mapa de Pobreza e Desigualdade
Municıpios Brasileiros’ DVD ROM available from IBGE at: http://www.ibge.gov.br/english/
presidencia/noticias/noticia_visualiza.php?id_noticia=1293&id_pagina=1
Ferre, Ferreira and Lanjouw 381
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
INSTAT (2004). Poverty and Inequality Mapping in Albania. (Tirana, INSTAT).
Kapadia-Kundu, N., and T. Kanitkar. 2002. “Primary Healthcare in Urban Slums”. Economic and
Political Weekly, December 21.
Kenya Central Bureau of Statistics. 2003. Geographic Dimensions of Well-Being in Kenya (Regal Press
Kenya, Nairobi).
Krugman, Paul. 1999. ‘The Role of Geography in Development’ International Regional Science Review,
22(2):142–61.
Lanjouw, Peter, and Jean O. Lanjouw. 2001. ‘The Rural Non-Farm Sector: Issues and Evidence from
Developing Countries’, Agricultural Economics 26 (2001) 1–23.
Lanjouw, Peter, and Rinku Murgai. 2009. ‘Agricultural Wage Labour, Non-Farm Employment and
Poverty in Rural India’, Agricultural Economics 40 (2009) 243–263.
Lanjouw, Peter, and Erika Rascon. 2010. ‘Town Size and Malnutrition in Mexico’, mimeo, DECRG,
the World Bank.
Lipton, Michael. 1977. Why Poor People Stay Poor: Urban Bias and World Development. London:
Temple Smith.
Lopez-Calva, L., A. Melendez, E. Rascon, L. Rodrıguez-Chammusy, and M. Szekely (2005) ‘Poniendo
al Ingreso de los Hogares en el Mapa de Mexico’ Working Paper EGAP-2005–04, Tecnologico de
Monterrey, Campus Ciudad de Mexico.
Luttmer, Erzo F.P. 2005. “Neighbors As Negatives: Relative Earnings And Well-Being,” Quarterly
Journal of Economics, V120 (3,Aug), 963–1002.
Molina, I., and J.N.K. Rao (2010) ‘Small Areas Estimation of Poverty Indicators’ The Canadian Journal
of Statistics 38 (3): 369–385
Prebisch, Raul. 1950. “The Economic Development of Latin America and its Principal Problems”,
CEPAL.
Ravallion, Martin, Shaohua Chen, and Prem Sangraula. 2007. “New Evidence on the Urbanization of
Global Poverty”. Population and Development Review 33(4), 667–701.
Rascon, Erika. 2010. “Cracking Malnutrition: an Resolved Challenge for Mexico?: Micro-Level
Estimation of Child Malnutrition Indicators”, mimeo, Development Economics Research Group, the
World Bank.
Rosenstein-Rodan, Paul. 1943. “Problems of Industrialization in Eastern and Southeastern Europe”,
Economic Journal, 53, pp. 202–12.
Tarozzi, A., and A. Deaton 2009. ‘Using Census and Survey Data to Estimate Poverty and Inequality for
Small Areas’ Review of Economics and Statistics 91(4) 773–792.
Thomas, T. (2008) ‘Is There Evidence of an Urban-Rural Spillover? An Empirical Analysis Using
Spatially Explicit Data from Three Countries’, mimeo, Development Economics Research Group, the
World Bank.
World Bank (2007) Measuring Poverty Using Household Consumption Brazil Poverty Assessment,
Poverty Reduction and Economic Management Network, The World Bank.
——— (2009) World Development Report: Reshaping Economic Geography (Washington D.C.: World
Bank).
——— (2011) Perspectives on Poverty in India: Stylized Facts from Survey Data, India Poverty
Assessment, Poverty Reduction and Economic Management Network, the World Bank.
382 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Impact of SMS-Based Agricultural Informationon Indian Farmers
Marcel Fafchamps and Bart Minten
This study estimates the benefits that Indian farmers derive from market and weatherinformation delivered to their mobile phones by a commercial service called ReutersMarket Light (RML). We conduct a controlled randomized experiment in 100 villagesof Maharashtra. Treated farmers associate RML information with a number of deci-sions they have made, and we find some evidence that treatment affected spatial arbi-trage and crop grading. But the magnitude of these effects is small. We find nostatistically significant average effect of treatment on the price received by farmers,crop value-added, crop losses resulting from rainstorms, or the likelihood of changingcrop varieties and cultivation practices. Although disappointing, these results are inline with the market take-up rate of the RML service in the study districts, whichshows small numbers of clients in aggregate and a relative stagnation in take-up overthe study period. JEL codes: O13, Q11, Q13.
The purpose of this study is to ascertain whether agricultural information dis-tributed through mobile phones generates economic benefits to farmers. We im-plement a randomized controlled trial of a commercial service entitled ReutersMarket Light (RML) offered by the largest and best-established private pro-vider of agricultural price information in India at the time of the experiment.Operating in Maharashtra and other Indian states, RML distributes price,weather, and crop advisory information through SMS messages. We offered aone-year free subscription to RML to a random sample of farmers to testwhether they obtain higher prices for their agricultural output.
Marcel Fafchamps is a professor at Oxford University, United Kingdom; his email address is marcel.
[email protected]. Bart Minten is a senior research fellow at the International Food Policy
Research Institute; his email address is [email protected]. The authors thank Dilip Mookherjee,
Shawn Cole, Sujata Visaria, three anonymous referees, and the editor for discussion and suggestions.
This research would not have been possible without the full cooperation of Thomson Reuters India. The
authors are particularly grateful to Raj Bhandari, Rantej Singh, Ranjit Pawar, and Amit Mehra for their
constant support, to Alastair Sussock, Sudhansu Behera, Gaurav Puntambekar, and Paresh Kumar for
their research assistance in the field, and to Grahame Dixie for his encouragements, suggestions, and
comments. Funding for this research was provided by the World Bank and Thomson Reuters. A
supplemental appendix to this article is available at http://wber.oxfordjournals.org/.
THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 383–414 doi:10.1093/wber/lhr056Advance Access Publication February 27, 2012# The Author 2012. Published by Oxford University Press on behalf of the International Bankfor Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions,please e-mail: [email protected]
383
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
We also investigate the channels through which the price improvementcomes: better arbitrage across space and time; ability to bargain with traders;and increased awareness about quality premium leading to better agriculturalpractices and postharvest handling. We present simple models of the first twochannels. Both models make testable predictions about farmer behavior in re-sponse to market price information. We test for the presence of these necessarychannels. Given that RML circulates weather and crop advisory information,we also examine whether profits increase thanks to better crop management,reduced losses, or improved quality.
There is a large interest and takeoff of similar private and public programsin India (Mittal, Gandhi, and Tripathi 2010) and elsewhere (Staatz, Kizito,Weber, and Dembele 2011), indicating that it is important to understand theimpact of these interventions.1 Based on this, we expected to find a large andsignificant effect of RML on treated farmers.
These expectations find less support than anticipated in our results. Manytreated farmers state that they use the RML information, and they are lesslikely to sell at the farm-gate and more likely to change the market at whichthey sell their output. These findings are consistent with the idea that treatedfarmers seek arbitrage gains from RML information. We do not, however, findany significant difference in price between beneficiaries of the RML service andnon-beneficiaries. This result is robust to the choice of estimator and method-ology. There is some evidence of heterogeneous effects: Treated farmers whoare young—and presumably less experienced—receive a higher price in someregressions, but the effect is not particularly robust.
Why we do not find a stronger effect of RML on the prices received byfarmers is unclear. Over the study period there were important changes inprices, which are part of a general phenomenon of food price inflation in India(see, for example, World Bank 2010). Although rapidly changing prices shouldin principle make market information more valuable, it is conceivable that themagnitude of the change blunted our capacity to identify a significant price dif-ference, given how variable prices are. Point estimates of the treatment effecton price received are, however, extremely small and sometimes negative. Wealso find no significant evidence of an effect on transaction costs, net price, rev-enues, and value-added.
In other channels by which RML may affect farmers, we find a statisticallysignificant but small increase of the likelihood of grading or sorting the crop.This is especially true for younger farmers, possibly explaining why theyachieve better prices. RML shows prices by grade to farmers and might havehelped to inform them about the benefit from grading. With respect to other
1. In India, apart from RML, other initiatives include the Indian Farmers Fertilisers Cooperative
Limited (IFFCO) Kisan Sanchar Limited (IKSL), a partnership between Bharti Airtel and IFFCO, as well
as the fisher friend program by Qualcomm and Tata Teleservices, in partnership with the MS
Swaminathan Research Foundation. In western Africa, Manobi and Esoko, private ICT providers, have
developed a number of SMS applications to facilitate agricultural marketing there.
384 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
RML information, we find no systematic change in behavior due to weather in-formation and no change on the crop varieties grown or cultivation practices.Farmers who changed variety or cultivation practice do, however, list RML asa significant source of information in their decision. To summarize, RML bene-fits appear minimal for most farmers, even though RML is associated with sig-nificant changes in where farmers sell their crops. There is some evidence thatyoung farmers benefit more from the service, but the effect is not robust.
I . D E S C R I P T I O N O F T H E I N T E R V E N T I O N A N D E X P E R I M E N T A L D E S I G N
The Indian branch of Thomson-Reuters distributes agricultural information tofarmers through a service called Reuters Market Light (RML). Subscribers areprovided with SMS messages in English or the local language of their choice—in total 75 to 100 SMS per month. Subscribers are offered a menu of threeagricultural markets and three crops to choose from. This menu is based onmarket research of farmers’ information needs and a pilot marketing programcarried out in 2006-07. Prior to the experiment, Reuters gathered encouraginganecdotal information from farmers regarding the usefulness of RML.
At the time of the experiment, Reuters had about 25,000 RML subscribersin Maharashtra. The RML content included market information, weather fore-cast, crop advisory tips, and commodity news. At the time of the study theprice for the service was about $1.50 per month. Thomson Reuters is a trans-national corporation specializing in market information services, its core busi-ness. The corporation has a long experience collecting and selling marketinformation and a reputation for accuracy. Although Thomson Reuters did notdisclose how it obtains market price information, we have no reason to suspectthat the provided information is inaccurate. Doubts about data accuracy wouldbe detrimental to the firm’s reputation and would hurt its effort to marketRML across India.
A randomized controlled trial (RCT) was organized by the authors to testthe effect of RML on the price received by farmers. The RCT was conductedin close collaboration with Thomson Reuters. Five crops were selected as thefocus for the study: tomatoes, pomegranates, onions, wheat, and soybeans. InMaharashtra these crops are all grown by smallholders primarily for sale.Whereas wheat and soybeans are storable, the other three crops are not andshould be subject to greater price uncertainty due to short-term fluctuations insupply and demand (Aker and Fafchamps 2011). We expect market informa-tion to be more relevant for perishable crops. Soybeans have long been growncommercially in Maharashtra, but tomatoes, onions, and pomegranates aremore recent commercial crops. We expect farmers to be more knowledgeableabout well-established market crops—such as wheat and soybeans—than aboutcrops whose commercial exploitation is more recent—such as tomatoes,onions, and pomegranates. Finally, pomegranate is a tree crop that is sensitiveto unusual weather and requires pesticide application. We expect the benefit
Fafchamps and Minten 385
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
from weather information and crop advisories to be more beneficial to pom-egranate farmers.
For each of the five crops, we selected one district where the crop is widelygrown by small farmers for sale: Pune for tomatoes, Nashik for pomegranates,Ahmadnagar for onions, Dhule for wheat, and Latur for soya. All five districtsare located in the central region of Maharashtra; we avoided the eastern partof the state where sporadic Maoist activity has been reported, and the westernpart of the state which is less suitable for commercial agriculture.
A total of 100 villages were selected for the study, 20 in each of the 5 dis-tricts. The villages were chosen in consultation with Thomson Reuters toensure they were located in areas not previously targeted by RML marketingcampaigns. Ten farmers were then selected from each village, yielding a totalintended sample size of 1,000 farmers.2
Two treatment regimes were implemented. In the first regime—treatment1—all 10 farmers in the selected village were offered free RML. In the secondregime—treatment 2—three farmers randomly selected among the villagesample were offered RML. The purpose of treatment 2 is to test whether thetreatment of some farmers benefit others as well. Farmers who are not signedup are used to evaluate the externalities generated if farmers share RMLinformation.
Bruhn and McKenzie (2009) show that, in RCTs, stratification improves effi-ciency. Randomization of treatment across villages was thus implemented byconstructing, in each district, triplets of villages that are as similar as possiblealong a number of dimensions that are likely to affect the impact of the treat-ment.3 A description of the process is provided in the supplemental appendix 1,available at http://wber.oxfordjournals.org/.
We also offered RML to randomly selected extension agents covering half ofthe treatment 1 and treatment 2 villages. In principle, extension agents coulddisseminate the relevant information they receive, making it unnecessary to dis-tribute it to individual farmers. Whether they do so in practice is unclear, giventhat extension agents visit villages infrequently.
The object of the RCT is to estimate the impact of RML on farmers whovoluntarily sign up for the service because they benefit from it. In villages tar-geted by RML, only a small proportion of farmers sign up. They tend to belarger farmers with a strong commercial orientation in RML crops. There is no
2. This sample size was determined as follows. The primary channel through which we expect SMS
information to affect welfare is the price received by producers. We therefore want a sample size large
enough to test whether SMS information raises the price received by farmers. Goyal (2010) presents
results suggesting that price information raises the price received by Indian farmers by 1.6 percent on
average. Based on this estimate and its standard error, a simple power calculation indicates that a total
sample size of 500 farmers should be sufficient to identify a 1.6 percent effect at a 5 percent significance
level. To protect against loss of power due to clustering, we double the sample size to 1,000. We did
not have sufficient information to do a proper correction of our power calculations for clustering.
3. More precisely, 6 triplets and one pair.
386 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
point estimating the effect of RML on people unlikely to benefit from it. Forthis reason, participants are limited to farmers growing the district-specificselected crop for sale. These farmers have a large enough marketed surplus toamortize the cost of information gathering and enough experience with thecrop to benefit from agricultural information. It remains that the farmers towhom we offered RML did not express an initial interest in it and may havedismissed the usefulness of something they did not pay for.
Since a mobile phone is required to receive RML messages, we limit partici-pants to farmers with a mobile phone at the time of the baseline survey. Weomit farmers who already were RML customers prior to baseline because theycould not be used for impact analysis. Because we carefully avoided areas inwhich RML had already been actively promoted, few farmers were eliminateddue to this condition. Further details on the sample and survey design arefound in supplemental appendix 2.
I I . C O N C E P T U A L F R A M E W O R K
Our primary objective is to test whether farmers benefit from the SMS-basedmarket information and, if so, how. Information gathered by Thomson Reutersand conversations with farmers and RML customers in villages of Maharashtrasuggest that farmers benefit in several ways. According to this information,timely access to market price information at harvest helps farmers decidewhere to sell, as in Jensen (2007). It also enables them to negotiate a betterprice with traders.
To illustrate how informed farmers may obtain a higher price through betterarbitrage, consider a farmer selecting where to sell his output. There are twopossible markets, i and j, with transport costs ti and tj. We assume tj . ti, thatis, market j is the more distant market. To focus on arbitrage, we assume thatthe distribution of producer prices F(p) is the same in both markets. In particu-lar, E[pi] ¼ E[pj] ¼ m. Given this, it is optimal for an uninformed farmer toalways ship his output to market i since EU (pi - tj) , EU (pj – ti) for anyutility function U(.).4 The average price received by farmers is thus m. A rele-vant special case is when i is selling at the farm-gate to an itinerant buyer and jis selling at the nearest market. In that case the farmer incurs no transactioncost, receives price pi, and does not learn the market price pj.
Now suppose that the farmer is given information on prices in i and j.Shipping to i remains optimal if pi – ti � pj – tj; otherwise, the farmer ships toj. The average farmer price now is:
E pi pi � ti � pj � tj
��� �Pr pi � ti � pj � tj
� �þ E pj pi � ti , pj � tj
��� �Pr pi � ti , pj � tj
� �� m
ð1Þ
4. Since p — ti . p — tj point wise. There is no role for risk aversion in this model.
Fafchamps and Minten 387
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
This equation holds with equality only if pi is always larger than pj — (tj —ti). This arises only if tj – ti is large relative to the variance of prices or ifprices in the two markets are strongly correlated. It follows that if price infor-mation allows farmers to arbitrage better across markets, the average farmerprice should rise and we should observe farmers now selling in different, moredistant markets.5 In the case of farm-gate sales, obtaining information about pj
induces farmers to sell at the market if pj . pi þ tj where pi is the price offeredby itinerant buyers.
There remains the issue of why pi = pj in the first place: If farmers can arbi-trage, so can traders. Let u be the trader shipment cost between i and j. Withperfect information, trader arbitrage yields:
pj þ u � pi � pj � uð2Þ
and thus j pi — pj j � u. Farmer arbitrage therefore arises whenever tj — ti ,
u, that is, when farmers have access to cheap transport to markets.6 However,if traders have a comparative advantage in transport—for instance, becausethey ship larger quantities and benefit from returns to scale—then it is possiblethat tj — ti . u for most if not all farmers. In this case, Pr(pi - ti , pj – tj) ¼0, which implies that farmers always sell at the closest market or location i,and the average farmer price is m, as in the case without information.
Even when price information does not trigger farmer arbitrage, it may facili-tate arbitrage by traders, thereby ensuring that jpi — pjj � u (Aker andFafchamps 2011). If farmers are risk averse, they would benefit from the reduc-tion in the variance of prices,7 irrespective of whether they receive market in-formation pi and pj or not prior to deciding where to ship their output.
The second way farmers can benefit from price information is when they sellto traders who are better informed about market prices—for example, whenselling at the farm-gate. To illustrate, consider price negotiation between aninformed trader who knows the market price realization pi, and an uninformedfarmer who only knows the price distribution F(pi). To demonstrate how infor-mation can benefit the farmer, imagine that the farmer mimics an auctionsystem and calls a decreasing sequence of selling prices until the trader acceptsit. In a competitive market with many buyers—which makes collusiondifficult—the selling price will be pi. In a one-on-one negotiation, as would
5. A similar reasoning applies to intertemporal arbitrage: Uninformed farmers may prefer to sell
immediately after harvest, whereas better informed farmers may choose to sell at a later date if the
anticipated price is higher. Since RML does not disseminate information about future prices, however,
we do not expect an intertemporal arbitrage effect, except perhaps in the immediate vicinity of harvest.
Several of the studied crops are perishable; this further limits opportunities for intertemporal arbitrage.
6. For instance, if transport costs per kilometer are identical for farmers and traders, condition tj —
ti , u holds generically on a plane, except when the farmer and the two markets are exactly in a
straight line. With many farmers distributed randomly on the plane, this has Lebesgue measure zero.
7. Except when they consume much of their output, something that is ruled out here since the
empirical analysis focuses on commercial crops
388 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
take place at the farm-gate, the buyer correctly anticipates that the farmer willcontinue calling lower prices below pi. He can thus wait for the farmer toreach his reservation price, which is the value of the farmer’s next best alterna-tive, namely, selling at the nearest market.
The expected payoff to an uninformed farmer of selling at the market isEU(pi — ti). Let ~pl ; pi � ti be the market price net of transport cost, andlet ~m ; E½~pi�. The farm-gate reservation price of a risk-averse farmer is theprice pr
i ¼ ~m� p that solves:
U ~m� pð Þ ¼ EU ~plð Þð3Þ
Using a standard Arrow-Pratt Taylor expansion, we get:
U ~mð Þ � U0 ~mð Þp � E U ~mð Þ þ U0 ~mð Þ ~m� ~plð Þ þ 1
2U00 ~mð Þ ~m� ~plð Þ2
� �ð4Þ
which we can solve for p:
p � � 1
2
U00
~mð ÞU0 ~mð Þ s
2 ¼ 1
2R CV2ð5Þ
where R is the farmer’s coefficient of relative risk aversion, s2 is the varianceof the market price, and CV ; s=~m is the coefficient of variation of price. Itfollows that a buyer can always buy from an uninformed farmer at pricepr
i ¼ ~m� p. Only if the realized market price pi , pri is the farmer unable to
find farm-gate buyers—in which case he must travel to the market and sell atpi , ~mþ ti � p but incur transport cost ti. The average price received by anuninformed farmer is:
m� ti � pð ÞPr pi � m� ti � pð Þ þ E pi pi , m� ti � pj½ �Pr pi , m� ti � pð Þ � m� tið6Þ
It follows that the larger R—and thus p—is, the lower the average farmer priceis. If risk aversion is negatively correlated with wealth, the above predicts thatpoor uninformed farmers receive a lower average price than non-poor unin-formed farmers. Similarly, the larger CV is—for instance, because the farmer isinexperienced and unsure about the price distribution—the lower the averagefarm-gate price is.
Once we introduce price information, the farmer’s farm-gate reservationprice becomes pi - ti and buyers are no longer able to exploit farmers’ risk aver-sion to buy below the market price. The expected price received by farmers ism — ti if they sell at the farm-gate, or m if they sell in the market. Hence, theaverage price received by informed farmers is unambiguously higher than thatof uninformed farmers. The difference is largest when uninformed farmers
Fafchamps and Minten 389
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
often sell at the farm-gate. If farmers do not sell at the farm-gate at all, infor-mation has no effect on the average price farmers receive.8
The two models above do not exhaust the possible channels by which priceinformation may affect farmer prices. In India one important possibility is ex-cessive fees collected by commission agents (Minten, Vandeplas, and Swinnen2012). However, since RML circulates no information about market fees, it isunclear why it should lead to their reduction.
Model predictions regarding prices can be summarized as follows. If priceinformation enables farmers to arbitrage across markets, treated farmersshould receive a higher price than control farmers, but only if treated farmersstart selling in distant markets. Otherwise, we expect no difference betweencontrol and treated farmers. The introduction of RML, however, could reducethe variance of prices for everyone through trader arbitrage, as in Aker andFafchamps (2011).
If price information helps farmers negotiate better prices with traders,treated farmers should receive a higher average price only if they were selling atthe farm-gate prior to treatment. For these farmers, we expect a stronger treat-ment effect for poor and inexperienced farmers. For farmers who were sellingprimarily if not exclusively through wholesale markets, we expect no effect ofthe treatment on price. But treatment may nevertheless induce farmers to sell atthe farm-gate for convenience reasons, or if traders have a comparative advan-tage in transporting produce from the farm-gate.
RML may benefit farmers in other ways, which we do not model since theyare more straightforward. Better knowledge of quality-driven price differentialsmay induce farmers to upgrade output quality, for instance by grading or treat-ing their crops. Weather information helps with farm operations. In particular,information about the probability of rainfall enables farmers to either delay(pesticide application) or speed up (harvest) certain farm operations.Information about air moisture is a good predictor of pest infestation andhence of the need to apply pesticide. Crop advisories assist farmers to choose amore appropriate technology (choice of variety, pesticide, and fertilizer).
I I I . T E S T I N G S T R A T E G Y
We now describe how we test the above predictions. Since the data arebalanced, we ascertain the effect of RML on outcome indicators by comparingcontrol and treatment in the ex post survey. Formally, let Yi represent an
8. Uninformed farmers would benefit if they could commit to sell at the market. If such a
commitment mechanism is unavailable, however, farmers can always be tempted to sell at the farm-gate
if offered a price above their reservation price. Of course, a sophisticated but uninformed farmer should
infer that if a trader is willing to buy from him at the farm-gate, the market price must be above his
reservation price, in which case he should sell at the market. If farmers are sophisticated, we should
therefore observe few if any farm-gate sales by uninformed farmers. In this case, providing market
information to farmers should make farm-gate sales more common.
390 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
outcome indicator—for example, price received—for farmer i. Let Wi ¼ 1 iffarmer i was offered a free subscription to the SMS-based market service, andSi ¼ 1 if farmer i signed up for the service. All treated farmers are in treated vil-lages but the converse is not true: only some farmers in treatment 2 villageswere offered RML. It is possible for treated farmers not to sign up for theservice—that is, for Si ¼ 0 even though Wi ¼ 1. Although control villageswere not targeted by RML marketing campaigns, it is also possible for non-targeted farmers to independently sign up—that is, for Si ¼ 1 even thoughWi ¼ 0.
We are interested in estimating the direct effect of RML on customers, thatis, those with Si ¼ 1. Since Si is subject to self-selection and Wi is not, we beginby reporting intent-to-treat estimates that compare control farmers to thosewho were offered the free subscription. The estimating equation is:
Yi ¼ uþ bWi þ eið7Þ
Next we investigate the effect of receiving the RML subscription. As we willsee, the likelihood of signing up is much higher among farmers who receivedthe offer of a free subscription. This means that Wi satisfies the inclusion re-striction and can be used as instrument for Si. We thus estimate an instrumentalvariable (IV) model of the form:
Yi ¼ uþ aSi þ eið8Þ
where ei is an error term possibly correlated with Si (self-selection effect) butuncorrelated, by design, with the instrument Wi.
Provided that there are no defiers, we can interpret IV estimates from equa-tion (8) as local average treatment effects (LATE). Assuming no defiers meansthat farmers who did not sign up for RML even though they were offered a freesubscription would not have signed up for it if they had not been offered a freeservice. In our setup, this assumption is unproblematic. We can therefore inter-pret a in equation (8) as the effect of RML for a farmer who would be inducedto sign up if offered the service for free. This is the IV-LATE approach.
Equation (7) can be generalized to investigate heterogeneous effects. Let Xi
be a vector of characteristics of farmer i thought to influence the effect of thetreatment. We expect larger RML benefits for commercial farmers who empha-sise crops for which RML information is useful. We also expect less experi-enced farmers to more benefit. The estimated model becomes:
Yit ¼ u þ aWi þ gXi þ hWiðXi � �XÞ þ eið9Þ
where �X denotes the sample mean of Xi. The average treatment effect is givenby a, whereas the heterogeneous effects of treatment on a farmer with charac-teristic Xi is a þ hðXi � �XÞ.
Fafchamps and Minten 391
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
When interpreting models (7), (8), and (9), one must remember that identify-ing the value of information is difficult because the value of informationchanges with circumstances. In particular, information is useful only when itcan be acted upon. Up-to-date price information is most useful around harvesttime. Crop advisory and input cost information are most useful at plantingtime. For information to be useful it must be provided in a timely manner.How valuable information is depends on the context: because information isnot useful in one year does not imply that it is never useful.
Second, information circulates through channels other than RML — farmersvisit markets and talk to each other and to commission agents. For models (7),(8), and (9) to identify the impact of RML information, the circulation of in-formation among farmers must not be so rapid and widespread that controlfarmers benefit from it as well. For this reason we regard the village as themost appropriate treatment unit, because information exchange is more likelyamong neighbors. We cannot, however, entirely rule out the possibility of spil-lovers across villages.
Third, price information may benefit farmers by improving their bargainingpower with traders and commission agents. Since the latter cannot easily distin-guish between RML and non-RML farmers, it is possible that they adapt theirbehavior toward all farmers, for instance, by making better price offers. If thisis the case, control farmers may benefit as much as treated farmers from theRML service. There is little we can do to protect against this form of contamin-ation, except check informally how agricultural wholesale prices change overtime as farmers become better informed.9
I V. T H E C O N T E X T A N D D A T A
Take-up of RML by Maharashtra farmers is a revealed preference measure ofthe benefits from the service. We report in Table 1 the number of agriculturalholdings in each study districts (2000/01 agricultural census) and the numberof RML subscribers over the study period. RML take-up has varied over time.Take-up increased rapidly in all five districts between 2007, the time at whichRML was introduced, and 2009, the time at which our experiment started.Take-up never exceeded 0.5 percent of the total farmer population, however.The table also shows that subscription levels have stabilized in recent years andhave even come down in some districts in 2010. It is only in Nashik that wesee a large increase in the number of subscribers between 2009 and 2010. Thismay be explained by Nashik having a nascent grape-growing and wine-making
9. It is also conceivable (albeit unlikely) that RML clients indirectly create a negative externality for
nonclients, for instance, because the selling behavior of RML clients indirectly lowers the price received
by nonclients, or because it raises the price for local consumers. If this were the case, we would
overestimate the effect of RML by comparing RML and non-RML farmers within the same village.
This is why we focus our analysis on comparisons across treatment and control villages.
392 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
industry that has been rapidly growing in recent years. Since grapes are grownprimarily by large farmers, they are not included in our study.
Next we report on contamination and noncompliance. Extensive contamin-ation could indicate that many farmers find RML beneficial and sought it outeven though it was not marketed locally. In contrast, extensive noncompliancecould suggest that treated farmers did not find the service useful. In Table 2 wecompare the experimental design, or intent to treat, in the two upper panels toactual RML usage in the lower panel. The uppermost panel describes the ori-ginal experimental design. This design assumes that 10 farmers would befound in each of the 100 villages selected for the study.
The middle panel of Table 2 describes how the experiment was implementedin practice. This represents what in the rest of the paper we call intent to treat.All farmers in treatment 1 villages were offered RML free of charge for oneyear. In control villages, no farmer was offered RML and no marketing ofRML was done by Thompson-Reuters. In treatment 2 villages, a randomlyselected subset of farmers (3 out of 10) were offered RML and the others werenot.
There was some attrition between the baseline and follow-up surveys: of the1,000 farmers interviewed in the baseline, 933 were revisited in the follow-upsurvey. There is some difference in attrition between the control and treatmentgroups, that is, 91 percent versus 96 percent versus 93 percent. To investigatewhether there is anything systematic about attrition, we regressed an attritiondummy on household characteristics.10 We find that onion producers(Ahmadnagar district) are more likely to drop out of the experiment, but none ofthe other variables is statistically significant. Triplet dummies are included asregressors throughout the analysis; they indirectly control for district/target crop.
TA B L E 1. Number of agricultural holdings and RML subscribers in the fivedistricts studied in Maharashtra
District:Crop followed
in surveyNumber of agricultural
holdings*
Number of RML subscribers**
2007 2008 2009 2010
Ahmadnagar onion 916,724 711 1,377 3,763 1,637Dhule wheat 230,216 108 1,296 1,028 840Latur soya 305,706 163 914 1,048 826Nashik pomegranate 591,763 2,176 1,561 3,934 6,514Pune tomato 667,365 392 653 3,495 781Total 2,711,774 3,550 5,801 13,268 10,598
*: Government of India, Agricultural Census, 2000/01.
**: Thompson-Reuters.
10. That is, household size, age of household head, education of household head, land owned, total
land cultivated of the selected crop in 2009, and target crop/district dummies.
Fafchamps and Minten 393
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
In the next panel we report actual RML usage, as depicted by the 2010survey and by the ex post phone interview. We note a significant proportion ofnoncompliers: only 59 percent of those farmers who were offered RML actual-ly used it. Non-usage has various proximate causes. Some subscribers simplyrefused the service. In the ex post phone interview, respondents were asked thereason for refusal. Some indicated that they believed they would be charged forservice later on; others were illiterate households who could not read SMS mes-sages and thus could not use the service anyway. Another reason for non-usagewas that subscribers never activated the RML service. To activate it, the sub-scriber had to select three crops and markets; some subscribers never com-pleted the activation sequence. Non-usage was also partly due to changes inphone number or to migration—for example, a household member leaving thefarm and taking the phone number with them. The RML service is tied to aspecific phone number, so if this phone number is no longer used by the house-hold, the service no longer reaches its intended target. Finally, a number of
TA B L E 2. Compliance and contamination
All villages Treatment 1 Treatment 2 Control
Numberof villages
RML RML RML RML
yes no yes no yes no yes no
Intended experimental design Number of householdsAll 100 455 545 350 0 105 245 0 300Tomato growers 20 91 109 70 0 21 49 0 60Pomegranate growers 20 91 109 70 0 21 49 0 60Onion growers 20 91 109 70 0 21 49 0 60Wheat growers 20 91 109 70 0 21 49 0 60Soya growers 20 91 109 70 0 21 49 0 60Realized design or
intent to treat
All 100 422 511 325 0 97 239 0 272Tomato growers 20 84 107 64 0 20 49 0 58Pomegranate growers 20 89 107 68 0 21 52 0 55Onion growers 20 88 105 68 0 20 49 0 56Wheat growers 20 86 102 67 0 19 47 0 55Soya growers 20 75 90 58 0 17 42 0 48RML usage (from 2010 survey and phone interview)
All households 100 247 686 181 144 56 280 10 262Tomato growers 20 44 147 35 29 9 60 0 58Pomegranate growers 20 65 131 42 26 19 54 4 51Onion growers 20 44 149 36 32 8 61 0 56Wheat growers 20 48 140 33 34 11 55 4 51Soya growers 20 46 119 35 23 9 50 2 46Extension agents:
Intended design 30 70 15 20 15 20 0 30Realized design 20 80 10 25 10 25 0 30
394 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Chinese-made phones could not display the Marathi script and householdswith such phones could not read the RML messages. Although there is vari-ation between them, all these proximate causes indicate a certain lack of inter-est in the service: if RML had been valuable, recipients would have made moreeffort to secure it—for example, by keeping the SIM card and getting anotherphone.
There is variation in noncompliance across districts: Noncompliance islowest in Nashik among pomegranate farmers (27 percent). This finding is con-sistent with the high take-up reported in Table 1, and indicates more interest inRML in that district. In contrast, the proportion of noncompliers is close tohalf among onion, tomato, and wheat growers. While noncompliance is high,contamination is low everywhere: only 3.7 percent of control farmers—10 outof 272 farmers—signed up for RML. This confirms that interest in the serviceamong study farmers is limited.
At the bottom of Table 2 we report variation between the intended experimen-tal design and the realized treatment for extension agents. The intent was to offerone year of RML service free of charge to the extension agents serving a random-ly selected subsample of 30 of the 70 treated villages. In practice, we onlymanaged to locate and offer RML to extension agents serving 20 of the treatedvillages. In order not to introduce contamination, RML was not offered to exten-sion agents serving control villages. This means that we can only measure theadditional effect that an informed extension agent may have over and above anindividual RML contract (treatment 1 villages) or in addition to treatment ofother farmers in the same village (control farmers in treatment 2 villages).
In Table 3 we compare control and intent-to-treat farmers in terms ofbalance. Columns 4 and 5 report the mean value of each variable for thecontrol group and their standard deviation, respectively. Columns 6 and 7report the coefficient of an intent-to-treat dummy in a regression of each vari-able on triplet fixed effects.11 Reported coefficients suggest good balance on allvariables, including area planted to the target crop, marketing, transactioncosts, past weather, and past technological innovation. We follow Deaton’s(2009) suggestion not to include unnecessary control variables in the analysisof randomized controlled trials (RCTs), as it may artificially inflate t-values.
V. E M P I R I C A L A N A L Y S I S
We now turn to the econometric analysis. Unless otherwise stated, all analysisis conducted in terms of intent to treat, that is, the treated are those who wereoffered a free one-year subscription to the RML, whether or not they used it.We also report local average treatment effect (LATE) results in which we in-strument actual RML usage with random assignment to treatment. We refer to
11. Bruhn and McKenzie (2009) indicate that fixed effects for each stratification cell should be
included in all regressions.
Fafchamps and Minten 395
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA B L E 3. Balancedness of treatment versus control in the 2009 baseline data
Unit
Numberof
observation
Control group Treatment*
Mean St. Dev. Coeff. t-value
Household characteristics
Education level head of household years 911 8.19 4.36 0.243 0.68Household size number 933 6.43 2.70 0.131 0.61Share of children in household share 933 0.26 - 20.008 20.72Share of elderly in household share 933 0.08 - 0.011 1.38Age head of household years 922 49.51 12.93 0.409 0.42Farm experience years 930 26.86 13.92 20.379 20.39Land ownership and cultivation
Land owned acres 933 9.62 8.62 0.677 0.64Land cultivated of tomato in Pune acres 191 1.79 3.01 20.154 20.40Land cultivated of pomegranate in Nashik acres 196 3.64 3.50 0.508 0.62Land cultivated of onions in Ahmadnagar acres 193 2.06 1.87 0.258 0.65Land cultivated of wheat in Dhule acres 188 5.76 3.95 21.073 21.65Land cultivated of soya in Latur acres 165 5.76 4.09 0.841 0.39Total crop area cultivated acres 933 14.78 13.04 1.340 0.88Marketing characteristics studied crop
Know market price of studied crop:2 the day before he sold it share 910 0.78 - 0.037 1.132 the week before he sold it share 912 0.38 - 0.070 1.582 a month before he sold it share 912 0.08 - 0.029 1.332 when he planted it share 912 0.06 - 0.020 1.67For each transaction:2 Prices obtained in each transaction Rs/kg 1563 13.22 10.20 20.149 20.442 Quantities sold per transaction log(kgs) 1563 7.11 1.57 20.067 20.752 Produce is sold in the village share 1554 0.15 - 20.016 20.572 Head of household made sale share 1561 0.85 - 20.021 20.582 Crop was graded/sorted before sale share 1561 0.70 - 0.036 0.842 Produce is sold through commission
agentshare 1555 0.40 - 0.032 0.83
Number of sale transactions per farmer number 894 1.74 1.19 20.001 20.01Transaction costs last transaction
Paid for transport of produce share 908 0.88 - 0.013 0.45Paid for personal transport share 797 0.11 - 0.022 0.90Sold through commission agent share 905 0.57 - 20.036 20.80Weather in 12 months prior to survey
Did not incur storm/heavy rainfall share 933 0.53 - 20.021 20.63Technology changes in 12 months prior to survey
Changed crop varieties share 933 0.34 - 20.020 20.57Changed cultivation practices share 933 0.28 - 20.004 20.11
All variables refer to 2009 data.
*village triplet code dummies and intercept included but not reported.
396 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
these results as IV or LATE estimates interchangeably. For much of the analysiswe use both treatment 1 and treatment 2 farmers to improve efficiency. Whenusing treatment 2 farmers, the intent-to-treat variable is set to 1 if a surveyedfarmer in a treatment 2 village was randomly assigned to treatment, and 0otherwise. All reported standard errors are clustered by village triplet (see ex-perimental design section).
RML Usage
We begin with RML usage as reported by farmers. In the baseline survey allrespondents were asked to list their main sources of information for agricultur-al prices, weather forecast, and advice on agricultural practices. Answers aretabulated in Table A.1 in the supplemental appendix. Own observation/experi-mentation is the main source of information reported by all respondents, fol-lowed by conversations with other farmers. Radio and television arementioned as a common source of information on the weather, less so for cropprices. RML is not mentioned by anyone.
In the top panel of Table 4 we report the average difference in the propor-tion of respondents who mention RML as a source of information in the expost survey. The average treatment effect on the treated (ATT) is calculatedusing the nearest neighbor-matching methodology described in Abadie at al.(2004), where matching is performed by triplet dummy. Reassuringly, treatedfarmers are significantly more likely to mention RML in all six categories. Thedifference is largest in magnitude for prices and weather, which are theprimary focus of the RML service: 24 percent and 23 percent more treatedrespondents mention RML as a source of information on crop prices andweather forecasts, respectively.
LATE-IV estimates are reported in the next panel of Table 4. These estimatesare obtained using regression analysis. Dummies are included for village selec-tion triplets. Since contamination is low (3.7 percent) but noncompliance ishigh (41 percent), we expected instrumented treatment effects to be larger thanthe intent-to-treat effect reported at the top of the table. This is indeed whathappens: We now find that farmers who were induced to use RML as a resultof treatment are 46 percent more likely to mention RML as a source of infor-mation on crop prices. The corresponding figures for weather prediction andfor input use are 44 percent and 39 percent, respectively. This suggests thatRML is seen as a source of information by a large proportion of participatingfarmers. Yet, the effect is not 100 percent, which means that, since non-usersdo not list RML, a sizable portion of treated respondents do not list RML as asource of information.
In the second part of the table we look for evidence of heterogeneous effectsby farmer age and farm size. We estimate regression (3) with triplet fixedeffects as suggested by Bruhn and McKenzie (2009). Farmers cultivating alarger area are significantly more likely to mention RML as a source of
Fafchamps and Minten 397
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
information. This effect is limited to treated large farmers for crop prices,weather predictions, and input use. Farmer age is never significant.
Next we examine whether treated farmers appear more knowledgeableabout crop prices. In the first four columns of results in Table 5, we presentATT estimates for knowing the sale price of the target crop before the day ofthe sale. Results show that treated farmers consider themselves more knowl-edgeable about crop prices in general. The difference is significant in all fourcases, that is, one day before sale as well as several months before sale. In thesecond panel of Table 5 we report IV-LATE estimates that, as for Table 4, are
TA B L E 4. Use of RML
Whole sample
Use RML as one of the sources of information for:
Cropprices
Weatherprediction
Crops toplant
Cultivationpractices
input use(d)
post-harvestpractices
Number ofobservations
925 931 925 925 918 924
Nearest neighbor matching (a)ATT Coeff 0.243 0.231 0.106 0.086 0.200 0.054
z-value 10.600 10.530 6.550 5.390 9.360 3.970
Regression results (b)1. IV-LATETreatment Coeff 0.463 0.439 0.206 0.172 0.386 0.113
t-value 10.460 10.530 5.230 4.710 8.940 3.220
Intercept Coeff 0.007 20.021 20.008 20.001 20.044 0.044t-value 0.840 22.530 21.000 20.150 25.080 6.300
2. Heterogeneous effects (c)Treatment Coeff 0.239 0.225 0.107 0.089 0.198 0.057
t-value 8.760 9.330 5.580 4.880 8.140 3.160
Intercept Coeff 20.034 20.470 20.034 20.026 20.066 0.027t-value 21.950 22.770 22.450 21.870 24.320 2.000
Dummy younghead ofhousehold
Coeff 0.024 0.022 0.007 0.005 0.014 20.003t-value 1.270 1.670 0.650 0.450 1.520 20.300
Total crop areacultivated
Coeff 0.001 20.000 0.001 0.001 0.000 0.001t-value 1.200 20.020 1.700 1.890 0.430 1.680
Interaction with treatmentDummy young
head ofhousehold
Coeff 0.045 0.018 0.057 0.019 0.007 0.034t-value 1.040 0.390 1.580 0.530 0.150 1.210
Total crop areacultivated
Coeff 0.002 0.004 20.000 20.001 0.002 20.000t-value 2.340 4.300 20.060 20.700 2.110 20.380
(a) Matching based on village triplet code dummies
(b) Village triplet code dummies included but not reported
(c) Mean value substracted from those control variables interacted with treatment
(d) fertilizers, pesticides, and herbicides
t-values based on standard errors clustered by village triplet code;
t-values in bold significant at the 10% level or better.
398 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E5
.K
now
ledge
and
info
rmat
ion
shari
ng
Know
pri
cebef
ore
sale
:Share
Collec
tpri
cein
fo
Whole
sam
ple
One
day
One
wee
kO
ne
month
atpla
nti
ng
Info
rmat
ion
farm
ing
No
of
peo
ple
consu
lted
Collec
tpri
ceat
pla
nti
ng
Num
ber
of
obse
rvat
ions
722
723
722
722
922
929
925
Nea
rest
nei
ghbor
mat
chin
g(a
)A
TT
Coef
f0.0
84
0.0
95
0.0
97
0.0
78
0.0
63
0.0
35
0.0
05
z-valu
e3.1
00
2.8
30
3.0
90
2.5
40
4.0
50
0.5
80
0.1
50
Reg
ress
ion
resu
lts
(b)
1.
IV-L
AT
E
Tre
atm
ent
Coef
f0.1
30
0.1
58
0.1
81
0.1
46
0.1
19
0.0
68
0.0
11
t-valu
e2.1
80
1.9
80
3.2
80
1.9
60
4.0
80
0.5
00
0.1
60
Inte
rcep
tC
oef
f0.7
17
0.3
04
-0.0
34
0.0
10
0.6
76
1.5
20
0.5
64
t-valu
e64.7
80
20.6
40
-3.2
80
0.7
20
116.3
30
55.3
90
39.0
10
2.
Het
eroge
neo
us
effe
cts
(c)
Tre
atm
ent
Coef
f0.0
65
0.0
73
0.0
84
0.0
68
-0.0
63
0.0
14
0.0
07
t-valu
e2.1
10
1.7
80
2.9
30
1.7
80
-4.0
80
0.2
00
0.1
90
Inte
rcep
tC
oef
f0.7
02
0.2
65
-0.1
04
-0.0
61
0.6
65
1.5
93
0.5
61
t-valu
e23.3
30
7.1
40
-3.5
10
-1.5
40
42.3
40
24.5
30
14.0
20
Dum
my
young
hea
dof
hh
Coef
f-0
.002
0.0
30
0.0
38
0.0
35
-0.0
04
-0.2
29
-0.0
41
t-valu
e-1
.220
0.6
90
1.0
90
0.8
90
-0.1
60
-2.4
00
-0.9
70
Tota
lcr
op
are
acu
ltiv
ated
Coef
f0.0
01
0.0
02
0.0
04
0.0
04
0.0
01
0.0
07
0.0
02
t-valu
e1.2
20
1.4
30
2.8
50
1.9
10
0.7
80
3.1
50
1.2
40
Inte
ract
ion
wit
htr
eatm
ent
Dum
my
young
hea
dof
hh
Coef
f-0
.010
-0.0
59
-0.0
48
0.0
13
0.0
37
0.0
21
0.0
89
t-valu
e-0
.220
-1.0
80
-0.8
30
0.2
20
1.1
80
0.1
50
1.5
00
Tota
lcr
op
are
acu
ltiv
ated
Coef
f-0
.001
-0.0
01
-0.0
00
-0.0
03
-0.0
01
-0.0
02
-0.0
02
t-valu
e-0
.960
-0.4
50
-0.4
50
-1.4
20
-0.6
80
-0.5
90
-1.1
30
(a)
Mat
chin
gbase
don
vil
lage
trip
let
code
dum
mie
s.
(b)
Villa
ge
trip
let
code
dum
mie
sin
cluded
but
not
report
ed.
(c)
Mea
nvalu
esu
bst
ract
edfr
om
those
contr
ol
vari
able
sin
tera
cted
wit
htr
eatm
ent.
t-valu
esbase
don
standard
erro
rscl
ust
ered
by
villa
ge
trip
let
code;
t-valu
esin
bold
signifi
cant
atth
e10%
leve
lor
bet
ter.
Fafchamps and Minten 399
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
larger in magnitude than the intent-to-treat ATT. There is no evidence of het-erogeneous effect along those two dimensions.
In the next column of Table 5 we investigate whether treated farmers reportsharing more information about farming with other farmers. If RML informa-tion is valuable, we expect treated farmers to be more likely to share it withothers. Results reported in Table 5 suggest that this is indeed the case, but theeffect is not large in magnitude: the intent-to-treat estimated ATT is a 6percent increase; the IV-LATE estimate is larger at 12 percent, but still relative-ly small. Both effects are statistically significant, however. There is no evidenceof heterogeneous effects by farm size or farmer age.
In the last two columns of Table 5 we check whether treated farmers econo-mize on search costs because of RML. To this effect, we examine whethertreated farmers make less effort gathering price information, either by consult-ing with others or by collecting price information in person at the time ofplanting. Contrary to expectations, results do not suggest this to be the case.The heterogeneous effect regression results reported at the bottom of thetable indicate that large farmers consult with more people and are more likelyto collect price information at planting time. For these farmers, the gain frommaking a better informed decision are larger, hence more effort is made togather relevant price information. But we find no significant evidence thatRML helps large farmers economize on these costs. This may only be tempor-ary, however: once farmers learn they can trust RML information they maydecide to rely on it more. Young farmers consult fewer people about prices,but there is no evidence of heterogeneous treatment effects by farmer age.
Price Received
There is considerable price variation within villages. Different crops have dif-ferent coefficients of variation: lower for nonperishable crops such as wheat(CV ¼ 0.07) and soya (CV ¼ 0.14), and higher for perishable crops such as to-matoes (CV ¼ 0.22), onions (0.44), and pomegranates (0.45). We thus expectRML to be particularly beneficial for more perishable crops since their pricesare more volatile and information is potentially more valuable.
This is what we investigate in Table 6. The dependent variable is the log ofthe unit price received by the respondent on average over all the sales transac-tions of the target crop during the 12 months preceding the survey. Similarresults are obtained if we use the price level instead of the log. The unit of ob-servation is the sales transaction. Most farmers report a single sale but somereport more than one, which explains why the number of observations exceedsthe number of participating farmers.
The first column of Table 6 reports the ATT obtained using nearest neighbormatching. Contrary to expectations, we find no beneficial effect of the treat-ment on price received: the treatment effect is negative and statistically signifi-cant. We worry that this may be due to the inclusion of treatment 2 villages inthe comparison. Indeed, in these villages, the small number of randomly
400 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA B L E 6. Prices obtained (expressed in log(Rs/kg))
ATT (a) ATT (b) IV-LATE OLS
Heterogeneouseffects (d)
long-model(c) OLS IV
For whole
sample(e)
No obs. 1480 688 1480 1425 1464 1457
Treatment Coeff 20.031 20.043 20.062 20.028 20.034 20.026t-value 22.000 20.520 21.670 21.510 21.860 21.430
Intercept Coeff 2.260 2.159 2.248 2.249t-value 309.620 21.250 93.64 99.080
Dummy younghead of hh
Coeff 0.021 20.013 20.013t-value 0.990 20.500 20.530
Total crop areacultivated
Coeff 0.005 0.001 0.001t-value 5.720 1.900 1.550
Dummy if soldto a trader
Coeff 20.011 20.006 20.008t-value 20.250 20.190 20.290
Treatmentextensionagent
Coeff 20.013t-value 20.500
Interaction with treatmentDummy young
head of hhCoeff 0.057 0.059t-value 1.750 1.850
Total crop areacultivated
Coeff 20.001 20.000t-value 20.590 20.240
Dummy if soldto a trader
Coeff 0.085 0.091t-value 1.750 1.830
For control/
treatment 1
village
No obs. 947 443 947 909 938 931
Treatment Coeff 20.015 0.031 20.079 20.017 20.046 20.032t-value 20.600 0.630 21.600 20.510 21.740 21.170
Intercept Coeff 2.211 2.071 2.218 2.209t-value 147.890 15.100 61.610 59.530
Dummy younghead of hh
Coeff 0.016 20.014 20.013t-value 0.490 20.410 20.360
Total crop areacultivated
Coeff 0.005 0.001 0.001t-value 2.940 0.950 0.800
Dummy if soldto a trader
Coeff 20.053 20.020 20.016t-value 20.970 20.770 20.510
Treatmentextensionagent
Coeff 20.048t-value 21.000
(Continued)
Fafchamps and Minten 401
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
treated farmers may circulate the RML information to untreated farmers, whowould then also benefit from it. This may blur the comparison between controland treated farmers due to a confounding externality between control andtreated farmers. To investigate whether this explains our result, we re-estimatethe ATT using only treatment 1 and control villages. The results are reportedin the second panel of Table 6. We again find a negative treatment effect onfarmer price, but it is not statistically significant. We also checked (results notreported here to save space) whether farmers in treatment 2 villages receivedhigher prices than in control areas—unsurprisingly, given the lack of result forstronger treatment 1, they do not. The next column reports dif-in-dif ATT esti-mates, using nearest neighbor matching. Point estimates are now slightly posi-tive, but nowhere near conventional levels of significance.
Next we examine whether the lack of effect is due to non-compliance. To in-vestigate this possibility, we instrument actual RML usage with theintent-to-treat dummy and report the results in the IV-LATE column ofTable 6. The estimated coefficient of receiving the RML service is still negative,but remains non-significant for the entire sample as well as for the samplewithout treatment 2 villages.
In Table A2 in the supplemental appendix, we repeat the ATT nearest neigh-bor matching and IV-LATE analysis for each crop separately. For the wholesample, ATT point estimates are negative for all crops, significantly so for
TABLE 6. Continued
ATT (a) ATT (b) IV-LATE OLS
Heterogeneouseffects (d)
long-model(c) OLS IV
Interaction with treatmentDummy young
head of hhCoeff 0.041 0.038t-value 1.040 0.910
Total crop areacultivated
Coeff 0.000 0.000t-value 0.050 0.180
Dummy if soldto a trader
Coeff 0.101 0.093t-value 2.210 1.820
(a) impact survey only; using nearest neighborhood matching; the reported coefficient on treat-ment is the ATT.
(b) diff-in-diff, nearest neighborhood matching; using average unweighted prices in baselineand impact survey.
(c) including but not reported dummies for graded, sold through commission agent, sold totrader, immediate payments, and quantity sold, years of education head of household, socialnetwork in village, land owned, years of farm experience, area cultivated of studied crop.
(d) Mean value substracted from those control variables interacted with treatment.
(e) village triplet code dummies included but not reported.
t-values based on standard errors clustered by village triplet code;
t-values in bold significant at the 10% level or better.
402 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
onions. IV-LATE point estimates remain negative, but are not statistically sig-nificant. When we restrict the analysis to control and treatment 1 villages, wefind negative ATT and IV point estimates for four out of five crops; except forone (IV for tomatoes), they are not significant.
We then examine whether intent-to-treat results may be affected by omittedvariable bias. This is unlikely because treatment is randomly assigned, but wecheck it anyway. To this effect, we add controls for farmer age and farm size,as well as dummies for type of sale (that is, whether sold in the village or to atrader, as opposed to sold in the local wholesale market or mandi). We alsoinclude a dummy equal to one if the extension agent serving the villagereceived the free RML service. Results are reported in Table 6 under the “OLSlong model” column. Other controls are included as well, as detailed at thebottom of the table, but their coefficients are not reported to save space. Againwe find no evidence of a significant treatment effect. The coefficient of the ex-tension agent treatment is similarly non-significant.
In the last two columns of Table 6 we investigate the possible existence ofheterogeneous effects. The OLS columns report the heterogeneousintent-to-treat effect, equation (9), with controls. We also estimate an heteroge-neous effect version of equation (8):
Yi ¼ u þ aSi þ gXi þ hSiðXi � �XÞ þ eið10Þ
Wooldridge (2002) recommends estimating IV models of this kind as follows.Let ~Sl be the predicted value of Si from the instrumenting equation. We con-struct a variable ~SlðXi � �XÞ and we estimate (10) using ~Sl and ~SlðXi � �XÞ asinstruments.
In the OLS (intent-to-treat) results we now find a negative average treatmenteffect but a positive heterogeneous effect on young farmers. Treated youngfarmers received a price that is about 6 percent higher on average. In the IVresults, the average treatment effect is non-significant, but the heterogeneousage effect remains. This suggests that less experienced farmers gain somethingfrom RML. These findings, however, are not robust to dropping treatment 2villages, as seen in the second panel of Table 6.
As robustness check, we correct for the possibility of non-random attribu-tion by adding an inverse Mills ratio as additional regressor in the IV-LATE re-gression. This Mills ratio is obtained from the attrition selection regressionmentioned in Section 3. Results, not shown here, are similar to those reportedin Table 6, and the Mills ratio is not statistically significant from 0 in the fullsample or when using treatment 1 only, suggesting that non-random attrition isunlikely to have affected our findings.
We also find that, in the OLS regression, farmers that grow more of thetarget crop get a significantly higher price on average. One possible explanationis that, for small crop sales, farmers make less effort to obtain price informa-tion and, hence, sell at a lower price. This effect, however, is not robust—it
Fafchamps and Minten 403
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
disappears in the IV regression or if we drop treatment 2 villages. Finally, con-sistent with expectations, treated households receive a price that is 8–9 percenthigher than control households when they sell to a trader as opposed to a com-mission agent. This is in line with the idea that better informed farmers can ne-gotiate better deals from buyers when they sell outside the relative safety of themandi.
We also examined whether treatment reduced the coefficient of variation ofthe price received by farmers in the same village. We expect price variationacross farmers to be less if they are better informed. Aker (2008), for instance,reports that the introduction of mobile phones in Niger facilitated price inte-gration and reduced price dispersion. We do not find a similar effect for RML:the coefficient of variation of prices in treatment 1 villages is 0.320; in controlvillages it is 0.228, that is, smaller than in treated villages. The difference,however, is not statistically significant: the t-value ¼ 1.52, with a p-value of0.135.
Costs and Revenues
RML may affect farmers in ways other than prices. Transaction costs per trans-action average 0.84 Rs/Kg. This compares to standard deviations for prices of2.2, 17.1, 4.6, 0.9, and 3.1 Rs/Kg for tomatoes, pomegranates, onions, wheat,and soya, respectively. There is therefore room for farmers to increase revenuesby reducing transaction costs.
In the first column of Table 7, we report ATT and IV estimates for totaltransaction costs on the farmer’s last crop sale. Transaction costs include trans-port, loading and off-loading, payment at checkpoints, personal transport, pro-cessing, and commissions. Point estimates are positive for the whole sample—suggesting that RML raises costs—but they become negative when we only usetreatment 1 villages. In both cases, however, point estimates are not significant.
In the next column we investigate whether farmers received a higher netprice (defined as the gross price minus the variable transaction costs in the lasttransaction). Mattoo, Mishra, and Narain (2007) estimate that transport costsper truck in India are between 0.09 to 0.13 Rs/kg/100 kilometers, which issmall relative to total transaction costs. It thus seems that, in transport cost atleast, arbitraging over space is not prohibitively expensive relative to othertransaction costs. If farmers use RML information to arbitrage across space,they may ship their crop to a more distant market and incur a higher transportcost, but obtain a higher price net of costs, as in Jensen (2007). This is notwhat we find: results remain resolutely non-significant whether we includetreatment 2 villages or not.
404 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Farmers may gain not on the unit price but on total revenue. This is investi-gated in column 3. We find large positive point estimates, but no significanteffect.12 If we use logs instead to limit the influence of outliers, we again findno significant effect. The last column reports similar results for value-added,that is, revenues minus monetary input costs such as fertilizer and pesticides. Ifweather information and crop advisories raise farmers’ technical and allocativeefficiency, we would expect value-added to rise. Results are similar to those for
TA B L E 7. Profitability measures
Transactioncost (c)
Net price(d)
Salerevenues
Valueadded
(e)
For whole sample
Number ofobservations
713 713 713 713
Nearest neighbor matching (a)ATT Coeff 0.078 20.760 48,247 46,352
z-value 1.420 21.480 0.580 0.580Regression results (b)IV-LATETreatment Coeff 0.146 21.450 87,074 84,530
t-value 1.050 21.730 0.880 0.910Intercept Coeff 1.576 8.906 66,545 59,235
t-value 59.060 55.350 3.500 3.320
For control/treatment 1 villages
Number ofobservations
458 458 458 458
Nearest neighbor matching (a)ATT Coeff 20.150 0.735 143,852 138,311
z-value 21.700 1.060 1.190 1.220Regression results (b)IV-LATETreatment Coeff 0.159 20.074 267,588 260,249
t-value 0.439 21.000 1.370 1.410Intercept Coeff 1.602 1.977 222,749 227,914
t-value 25.230 84.880 20.370 20.480
(a) Matching based on village triplet code dummies.
(b) Village triplet code dummies included but not reported.
(c) last transaction only; includes costs for transport, loading, off-loading, payments at check-point/toll or road-block, personal transport, processing, commission expressed in Rs/kg.
(d) last transaction only; gross price minus transaction costs expressed in Rs/kg.
(e) sales minus monetary input costs (fertilizer, pesticides, spray, purchased seeds, manure).
t-values based on standard errors clustered by village triplet code;
t-values in bold significant at the 10% level or better.
12. Sale values are large because quantities sold are large. This is especially true for onions and
pomegranates where the average size of a single transaction is 10 metric tons. There is, however, a lot
of variation around this average.
Fafchamps and Minten 405
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
sale revenues: large positive point estimates, but nothing statistically significant.Similar findings obtain if we use logs instead.
Marketing
In the conceptual section we argued that if RML is used by farmers to increasethe price they receive, we should observe differences in marketing practices. Ifprice information makes enables farmers to arbitrage across markets, weshould observe systematic changes in where farmers sell.
We first note that most sales take place in a market, nearly always a whole-sale market or mandi. The only exceptions are pomegranates for which, atbaseline, 44 percent of sales are conducted at the farm-gate and, to a lesserextent, wheat, with 7 percent of farm-gate sales. For the other crops, farm-gatesales represent less than 2 percent of recorded sales. Second, market diversifica-tion varies from crop to crop. Sales of perishable crops are geographically con-centrated: 98 percent and 81 percent of all market sales of tomatoes andpomegranates, respectively, occur at a single district market. Concentration isless for other crops: for onions, 51 percent of sales go to one district market.Corresponding figures for wheat and soya are 54 percent and 57 percent,respectively.
To investigate whether treatment changed where farmers sell their crops, weconstruct an overlap index that captures the extent to which a farmer sold tothe same location in the baseline and follow-up surveys. There are 39 whole-sale markets listed in the data, with farm-gate sales treated as a separate loca-tion. The index is weighted by quantity. An index value of 1 means the farmsold in the same location in the two survey rounds; a value of 0 means thatnothing was sold at the same place. We also construct an added marketdummy, which takes value 1 if the farmer sold in a new market or location inthe follow-up survey, and a dropped market dummy equal to 1 if the farmerstopped selling in a specific location in the second round.
Average treatment effects for the market overlap index and for the addedand dropped market dummies are reported in Table 8. In the top panel we usethe entire sample; in the second panel we only use the treatment 1 and controlsamples. With the entire sample treatment has a significant effect: treatedfarmers are 10 percentage points more likely to add a new sales location(market or farm-gate) and 9 percentage points more likely to drop one sales lo-cation. Treatment also reduces the overlap index by 10 percent on average.When we instrument RML usage with assignment to treatment, point estimatesdouble and remain significant. These results are consistent with the predictionsof the arbitrage model although, as we have seen in the previous two subsec-tions, changing sales location does not appear to have resulted in a higher priceon average. Point estimates are also slightly smaller when we limit the sampleto treatment 1 and control villages (second panel of Table 8), but they are nolonger statistically significant at the 10 percent level, perhaps because of the re-duction in sample size.
406 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
We continue our investigation of crop marketing in Table 9. The unit ofanalysis is an individual sale transaction. We first examine whether farmers sellat a wholesale market or mandi. As we have discussed earlier, farmers maychoose to sell at the mandi because it is the only way to obtain accurate priceinformation, even though doing so raises transaction costs relative to farm-gatesales. If this is the case, the RML service may give farmers the confidence notto sell at the wholesale market, for instance, because they can better negotiatewith a farm-gate buyer.
To investigate this possibility, we test whether treated farmers are less likelyto sell at the mandi. Results, reported in the first column of Table 9, indicatethat this is not the case: The intent-to-treat ATT, reported at the top of column1, raises the likelihood of selling at the mandi. In the rest of column 1 weexamine whether the results are different when we use IV-LATE instead, orwhen we allow for heterogeneous effect by firm size and farmer age. Resultsare qualitatively similar. The magnitude of the effect, however, is small,
TA B L E 8. Spatial arbitrage and market changes
Number of markets
Added Dropped Overlap index (c)
For whole sample
Number of observations 691 691 691Nearest neighbor matching (a)ATT Coeff 0.099 0.087 20.095
z-value 2.980 2.680 23.030
Regression results (b)IV-LATETreatment Coeff 0.208 0.194 20.197
t-value 2.120 2.080 22.090
Intercept Coeff 0.575 0.463 0.493t-value 30.430 25.850 27.290
For control/treatment 1 villages
Number of observations 445 445 445Nearest neighbor matching (a)ATT Coeff 0.045 0.074 20.077
z-value 0.880 1.560 21.650Regression results (b)IV-LATETreatment Coeff 0.187 0.189 20.198
t-value 1.260 1.320 21.400Intercept Coeff 0.629 0.503 0.489
t-value 13.540 11.230 11.110
(a) Matching based on village triplet code dummies.
(b) Village triplet code dummies included but not reported.
(c) overlap index of sales location between years, weighted by quantity – see text for details
t-values based on standard errors clustered by village triplet code.
Fafchamps and Minten 407
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
probably because most farmers already sell at the mandi. When we differentiatethe effect by crop, it is significant for pomegranates (point estimate 0.157 withz-value of 2.35) and—less so—for soya (point estimate 0.079 with z-value of1.73); it is not significant for the other three crops. That pomegranates aremost affected is hardly surprising given that pomegranates are the only cropwith a sizable proportion of farm-gate sales at baseline. Thus, if anything,RML makes farmers more likely to sell at the mandi.
Among farmers who sell at the market, however, Table 9 has shown achange in crop destination. To verify this further, we asked farmers who sell ata particular wholesale market whether they do so because it is the closestmarket. We see from the second column of Table 9 that treated farmers are
TA B L E 9. Other marketing characteristics, all transactions
Sold inwholesale
market
if whole-salemarket, chosenbecause closest
Sold through acommission
agentSold totrader
Crop wasgraded/sorted
before sale
For whole sample
Number ofobservations
1477 1352 1482 1470 1478
Nearest neighbor matching (a)ATT Coeff 0.030 20.078 0.006 0.046 0.033
z-value 2.540 23.220 0.230 1.740 2.260
Regression results (b)1. IV-LATETreatment Coeff 0.063 20.131 0.539 0.084 0.055
t-value 1.750 20.940 0.844 1.050 1.120Intercept Coeff 0.923 0.199 0.933 0.450 0.925
t-value 132.800 6.940 89.080 28.350 98.140
2. Heterogeneous effects (c)Treatment Coeff 0.032 20.064 0.054 0.039 20.029
t-value 1.820 21.010 0.620 1.010 21.120Intercept Coeff 0.955 0.277 0.898 0.355 0.045
t-value 57.990 4.760 19.650 6.900 2.830
Dummy younghead of hh
Coeff 20.013 20.091 0.049 0.140 20.024t-value 21.080 22.100 0.690 2.950 21.180
Total crop areacultivated
Coeff 20.002 20.001 20.001 0.001 20.002t-value 21.620 20.400 20.580 0.570 21.410
Interaction with treatmentDummy young
head of hhCoeff 0.011 0.094 20.015 20.159 0.052t-value 0.590 0.460 20.120 22.330 1.950
Total crop areacultivated
Coeff 0.003 0.002 20.008 20.002 0.002t-value 1.790 0.460 21.330 21.130 1.380
(a) Matching based on village triplet code dummies.
(b) Village triplet code dummies included but not reported.
(c) Mean value substracted from those control variables interacted with treatment.
t-values based on standard errors clustered by village triplet code;
t-values in bold significant at the 10% level or better.
408 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
less likely to say they sell at a market because it is closest. Taken together, theevidence therefore suggests that treated farmers are more likely to sell fartheraway from home—either by switching from farm-gate to market sale or byswitching to a more distant mandi.
To investigate this further, we test whether treated farmers are more likelyto sell directly to a trader (typically at the farm-gate) or without the help of acommission agent. If RML improves price information, farmers may be less re-luctant to sell to a trader, knowing they can insist on a price commensurate tothe price at the nearest mandi. By a similar reasoning, they may rely less oncommission agents who are contractually obliged to help farmers get the bestprice but to whom a fee must be paid. Table 9 shows this is not the case forcommission agents—the ATT is not significantly different from 0 in any of thethree methods we report. For selling to a trader, we find a weakly significantATT, but significance disappears when we use IV-LATE or allow for heteroge-neous effects. In the heterogeneous effect regression reported in the last panelof Table 9, we see that young farmers are more likely to report selling to atrader, but this relationship disappears with treatment, suggesting that youngfarmers learn not to sell to traders.
Taking columns 1 to 4 together, the evidence suggests that RML helpedsome farmers realize that they could obtain a higher price by going to a moredistant mandi rather than selling at a closer market or at the farm-gate. It ispossible that some farmers choose to sell locally because of uncertainty regard-ing the return from traveling to the more distant mandi. Providing informationabout the mandi price reduces the risk of traveling to the mandi, and the reduc-tion in uncertainty may be what induced some farmers to incur the additionalcost of traveling. In contrast, the evidence does not support the hypothesis thatbetter informed farmers are emboldened to sell in local markets or at the farm-gate because they can insist on receiving a price more in line with the regionalwholesale price.
Finally, RML provides information on the price spread due to crop quality,that is, it shows prices by grade. Consequently, we expect treated farmers topay more attention to quality, for instance, by grading or sorting their outputinto separate categories to obtain a better price on the top quality. This is whatwe find (see the last column of Table 9) for the ATT where the effect is statis-tically significant. The magnitude of the effect, however, is small: treatmentraises the proportion of farmers grading or sorting their output by 3 percentagepoints. The effect also disappears in the IV-LATE regression; it resurfaces inthe heterogeneous effect regression, but only when interacted with farmerage—that is, young farmers are slightly more likely to grade or sort theiroutput as a consequence of treatment.
Weather Information
RML provides weather forecasts that are spatially disaggregated—and hencepresumably more accurate than those publicized on the radio. Do RML
Fafchamps and Minten 409
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
forecasts help farmers improve yields, for instance, because farmers can takebetter anticipative action?
We investigate this question in Table A.3 in the supplemental appendix.Farmers were asked whether or not they incurred unusually high rainfallevents, such as a storm or heaving downpour. Some 58 percent said they did.The likelihood of reporting a storm is correlated with treatment in the IV andheterogeneous effect regressions: treated farmers are more likely to report in-curring a storm. Since we have no reason to believe that the weather is corre-lated with treatment, this is most likely due to response bias: farmers whoreceive regular weather information become more aware of unusual rainfallevents, and are more likely to report them to enumerators.
We test whether treated farmers were able to reduce output loss or increaseoutput following a storm. We find no evidence that this is the case. We alsofind little evidence of beneficial heterogeneous effects: young farmers reportmore output loss at harvest following a storm, not less.
Agricultural Technology and Practices
In addition to price and weather information, RML provides crop advisorymessages relaying information on crop varieties, pesticide use, and cultivationpractices. This information may be particularly valuable for sample farmersbecause some of our target crops are relatively new to them.
In Table 10 we examine whether farmers changed the variety of the targetcrop that they grow. Some 31 percent of respondents stated they did changevariety between the two survey years, but this proportion is the same irrespect-ive of treatment. Of those who changed variety, 65 percent stated they did soto improve profitability. Again we find no statistical relationship with treat-ment. Farmers who stated they changed crops to improve profitability wereasked whether they did so because of RML. Here we find a statistically positivetreatment effect: depending on the estimation method, treated farmers are 14–20 percent more likely to list RML as inspiration for the change. This is re-assuring, but not necessarily conclusive given that treatment is found to haveno effect on the propensity to change variety or on the reason for changingvariety.
In the last two columns of Table 10 we turn to cultivation practices. In2010 farmers were asked whether they changed anything about their cultiva-tion practices since the previous year; 22 percent of respondents stated they didso. We find no evidence that treated farmers were more likely to change culti-vation practices.
Those who did change were asked what made them change practices. Ofthose farmers who report a change, a large proportion mentions RML as thereason for the change. The effect is statistically significant and large inmagnitude—a 20–41 percent higher likelihood of listing RML, depending onthe estimator. As for crop variety, this evidence is reassuring but not conclusive
410 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
given that treatment has no noticeable effect on changing crop practicesthemselves.
V I . C O N C L U S I O N
We have reported the results of a randomized controlled trial of the impact ofan SMS-based agricultural information service in Maharashtra. This informa-tion service, called Reuters Market Light (RML), sends SMS to farmers withinformation on prices, weather forecasts, crop advice, and news items. Theprice information is expected to improve farmers’ ability to negotiate with
TA B L E 10. Crop varieties grown and cultivation practices
For wholesample
Change of cropvariety since
last year
If yes,reason isprofita-bility
Ifprofitability,
because of RML
Change incultivation
practices last year
If change,because of
RML
Number ofobservations
895 240 156 911 203
Nearest neighbor matching (a)ATT Coeff 0.029 0.020 0.155 20.027 0.211
z-value 1.100 0.460 2.830 21.110 3.990
Regression results (b)1. IV-LATETreatment Coeff 0.043 0.006 0.200 20.045 0.410
t-value 0.970 0.090 2.060 21.240 5.170
Intercept Coeff 0.525 0.374 20.033 0.476 20.016t-value 59.350 30.950 22.060 65.780 20.950
2. Heterogeneous effects (c)Treatment Coeff 0.021 20.003 0.140 20.025 0.199
t-value 0.910 20.080 2.220 21.270 3.580
Intercept Coeff 0.518 0.408 20.112 0.432 20.051t-value 17.210 9.280 25.680 14.940 21.270
Dummy younghead of hh
Coeff 20.001 20.079 0.103 0.042 0.036t-value 20.030 21.280 2.350 0.890 0.850
Total crop areacultivated
Coeff 0.001 0.002 0.001 0.001 0.000t-value 0.770 1.640 0.980 1.310 20.030
Interaction with treatmentDummy young
head of hhCoeff 20.091 0.053 20.001 20.009 0.033t-value 21.760 0.680 20.010 20.150 0.220
Total crop areacultivated
Coeff 20.002 0.000 0.000 20.001 0.001t-value 21.240 20.120 0.100 20.910 0.210
(a) Matching based on village triplet code dummies.
(b) Village triplet code dummies included but not reported.
(c) Mean value substracted from those control variables interacted with treatment.
t-values based on standard errors clustered by village triplet code;
t-values in bold significant at the 10% level or better.
Fafchamps and Minten 411
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
buyers and to enable them to arbitrage better across sales outlets. Weather in-formation should help farmers reduce crop losses due to storms. Crop advisoryinformation should induce farmers to adopt new crop varieties and improvecultivation practices.
The trial was conducted in collaboration with Thomson-Reuters, the pro-vider of RML. The experiment involved 933 farmers in 100 villages of centralMaharashtra. Treatment was randomized across villages and, in some cases,across farmers as well. Participating farmers were surveyed twice in face-to-faceinterviews. We also conducted a follow-up telephone survey to gather informa-tion on reasons for non-compliance. Randomization appears to have workedwell in the sense that the control and treatment samples are balanced on mostrelevant variables. Although contamination is limited, non-compliance iscommon, which is why we reported intent-to-treat estimates throughout. Wealso reported instrumental variable (IV) estimates in which selection into treat-ment is used to instrument RML usage.
We find no statistically significant average treatment effect on the pricereceived by farmers, crop losses resulting from rainstorms, or the likelihood ofchanging crop varieties and cultivation practices. Treated farmers appear tomake use of the RML service and they associate RML information with anumber of decisions they have made. But, based on the available evidence, onaverage they would have obtained a similar price or revenue, with or withoutRML.
Although disappointing, our results are in line with the RML take-up rate inthe study districts. After a rapid expansion following the introduction of theservice in 2007–09, take-up shows a relative stagnation in 2009–10, suggest-ing a possible loss of interest. We cannot, however, rule out that supply-sidefactors played a role. We also suspect that some farmers do not know how torenew the service.13
Although the absence of positive effect on price may surprise and disap-point, we find evidence of an RML information effect on where farmers selltheir crop: they are less likely to sell at the farm-gate—especially youngfarmers—and more likely to sell at a different, more distant wholesale market.These results contradict the idea that RML information enables farmers to ne-gotiate better prices with itinerant traders, but are consistent with using RMLinformation to arbitrage across sales outlets. Why arbitrage does not translateinto higher prices is unclear, but some possible explanations arise from thedata. First, few farmers sold at the farm-gate at baseline—except for pomegra-nates—thereby limiting the number of farmers who could realize that selling atthe market was more beneficial than selling at the farm-gate, as a few did.Second, before treatment crop sales were concentrated on a single wholesalemarket in each district. Spatial concentration probably limits the range of
13. The provider has indeed encountered difficulties in setting up a reliable system for enabling
clients to easily and reliably make repeat purchases of the RML service.
412 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
alternative market destinations nearby—and thus opportunities for arbitragingby farmers.
We find little evidence of other RML effects. If RML information helpsfarmers improve crop quality, we should observe treated farmers changing agri-cultural practices, especially for crop varieties and grading. We do not, exceptfor grading but the magnitude of the effect is small. We also find no significanteffect on transaction costs, revenues, and value-added.
Taken together, the evidence is consistent and compelling. Surveyed farmerssell almost exclusively to a wholesale agricultural market nearby. If tradershave a comparative advantage in transport, for example because farmers donot know anyone they can trust in other markets (Gabre-Madhin 2001), traderarbitrage across markets should ensure that farmers cannot fetch a more remu-nerative price by selling elsewhere. Hence it is optimal for farmers to sell to thenearest market. Similarly, if farmers fear being cheated when they sell at thefarm-gate, it is optimal for them not to do so. Given this, it is perhaps not sosurprising that better price information did not translate into higher farmerprices.14
If the above interpretation is correct, it has a number of implications for theexternal validity of our findings. Price information could help if spatial arbi-trage across agricultural markets does not hold, for example because marketsare disorganized, segmented, or too thin to attract a steady flow of buyers—orbecause producers have a comparative advantage in transport, as in Jensen(2007). Even in such a case, however, price information is likely to be used firstby traders, as documented by Aker (2008). Price information could also helpfarmers who sell at the farm-gate, such as the coffee growers studied byFafchamps and Hill (2008). A stronger effect on crop quality may be obtainedif price information is detailed by variety and grade and if farmers are providedwith complementary information on how to produce high-price varieties andgrades. These suggestions should help steer policy intervention toward regionsand markets where the effect of price information may be beneficial, and avoidwasting resources on markets where it is unlikely to matter.
RE F E R E N C E S
Abadie, A., D. Drukker, J. L. Herr, and G. W. Imbens. 2004. “Implementing Matching Estimators for
Average Treatment Effects in Stata.” Stata Journal 4 (3): 290–311.
Aker, J.C. 2008. Does Digital Divide or Provide? The Impact of Mobile phones on Grain Markets in
Niger. Berkeley: University of California Press.
14. Nothing in this argument relies on buyer competition within each market: Even if buyers act in
a monopsonistic fashion, as documented by Banerji and Meenakski (2004) for Delhi wheat markets,
giving farmers information about market prices would not improve farmer prices—unless buyers price
discriminate across markets and spatial arbitrage does not hold.
Fafchamps and Minten 413
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Aker, J. C., and M. Fafchamps. 2011. Mobile Phones and Farmers’ Welfare in Niger. Berkeley:
University of California Press.
Banerji, A., and J. V. Meenakski. 2004. “Buyer Collusion and Efficiency of Government Intervention in
Wheat Markets in Northern India: An Asymmetrical Structural Auction Analysis.” American
Journal of Agricultural Economics 86 (1): 236–253.
Bruhn, M., and D. McKenzie. 2009. “In Pursuit of Balance: Randomization in Practice in Development
Field Experiments.” American Economic Journal: Applied Economics 1 (4): 200–232.
Deaton, A. 2009. Instruments of Development: Randomization in the Tropics, and the Search for the
Elusive Keys to Economic Development. NBER Working Paper #14690.
Fafchamps, M., and R. V. Hill. 2008. “Price Transmission and Trader Entry in Domestic Commodity
Markets.” Economic Development and Cultural Change 56 (4): 729–766.
Gabre-Madhin, E. 2001. “The Role of Intermediaries in Enhancing Market Efficiency in the Ethiopian
Grain Market.” Agricultural Economics 25 (2–3): 311–320.
Goyal, A. 2010. “Information, Direct Access to Farmers, and Rural Market Performance in Central
India.” American Economic Journal: Applied Economics 2 (July): 22–45.
Jensen, R. 2007. “The Digital Provide: Information (Technology), Market Performance, and Welfare in
the South Indian Fisheries Sector.” Quarterly Journal of Economics 127 (3): 879–924.
Mattoo, A., D. Mishra, and A. Narain. 2007. From Competition at Home to Competing Abroad.
Washington, DC: World Bank.
Minten, B., A. Vandeplas, and J. Swinnen. 2012. “Regulations, Brokers, and Interlinkages: The
Institutional Organization of Wholesale Markets in India.” Journal of Development Studies,
forthcoming.
Mittal, S., S. Gandhi, and G. Tripathi. 2010. Socio-Economic Impact of Mobile Phone on Indian
Agriculture. ICRIER Working Paper no. 246. New Delhi: International Council for Research on
International Economic Relations.
Staatz, J.M., A.M. Kizito, M.T. Weber, and N.N. Dembele 2011. Evaluating the Impact on Market
Performance of Investments in Market Information Systems: Methodological Challenges. MSU Staff
Paper 2011-08. East Lansing: Department of Agricultural, Food and Resource Economics, Michigan
State University.
Wooldridge, J. M. 2002. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT
Press.
World Bank. 2010. India Economic Update, June 23 2010.
414 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Crises, Food Prices, and the Income Elasticityof Micronutrients:Estimates from Indonesia
Emmanuel Skoufias, Sailesh Tiwari, and Hassan Zaman
The 2008 global food price crisis and more recent food price spikes have led to agreater focus on policies and programs to cushion the effects of such shocks onpoverty and malnutrition. Analysis of the income elasticities of micronutrients andtheir changes during food price crises can shed light on the potential effectiveness ofcash transfer and nutrition supplement programs. This article examines these issuesusing data from two cross-sectional household surveys in Indonesia, taken before(1996) and soon after (1999) the 1997–98 economic crisis, which led to a sharp in-crease in food prices. First, using nonparametric and regression methods, the articleexamines how the income elasticity of calories from starchy staples as a share of totalcalories differs between the two survey rounds. Second, the article estimates incomeelasticities of important nutrients in Indonesia. The analysis finds that, althoughsummary measures such as the income elasticity of the starchy staple ratio might notchange during crises, this stability masks important differences across individual nutri-ents. In particular, income elasticities of some key micronutrients, such as iron,calcium, and vitamin B1, are significantly higher in a crisis year than in a normalyear, yet the income elasticities of others—such as vitamin C—remain close to zero.These results suggest that cash transfer programs might be even more effective duringcrises to ensure the consumption of essential micronutrients. But to ensure that all keymicronutrients are consumed, nutrition supplement programs are also likely required.JEL classification codes: I12, O12, D12, E31 keywords: Income elasticity, crisis,prices, micronutrients, vitamin C, starchy staple ratio
Global food prices rose sharply in 2008 and, in real terms, reached levels notseen since the early 1970s. The Food and Agriculture Organization (FAO2008) Food Price Index grew by 73 percent between September 2006 andmid-2008, driven by unprecedented increases across all food categories. Duringthe same period meat and fish prices increased 25 percent, eggs and milk 91
THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 415–442 doi:10.1093/wber/lhr054Advance Access Publication November 22, 2011# The Author 2011. Published by Oxford University Press on behalf of the International Bankfor Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions,please e-mail: [email protected]
Emmanuel Skoufias (corresponding author; [email protected]) is lead economist in the
Poverty Reduction and Equity Group of the World Bank. Sailesh Tiwari ([email protected]) is in
the Young Professionals Program of the World Bank and a Ph.D. candidate in the Economics
Department of Brown University. Hassan Zaman ([email protected]) is lead economist in the
Poverty Reduction and Equity Group of the World Bank. The authors are grateful to Harold Alderman
for useful comments on an earlier version of this article. A supplemental appendix to this article is
available at http://wber.oxfordjournals.org.
415
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
percent, oils and fats 149 percent, and cereal grains 123 percent. After a down-ward trend in 2009, global food prices rose again in late 2010, with prices inDecember 2010 close to their 2008 peak. Such food price volatility, particularlysharp spikes, needs to be monitored closely for its impact on the poor. Areview of the literature on the effects of the 2008 food price increases suggeststhat they are likely to have had a significant impact on the incidence of poverty(Ivanic and Martin 2008) and undernourishment (Tiwari and Zaman 2010)throughout the developing world.
Soaring food prices and their adverse effects have not only heightened con-cerns about food security and malnutrition in parts of the developing world,but have also sparked a renewed interest in the design of effective policyresponses. From the household point of view, such price increases have twomain consequences: they reduce the purchasing power of household income,especially among poorer households, which spend a larger share of theirincomes on food, and they result in a relative price effect that induces house-holds to substitute away from more expensive foods.1
Government interventions during rising food prices have almost always beenmotivated by the need to compensate poor households for their lost purchasingpower. These interventions—commonly known as social safety net programs—are aimed at smoothing consumption and protecting the caloric availability ofhouseholds to prevent increases in poverty and hunger. A review of the safetynet programs used during the 2008 food price crisis shows that they typicallytook the form of income support, cash transfers, price subsidies, and supple-mentary feeding programs or in-kind transfers of staple foods (Wodon andZaman 2010).
Yet even if income support or in-kind staple food distribution is successfulat preventing available calories from reaching dangerously low levels, there arevalid concerns about dietary diversity and the consequent risk of malnutrition.When household income drops, households may keep calorie levels more orless constant through substitutions within and between food groups, while theconsumption of essential micronutrients might decrease significantly as house-holds consume less meat, vegetables, eggs, and milk (Behrman 1995). Theextent to which the consumption of micronutrients responds to decreases inincome among poor households is of particular concern given the long-runconsequences that a diet poor in micronutrients can have on child developmentbefore and after birth.
Though there is ample evidence on the income elasticity of calories (Straussand Thomas 1995; Subramanian and Deaton 1996; Skoufias 2003), empiricalevidence on the income elasticity of micronutrients is sparse. The evidence thatdoes exist suggests substantial differences in the income elasticities of
1. Purchasing power improves for net sellers of agricultural commodities whose prices increase. In
addition, for households close to subsistence and already consuming the cheapest sources of calories,
substitution possibilities are much more limited.
416 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
micronutrients (Behrman and Deolalikar 1987; Bouis 1991). Pitt andRosenzweig (1985), focusing on Indonesia and using data from farm house-holds, report very low income elasticities (less than 0.03) for a set of nutrientsthat included calories, proteins, fats, carbohydrates, calcium, phosphorus, iron,and vitamins A (carotene) and C (ascorbic acid). Chernichovsky and Meesook(1984), using data from rural and urban areas, report much higher income elas-ticities for nutrients—for example, from 0.7 to 1.2 for the poorest 40 percent(by expenditure) of the population on Java. Similarly diverse estimates havebeen reported for other countries.2
Most of the empirical evidence sheds light on whether the price sensitivityof the demand for food and nutrients varies with income level (Timmer andAlderman 1979; Timmer 1981; Pitt 1983) or whether the income elasticity ofcalories varies with income level (Behrman and Deolalikar 1987; Ravallion1990; Strauss and Thomas 1995; Subramanian and Deaton 1996). But evi-dence is lacking on whether the income elasticity of nutrients varies significantlydepending on changes in the relative prices faced by households, such aschanges experienced during food price spikes. This is an important gap in theliterature and perhaps a large contributor to the significant divergence in esti-mates of the income elasticity of nutrients. When income elasticities are esti-mated using cross-sectional data, price variations necessarily come from thecross-section as well. But despite being pointed out by Deaton and Muellbauer(1980), the possibility that these estimated elasticities could be sensitive to thedegree of variability of prevailing relative prices is something that has neitherbeen tested explicitly nor been used to qualify the policy recommendations thatemerge from existing studies.
In 1998, during the financial crisis in Indonesia, the value of the rupiahdepreciated dramatically, falling from around 2,400 per US dollar in June 1997to just under 15,000 per dollar in June 1998 and finally settling at 8,000-9,000per dollar in December 1998. These fluctuations in the exchange rate led tolarge increases in the prices of tradable commodities in domestic markets.Indonesia’s consumer price index rose 107 percent between February 1996 andFebruary 1999. During the same period the food price index jumped 188percent. In addition, subsidies for consumer goods such as rice, oil, and fuelwere removed in 1998. So it is questionable whether estimates of the incomeelasticity of nutrients obtained from a sample of households observed during
2. Behrman and Deolalikar (1987), using data from villages covered by the International Crops
Research Institute for the Semi-Arid Tropics (ICRISAT), report income elasticity estimates of 0.06 to
0.19 for proteins (depending on whether level estimates or differences over time are used), 0.30 to
–0.22 for calcium, –0.11 to 0.30 for iron, 0.19 to 2.01 for vitamin A, –.08 to 0.18 for thiamine, 0.69
to 0.01 for riboflavin, –0.15 to 0.21 for niacin, and 0.15 to 1.25 for vitamin C. A Nicaraguan study by
Behrman and Wolfe (1984) reports significant income elasticity estimates in the range of 0.04 to 0.11
for calories, proteins, iron, and vitamin A (with statistically significant, but quantitatively small,
nonlinearities). A Philippine study by Bouis (1991) reports an income elasticity of 0.44 for iron, an
income elasticity of 0.16 for calories, and insignificant income elasticities for vitamin A and vitamin C.
Skoufias, Tiwari, and Zaman 417
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
precrisis years can provide any guidance on how caloric and micronutrientavailability might respond to additional income (other things being equal)during a period with a different set of relative prices. From a policy perspec-tive, the sensitivity of the income elasticity of nutrients to relative prices impliesthat policies aimed at increasing household income—such as employment andcash transfer programs—might be less effective in protecting nutritional out-comes under some economic conditions.
This article uses shocks to food prices in Indonesia to examine the relation-ship between nutrient consumption and prices, with the analysis conducted attwo levels. First, using the starchy staple ratio (the calories from starchy staplefoods such as cereals and tubers as a share of total calories) as a summarymeasure of household nutritional welfare, it assesses the impact of dramaticchanges in food prices on household dietary composition. The SSR is definedas the share of calories consumed obtained from starchy staple foods such ascereals and tubers. According to Bennett’s Law this ratio is inversely related tothe importance of starches relative to higher-quality, more expensive,micronutrient-rich foods such as meat, fish, fruits, and vegetables. The focus ison how the income elasticity of the starchy staple ratio differs between surveyrounds in 1996 and 1999, when relative prices were very different for cerealsand the other major food groups.3 The analysis is conducted for Indonesia’sentire population and for the poor in urban and rural Central Java (one of thecountry’s poorest provinces) in 1996 and 1999. Results are reported using bothnonparametric and regression methods.
This analysis is supplemented by updated estimates of the income elasticitiesof important nutrients in Indonesia, such as calories, proteins, fats, carbohy-drates, calcium, phosphorous, iron, and vitamins A, B1 (thiamin), and C. Intimes of crisis, cash transfers may be the fastest and least costly way of reachinghouseholds most likely to be adversely affected, if the delivery infrastructure isin place and leakage is low. Reliable elasticity estimates can help policymakersdetermine beforehand whether cash transfers are likely to increase nutrientavailability among poor households or if other interventions may be needed.
If estimated elasticities are high and significantly different from zero, policiessuch as cash transfers that aim to compensate for price increases are likely tobe effective. But if estimated elasticities are indistinguishable from zero, alter-native interventions may be needed. Particular emphasis is placed on the sensi-tivity of the elasticity estimates to biases due to measurement errors inconsumption and nutrient availability at the household level.
The article also presents a test of whether the income elasticity of nutrientsvaries with the economic conditions facing households. Other things being
3. As a measure of dietary composition, the starchy staple ratio is useful because it summarizes
changes in nutritional welfare over time. There is also evidence from Indonesia that a closely related
variable—the share of total food expenditure going to nongrains, such as animal and plant sources—is
positively associated with lower incidence of stunting among children younger than five (Sari and others
2010).
418 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
equal, changes in the relative prices of staple foods may result in unexpectedresponses of how the demand for nutrients responds to cash transfers. Forexample, if total calorie availability is already low and the price of a stapleincreases during a crisis, households receiving cash transfers might spend theadditional income on that same staple if it continues to be the cheapest sourceof calories (Behrman 1988; Behrman and Deolalikar 1989).4
This article is structured as follows. Section I describes the data used for theanalysis and the construction of key variables and presents background infor-mation on the changes in prices and nutrient availability between 1996 and1999 in Indonesia. Section II discusses the empirical strategy and the results ofestimation using both nonparametric and regression methods. Section III sum-marizes the findings and presents some policy implications.
I . B A C K G R O U N D A N D D A T A
The analysis in this article is based on the detailed consumption module of theNational Socio-Economic Survey (SUSENAS) conducted every three years bythe Central Statistical Agency of the Government of Indonesia. The consump-tion module is nationally representative of urban and rural areas in each of thecountry’s 26 provinces and the Jakarta metropolitan area.5 The survey included60,678 households in 1996 and 62,217 households in 1999.
In addition to the detailed nature of the survey, one of the main advantagesof comparing the income elasticity of calories in these two years is the oppor-tunity to examine economic behavior under dramatically different relative priceregimes. In February 1999, when the 1999 SUSENAS was conducted, inflationin Indonesia had reached its peak since the start of the financial crisis in late1997 and its intensification in mid-1998. An additional benefit is that the samequestionnaire was used at the same point in time in both survey years. Thusthe possible influence of seasonal factors in the relationship of income tocalories—as emphasized by Behrman, Foster, and Rosenzweig (1997)—can becontrolled for.6 A detailed discussion of the SUSENAS consumption moduleand the construction of key variables used in the analysis is presented inappendix A.
To make meaningful comparisons across two cross-sectional surveys that arethree years apart, the nominal income of households in 1999 must be expressed
4. This statement is not intended to compare the effectiveness of cash transfers relative to other
possible alternatives of increasing nutrient availability within households. Alternatives include in-kind
food transfers and employment creation programs.
5. The analysis also uses some variables from the larger SUSENAS (core survey) containing
observations for about 205, 000 households.
6. The Idul Fitri-Lebaran holiday following the fasting month (Ramadan) is a moving holiday—in
1999 it fell in late January. Central Statistical Agency officials confirmed that the survey was conducted
two weeks after the holiday, so the value of household food consumption has little chance of appearing
unusually high due to the holiday.
Skoufias, Tiwari, and Zaman 419
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
in 1996 rupiah. A critical point for the construction of real income in 1999 isthe fact that changes in food prices affect households differently depending onthe share of their budgets spent on food. Typically, poorer households spend amuch larger share of their incomes on food—nearly 60 percent for poor ruralhouseholds in Indonesia, compared with 40 percent for urban households atthe top of the expenditure scale.
The SUSENAS consumption module includes data on the value and quantityof food items consumed, allowing calculations of unit values at the householdlevel. A deflator was constructed combining the unit values calculated from theconsumption module with province-specific prices reported for nonfood itemsby the Central Statistical Agency.7 First, because expenditures are only col-lected for nonfood items, a deflator for nonfood items was constructed usingthe mean shares of major groups of nonfood items in the February 1999 surveyas weights and the province-specific price indexes for these groups.8 Second, ahousehold-specific food deflator was calculated from a weighted average of the52 food items used in the calculation of the poverty line in Indonesia.Specifically, the household-specific food deflator was calculated using theformula
PhFð99Þ ¼
X52
i¼1Sh
i ð99ÞPiðR; 96ÞPiðR; 99Þ
� �� ��1
ð1Þ
which is the standard formula for calculating a Paasche price index (Deatonand Zaidi 1999). Si
h denotes the share for food item i of the total amountexpended on the 52 food items, and the superscript h indicates that the sharevaries by household. The second term is the ratio of the median unit value offood item i in region R in 1996 to the corresponding unit value in 1999.Household-specific unit values of food items are replaced by median unitvalues for each of the urban and rural areas of the 26 provinces and theJakarta metropolitan area (a total of 53 regions) to minimize the influence ofmeasurement errors and differences in the quality of food consumed by wealth-ier households (Deaton 1988). With these price deflators for food andnonfood, the overall price deflator for household h in 1999, Ph(99), can beexpressed as
Phð99Þ ¼ WhFð99ÞPh
Fð99Þ þ 1� WhFð99ÞPh
NFðR; 99Þ� �
: ð2Þ
Note that the weights applied to food and nonfood also vary across house-holds. The weight for each household is calculated from the predicted value of
7. More details on the construction of the price indexes can be found in Skoufias (2003). Suryahadi
and others (2000), and Levinsohn, Berry, and Friedman (1999), all of whom take a similar approach to
constructing household-specific price indexes for Indonesia.
8. The province-specific price indexes for food and nonfood groups reported by the Central
Statistical Agency are based on prices for 27 cities in 1996 and 44 cities in 1999.
420 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
the regression of household food share in 1999, WhFð99Þ, on the logarithm of
per capita consumption and the logarithm of household size. This approacheliminates the influence of household-specific, unobserved components (such astaste preferences) on the share of food.
To provide more concrete evidence about the relative price regimes prevail-ing in the two survey years, figures 1a and 1b show the changes in mean pricesper 1,000 calories (1 kilocalorie) paid by rural and urban households between1996 and 1999 in Central Java, a densely populated province with a high con-centration of poor people. Prices per kilocalorie are calculated by dividing thenominal value of household consumption for each food group by the
FIGURE 1. Changes in Price of 1,000 Calories by Food Group Relative toCereals and Tubers in Rural and Urban Central Java, 1996 and 1999
Source: Authors’ calculations based on 1996 and 1999 SUSENAS consumption modules usingexpenditure quartiles.
Skoufias, Tiwari, and Zaman 421
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
kilocalories provided by all the items in the group.9 Poorer households mayconsume food items of lower quality and, as a result, the prices per kilocaloriepaid by these households may be lower than those paid by richer households.To investigate for this possibility, prices per kilocalorie are calculated separate-ly for households at the bottom and top 25 percent of the distribution of totalconsumption per capita in each year.10
Figures 1a and 1b confirm that the relative prices faced by householdschanged considerably between 1996 and 1999.11 In general, the absolute pricesof calories from all food groups seems to have increased dramatically betweenthe two years in both urban and rural areas. Relative to cereals, food groupssuch as meat and fish, fruits and vegetables, and eggs and milk became moreexpensive in both rural and urban areas. There is also considerable heterogen-eity between the rich and the poor in the magnitude of the relative pricechanges. Price increases appear more pronounced for the poor, particularly inurban areas. For example, in urban areas the price of eggs and milk relative tocereals increased 23 percent for the poorest consumers but only 4 percent forthe richest consumers.
Price changes of this magnitude undoubtedly had a large impact on calorieand nutrient availability at the household level. Figures 2a and 2b present thechange in average daily calories per capita between 1996 and 1999 for ruraland urban Central Java. There was a significant reduction in both total andproportional calorie availability. Calorie availability at the household level inCentral Java declined by 8 percent in rural areas and 6 percent in urban areas.There was a much larger reduction in calories derived from food groups richerin micronutrients—meat and fish, fruits and vegetables, and eggs and milk.The figure also shows considerable heterogeneity in the reduction between therich and the poor, with larger reductions for the poor in almost all cases inboth rural and urban Central Java.
The change in calories sourced from cereals and tubers is interesting. Inrural Central Java calories obtained from cereals and tubers fell 12 percentamong the poorest households, but only 1 percent among the richest house-holds. By contrast, roughly the same level of decline was experienced by thepoorest and richest households in urban areas. This discrepancy is largelybecause rich rural consumers are likely to be landowners, possibly engaged inthe production of some of these staples. Thus price increases for these staples
9. The calorie prices reported are derived by dividing expenditures by total calories in the food
group in each year. Thus the price of calories in 1999 may be biased downward depending on the
extent to which households substituted away from more expensive food items within and between
groups.
10. In 1999 the percentiles of total consumption per capita are estimated after dividing
consumption by the deflator discussed earlier.
11. For a related analysis of the impact of the Indonesian crisis on budget shares, with repeated
observations on sampled households, see Thomas and others (1999).
422 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
may have made those rich consumers better off and not affected their con-sumption of cereals and tubers by much.
I I . E M P I R I C A L F R A M E W O R K , A N A L Y S I S , A N D R E S U L T S
Economic theory provides little guidance on how the income elasticity of agiven commodity might change with changes in prices. Given a Marshalliandemand function for any food item xi, summarized by the functionxi ¼ xð�p;M;ZÞ where �p is a vector of relative prices, M is real income, and Z isa vector of preference shifters (such as household demographic characteristics),it follows that, in general, the response of demand to income changes depends
FIGURE 2. Changes in Per Capita Calorie Consumption by Food Group inRural and Urban Central Java, 1996 and 1999
Source: Authors’ calculations based on 1996 and 1999 SUSENAS consumption modules usingexpenditure quartiles.
Skoufias, Tiwari, and Zaman 423
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
on the same set of variables:
qxi
@Mð�p;M;ZÞ:
Without strong (even arbitrary) assumptions about the separability of prefer-ences between and within specific food groups, little can be said (at least froma theoretical perspective) about how changes in prevailing relative prices mightaffect the demand for a commodity in response to income changes. Underthese circumstances this issue can be addressed only empirically. This is preciselythe gap that this article aims to fill. The question addressed empirically iswhether income elasticity differs significantly between a noncrisis year (1996)and a crisis year (1999), which have very different relative price vectors, �p96
and �p99:
qxi
@Mð�p96;M;ZÞ= qxi
@Mð�p99;M;ZÞ ð3Þ
The analysis uses both nonparametric and regression methods that take intoaccount the role of measurement error in total outlay.
The “almost ideal demand system” proposed by Deaton and Muellbauer(1980) provides the empirical framework for the analysis because it makestransparent the dependence of income elasticity on relative prices.12 Startingwith the Working-Leser formulation that relates value shares to the logarithmof total expenditures
wi ¼ ai þ bilogðxÞ ð4Þ
Deaton and Muellbauer propose a way of making the coefficient of totalexpenditures,bi, a function of prices. Given a cost function
logcðu; pÞ ¼ að pÞ þ ubð pÞ ð5Þ
and choosing the functions a(p) and b(p) to be of the form
að pÞ ¼ a0 þX
kaklogpk þ
1
2
Xk
Xlgkl�logpklogpl ð6Þ
bð pÞ ¼ b0Pkpbk
k ð7Þ
the Engle curve can be expressed as
wi ¼ ai þX
jgijlogpij þ bilogðx=pÞ ð8Þ
12. The quadratic Engel curve proposed by Banks, Blundell, and Lewbel (1997) is an extension of
the almost ideal demand system specification that also explicitly recognizes the dependence of the
income elasticity to prices (see equation 11 in their paper).
424 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
where P is the price index defined by
log P ¼ a0 þX
ak log pk þ1
2
Xk
Xlgkllogpklogpl ð9Þ
and the parameters gijS are defined as
gij ¼1
2gij� þ g ji�� �
¼ g ji: ð10Þ
Based on this framework, the question addressed in this article translates towhether the b
0s of equation (7), estimated using cross-sectional data from
1996—the normal year—as the coefficients of the logarithm of total expendi-tures in equation (8), are statistically and economically similar to those esti-mated in 1999—the crisis year.
Nonparametric Regression and the Starchy Staple Ratio
The analysis begins with an investigation of how the starchy staple ratio (SSR)and its income sensitivity vary in a cross-section of households between 1996and 1999. SSR, calculated as the share of total calories obtained from cerealsand tubers, is considered a more useful aggregate measure of householdwelfare than is total caloric availability per capita because SSR captures diver-sity in diets. According to Bennett’s Law, SSR declines with householdincome.13 This analysis adopts a flexible approach that examines Indonesianhouseholds by aggregating the caloric contents of the more than 200 fooditems included in the SUSENAS survey. Given that the income elasticity of cal-ories in Indonesia is known to be nonlinear (Skoufias 2003), nonparametricmethods are used that also allow the income elasticity of the SSR to vary withincome level.
Using y to denote the logarithm of SSR and x the logarithm of per capitatotal household consumption, the regression function is:
mðxÞ ¼ Eðy=xÞ: ð11Þ
Following Subramanian and Deaton (1996) and Deaton (1997), m(x) is esti-mated using a smooth local regression technique proposed by Fan (1993).14 Atany given point x, a weighted linear regression is run of the logarithm of SSRon the logarithm of per capita consumption. The weights are chosen to be thelargest for sample points close to x and to decrease with distance from x. Thedistribution of the logarithm of per capita consumption is divided into 100evenly spaced grids, and local regressions are estimated for each grid instead of
13. Bennett’s law involves the relation between household diet and income, while Engel’s law relates
the share of food expenditures in a household budget to household income. Timmer, Falcon, and
Pearson (1983) provide a detailed discussion of Bennett’s law.
14. Fan (1993) demonstrates the superiority of the smooth local regression technique over kernel
and other methods.
Skoufias, Tiwari, and Zaman 425
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
for each point x in the sample. For the local regression at x, observation i hasthe quartic kernel weight
wiðxÞ ¼15
161� x� xi
h
� �2� �2
ð12Þ
if 2h � x2xi � h and zero otherwise. The quantity h is a bandwidth set totrade off bias and variance, which tends to zero with increasing sample size.This analysis uses a bandwidth of 0.8.
Figures 3a and 3b compare logarithms of SSR of dietary content in 1996and 1999 with logarithms of per capita expenditure (lnPCE), using 1996prices, for rural and urban Central Java. With the income level in 1999 madecomparable to that in 1996, these figures make it possible to examine theeffects of changes in food prices between 1996 and 1999 on household dietarycomposition (as summarized by SSR), assuming that family size did not changesignificantly during that period.
An advantage of these nonparametric figures is that potential biases in themeasurement of calorie availability among higher-income households do notaffect estimates of nutrient income elasticity among poorer ones. For example,figures 3a and 3b make it possible to obtain a sense of the extent to which theexclusion of nutrients obtained from prepared foods (consumed more bywealthier households) affects the nutrient income elasticity of wealthierhouseholds.
In rural Central Java in 1999 SSR was just below the median level of percapita expenditure. SSR was slightly below its level in 1996 for households inthe lower half of the distribution and above it for households in the upper half(see figure 3a). This suggests that in the crisis year the availability of caloriesfrom cheaper food sources (such as cereals and tubers) generally increased forhigher-income groups in rural areas.15 This pattern is even more obvious inurban areas, where the 1999 SSR line lies almost uniformly above the 1996SSR line. Thus for the urban poor there is a rather clear shift to the right in1999, indicating a larger reliance on starchy staples for calories in times ofcrisis. Among rural households the pattern is less clear.
The differences in the slopes of the SSR lines in 1999 and 1996, in bothrural and urban areas, suggest that the responsiveness of SSR to increases inincome varies in crisis and noncrisis years (figures 3c and 3d). In rural areasthe elasticity of SSR relative to income is higher during the crisis year andinvariant to household income at about –0.27 percent. Thus it appears that anincrease in income during a crisis year, such as cash transfers to poor ruralhouseholds, is likely to be more effective in increasing dietary diversity (that is,lowering SSR) than in a noncrisis year.
15. Studdert, Frongillo, and Valois (2001) corroborate the increase in household food insecurity and
compromised diet during the crisis in three Java provinces, including Central Java.
426 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
In contrast, in urban areas, where there was a larger reliance on starchystaples for calories in the crisis year, the income elasticity of SSR was about thesame in 1996 and 1999 for poorer households, becoming smaller than the elas-ticity in the noncrisis year.16 Note, however, that as bivariate plots of calorie
FIGURE 3. Features of the Starchy Staple Ratio in Urban and Rural CentralJava, 1996 and 1999
Source: Authors’ calculations based on 1996 and 1999 SUSENAS consumption modules.
16. Supplemental appendix S3, available at http://wber.oxfordjournals.org, examines whether the
elasticity estimates are significantly different at different levels of outlay by checking whether the
standard error bands for the 1996 estimate overlap those for the 1999 estimate. The figures do not
reveal any significant differences in the income elasticity estimates for the crisis and noncrisis years, a
result confirmed by the regression methods used in the latter part of this article.
Skoufias, Tiwari, and Zaman 427
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
shares at the household level these are descriptive only. The next section per-forms a more detailed analysis of SSR and its elasticity.
Regression Analysis
The analysis in the previous section focused on the bivariate relationshipbetween SSR and per capita expenditure. Though being informative about thegeneral shape of that relationship and how it changed between 1996 and 1999,that analysis cannot account for a number of critical factors—primarily the dif-ferences in the age and gender composition of households, as well as theproblem of correlated errors between household consumption and nutrients.This section examines the elasticity of SSR relative to household consumption,controlling for a set of household characteristics.
The focus on the effects of relative price vectors on the income elasticity ofdemand faces some constraints.17 A typical cross-sectional household surveycollects data within a short timeframe. As a result, most of the price variationfor any household commodity comes from differences in transport costs,market segmentation, the quality of the commodity, and other transactioncosts that can prevent the equalization of prices paid by consumers for thesame commodity.
To the extent that households in different locations are surveyed at differenttimes of year, a survey might also capture seasonal variability in prices. Evenso, it is doubtful that seasonal variation in prices provides an adequate repre-sentation of the change in relative prices that consumers face during an eco-nomic crisis. Household panel data provide an opportunity to circumvent someof these shortcomings. Behrman and Deolalikar (1987), for example, analyzethe relationship of calories to income using data from village surveys conductedby the International Crops Research Institute for the Semi-Arid Tropics(ICRISAT). But even these data shed little light on the relationship betweenhousehold food consumption and spending during a crisis, because the eco-nomic environment was relatively stable during the period of that study.
To compensate for the fact that these are cross-sectional data from surveystaken in different years, not longitudinal data, this analysis uses a flexible speci-fication that provides an explicit test of the difference between the elasticitycoefficients in the precrisis and crisis years. To implement this, the two sets ofcross-sectional data were pooled, then the following regression was run:
ln sSRjkt ¼ a96 þ b96ln PCEjkt þ g 096Xjkt þ m96
� þD99� a99 þ b99ln PCEjkt þ g 096Xjkt
� þ m99 þ 1jkt ð13Þ
17. In much of the literature on the nutrient-income relationship (see Strauss and Thomas 1995),
prices are typically left out of the specification of the Engel curve estimated using cross-sectional data.
This is based on the ad hoc assumption that all households face the same prices. Notable exceptions are
Subramanian and Deaton (1996), Behrman and Deolalikar (1987), Bouis and Haddad (1992), and
Banks, Blundell, and Lewbel (1997).
428 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
where lnSSRjkt is the natural logarithm of SSR for household j that lives incluster k in year t, m96 and m99 are vectors of binary variables identifying thecluster fixed effects in the 1996 and 1999 rounds,18 and D99 is a binarydummy variable equal to one for observations in 1999 and zero otherwise. Thevariable denotes the natural logarithm of real per capita consumption expendi-tures for household j that lives in cluster k in year t. 1jkt is an error term sum-marizing the influence of random disturbances. The set of control variablesXjkt is a vector of household characteristics indicating the logarithm of house-hold size and variables characterizing the age and gender composition of thehousehold, all expressed as ratios of the total family size (the number of chil-dren ages 0–5, the number of children ages 6–12, the number of male andfemale household members ages 13–19 and 20–54, and the number of menolder than 55). Additional binary variables include whether the householdhead is a woman, dummy variables on the education levels of the householdhead and his or her spouse (completed primary, junior high, or senior highschool), and the sectors of employment of the household head and his or herspouse (self-employed, unemployed, or a wage worker).
With this specification, cluster-level fixed effects are allowed to differ acrossthe two survey years, providing control over differing relative prices betweenyears.19 In addition, the coefficients of all the control variables and lnPCEjkt
are allowed to vary across the two years, providing an explicit test of the differ-ence in the income elasticity of the various dependent variables between 1996and 1999. The dummy variable D99 is also able to absorb any other aggregateeffects (aside from relative prices) that might have changed between 1996 and1999.
As first pointed out by Bouis and Haddad (1992), a food expenditure surveycan overstate the nutrient availability in wealthier households, since it iscommon for these households to provide meals to employees and domesticservants.20 In addition, following the 1997–98 crisis, it is plausible that therewas an increase in the frequency of this practice. To minimize potential pro-blems introduced by the fact that the level (and thus the elasticity) of SSR andnutrients may be less accurately measured for wealthier households, the estima-tion limits the sample to the lower half of the distribution of consumption percapita in rural and urban Central Java. Robust standard errors are estimated tocontrol for unknown forms of heteroskedasticity.
18. The SUSENAS survey is a clustered survey with at most 16 households surveyed per cluster each
year. Clusters in 1996 and 1999 have the same code, but it is unclear whether they represent the same
clusters across the two years. Because of this limitation, clusters with the same code in different years
are treated as different clusters.
19. Subramanian and Deaton (1996) use the same approach with cross-sectional data. Cluster-level
fixed effects also take into account other time-invariant local characteristics that determine dietary
intakes and preferences.
20. In the SUSENAS survey, domestic servants are counted as household members.
Skoufias, Tiwari, and Zaman 429
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
In defense of the ordinary least squares (OLS) estimates, it is important to bearin mind that the SUSENAS survey is a seven-day food intake and consumptionsurvey that carefully collects information about food consumed outside the house-hold as well as food received in kind from outside sources. The survey also cap-tures food received as payment for services, food gifts, and food transfers. Sincethe same questionnaire was applied at the same point in time each survey year,there is no reason to believe that biases exist due to these factors.
Yet the possibility remains that correlated measurement errors in total foodconsumption (and thus calorie and nutrient availability) are potential sources ofbias in estimates of income elasticities of nutrients. As first noted by Bouis andHaddad (1992) in a linear version of equation (13), the likelihood that measure-ment errors in nutrient availability are positively correlated with measurementerrors in household consumption implies that this type of measurement error isnot the standard errors-in-variables problem, where coefficients are likely to bebiased toward zero. In the context of correlated measurement errors in the de-pendent and independent variables of a regression, it is unclear whether theupward bias from the correlated errors outweighs the standard downward attenu-ation bias from the measurement error in total consumption. So the direction ofnet bias in income elasticity estimates obtained using OLS methods will generallydepend on the relative size of the correlation between the measurement errors andthe variance of the measurement error in household consumption.
In the case of a log-linear equation such as that of equation (13), Deaton(1997) notes that elasticity estimates using the log of nonfood expenditures asthe sole instrumental variable (IV) are likely to be biased downward, implyingthat the elasticities estimated by OLS and IV methods provide upper and lowerbounds, respectively, for their true value. To address these considerations, theanalysis uses an index of household assets, constructed using the principal com-ponents method, as an instrumental variable for lnPCEjkt. Specifically, in eachsurvey year the index of household assets is estimated using variables for thenumber of cows and buffaloes, sheep and goats, chickens and ducks, and pigsowned by the household.
In addition, a series of dummy variables summarizes the household resi-dence and its environment, such as whether the roof is concrete or tile, thewalls are brick or wood, the floors are tile or cement, the toilet is private orshared, drinking water is accessed through a public network or purchased, andenergy for cooking, lighting, and heating is obtained through the public electricor gas utility. Because the consumption variable interacting with the 1999dummy variable enters into the estimated equation, the instrument used forthis interaction term is the asset index based on 1999 data interacting with the1999 year dummy variable. The results of the first-stage regressions are inappendix B.21 To examine whether the estimated elasticities vary significantly
21. In all cases, a Hausman-type test (Hausman 1978; Holly 1982) for the absence of measurement
error in the consumption variable rejected the null hypothesis.
430 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
for the poor, separate analyses are performed for the entire sample and for thesample of households with per capita expenditure below the median.
Table 1 presents the elasticity estimates for SSR obtained from these regres-sions, fitted separately for the entire sample and for the subset of poorer house-holds. As expected, the negative sign on the point estimates for SSR incomeelasticity is consistent with the nonparametric observation that the share of cal-ories derived from starchy staples declines with income. Further, the compari-son of the elasticity estimates of the IV estimates between urban and ruralareas suggests that SSR elasticity is marginally higher in rural Central Java.This implies that the rate at which households, as their incomes rise, switchfrom cheap sources of calories to food groups that are better sources of micro-nutrients is higher in rural than in urban areas.
The estimated coefficient for b99 in the regression is of particular interestbecause it contains information on how different the elasticities were in 1999(the crisis year) relative to 1996 (the reference year). The estimates for b99 inthe IV specification reported in table 1 suggest that SSR elasticity in 1999 waspractically identical to the elasticity in 1996. The size of the marginal effect is
TA B L E 1. Elasticities of the Starchy Staple Ratio for Rural and UrbanHouseholds in Central Java, 1999
Area and variable
All households Poorer householdsa
Ordinary leastsquares
Instrumentalvariable
Ordinary leastsquares
Instrumentalvariable
RurallnPCE –0.25*** –0.24*** –0.27*** –0.31***
(0.01) (0.01) (0.02) (0.04)Marginal effect
(lnPCE*D99)0.01 0.00 0.04 -0.01
(0.02) (0.02) (0.03) (0.07)UrbanlnPCE –0.26*** –0.22*** –0.31*** –0.21***
(0.02) (0.02) (0.02) (0.07)Marginal effect
(lnPCE*D99)–0.02 0.00 0.05 0.10
(0.02) (0.04) (0.04) (0.11)
*** Significant at the 1 percent level.
Note: Numbers in parentheses are robust standard errors, corresponding to the elasticity esti-mates. Each column represents a separate regression using a wide range of household-level eco-nomic and demographic control variables. Instrumental variable estimates were obtained byinstrumenting the natural logarithms of per capita expenditure with household-specific assetindexes.
a. Defined as the lower half of the distribution based on consumption per capita.
Source: Authors’ analysis based on SUSENAS 1996 and 1999 data.
Skoufias, Tiwari, and Zaman 431
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
small and not statistically significant—implying that the income elasticity ofSSR was invariant to the different relative prices prevailing in 1999.
Comparing these results with those for the subset sample of the poor inCentral Java shows a generally higher elasticity among the poor. This indicatesthat cash transfers could be more effective in reducing the SSR for the poorrelative to the entire sample population. But the fact that the coefficients of theinteraction term are not significantly different from zero implies that cashtransfers may have been ineffective during the crisis year.
Even though the invariance of SSR elasticity to price increases suggests littlesubstitution toward or away from cereals and tubers, the complex pattern ofprice changes in Indonesia during the financial crisis of the late 1990s couldhave induced households to substitute within and among other food groups.To probe deeper into the consequences of such substitutions for the elasticityof micronutrients, the estimation in equation (13) is repeated using the naturallogarithm of the consumption of specific nutrients as dependent variables. Theelasticity coefficients obtained from these regressions for the entire sample arereported in table 2.
All the OLS estimates of the nutrient elasticities for both rural and urbanCentral Java are positive and statistically significant for 1996. For example, inrural areas the estimates for the elasticity range from 0.18 for carbohydrates to0.63 for fat. In urban areas the spread is narrower, ranging from 0.13 for car-bohydrates to 0.44 for fats. The IV estimates, on the other hand, while general-ly statistically significant, appear to be smaller than the corresponding OLSestimates for both urban and rural areas. This suggests that the upward biasfrom the correlated errors may outweigh the standard downward attenuationbias in the OLS estimates. It is also noteworthy that vitamin C is the only nu-trient for which the IV estimates of elasticity are not significant in both urbanand rural areas—suggesting that vitamin C consumption may not be responsiveto income in Central Java, whether in a crisis year or not.
The estimates for b99, the coefficient on the interaction between the loga-rithm of per capita expenditure and the year dummy for 1999, are statisticallysignificant for calories, proteins, fats, carbohydrates, phosphorous, iron, andvitamin B for rural areas and for calories, proteins, fats, calcium, phosphorous,and iron for urban areas. This is evidence of a difference in the elasticity ofthese nutrients in the crisis year relative to the reference year. Moreover, thesedifferences are quite large, particularly for urban areas. The income elasticitiesof calories and iron appear to have doubled in urban areas in 1999. The elasti-cities for proteins, calcium, and phosphorous also nearly doubled. The incomeelasticity of fats more than doubled in 1999 for urban households.
But there are some nutrients for which income elasticity did not changesignificantly between 1996 and 1999: calcium and vitamin A in rural areas andcarbohydrates, vitamin A, and vitamin B in urban areas. An increase in incomeelasticity of any specific nutrient in the crisis year—particularly of the magni-tudes seen in urban areas—may be considered an indicator of deteriorating
432 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E2
.In
com
eE
last
icit
ies
of
Nutr
ients
for
Rura
land
Urb
an
House
hold
sin
Cen
tral
Java,
1999
Are
a,
esti
mat
or,
and
vari
able
Calo
ries
Pro
tein
sFat
sC
arb
ohydra
tes
Calc
ium
Phosp
horu
sIr
on
Vit
am
inA
Vit
am
inB
Vit
am
inC
Ru
ral
Ord
inar
yle
ast
squar
esln
PC
E0.2
5***
0.3
1***
0.6
3***
0.1
8***
0.3
7***
0.2
2***
0.2
8***
0.2
7***
0.1
9***
0.2
0***
(0.0
2)
(0.0
2)
(0.0
4)
(0.0
1)
(0.0
3)
(0.0
2)
(0.0
3)
(0.0
6)
(0.0
2)
(0.0
6)
Marg
inal
effe
ct(l
nP
CE
*D
99)
0.0
5**
0.0
9***
0.0
30.0
7***
0.1
0**
0.1
0***
0.1
6***
–0.0
80.1
1***
–0.0
7(0
.03)
(0.0
3)
(0.0
6)
(0.0
2)
(0.0
4)
(0.0
3)
(0.0
4)
(0.0
8)
(0.0
4)
(0.0
9)
Inst
rum
enta
lva
riab
leln
PC
E0.1
4***
0.1
9***
0.5
0***
0.0
7***
0.2
2***
0.1
0***
0.1
6***
0.1
3***
0.0
7***
–0.0
3(0
.02)
(0.0
2)
(0.0
4)
(0.0
2)
(0.0
3)
(0.0
2)
(0.0
3)
(0.0
5)
(0.0
2)
(0.0
5)
Marg
inal
effe
ct(l
nP
CE
*D
99)
0.0
7**
0.1
0**
0.1
9***
0.0
6**
0.0
70.1
0***
0.1
4***
–0.1
50.1
3**
–0.1
0(0
.03)
(0.0
4)
(0.0
7)
(0.0
3)
(0.0
5)
(0.0
4)
(0.0
4)
(0.0
9)
(0.0
5)
(0.0
9)
Urb
an
Ord
inar
yle
ast
squar
esln
PC
E0.1
9***
0.2
6***
0.4
4***
0.1
3***
0.3
6***
0.2
2***
0.2
5***
0.3
0***
0.1
9***
0.3
1***
(0.0
1)
(0.0
2)
(0.0
4)
(0.0
1)
(0.0
3)
(0.0
2)
(0.0
3)
(0.0
5)
(0.0
2)
(0.0
5)
Marg
inal
effe
ct(l
nP
CE
*D
99)
0.0
6*
0.0
60.2
3***
0.0
40.1
1**
0.0
50.1
1**
0.0
60.0
70.0
5(0
.03)
(0.0
4)
(0.0
6)
(0.0
3)
(0.0
5)
(0.0
4)
(0.0
5)
(0.0
8)
(0.0
5)
(0.0
8)
Inst
rum
enta
lva
riab
leln
PC
E0.1
2***
0.1
8***
0.3
4***
0.0
7***
0.2
6***
0.1
5***
0.1
7***
0.2
2***
0.1
3***
0.0
9(0
.02)
(0.0
2)
(0.0
4)
(0.0
2)
(0.0
3)
(0.0
2)
(0.0
3)
(0.0
6)
(0.0
3)
(0.0
6)
Marg
inal
effe
ct(l
nP
CE
*D
99)
0.1
2**
0.1
6**
0.3
8***
0.0
80.2
3***
0.1
4**
0.1
7**
0.0
70.0
80.1
0(0
.05)
(0.0
7)
(0.1
0)
(0.0
5)
(0.0
8)
(0.0
7)
(0.0
7)
(0.1
2)
(0.0
8)
(0.1
1)
***
Sig
nifi
cant
atth
e1
per
cent
leve
l;**
signifi
cant
atth
e5
per
cent
leve
l;*
signifi
cant
atth
e10
per
cent
leve
l.
Note
:N
um
ber
sin
pare
nth
eses
are
robust
standard
erro
rs,
corr
espondin
gto
the
elast
icit
yes
tim
ates
.E
ach
colu
mn
repre
sents
ase
para
tere
gre
ssio
nusi
ng
aw
ide
range
of
house
hold
-lev
elec
onom
icand
dem
ogra
phic
contr
ol
vari
able
s.A
llsp
ecifi
cati
ons
contr
ol
for
clust
erfixed
effe
cts
and
year
dum
my
vari
able
s.In
stru
men
tal
vari
able
esti
mat
esw
ere
obta
ined
by
inst
rum
enti
ng
the
nat
ura
llo
gari
thm
sof
per
capit
aex
pen
dit
ure
wit
hhouse
hold
-spec
ific
ass
etin
dex
es.
Sourc
e:A
uth
ors
’analy
sis
base
don
SU
SE
NA
S1996
and
1999
dat
a.
Skoufias, Tiwari, and Zaman 433
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
nutritional quality and availability. Consider the income elasticity of fats,which have the highest elasticity of all nutrients for both rural and urban con-sumers. Among households in Central Java, this reflects the “luxury good”status of foods such as meat and fish. The price increases in 1999 also appearto have severely curtailed households’ ability to obtain other basic nutrients,such as protein, calcium, phosphorous, and iron, as well as overall calories.
Table 3 shows elasticity estimates for a restricted sample of the poor (percapita expenditure below the median), which are uniformly higher than for theentire sample population. The IV estimates reveal four broad types of nutrients:those for which income elasticities are statistically indistinguishable from zeroin 1996 and 1999, those for which elasticities are statistically indistinguishablefrom zero in 1996 but significant in 1999, those for which elasticities are statis-tically different from zero but not statistically different between 1996 and1999, and those for which elasticities are statistically different from zero andstatistically different between the two years.
Table 4 categorizes nutrients based on the IV results for Central Java. Itshows that, unlike for the entire sample, the elasticity estimates for poorerhouseholds suggest considerable heterogeneity in the elasticities of nutrients.For the urban poor in Central Java, for example, elasticity in the normal year(1996) is significantly different from zero only for fats. But in the crisis year(1999), elasticity is significantly larger for fats as well as for calories, proteins,carbohydrates, phosphorous, and vitamin B. Elasticities for calcium, iron, andvitamins A and C are never significant, indicating that cash transfers may notbe effective vehicles for protecting these nutrients in poorer households.22
Robustness Check: Is It Prices or Other Factors?
The empirical specification of equation (13) relies on cross-sectional variation inrelative prices to identify income elasticities. Pooling the cross-sectional datafrom the two survey years allowed the construction of a basic test for whetherthe estimated elasticities were different in 1996 and 1999. Although a dummyvariable for year is included in the specification, which would presumably absorbeverything else that changed between the two survey years, there could be linger-ing questions about whether the findings are due to changes in food pricesspecifically or perhaps driven by other, unrelated effects of the economic crisis. Iffood prices could be observed at the household level, then the issue could betested directly. But prices are not observed at the household level. Rather, unitvalues are observed, and these are known to be contaminated by variations inquality. To confirm that changes in food prices are indeed driving these results,
22. A simple calculation based on IV estimates for all households in urban Central Java shows that
a cash transfer of 25 percent of income would increase the consumption of nutrients considered in this
article by from 6–18 percent in the 1999 crisis year—twice the impact range of 3–9 percent for the
same set of nutrients in the 1996 noncrisis year.
434 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E3
.In
com
eE
last
icit
ies
of
Nutr
ients
for
Poore
rH
ouse
hold
sin
Rura
land
Urb
anC
entr
al
Java,
1999
Are
a,
esti
mat
or,
and
vari
able
Calo
ries
Pro
tein
sFat
sC
arb
ohydra
tes
Calc
ium
Phosp
horu
sIr
on
Vit
am
inA
Vit
am
inB
Vit
am
inC
Ru
ral
Ord
inar
yle
ast
squar
eln
PC
E0.2
9***
0.3
6***
0.7
4***
0.2
2***
0.4
5***
0.2
4***
0.3
5***
0.2
8***
0.1
9***
0.2
3**
(0.0
2)
(0.0
2)
(0.0
6)
(0.0
2)
(0.0
4)
(0.0
3)
(0.0
4)
(0.0
9)
(0.0
3)
(0.1
0)
Marg
inal
effe
ct(l
nP
CE
*D
99)
0.1
2***
0.1
4***
0.0
80.1
4***
0.1
1*
0.1
6***
0.2
3***
–0.0
40.2
2***
–0.1
5(0
.03)
(0.0
4)
(0.0
9)
(0.0
3)
(0.0
6)
(0.0
5)
(0.0
5)
(0.1
3)
(0.0
5)
(0.1
5)
Inst
rum
enta
lva
riab
leln
PC
E0.0
30.1
2***
0.4
1***
–0.0
30.1
1*
–0.1
40.0
5–
0.0
5–
0.1
2**
0.3
7***
(0.0
3)
(0.0
4)
(0.0
7)
(0.0
4)
(0.0
6)
(0.0
8)
(0.0
5)
(0.1
1)
(0.0
5)
(0.1
2)
Marg
inal
effe
ct(l
nP
CE
*D
99)
0.0
40.0
60.3
6**
0.0
10.1
70.0
30.1
9*
–0.0
50.1
8–
0.1
6(0
.08)
(0.0
9)
(0.1
6)
(0.0
8)
(0.1
1)
(0.1
7)
(0.1
0)
(0.2
4)
(0.1
3)
(0.2
4)
Urb
an
Ord
inar
yle
ast
squar
esln
PC
E0.2
2***
0.3
1***
0.7
9***
0.1
3***
0.4
9***
0.2
4***
0.3
6***
0.4
1***
0.1
9***
0.4
4***
(0.0
3)
(0.0
3)
(0.0
7)
(0.0
3)
(0.0
6)
(0.0
3)
(0.0
5)
(0.1
1)
(0.0
4)
(0.1
3)
Marg
inal
effe
ct(l
nP
CE
*D
99)
0.1
9***
0.1
4***
0.2
3**
0.2
0***
0.1
00.1
7***
0.1
6**
–0.0
40.2
1***
–0.1
0(0
.05)
(0.0
5)
(0.1
0)
(0.0
5)
(0.0
9)
(0.0
5)
(0.0
7)
(0.1
7)
(0.0
6)
(0.1
9)
Inst
rum
enta
lva
riab
leln
PC
E–
0.0
20.0
50.4
7***
–0.1
10.2
2–
0.0
00.0
60.0
6–
0.0
8–
0.1
7(0
.07)
(0.0
8)
(0.1
7)
(0.0
8)
(0.1
5)
(0.0
8)
(0.1
3)
(0.3
1)
(0.1
0)
(0.3
3)
Marg
inal
effe
ct(l
nP
CE
*D
99)
0.3
8***
0.3
6**
0.7
6***
0.3
5**
0.1
50.3
5**
0.1
80.1
80.3
5*
–0.0
5(0
.13)
(0.1
5)
(0.2
7)
(0.1
5)
(0.2
4)
(0.1
5)
(0.2
1)
(0.5
1)
(0.1
9)
(0.5
2)
***
Sig
nifi
cant
atth
e1
per
cent
leve
l;**
signifi
cant
atth
e5
per
cent
leve
l;*
signifi
cant
atth
e10
per
cent
leve
l.
Note
:N
um
ber
sin
pare
nth
eses
are
robust
standard
erro
rs,
corr
espondin
gto
the
elast
icit
yes
tim
ates
.E
ach
colu
mn
repre
sents
ase
para
tere
gre
ssio
nusi
ng
aw
ide
range
of
house
hold
-lev
elec
onom
icand
dem
ogra
phic
contr
ol
vari
able
s.A
llsp
ecifi
cati
ons
contr
ol
for
clust
erfixed
effe
cts
and
year
dum
mie
s.In
stru
men
tal
vari
able
esti
mat
esw
ere
obta
ined
by
inst
rum
enti
ng
nat
ura
llo
gari
thm
sof
per
capit
aex
pen
dit
ure
wit
hhouse
hold
-spec
ific
ass
etin
dex
es.
Poore
rhouse
hold
are
defi
ned
as
the
low
erhalf
of
the
dis
trib
uti
on
base
don
per
capit
aex
pen
dit
ure
.
Sourc
e:A
uth
ors
’analy
sis
base
don
SU
SE
NA
S1996
and
1999
dat
a.
Skoufias, Tiwari, and Zaman 435
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
regressions of the following form are analyzed for each nutrient:
ln Yj ¼ b0 þ b1 ln PCEj þX
ibiln
�Pl
PCT
�þX
igi ln
�Pl
PCT
�� ln PCEj
þ d 0Xj þ 1j ð14Þ
where i indexes four food groups: meat and fish, fruits and vegetables, eggs andmilk, and others, PCT is the average of the unit values of cereals and tubers at thevillage (desa) level, and Pi is the average unit value of each of the four foodgroups such that the ratios in equation (14) represent the price of each of the fourfood groups relative to the price of cereals and tubers. As before, Xj representsthe set of controls used in previous regressions. Using village-level averages ofunit values avoids the problem of quality differences in household-level purchasesof food.23
These regressions are estimated for 1996 with the primary goal of assessingwhether or not the coefficients on the interaction terms (the gi
0s) are significantly
different from zero. These estimations are made for each of the micronutrientsand separately for urban and rural central Java. These estimations reveal thatat least one interaction term is always statistically different from zero. In add-ition, an F-test of the joint significance of these interactions rejects the null hy-pothesis in almost all instances (see Appendix S4, a supplemental appendixavailable at http://wber.oxfordjournals.org). This makes an explicit connectionbetween the estimated income elasticity of any one micronutrient and the
TA B L E 4. Nutrient Classifications for Poorer Households in Central JavaBased on the Effectiveness of Cash Transfers
Area
Category 1: cashtransfers have no
effect
Category 2: cashtransfers have no
effect in normal years,but useful in crisis
years
Category 3: cashtransfers are effective,but no more or less so
in crisis years
Category 4: cashtransfers are generallyeffective, but no more
so in crisis years
Rural Calories,Carbohydrates,Phosphorous,Vitamin A
Iron Proteins, Calcium,Vitamin B, VitaminC
Fats
Urban Calcium, Iron,Vitamin A,Vitamin C
Calories, Proteins,Carbohydrates,Phosphorous,Vitamin B
Fats
Source: Authors’ analysis based on SUSENAS 1996 and 1999 data. Poorer household aredefined as the lower half of the distribution based on per capita expenditure.
23. Deaton (1988) notes that using cluster means of unit values in regressions of this form is
essentially the same as using cluster dummy variables to instrument for individual unit values in a
regression of shares and prices, where prices are expressed as unit values.
436 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
prevailing relative prices of food in the economy, corroborating the finding thatelasticities estimated under one set of relative prices do not necessarily remainvalid when relative prices change dramatically. In particular, this also providesadditional confidence in the attribution of differences in the income elasticitiesof micronutrients between the normal year and the crisis year to food prices.
I I I . I M P L I C A T I O N S
There is considerable heterogeneity in the income elasticity of demand fornutrients over time based on analysis of household data from the 1996 and1999 consumption modules of the SUSENAS in Indonesia. A comparison ofOLS and IV estimates of the demand for nutrients suggests that OLS estimatesare likely to be misleading due to bias from correlated errors in consumptionand nutrient content. In particular, the finding that IV estimates are generallylower than OLS estimates suggests that the upward bias due to correlatedmeasurement errors in nutrient intake and household-level consumption mayoutweigh the possible attenuation bias.
The analysis also shows that for most nutrients, including micronutrientssuch as phosphorous, iron, and calcium, income elasticity estimates are signifi-cantly higher in a crisis year. Moreover, the magnitude of the increase appearsto be larger in urban than in rural areas. On the other hand, for some micronu-trients, such as vitamin C, the income elasticity estimates obtained from the IVspecification are statistically indistinguishable from zero. This suggests thatincome may have limited leverage to increase or protect the consumption ofvitamin C, whether in a crisis year or not. The separate analysis of nutrientelasticity for poor households reinforces the message that, some nutrients couldbe effectively protected using cash transfers during crisis years (such as ironand fats in rural Central Java) while others are unlikely to be responsive toincome supplements (such as vitamin A in rural Central Java).
These results have two specific policy implications. First, given the signifi-cant increases in the income elasticities of both micronutrients and macronutri-ents during economic crises suggests that cash transfer programs can helphouseholds protect their consumption of essential nutrients, with importantdifferences between the urban and rural poor. To the extent that delivery infra-structure is already in place exists and leakage is low, cash transfers are widelyaccepted as the quickest and cheapest intervention to scale up to reach house-holds most likely to be adversely affected. This research shows that they canalso be more effective in protecting the consumption of some key nutrientsduring economic crises than in normal economic conditions.
Second, a complete reliance on cash transfers may be insufficient if the policygoal of policy in response to economic crises is to protect all important nutrients.For example, consumption of vitamin C, an important micronutrient, was unre-sponsive to income in both rural and urban Central Java. This suggests that tar-geted micronutrient supplementation programs, designed with careful attention to
Skoufias, Tiwari, and Zaman 437
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
differences between the urban and rural poor, might have to accompany cashtransfers to ensure that key micronutrients are not sacrificed during crises. Futureresearch could be directed at understanding and identifying specific nutrients thathouseholds are likely to sacrifice during a crisis in different settings.
A P P E N D I X A . D A T A D E S C R I P T I O N
Data from the consumption module of SUSENAS, collected every three years,included 216 food items in 1996 and 214 food items in 1999.24 The seven-dayfood intake survey makes a very good effort to capture the total value of the foodconsumed by households. In both years, households were asked to report thequantity and value of food purchased, given to them as gifts, or consumed out oftheir own production during the previous week.25 Items are valued by local inter-viewers using the prevailing market prices in the villages where households reside.
The micronutrient content of each food item is calculated using conversionfactors published by the Nutrition Directorate in the Ministry of Health ofIndonesia (Direktorat Gizi, Departemen Kesehatan 1988). That publicationcontains the micronutrient content of a comprehensive list of foods; each wasmatched with one or more of the food items captured in the SUSENAS con-sumption module. Because both the quantities and calories of each food itemare available in the SUSENAS dataset, either may be used to derive the micro-nutrient content. Additional investigation determined that it is preferable torely on the calorie data rather than the quantity data. In 1996, for example,the quantity of various food items was recorded in kilograms rather than ingrams as the questionnaire specified.26 In addition, for a number of food items,quantity was coded in pieces, such as number of eggs, rather than in weight,but calories were provided per unit of weight.27 Similar problems were notedwith the coding of quantities in 1999.
Because of these issues, the analysis in this article used the calorie informa-tion provided by the BPS to derive a more reliable measure of the quantity ofeach food item and micronutrient consumed. First, the standardcalories-to-quantity conversion formula (also applied by the Central StatisticalAgency) was used to derive a new quantity for each food item. Second, thequantity-to-micronutrient formula obtained from the Ministry of Health wasapplied to derive the quantity of micronutrients for each food item. Thisapproach implicitly assumes that the calorie data are more reliable than the
24. The difference arises from the fact that “high quality” and “imported” rice were treated as
separate food items in the cereals category in 1996, but not in 1999.
25. Van de Walle (1988) provides a guide to the SUSENAS consumption module that is still very
useful despite some changes in the questionnaire.
26. In 1996 this was the case for food items with codes 45–52, 95, 102, 110–114, 126–127,
158–166, 171–180, 184–185, 187–194, 208, 212, and 215.
27. In 1996 this was the case for food items with codes 75, 76, 81, 82, 84, 157, 167, and 203.
438 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
original quantity data—a reasonable assumption because the quantity datamust have been processed in some way to apply the standard conversionfactors to calculate calories.
The value of food consumption is the sum of spending on grains, meat andfish, eggs and milk, vegetables, pulses, fruits, seasonings, fats and oils, softdrinks, prepared foods and other food items, and alcohol.28 The referenceperiod for consumption of these items is the seven days preceding the day ofthe survey interview. Weekly consumption was transformed into monthly con-sumption transformed into monthly consumption by multiplying by (30/7) Fornonfood expenditures the survey collects two measures, one for the priormonth and one for the previous 12 months. To avoid exclusion errors, averageexpenditures per month were calculated from the reported expenditures of thelast 12 months. Expenditures on nonfood items include tobacco, housing,clothing, health and personal care, education and recreation, transportationand communication, taxes and insurance, and ceremonial expenses.Expenditures on durables such as household furniture, electric appliances, andaudiovisual equipment are excluded for aggregate household consumption. Ahousehold’s income is measured by monthly per capita consumption, denotedby and calculated by dividing the monthly total of food and nonfood consump-tion in survey period t by the size of the household in the period.29
A P P E N D I X B . F I R S T - S T A G E R E G R E S S I O N S
TA B L E B-1. First-stage Regressions
(1) (2) (1) (2)Variable lnPCE lnPCE*D99 Variable lnPCE lnPCE*D99
Log of householdasset index
0.24*** –0.02*** Household headself-employed withpermanentassistance
0.20*** 0.01
(0.01) (0.00) (0.04) (0.02)Log of household
asset index*D99–0.12*** 0.14*** Household head
working withoutpay
0.00 0.00
(0.01) (0.01) (0.02) (0.01)Number of boys
under age 50.03 0.04 Household head
literate0.00 0.04*
(0.08) (0.03) (0.07) (0.02)
(Continued)
28. Unlike SUSENAS, this article does not include tobacco expenditures in the food consumption
total.
29. It is implicitly assumed that there are no economies of scale at the household level. For
comparing income elasticity over time, this assumption is not overly limiting. In any case, the regression
analysis controls for the gender and age composition of families in each survey year.
Skoufias, Tiwari, and Zaman 439
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TABLE B-1. Continued
(1) (2) (1) (2)Variable lnPCE lnPCE*D99 Variable lnPCE lnPCE*D99
Number of girlsunder age 5
–0.11 0.02 Spouse has noeducation
–0.04 0.01
(0.08) (0.03) (0.03) (0.01)Number of boys ages
6–120.07 0.02 Spouse has not
completed primaryschool
–0.09*** 0.03***
(0.08) (0.03) (0.03) (0.01)Number of girls ages
6–120.06 0.02 Spouse has
completed primaryschool
–0.10*** 0.02**
(0.08) (0.03) (0.03) (0.01)Number of boys ages
13–190.26*** 0.01 Spouse has
completed junioror senior highschool
–0.02 0.01
(0.08) (0.03) (0.03) (0.01)Number of girls ages
13–190.27*** 0.03 Spouse
self-employedwithout assistance
0.00 0.01
(0.08) (0.03) (0.02) (0.01)Number of men ages
20–540.28*** –0.01 Spouse
self-employed withnonpermanentassistance
0.04 0.02*
(0.07) (0.02) (0.03) (0.01)Number of women
ages 20–540.14** 0.01 Spouse
self-employed withpermanentassistance
–0.07 –0.09**
(0.06) (0.02) (0.06) (0.05)Household head has
no education–0.20*** 0.05** Spouse working
without pay0.08** 0.03*
(0.07) (0.02) (0.04) (0.01)Household head has
not completedprimary school
–0.14*** 0.00 Spouse with wageemployment
0.01 0.00
(0.03) (0.01) (0.02) (0.01)Household head has
completed primaryschool
–0.09*** 0.01 Spouse literate 0.07** –0.01
(0.02) (0.01) (0.03) (0.01)Household head has
completed junioror senior highschool
–0.06** –0.01 Household size –0.47*** 0.00
(0.02) (0.01) (0.02) (0.01)(Continued)
440 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TABLE B-1. Continued
Household headself-employedwithout assistance
–0.01 0.01 D99 (¼1 if year is1999)
0.58*** 12.52***
(0.02) (0.01) (0.14) (0.10)Household head
self-employed withnonpermanentassistance
0.04* 0.01 R-squared 0.705 0.999
(0.02) (0.01) F-statistic 148.3 24566
*** Significant at the 1 percent level; ** significant at the 5 percent level; * significant at the10 percent level.
Note: These regressions also include the control variables that interact with the 1999 yeardummy variable.
Source: Authors’ analysis based on equation (13).
RE F E R E N C E S
Banks, J., R. Blundell, and A. Lewbel. 1997. “Quadratic Engel Curves and Consumer Demand.” The
Review of Economics and Statistics 79 (4): 527–39.
Behrman, J. 1988. “Nutrition and Incomes: Tightly Wedded or Loosely Meshed?” Pew/Cornell Lecture
Series on Food and Nutrition Policy. Cornell University, Ithaca, NY.
———. 1995. “Micronutrients and Household Behavior: What We Know and What We Don’t Know.”
Working Paper 2 on Agricultural Strategies for Micronutrients. International Food Policy Research
Institute, Washington DC.
Behrman, J., and A. Deolalikar. 1987. “Will Developing Country Nutrition Improve With Income? A
Case Study for Rural South India.” Journal of Political Economy 95 (3): 492–507.
———. 1989. “Is Variety the Spice of Life? Implications for Calorie Intake.” The Review of Economics
and Statistics 71 (4): 666–72.
Behrman, J., and B. L. Wolfe. 1984. “More Evidence on Nutrition Demand: Income Seems Overrated
and Women’s Schooling Underemphasized.” Journal of Development Economics 14 (1–2): 105–26.
Behrman, J., A. Foster, and M. R. Rosenzweig. 1997. “The Dynamics of Agricultural Production and
the Calorie-Income Relationship: Evidence from Pakistan.” Journal of Econometrics 77 (1):
187–207.
Bouis, H. 1991. “The Determinants of Household-level Demand for Micronutrients: An Analysis for
Philippine Farm Households.” International Food Policy Research Institute, Washington, DC.
Bouis, H., and L. Haddad. 1992. “Are Estimates of Calorie-Income Elasticities Too High? A
Recalibration of the Plausible Range.” Journal of Development Economics 39 (2): 333–64.
Chernichovsky, D., and O. Meesook. 1984. “Urban-Rural Food and Nutrition Consumption Patterns in
Indonesia.” World Bank Staff Working Paper 670. World Bank, Washington DC.
Deaton, A. 1988. “Quantity, Quality, and the Spatial Variation of Price.” American Economic Review
78 (3): 418–30.
———. 1997. The Analysis of Household Survey Data. Baltimore, Md.: The Johns Hopkins University
Press.
Deaton, A., and J. Muellbauer. 1980. “An Almost Ideal Demand System.” American Economic Review
70 (3): 312–26.
Deaton, A., and S. Zaidi. 1999. “Guidelines for Constructing Consumption Aggregates for Welfare
Analysis.” Princeton University, Princeton, NJ.
Skoufias, Tiwari, and Zaman 441
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Gizi, Direktorat, and Departemen Kesehatan. 1988. Daftar Komposisi Bahan Makanan. Jakarta:
Penerbit Bhratava karya Aksra.
Fan, J. 1993. “Local Linear Regression Smoothers and Their Minimax Efficiencies.” Annals of Statistics
21 (March): 998–1004.
FAO (Food and Agriculture Organization). 2008. “Food Outlook.” Rome. http://www.fao.org/docrep/
011/ai474e16htm.
Hausman, J. A. 1978. “Specification Tests in Econometrics.” Econometrica 46 (6): 1251–71.
Holly, A. 1982. A Remark on Hausman’s Specification Test.” Econometrica 50: 749–59.
Ivanic, M., and W. Martin. 2008. “Implications of Higher Global Food Prices for Poverty in
Low-income Countries.” Agricultural Economics 39 (s1): 405–16.
Levinsohn, J., S. Berry, and J. Friedman. 1999. Impacts of the Indonesian Economic Crisis: Price Changes
and the Poor. NBER Working Paper 7194. Cambridge, MA: National Bureau of Economic Research .
Pitt, M. M. 1983. “Food Preferences and Nutrition in Rural Bangladesh.” The Review of Economics
and Statistics 65 (1): 105–14.
Pitt, M. M., and M. R. Rosenzweig. 1985. “Health and Nutrient Consumption across and within Farm
Households.” The Review of Economics and Statistics 67 (2): 212–23.
Ravallion, M. 1990. “Income Effects on Undernutrition.” Economic Development and Cultural Change
38 (3): 489–515.
Sari, M., S. de Pee, M.W. Bloem, K. Sun, A. L. Thorne-Lyman, R. Moench-Pfanner, N. Akhter,
K. Kraemer, and R. D. Semba. 2010. “Higher Household Expenditure on Animal-Source and
Nongrain Foods Lowers the Risk of Stunting among Children 0-59 Months Old in Indonesia:
Implications of Rising Food Prices.” Journal of Nutrition 140 (1): 195–200.
——. 2003. “Is the Calorie-Income Elasticity Sensitive to Price Changes? Evidence from Indonesia.”
World Development 31 (7): 1291–1307.
Strauss, J., and D. Thomas. 1995. “Human Resources: Empirical Modeling of Household and Family
Decisions.” In Handbook of Development Economics vol. 3, ed. J. Behrman, and T. N. Srinivasan.
Amsterdam: North-Holland.
Studdert, L., E. Frongillo, and P. Valois. 2001. “Household Food Insecurity Was Prevalent in Java
during Indonesia’s Economic Crisis.” Journal of Nutrition 131 (10): 2685–91.
Subramanian, S., and A. Deaton. 1996. “The Demand for Food and Calories.” Journal of Political
Economy 104 (1): 133–62.
Suryahadi, A., S. Sumarto., Y. Suharso, and L. Pritchett. 2000. “The Evolution of Poverty during the
Crisis in Indonesia, 1996 to 1999 (Using Full SUSENAS Sample).” SMERU Working Paper. Social
Monitoring and Early Response Unit, Jakarta.
Thomas, D., E. Frankenberg, K. Beegle, and G. Teruel. 1999. “Household Budgets, Household
Composition, and the Crisis in Indonesia: Evidence from Longitudinal Household Survey Data.”
Paper presented at the 1999 Population Association of America Meetings, RAND, Santa
Monica, CA.
Timmer, P. 1981. “Is There Curvature in the Slutsky Matrix?” The Review of Economics and Statistics
62 (3): 395–402.
Timmer, P., and H. Alderman. 1979. “Estimating Consumption Parameters for Food Policy Analysis.”
American Journal of Agricultural Economics 61 (S): 982–94.
Timmer, P., W. P. Falcon, and S. Pearson. 1983. Food Policy Analysis. Published for the World Bank.
Baltimore, Md.: The Johns Hopkins University Press.
Tiwari, S., and H. Zaman. 2010. “The Impact of Economic Shocks on Global Undernourishment.”
Policy Research Working Paper 5215. World Bank, Washington, DC.
Van de Walle, and Dominique. 1988. “On the Use of the SUSENAS for Modeling Consumer Behavior.”
Bulletin of Indonesian Economic Studies 24 (2): 107–22.
Wodon, Q., and H. Zaman. 2010. “Higher Food Prices in Sub-Saharan Africa: Poverty Impact and
Policy Responses.” The World Bank Research Observer 25 (1): 157–76.
442 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Economic Geography and Economic Developmentin Sub-Saharan Africa
Maarten Bosker and Harry Garretsen
Sub-Saharan Africa’s (SSA) physical geography is often blamed for its poor economic per-formance. A country’s geographical location does, however, not only determine its agri-cultural conditions or disease environment. It also pins down a country’s relative positionvis-a-vis other countries, affecting its ease of access to foreign markets. This paper assessesthe importance of market access for manufactures in explaining the observed income dif-ferences between SSA countries over the period 1993–2009. We construct yearly, theory-based measures of each SSA country’s market access using the information contained inbilateral manufacturing trade flows. Using these measures, we find a robust positive effectof market access on economic development that has increased in importance during thelast decade. Interestingly, when further unraveling this finding, access to other SSAmarkets in particular turns out to be important. JEL codes: O10, O19, O55, F1
Sub-Saharan Africa (SSA) is home to the world’s poorest countries. Alongsidefactors such as poor institutional quality, low (labour) productivity and lowlevels of human capital, the region’s geographical disadvantages are oftenviewed as an important determinant of its dismal economic performance. Acountry’s geography directly affects economic development through its effecton disease burden, agricultural productivity, the availability of naturalresources, or its accessibility (see Gallup and others 1999; Collier andGunning, 1999; Ndulu, 2007; Nunn and Puga, 2011).
Recently, the new economic geography (NEG) literature (see Krugman,1991; Fujita and others 1999; World Bank 2008) has highlighed anothermechanism through which geography affects a country’s prosperity. A country’s
Maarten Bosker ([email protected]; corresponding author) is assistant professor at the Erasmus
University Rotterdam. He is also affiliated with the CEPR, the Tinbergen Institute and Utrecht
University, The Netherlands. Harry Garretsen ( [email protected]) is professor at the University of
Groningen, The Netherlands. He is also affiliated with Cambridge University and CESifo. The authors
thank Rob Alessie, Bernard Fingleton, Henri Overman, Giacomo Pasini, Joppe de Ree, Steve Redding,
Marc Schramm and seminar participants in Cambridge, Glasgow, Milan, Oxford, Rome, Rotterdam,
Savannah, and Utrecht for useful comments and discussions on an earlier version of this paper. In
particular we thank the editor, Elisabeth Sadoulet, and three anonymous referees for very helpful
remarks that have significantly improved our paper.
THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 443–485 doi:10.1093/wber/lhs001Advance Access Publication February 14, 2012# The Author 2012. Published by Oxford University Press on behalf of the International Bankfor Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions,please e-mail: [email protected]
443
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
location not only determines its physical geography; it also pins down its pos-ition on the globe vis-a-vis all other countries (its relative geography). TheNEG literature in particular emphasizes the important role of relative geog-raphy in determining a country’s access to international markets. It predictsthat the better this market access, the higher a country’s level of income.
Redding and Venables (2004) were the first to establish empirically thatmarket access indeed matters for economic development. Based on the estima-tion results for a worldwide sample of 101 countries, they find for examplethat if Zimbabwe were located in central Europe, the resulting improvement inits market access would ceteris paribus increase its per capita income by almost80 percent. Subsequently, several other studies confirmed the positive effect ofmarket access on economic development (e.g. Knaap 2006; Breinlich 2006; orMayer, 2008). More recently, it has also been found to hold for developingcountries (see Deichmann, Lall, Redding and Venables, 2008 for a good over-view). Amiti and Cameron (2007) show that wages are higher in Indonesiandistricts that enjoy better market access, Hering and Poncet (2010) and Boskerand others (2010) find similar evidence in case of Chinese cities, and Fally andothers (2010) do so for Brazilian states.1 The importance of relative geographyin shaping global and regional patterns of economic development has also notgone unnoticed in policy circles: it was the main topic of the World Bank’s2009 World Development Report (World Bank, 2008).
Despite the attention given to the role of economic geography in shapingpatterns of economic development in both the developing and developedworld, we are unaware of a study that empirically establishes whether, and ifso, to what extent, it can help to explain the observed differences in economicdevelopment between SSA countries.2
SSA is only a marginal player in the world’s export and import markets.Since 1970, the region’s share in global trade has declined from about4 percent to a mere 2 percent in 2005 (IMF, 2007). Through their detrimentaleffect on market access, high trade costs are generally viewed as one of themain causes for SSA’s poor trade performance (see Freund and Rocha, 2011;Collier, 2002; Foroutan and Pritchett, 1993; Coe and Hoffmaister, 1999;Limao and Venables, 2001; Amjadi and Yeats, 1995; Portugal-Perez andWilson, 2008). Increasing SSA participation in world markets, as well as
1. Moreover, Amiti and Javorcik (2008) find that market access positively affects the amount of FDI
in Chinese provinces and Lall, Shalizi and Deichmann (2004) show that market access is an important
determinant of firm level productivity in India.
2. The only paper we know of focusing on the role of market access in SSA is Elbadawi, Mengistae
and Zeufack (2006) that shows that differences in terms of export performance between firms in 10 SSA
countries and firms in other developing countries (e.g. India, China, Malaysia or Peru) can partly be
explained by SSA’s poor market access. Their paper does not link export performance — or market
access — to income per capita. Another paper that is similar in spirit to ours is that by Arora and
Vamvakidis (2005), which looks at how South Africa’s economy influences development in the rest of
SSA.
444 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
stimulating trade relations between SSA countries, is viewed as very importantto its future economic success (IMF, 2007; World Bank 2007; or UNCTAD2010). Especially, expanding the (exporting) manufacturing sector is seen ascrucial to the region’s chances on future economic success (Collier andVenables, 2007; IMF, 2007; World Bank, 2007). It has been one of the keyingredients of the sustained growth witnessed in the rapidly developing Asiancountries (see e.g. Johnson et al., 2006; or Jones and Olken, 2008). Developingan exporting manufacturing sector will not only help to diversify SSA coun-tries’ export portfolio, making them less vulnerable to price fluctuations onworld commodity markets, it is also expected to increase overall productivitythrough increased knowledge spillovers and learning by doing (VanBiesebroeck, 2005; Bigsten and Soderbom, 2006).
As a result, improving the region’s market access by investing in infrastruc-ture, stimulating regional integration, or providing preferential access toEuropean and U.S. markets are all seen as a vital ingredients for improvingSSA’s trade potential and its overall economic performance (Buys et al., 2010;UNCTAD, 2009, 2010; Frazer and van Biesebroeck, 2010; Freund and Rocha,2011).
Against this background the main contribution of our paper is to empiricallyestablish the importance of SSA market access, and market access for manufac-tures in particular, for its economic development over the last two decades.3
To do this, we follow the empirical strategy introduced by Redding andVenables (2004) that is firmly based in the theoretical new economic geography(NEG) literature. We first construct yearly measure(s) of each SSA country’smarket access over the period 1993–2009, making use of bilateral manufactur-ing export data involving at least one SSA country. Next, and by using our con-structed measure(s) of market access, we estimate the impact of market accesson GDP per worker. We do adapt the Redding and Venables (2004) strategy inthree different ways. First, when constructing our market access measure(s)using the information contained in bilateral trade flows, we take explicitaccount of the fact that most SSA countries only trade with a fraction of theirpossible trade partners. Second, our 17-year sample period allows us to usepanel data methods and control for time-invariant unobserved heterogeneity inall our estimations (hereby most notably capturing all possible differences inSSA countries’ physical geography). Finally, we distinguish explicitly betweenthe importance of access to other SSA markets and to markets in the rest of theworld (ROW).
Overall, our main finding is that economic geography is an important deter-minant of SSA’s economic development. Even after controlling for many otherposited explanations of SSA’s poor economic performance such as its physicalgeography, education levels, or institutional quality, market access for
3. Throughout the paper, unless explicitly noted otherwise, market access refers to a country’s
market access for manufactures.
Bosker and Garretsen 445
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
manufactures has a significant positive effect on GDP per worker. The effect ofmarket access that we find for SSA countries is, however, significantly lowerthan that found in comparable studies looking at Brazil (Fally and others2010), Indonesia (Amiti and Cameron, 2007), or China (Hering and Poncet,2010). But, although lower than those found in other parts of the world, ourresults do show that the effect of market access has increased markedly in SSAover the last two decades: the positive relationship between market access andeconomic development is strongest in the second half of our sample period.
Interestingly, when further unraveling this finding by distinguishing betweenthe importance of access to other SSA markets and to markets in the rest of theworld (ROW), we find that it is the variation in access to other SSA markets inparticular that drives our findings. ROW market access loses its significanceafter controlling for other (more standard) explanations for SSA’s poor eco-nomic performance. It confirms the (increased) importance of SSA markets forSSA’s own economic development (see also Easterly and Reshef, 2010, foot-note 2; UNCTAD, 2009, 2010).
Finally, based on our estimation results, we tentatively ‘decompose’ the con-tribution of policy-relevant variables to overall market access, and look at thepredicted effect of several (policy induced) changes aimed at improving SSAmarket access. This shows that improving SSA infrastructure (see also Buys andothers, 2010), alleviating the burden of landlocked countries, and increasing re-gional economic integration, all positively affect a country’s market access invarying degrees, carrying important benefits for its economic development.
I . E C O N O M I C D E V E L O P M E N T A N D M A R K E T AC C E S S : T H E O R E T I C A L
F R A M E W O R K A N D E M P I R I C A L S T R A T E G Y
At the heart of our analysis lies the theoretical relationship between marketaccess in manufactures and income levels that follows from standard economicgeography theory. Referring to Appendix C for a more formal exposition ofthe NEG model that underlies our analyses,4 this relationship is shown in log-linear form in equation (1) below (corresponding to equation (C5) in AppendixC, with an added subscript t to denote years):
ln wit ¼ gt þ x1 ln cit þ x2 lnXR
j
mjtTð1�sÞijt
zfflfflfflfflfflffl}|fflfflfflfflfflffl{MAijt
|fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl}MAit
¼ gt þ x1 ln cit þ x2 ln MAit ð1Þ
4. See Fujita, Krugman, and Venables (1999), Puga (1999), Head and Mayer (2004) for more
detailed expositions of various basic NEG models. See also Head and Mayer (2010), who show that the
relationship between market access and economic development not only follows from NEG models but
can be derived from a more general class of models.
446 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Equation (1) is the so-called wage equation that lies at the heart of virtually allempirical NEG studies (see e.g. Hanson, 2005; Redding and Venables, 2004;Knaap, 2006 and Amiti and Cameron, 2007; Hering and Poncet, 2010). It pre-dicts that wages in country i in year t, wit, are higher the better a country’s pro-duction efficiency, cit, and, most importantly for our present purposes, thebetter its so-called real market access MAit.
5
This market access is a trade cost (Tij) weighted sum of all countries’ marketcapacities (mj). Each country j’s contribution to country i’s market access con-sists of country j’s market capacity, a reflection of its real spending power,weighted by the level of trade costs incurred when shipping goods fromcountry i to country j, i.e. MAij ¼mj /Tij
s21. The closer (or better connected) acountry is to world markets, the better its market access. It is equation (1) thatconstitutes the backbone of our empirical analysis into the relevance of marketaccess for SSA economic development (see section III).
Estimating the Wage Equation: The Redding and Venables (2004) Approach
Estimating equation (1) is not as straightforward as it may seem. The difficultycomes from the fact that the market access term, MAit, is not directly observed.Two estimation strategies have been proposed to deal with this issue.
The first strategy follows Hanson (2005) and estimates equation (1) directly.A drawback of this method is that it requires additional assumptions on howto proxy each country’s market capacity, mj. Particularly problematic in thisrespect is the fact that a country’s price index of manufacturing varieties, oneof the two main components of mj (see Appendix C) is not directly observed.6
This is why we base our empirical analysis on the second proposed strategyto estimate (1). This method does not face the problem of having to makead-hoc assumptions on how to proxy a country’s market capacity. It was firstintroduced by Redding and Venables (2004)7 and involves a two-step
5. In Redding and Venables (2004) and Knaap (2006), each firm also uses a composite intermediate
input (made up of all manufacturing varieties) in production, allowing them to also look at the
relevance of so-called supplier access for income levels. Since our goal is to establish the relevance of
market access we, in line with Breinlich (2006), skip intermediate inputs and thereby ignore supplier
access [this also has the advantage that we avoid the multicollinearity problems when including both
market and supplier access in the estimations, see Redding and Venables (2004) and Knaap (2006)]. In
this respect our derivation and application of the wage equation is similar to Hanson (2005), see also
Head and Mayer (2004, pp. 2622–2624), or Head and Mayer (2010). As a robustness check, Table 4c
shows results when also including constructed measures of supplier access to (1).
6. Moreover, this direct estimation strategy jointly identifies the relative importance of the different
components making up a country’s overall market access and the overall effect of market access on
income levels. It does so solely from the spatial distribution of GDP (per capita) across countries. The
nonlinear nature of (1) makes this an impossible task without putting a priori restrictions on (some) of
the parameters (see e.g. Amiti and Cameron, 2007). Econometrically, the parameter on market access,
x2, and the parameters within the market access term (e.g. s) are not separately identified when directly
estimating (1).
7. Other papers using this strategy include inter alia Knaap (2006), Breinlich (2006), Mayer (2008),
Head and Mayer (2006, 2010), Hering and Poncet (2010) or Bosker and Garretsen (2010).
Bosker and Garretsen 447
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
procedure. In a first step, additionally collected information contained in (bilat-eral) trade data is used to provide estimates of the (relative) role of trade costs,Tij, and market capacity, mj, in determining a country’s market access. Theway this is done is firmly based on NEG theory. As derived in Appendix C,equation (C6) shows that the connection between bilateral exports and marketaccess follows directly from a standard NEG model (where we have againadded a subscript t to denote years).
EXijt ¼ sit ½mjtTð1�sÞijt �|fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl}
MAijt
ð2Þ
Exports EXij from country i to country j depend on the ‘supplier capacity’ ofthe exporting country, si (see (C6) for its definition), the market capacity of theimporting country, mj, and the magnitude of bilateral trade costs Tij betweenthe two countries. Comparing MAij in (1) and (2) immediately shows that wecan estimate (2), and use the resulting predicted values of MAijt to constructyearly measures of each country’s market access. More formally:
A. Estimate the bilateral export equation (2) in log-linear form using informa-tion on bilateral export flows and trade costs, capturing each country’s supplierand market capacity by a full set of importer-year and exporter-year dummies:
ln EXijt ¼ rit þ m jt þ b ln Tijt þ 1ijt ð3Þ
B. Use m jt and b , the estimated parameters on the included importer-yeardummies and on trade costs respectively, to construct each country’s market accessbased on the direct relationship between bilateral exports and market access [againcompare (1) to (2)]:
MAit ¼X
j
expðm jtÞTbijt|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}
MAijt
ð4Þ
The constructed measure of each country’s market access shown in (4) is then usedin the second step to get an estimate of the impact of market access on incomelevels:
ln wit ¼ gt þ x2 ln MAit þ hit ð5Þ
where the error term hit in (5) captures a country’s level of technological efficiency[cit in (1)]. The estimated value of x2, together with its standard deviation, is themost important parameter for the purpose of our paper. It provides us with an in-dication of the size, sign, and significance of the effect of market access on incomelevels.
In the next two sections we implement the above-described two-step methodin order to verify the importance of market access for SSA economic
448 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
development over the last two decades (1993–2009). In section II, we focus onestimating the trade equation and constructing our measure(s) of marketaccess. Next, in section III, we estimate the wage equation making use of ourconstructed measures of market access and show that market access is of in-creasing importance in understanding the observed differences in economic de-velopment between SSA countries. Moreover, we decompose each country’smarket access into access to other SSA markets and to markets in the ROW,and show that having good access to other SSA markets has become particular-ly important over the last two decades.
I I . E S T I M A T I N G T H E T R A D E E Q U A T I O N A N D C O N S T R U C T I N G
M A R K E T AC C E S S
The starting point of our empirical analysis is the trade equation (3) capturingeach country’s supplier and market capacity by an exporter-year and importer-year dummy respectively. In order to estimate (3), we collected information onyearly bilateral manufacturing exports to and from SSA countries over theperiod 1993–2009.8 We take this data from the UN COMTRADE database,focusing on manufacturing goods as defined by the Standard InternationalTrade Classification (SITC Rev.3). It contains information on bilateral manu-facturing export flows from each SSA country to and from 47 other SSA coun-tries and 153 countries in the rest of the world.
We think a particular focus on manufacturing exports is warranted for twoimportant reasons. First, it most closely follows the NEG theory that underliesour analysis. NEG theory only predicts a relationship between market access inthe manufacturing sector and income levels (see Appendix C). It is not evidentfrom theory that a similar relationship should hold for primary goods’ marketaccess (trade patterns of which are more likely to be dominated by more stand-ard comparative advantage or Heckscher-Ohlin type forces). Taking totalexport flows when estimating (3) is likely to give biased estimates of the para-meters needed to build our measures of market access, particularly in SSAwhere overall exports are dominated by exports of natural resources and/oragriculture (although this dominance varies substantially between SSA coun-tries, see Figure B2 and Table B2 in Appendix B).9
8. See Appendix A for a full list of variables (including data sources) that we use in our analysis.
9. This problem is much less present when looking at different samples of countries (e.g. Europe,
North America and even South-East Asia and parts of Latin America) where trade is dominated by
manufacturing goods. We have also done our analysis using total bilateral SSA exports as the basis for
constructing our market access measures. When using these measures we do not find a significant effect
of market access on economic development (results are available upon request). This could be an
indication that relative location to markets for a country’s natural resources (which dominate SSA
exports to the rest of the world) does not matter. However, given that (as stressed in the main text) a
theoretical underpinning of a relationship between market access for primary products and economic
development is lacking, we decided not to particularly stress this finding. We do add controls for a
country’s economy’s dependence on natural resources when estimating (5).
Bosker and Garretsen 449
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Second, developing the (exporting) manufacturing sector is viewed by manyas crucial to the region’s chances on future economic success (Collier andVenables, 2007; IMF, 2007; World Bank, 2007). Previous spells of sustainedgrowth (mostly experienced by Asian countries) were all accompanied by arapid expansion of international trade, and trade in manufacturing goods inparticular (see e.g. Johnson and others 2006 or Jones and Olken, 2008). In thisrespect it is interesting to note that manufacturing goods already dominate SSAexports to the rest of Africa (see UNCTAD, 2009; 2010 [see Annex 4.5 for acountry-by-country overview]). They constitute an average 40 percent of totalintra-SSA exports. When disregarding primary exports (fuel, ores, minerals,etc.), that account for roughly 75 percent of total exports to the ROW, thesame is true of SSA exports to the ROW.
Next, we need to decide on how to measure trade costs, Tijt. The NEG-modeldoes not specify trade costs in any way (except that they are of the iceberg type).In the absence of actual trade cost data and following the modern empiricaltrade and economic geography literature (see e.g. Anderson and van Wincoop,2004; Limao and Venables, 2001; Redding and Venables, 2004; Bosker andGarretsen, 2010), we specify Tijt to be a multiplicative function10 of the follow-ing observable variables that are commonly used in the literature11: bilateral dis-tance (Dij), sharing a common border (Bij), a common language (CLij), or acommon colonial heritage [distinguishing between sharing a common colonizer(CCij) and having had a colony-colonizer relationship (CRij)], and finally adummy variable indicating membership of the same African regional or freetrade agreement (RFTAijt) in year t (see Appendix A for a full list of the RFTAsincluded in this definition). In loglinear form this amounts to substituting thefollowing trade costs specification for b ln Tijt in (3):
b ln Tijt ¼ d1 ln Dij þ d2 ln Bij þ d3 ln CLij þ d4 ln CCij þ d5 ln CRij
þ d6RFTAijt ð6Þ
Overall, this results in the following bilateral export equation that we estimatefor each year of our sample period separately (which explains the added sub-script t to all coefficients):
ln EXijt ¼ rit þ m jt þ d1t ln Dij þ d2t ln Bij þ d3t ln CLij þ d4t
ln CCij þ d5t ln CRij þ d6tRFTAijt þ 1ijt
ð7Þ
10. This is the usual choice in the gravity literature (see e.g. Limao and Venables, 2001;
Subramanian and Tamarisa, 2003). See Hummels (2001) for a critique on this, arguing in favor of an
additive specification instead.
11. Tariffs are also an important component of trade costs. However, using the available SSA tariff
data available in UN TRAINS, reduces our sample from 16560 to 546, 1251 or 4859 for 1993, 2000
and 2009 respectively. For this reason we excluded tariffs from our trade cost specification.
450 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Equation (7) forms the basis for constructing our yearly measures of each SSAcountry’s market access over the 1993–2009 period.
Estimating the Trade Equation. Dealing with the ‘Zeroes’ in Bilateral SSATrade
The actual estimation of (7) raises a number of issues of its own. In particular,the presence of zero trade flows complicates matters. The average SSA countryexports manufacturing goods to only 20 percent of possible partner countries,so that about 80 percent of bilateral SSA manufacturing export flows arezeroes. And, although the number of zeroes drops over our sample period(from 89 percent in 1993 to 77 percent in 2003), it does complicate matterswhen estimating the parameters of (7) that we need to construct our measuresof market access. Failing to adequately take account of these zeroes results ininconsistent estimates of these parameters, and thus in wrongly constructedmarket access measure(s).
To deal with these zero observations, several estimation strategies have beenproposed that each have their (dis)advantages. We follow Helpman and others(2008) and use a Heckman 2-step estimation strategy to estimate the para-meters of (7). This method has the virtue of not having to impose exogenoussample selection, that is, assuming that there is no unobserved variable relatedto both the probability to trade and the amount of trade [as e.g. discarding thezero observations and applying OLS on the non-zeroes only, or applyingzero-inflated Poisson or negative binomial methods do]. Nor do we have toassume a priori that the exact same model explains both the zero and thenon-zero bilateral trade flows [as e.g. using Tobit, NLS or pseudo-Poisson(PPML) techniques would imply].12
The Heckman 2-step procedure amounts to first estimating, using probit,how each of the included explanatory variables affects the probability to trade.Next, in the second stage, the effect of each variable on the amount of trade isestimated, including the inverse Mills ratio (constructed using the results fromthe first step) to control for the endogenous selection bias that would plaguethe results when simply discarding the non-zero observations (see for instancech. 17 in Wooldridge, 2003). However, using the Heckman 2-step procedure isalso not free of assumptions: results are only convincing when one can rely ona valid exclusion restriction: a valid exclusion restriction: that is, having at least
12. Note that, due to the assumed CES utility function, the NEG model set out in Appendix C in
principle implies that each country trades at least something with every other country. This implies that
using the NEG trade equation in explaining both the zero and the non-zero trade flows ascribes the zero
observations to the error term only (relying on arguments of measurement error or reporting errors, see
also Santos Silva and Tenreyro, 2006, p. 643). Although maybe defendable when looking at samples
with a limited amount of ‘zeroes’, we think this is very unlikely in our SSA case, where about
80 percent of the observations are zeroes.
Bosker and Garretsen 451
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
one variable that determines the probability to trade but not the amount oftrade (see Wooldridge, 2003, p. 589).13
The choice of such a variable is generally quite difficult. However, in ourcase we can build on a recent paper by Helpman and others (2008), and usetheir suggested measure of the religious similarity of two countries as the vari-able explaining the probability to trade but not the amount of trade condition-al upon trading. The economic rationale behind the use of this variable isfirmly based on recent trade models that show that in order to trade at all,exporters have to be able to cover the fixed costs of exporting. The higherthese costs between two countries, the higher the probability of not observingany bilateral trade between them. Helpman and others (2008, p. 466) showthat religious (dis)similarity serves as a useful proxy of these fixed costs,14 andthey moreover show convincing evidence that, econometrically, it can not berejected as a valid ‘instrument’.15
Using Helpman and others’s (2008) religious similarity variable to fulfill the(necessary) exclusion restriction, we estimate (7) for each year in our sampleperiod separately.16 To explicitly allow intra-SSA trade to be differently influ-enced by our included variables, we interact all variables, including allimporter- and exporter-dummies, with an intra-SSA trade dummy-variable.Table 1 shows the results. In Table 1, the postfix “- SSA” denotes that a vari-able is interacted with this intra-SSA trade dummy. Significance of an“SSA”-variable indicates a significantly different effect of that particular
13. Another disadvantage of the Heckman two-step method is that it does not adequately take
account of the heteroscedasticity inherently present in bilateral trade data (see Santos Silva and
Tenreyro, 2006). However, we think that the disadvantages of the current methods available that do do
this (see the discussion in the text and also footnote 12), that is, either assuming exogenous sample
selection (OLS on the non-zeroes or zero-inflated Poisson) or imposing that the zero trade flows are the
result of measurement or reporting errors (PPML), do not outweigh the ability of the Heckman
two-step procedure to take account of endogenous sample selection.
14. We use religious similarity in all main results presented in this paper. We have also looked into
the possibility of using the other two ‘instruments’ proposed by Helpman and others (2008): “the
number of days and procedures needed to start a business” and “the costs incurred when starting a
business.” We constructed the same two variables using the available data in the World Bank’s “Doing
Business Survey”. However, using these two alternative variables in the first stage reduces our sample
size significantly (data are missing for the period up to 2003, and for the 2003–2009 period using them
reduces the average yearly sample size from 16560 to 12781). Moreover, we find that, in case of our
sample, these two variables are poor predictors of the probability to trade in the first stage of our
Heckman-estimation strategy. Results are available upon request.
15. Helpman and others (2008) also show the importance of taking explicit account of firm
heterogeneity when estimating the trade equation. We decided not to do this in this paper because it is
not clear what the consequences are of introducing firm heterogeneity in the NEG model that we use,
and for the wage equation (5) in particular. It lies beyond the current scope of the paper to develop a
fully-fledged NEG model incorporating firm heterogeneity. As such, we decided to stick to the more
standard NEG model used in all previous empirical work looking at the relationship between market
access and economic development, and refrain from explicitly incorporating firm heterogeneity into our
analysis. This is certainly not to deny that it would be a very interesting avenue for future research.
16. Although religious similarity itself does not change over time, we hereby do allow its effect on
the probability to trade to differ between each year in our sample.
452 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
variable on intra-SSA trade than on SSA trade with the ROW. The coefficientsgive the overall effects of each of the included variables on the amount of trade(after taking the first stage into account) and the results for 0/1 trade refer tothe estimated coefficients in the first stage probit estimations.
First of all, the final rows of Table 1 show that also in case of our tradesample restricted to SSA bilateral exports, the usefulness of the Helpman andothers (2008) approach can not be rejected:17 religious similarity has a
TA B L E 1. Trade Equation with Importer and Exporter Dummies
Dep: ln Manuf Exports
1993–2009
Coefficients eq. (7) 0/1 trade
ln dist 21.96*** 20.97***[0.00] [0.00]
ln dist - SSA 20.44* 20.38[0.24] [0.14]
Contiguity 0.92 1.16*[0.38] [0.18]
Contiguity - SSA 0.33 20.56[0.57] [0.36]
Com. lang. 0.81*** 0.45***[0.00] [0.00]
Com. lang. - SSA 0.23 0.02[0.42] [0.50]
Com. col. 0.76*** 0.30**[0.003] [0.03]
Com. col. – SSA 20.18 0.42[0.51] [0.11]
Colonizer 1.55*** 0.80[0.00] [0.19]
RFTA 0.73* 20.25[0.32] [0.43]
RFTA - SSA 0.62 0.44[0.32] [0.60]
Religion HMR 2 0.613**2 [0.034]
P-value Mills’ ratio [0.000]*** 2
nr. obs. 16560Censored 13344 (80%)Uncensored 3216 (20%)
Notes: We report estimated coefficients and not marginal effects. Marginal effects are avail-able upon request. The coefficients are used as input in the construction of our market accessmeasures. These coefficients actually differ by year, the numbers reported in the Table are themean coefficients, mean p-values (in brackets), and mean number of observations over the periodspecified. *, **, *** denotes significant at the 5 percent level in at least 50, 80, or 100 percent ofthe years.
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
17. Note that, as in case of two stage least squares, one can never fully test the validity of religious
similarity as our ‘instrument’. This ultimately hinges upon believing the arguments put forward by
Helpman and others (2008) in favour of using this variable to satisfy the needed exclusion restriction.
Bosker and Garretsen 453
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
significantly positive effect on the probability to trade18 and, moreover, theinverse Mills’ ratio is significant in the second stage (hereby not rejecting theneed to take account of endogenous sample selection).
Turning to the results on the importance of our included trade costs vari-ables,19 we confirm the standard result that distance negatively affects theamount of trade between countries. We do not find convincing evidence thatthe penalty on distance is significantly higher for intra-SSA trade (see alsoForoutan and Pritchett, 1993).20 We only find a significantly higher distancepenalty for intra-SSA trade in 50 percent of our sample years. Interestingly, wefind this significantly higher distance penalty during the later years in oursample in particular, suggesting that the improvements in SSA trade costs thathave been made in recent years have been biased towards improving trade costswith the ROW instead of better connecting the sub-continent.
Second, we do not find evidence of a border effect in SSA trade. For SSAtrade with the ROW this may not be that surprising. The only SSA countriesthat border non-SSA countries are those bordering North African countries,and SSA countries trade less with these countries than with other non-Africancountries (see e.g. IMF, 2007). The lack of a “border-effect” is arguably moresurprising for intra-SSA trade given that studies looking at other parts of theworld (e.g. Europe or the United States) usually find strong evidence that neigh-bors trade disproportionately more with each other.
By contrast, we do find strong effects of language and colonial history ontrade volumes of SSA countries. Sharing a colonial history has a strong positiveeffect on the amount of trade. Especially SSA trade with its former colonizer(s)is much higher than trade with other countries in the world. Having a commoncolonizer also boosts bilateral trade, and this effect is not significantly differentfor intra-SSA trade compared to trade with other former colonies in the ROW.Sharing a common language also stimulates both intra-SSA and SSA-ROWtrade in largely the same way (see also e.g. Foroutan and Pritchett, 1993; orCoe and Hoffmaister, 1999).
Finally, we do not find very convincing evidence that SSA trade in manufac-tures benefits significantly from the many RFTAs that are in existence on thesub-continent. We only find a significant positive effect of having an RFTA onexport volumes in 50 percent of the years in our sample period. It is interesting
18. Religious similarity is significant at the 1 percent level in all years, except for 1993 (p-value ¼
0.43), 1994 and 1995 (in both years it is significant at the 10 percent level).
19. Given our main interest in the estimated coefficient of the trade equation in the second stage for
our main purpose to construct various market access measure(s), the results of the first stage probit
estimation are not explicitly discussed.
20. Note that comparing our findings to other studies looking at SSA-trade (e.g. Foroutan and
Pritchett, 1993; Coe and Hoffmaister, 1999; Subramanian and Tamirisa, 2003 or Limao and Venables,
2001) is difficult due to the difference in estimation strategy used. These other studies use, for example,
Tobit or NLS techniques to estimate the trade equation. Moreover, they usually do not include
importer-year and exporter-year dummies in their regressions.
454 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
to note however, that we find these significant positive effects of RFTAs onexport volumes for the latest years in the sample (2006–2009) in particular. Atentative indication that African RFTAs, many of whom often only exist(ed) onpaper, could be becoming more effective in coordinating policies favorable totrade (see also UNCTAD, 2009).
Constructing Market Access, Distinguishing Between Access to SSA and to theROW
Using the yearly-estimated coefficients of (7) and the relationship between thetrade equation in (2) and market access in (1), the next step is to constructmarket access using (4). Besides calculating overall market access, we also lookat three different subcomponents of market access. Following among othersRedding and Venables (2004), Breinlich (2006) or Head and Mayer (2010), wedistinguish explicitly between the respective contribution of domestic marketaccess (DMA) and foreign market access (FMA). Furthermore, in order to beable to distinguish between the relevance of access to other SSA markets and tomarkets in the rest of the world (ROW) respectively, we in turn split foreignmarket access (FMA) into access to other SSA markets and access to ROWmarkets:
MAit ¼ DMAit þMASSAit þMAROW
it|fflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}FMA
it
; ð8Þ
where MASSAit ¼
PRj[SSA;j=i
MAijt, MAROWit ¼
PRj�SSA
MAijt and DMAit ¼MAiit. We
construct the different components of a country’s total market access for eachyear separately according to (4), adapted to take into account the estimatedparameters of (7):
MAROWit ¼
Xj�SSA
expðm jtÞDd1t
ij expðd2tBij þ d3tCLij þ d4tCCij
h
þd5tCRij þ d6tRFTAijtÞi ð9Þ
MASSAit ¼
Xj[SSA;j=i
hexpðmSSA
jt ÞDd SSA
1t
ij expðd SSA2t Bij þ d SSA
3t CLij þ d SSA4t CCij
þd SSA5t CRij þ d SSA
6t RFTAijtÞi
DMAit ¼ expðmSSAit ÞD
d SSA1t =2ð Þ
ii expðd SSA2t þ d SSA
3t þ d SSA4t þ d SSA
6t Þ;
Bosker and Garretsen 455
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
where coefficients with superscript “SSA” denote the estimated effect of a vari-able on intra-SSA trade (i.e. the coefficient on a variable plus the coefficient onthat variable interacted with the intra-SSA dummy). We construct domesticmarket access (DMA) in Redding and Venables’ (2004) preferred way. That is,we use a country’s internal distance (see Appendix A for its exact definition) asinput in the trade cost function, assuming that the speed at which internaltrade decays with distance is half as strong as for trade with other countries
(i.e. we divide d SSA1t by two). Moreover, we take each country as sharing a
common border, a common language, a common colonial history, and havingan RFTA with itself.
Having constructed the three different components of a country’s totalmarket access using (9), it is straightforward to obtain both total market access(MA) and access to foreign markets (FMA) according to equation (8).
I I I . E V I D E N C E O N T H E R O L E O F E C O N O M I C G E O G R A P H Y I N S S AE C O N O M I C D E V E L O P M E N T
Having constructed yearly market access measures for all 48 SSA countries inour sample, we are finally in a position to assess the effect of market access oneconomic development.
Market Access and Economic Development
We start by visualizing the relationship between market access and incomelevels. Figure 1a plots market access (MA) against GDP per worker (our pre-ferred proxy of wages, see below for more on this choice) over our entiresample period. It shows an overall positive relationship between GDP perworker over our sample period.
Furthermore, when looking for possible differences over our 17 year sampleperiod, Figure 1b suggests that the relationship between income levels andmarket access has strengthened in the most recent years of our sample.
FIGURE 1a. Market Access and GDP Per Worker in SSA
Notes: The raw correlation between ln MA and ln GDP per worker is 0.22 [p-value: 0.00].Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
456 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
To look at this in a more rigorous way, we turn to the estimation of theNEG wage equation, (5). In the absence of reliable wage data for all SSA coun-tries in all years of our sample, we need to proxy wages. Since many SSA facehigh unemployment rates, we decided not to use GDP per capita (as e.g.Redding and Venables, 2004; Breinlich, 2006; or Head and Mayer, 2010 do),but take GDP per worker as a more appropriate measure.21 The error term hit
in (5) captures cit, a country’s level of production efficiency. Again followingRedding and Venables (2004), we start by assuming that these cross-countrydifferences in technology are captured by an idiosyncratic error term and esti-mate (5) using pooled OLS (implicitly only allowing for other variables deter-mining technological efficiency that are uncorrelated with our market accessmeasure). The result is shown in the first column of Table 2 below.22 We findthat the estimated market access coefficient is positive and significant. A 1percent increase in a country’s market access increases GDP per worker by0.07 percent.
This conclusion is, however, somewhat premature. It is only valid under theearlier-mentioned assumption of idiosyncratic differences in countries’ produc-tion efficiency, cit, that are uncorrelated with market access. As this assumptionis likely to be violated, we subsequently make use of the panel data nature ofour data set. We include country fixed effects to capture all time-invariantcountry-specific variables that affect a country’s production efficiency. Most
FIGURE 1b. Market Access and GDP Per Worker in SSA, Changes OverSample Period
Notes: The raw correlation between ln MA and ln GDP per worker is 0.05 [p-value: 0.35] forthe 1993–2000 period and 0.26 [p-value: 0.00] for the 2001–2009 period. Source: Authors’analysis based on data sources discussed in the main text or Appendix A.
21. All results in this paper also hold when using GDP per capita instead. They are available upon
request.
22. We only show bootstrapped standard errors for all our estimation results (they are based on 200
replications). The bootstrapped standard errors take explicit account of the fact that our measures of
market access are all generated regressors. See Redding and Venables (2004, p. 64) for more details.
Results only become stronger when using robust standard errors instead.
Bosker and Garretsen 457
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E2
.M
ark
etA
cces
sand
Eco
nom
icD
evel
opm
ent
inSSA
Dep
:ln
GD
PPer
Work
er1
23
45
lnM
A0.0
67***
0.0
31***
0.0
21***
0.0
11
0.0
31**
[0.0
00]
[0.0
03]
[0.0
06]
[0.1
29]
[0.0
39]
Polity
IV2
22
0.0
02
0.0
02
0.0
004
22
[0.5
59]
[0.6
33]
[0.9
64]
Urb
aniz
atio
nra
te2
22
0.0
09
0.0
12
0.0
05
22
[0.3
33]
[0.4
73]
[0.8
12]
Gr.
pri
m.
enro
llm
ent
22
20.0
02
20.0
01
0.0
004
22
[0.1
79]
[0.6
59]
[0.8
15]
%O
ilin
GD
P2
20.0
18***
0.0
26**
0.0
01
22
[0.0
00]
[0.0
18]
[0.8
87]
Civ
ilw
ar
22
20.1
59**
20.2
18**
0.0
15
22
[0.0
15]
[0.0
44]
[0.8
09]
Civ
ilco
nflic
t2
22
0.0
55
20.0
32
0.0
53
22
[0.1
12]
[0.3
74]
[0.1
63]
%A
gri
cult
ure
ingdp
22
20.0
14***
20.0
12**
20.0
2***
22
[0.0
01]
[0.0
12]
[0.0
01]
lnw
ork
ing
pop.
den
s.2
22
0.0
22
20.2
75
1.0
43
22
[0.9
43]
[0.6
23]
[0.1
14]
Nr.
obs
775
775
583
268
315
Tim
e-per
iod
1993
–2009
1993
–2009
1993
–2009
1993
–2000
2001
–2009
P-v
alu
eco
untr
yFE
2[0
.000]
[0.0
00]
[0.0
00]
[0.0
00]
P-v
alu
eye
ar
FE
2[0
.000]
[0.0
00]
[0.0
00]
[0.0
01]
Note
s:P-v
alu
esin
bra
cket
s.B
oots
trapped
p-v
alu
eson
the
basi
sof
200
replica
tions.
***,**,*
den
ote
ssi
gnifi
cance
at1,5,or
10
per
cent
resp
ecti
vely
.So
urc
e:A
uth
ors
’analy
sis
base
don
dat
aso
urc
esdis
cuss
edin
the
main
text
or
Appen
dix
A.
458 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
notably, we hereby control for differences in physical geography that is oftenblamed for Africa’s poor development (climate, primary resource endowments,soil quality, etc). By including time (year) fixed effects as well, we also takeinto account any shocks that are affecting all countries similarly. Examples arethe introduction of new technological innovations made in developed countries(a prime example here are mobile phones, which have rapidly spread all overSSA) or worldwide economic shocks such as changes in the world price of agri-cultural products or natural resources. The second column of Table 2 showsthat the inclusion of fixed effects is quite important (corroborating findings byHead and Mayer, 2010): the effect of total market access on GDP per workeris still positive and significant, but it is about half that found in column 1: a1 percent increase in a country’s total market access increases GDP per workerby 0.03 percent.
However, the inclusion of country- and year-fixed effects may still notprovide us with accurate estimates of the effect of market access. They onlycontrol for time-invariant country-specific or country-invariant time-specificvariables. It is not unlikely that a country’s production efficiency is also deter-mined by time- and country-varying variables that are correlated with marketaccess. If this is the case, we would still obtain biased estimates of the coeffi-cient on market access, even when allowing for country- and year fixed effects.We therefore include several additional control variables to our regressions (seealso Breinlich, 2006; Redding and Venables, 2004; Hering and Poncet, 2010;Fally and others, 2010). They are related to a country’s quality of institutions(polity IV), its human capital (gross primary enrolment), its scope for econom-ics of density (working population per km2 urbanization rate), and whether ornot it is in a state of civil war or conflict.
Moreover, given that our market access measures all focus on the import-ance of manufactures, whereas SSA countries vary widely in the importance ofthis sector in their overall exports (see Figure B2 and Table B2 in Appendix B),we control for a country’s economy’s dependence on natural resources by con-trolling for the importance of oil, and of agriculture respectively in its overalleconomy (note that, given its time-invariant nature, the presence of oil, or anyother natural resource, is already controlled for by the included country fixedeffects). Column 3 of Table 2 shows the corresponding estimation results.
Adding these additional controls further lowers the effect of market access23
but it remains significantly related to GDP per worker: a 1 percent increase ina country’s market access raises its income level by 0.02 percent. As to thecontrol variables, we find that three of them are significant.24 SSA countries
23. This is also partly driven by the significant reduction in sample size resulting from the fact that
not all controls are available for all countries in all years (using the reduced sample in column three
without including any of the controls but only country- and year-FE gives an estimated coefficient of
0.23 [p-value: 0.036]).
24. Note that the non-significance of some of our included controls may also be the result of them
having very little within-variation, leaving us with a danger of making type II errors on these variables.
Bosker and Garretsen 459
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
that are more dependent on agriculture, and those plagued by civil war tend tohave lower levels of GDP per worker. Also, we find that the more oil-dependent SSA countries have higher levels of income per worker.
Finally, column 4 and 5 show the results of estimating the relationshipbetween market access and GDP per worker for the first and second half of thesample respectively (always including fixed effects and the eight above-mentioned control variables). This confirms the preliminary evidence shown inFigure 1b: the positive relationship between market access and economic devel-opment is strongest in the second half of our sample period (2001–2009).Although positive, we do not find a significant effect of market access onincome levels for the early years in our sample.25
Columns 3–5 constitute our baseline results. They show that market accessis a significantly positive determinant of a SSA country’s economic develop-ment. Moreover, the importance of market access has increased over the lasttwo decades. Based on the results shown in column 5, a 1 percent increase intotal market access increases GDP per worker by 0.031 percent. When com-paring this result to similar studies using samples encompassing both developedand developing countries (e.g. Redding and Venables, 2004; or Head andMayer, 2010), but also to other studies looking at developing economies likeBrazil (Fally and others, 2010), China (Hering and Poncet, 2010) or Indonesia(Amiti and Cameron, 2007), we find a substantially lower effect of marketaccess on economic development.26 Given the fact that the manufacturingsector is still relatively undeveloped in SSA compared to many countries con-sidered in these other studies, this finding may not be that surprising.Nevertheless, our results show that economic geography matters, also in SSA.Moreover, its importance has increased in recent years.
Decomposing the Importance of Access to SSA Markets and to Markets in theROW
Having established the importance of market access for SSA economic develop-ment, we further decompose this finding in this section. In particular, we lookat the relative importance of access to other SSA markets and markets in theROW respectively. To do so, we use the decomposition of a country’s marketaccess in domestic market access (DMA) and foreign market access (FMA) [see(9)], further decomposing the latter into access to other SSA markets (MAit
SSA)and access to markets in the ROW (MAit
ROW).
25. But we note that it is also not significantly different from the effect we find during the later
years.
26. The closest estimate we found in these studies is the one reported on China by Hering and
Poncet (2010). In their specification that includes most possible other controls related to wages, they
find a positive effect of 0.05 percent in response to an increase in market access of 1 percent.
460 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
SSA Market Access and ROW Market Access Compared
In Figures 2a and 2b, we start by showing some stylized facts about the twomost interesting components of market access, ROW and SSA market access.The left-hand panel of Figure 2a shows that having good access to SSA marketsis virtually uncorrelated with good access to markets in the ROW. The right-hand panel adds to this by plotting the mean relative importance of SSA in acountry’s FMA against mean GDP per worker over our sample period. Thisshows that ROW market access dominates FMA for SSA’s island nations,countries located in SSA’s north east (e.g. Sudan Djibouti, or Eritrea), andSouth Africa (SSA’s economic powerhouse27). By contrast, for countries closeto South Africa or Nigeria (e.g. Botswana, Swaziland; Togo or Benin), SSA
FIGURE 2a. Decomposing Market Access – variation in access to SSA andaccess to ROW
Notes: The overall mean share of SSA in FMA is 56% (s.d. 0.31). Source: Authors’ analysisbased on data sources discussed in the main text or Appendix A.
FIGURE 2b. Market Access and Distance to Major Markets
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
27. We note that all results presented in the paper are robust to the exclusion of South Africa from
the sample.
Bosker and Garretsen 461
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
market access dominates their overall FMA. But, despite the substantial differ-ence between countries in the degree to which the ROW dominates their FMA,Figure 2a shows no clear relationship between the relative importance of SSAor the ROW in FMA, and income levels.
Figure 2b shows part of the reason why SSA and ROW market access arelittle correlated. Given the strong distance decay effects we found when esti-mating the trade equation (see Table 1), countries located closest to Europe(Africa’s largest export market) have the best ROW market access.28 On thecontrary, countries closest to South Africa, but also those in West Africa (seeFigure B1 in Appendix B), tend to have the best SSA market access.29
SSA market access, ROW market access and SSA economic development
Next, Figure 3 depicts the relationship between the various (sub)components oftotal market access and economic development. We separately plot DMA,ROW market access and SSA market access against GDP per worker. Givenour finding of an increased importance of market access over the years (seeTable 2), Figure 3 distinguishes between the first and second half of our sample(1993–2000 and 2001–2009). These simple scatterplots show that the rela-tionship of all three different (sub)components of a country’s market accesswith GDP per worker appears strongest during the 2nd half of our sampleperiod.30
To more formally assess the importance of market access’s different compo-nents, we re-estimate equation (5) replacing total market access by DMA,access to SSA markets (denoted as FMA-SSA in Table 3) and ROW marketaccess (denoted as FMA-ROW in Table 3). The estimation results are shown inTable 3 below, always distinguishing between the first and second half of oursample period. All regressions include the same control variables as in columns3–5 of Table 2 and a full set of country- and year fixed effects.31
First of all, the results confirm our earlier finding that market access hasbecome increasingly important for SSA countries. In the first half of oursample period we only find some evidence that, of the various market accesscategories, domestic market access is weakly significant (at 10 percent). FMA,also when further split between SSA and ROW market access, is not
28. We find a very similar picture when plotting ROW market access against distance to the USA.
See Figure B1 in Appendix B.
29. Congo (COG) and the Democratic Republic of Congo (ZAR), are two exceptions here. These
countries’ SSA market access is the best in our sample, an artifact of the fact that the two main cities in
these two countries (used to calculate the distance between them) are located only 10.5 km apart (the
next smallest distance is that between Nigeria and Benin: 105km). Leaving these two countries out of
our sample does not change any of the results presented in our paper.
30. Note that Equatorial Guinea (GNQ) shows as somewhat of an outlier in these scatterplots. This
is due to its rapid economic growth over our sample period following the discovery of large oil and gas
reserves. All results in our paper our fully robust to leaving this country out of the sample.
31. The results on these control variables are very similar to those in Table 2. They are available
upon request, as are the results when doing the estimations using our entire sample period.
462 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
FIG
UR
E3
.M
ark
etA
cces
s’D
iffe
rent
Com
ponen
tsand
Eco
nom
icD
evel
opm
ent
Note
s:A
llfigure
splo
tdem
eaned
lnG
DP
per
work
eragain
sta
dem
eaned
subco
mponen
tof
mark
etac
cess
.D
emea
ned
mea
nin
gth
atw
esu
bst
ract
aco
untr
y’s
mea
nln
GD
Pper
work
eror
mea
nsu
bco
mponen
tof
mark
etac
cess
from
each
obse
rvat
ion.
Doin
gth
isalr
eady
rem
ove
sth
ein
fluen
ceof
any
unobse
rved
tim
e-in
vari
ant
from
the
dat
a,
makin
gour
plo
tsm
ore
dir
ectl
yre
late
dto
our
findin
gs
inT
able
2th
atalw
ays
incl
ude
afu
llse
tof
countr
yfixed
effe
cts.
Itis
als
odone
tosh
ow
that
ther
eex
ists
subst
anti
al
wit
hin
countr
y-v
ari
atio
nin
both
each
resp
ecti
vesu
bco
mponen
tof
mark
etac
cess
as
wel
las
inln
GD
Pper
work
er.
Sourc
e:A
uth
ors
’analy
sis
base
don
dat
aso
urc
esdis
cuss
edin
the
text.
Bosker and Garretsen 463
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E3
.Fore
ign
Mark
etA
cces
sand
Eco
nom
icD
evel
opm
ent
inSSA
Dep
:ln
GD
PPer
Work
er1
23
45
67
8
lnD
MA
0.0
04*
0.0
04*
0.0
20.0
22
22
2
[0.0
87]
[0.0
87]
[0.1
51]
[0.1
95]
22
22
lnFM
A0.0
09
20.0
48**
20.0
11
20.0
53**
2
[0.6
79]
2[0
.036]
[0.6
06]
2[0
.019]
2
lnFM
A-
RO
W2
20.0
87
22
0.0
39
22
0.0
75
22
0.0
39
2[0
.396]
2[0
.655]
2[0
.435]
2[0
.665]
lnFM
A-
SSA
20.0
14
20.0
49**
20.0
16
20.0
54***
2[0
.370]
2[0
.023]
2[0
.311]
2[0
.004]
Contr
ols
:se
eT
able
2
nr.
obs
268
268
315
315
268
268
315
315
tim
e-per
iod
1993
–2000
1993
–2000
2001
–2009
2001
–2009
1993
–2000
1993
–2000
2001
–2009
2001
–2009
p-v
alu
eco
untr
yFE
[0.0
00]
[0.0
00]
[0.0
00]
[0.0
00]
[0.0
00]
[0.0
00]
[0.0
00]
[0.0
00]
year
FE
[0.0
00]
[0.0
00]
[0.0
01]
[0.0
01]
[0.0
00]
[0.0
00]
[0.0
01]
[0.0
01]
Note
s:P-v
alu
esin
bra
cket
s.B
oots
trapped
p-v
alu
eson
the
basi
sof
200
replica
tions.
***,**,*
den
ote
ssi
gnifi
cance
at1,5,or
10
per
cent,
resp
ecti
vely
.So
urc
e:A
uth
ors
’analy
sis
base
don
dat
aso
urc
esdis
cuss
edin
the
main
text
or
Appen
dix
A.
464 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
significantly related to income levels in these years (see column 1 and 2). Thischanges markedly during the latest years in our sample (2001–2009). In par-ticular, we find that access to foreign markets has become much more import-ant in explaining the differences in economic development between SSAcountries (see column 3). Even more interesting, when further decomposingthis effect of FMA in column 4, we find that it is access to SSA markets in par-ticular that drives these findings: A 1% increase in a country’s access to otherSSA markets is associated with a 0.05% increase in GDP per worker. Theeffect of ROW market access instead is insignificant.
Some care has to be taken in interpreting these results. Our findings do goagainst those proclaiming that intra-SSA economic linkages are too weak andunder-developed to be of importance to SSA countries. They support the viewthat stimulating intra-SSA manufacturing export markets is very important forthe future viability of SSA countries that want to become less dependent onnatural resource revenue (Collier and Venables, 2007). Indeed, manufacturingalready dominates intra-SSA exports (UNCTAD, 2009). Moreover, intraSSA-exports have increased faster than SSA-exports to the rest of the world inrecent years (Easterly and Reshef, 2010; ARIA IV). However, given that wefocus on market access for manufacturing goods (see also our discussion at thebeginning of section II), our results should not be taken as saying that access tomarkets in the ROW does not matter for SSA. SSA exports to the ROW aredominated by natural resources (for more than 75 percent). The fact thatmarket access to the ROW for SSA manufacturing products is not significant,does not say much about the importance of, for example, lowering tariffs onSSA agricultural products for SSA countries’ economic development. It canalso be taken as an indication that, to date, most SSA manufactures are not yetfinding their way to markets outside the (sub)continent.32
Finally, also note that DMA loses its (weak) significance in the second halfof our sample period. It confirms the idea that for most SSA countries theirown domestic market size is too small to be of significant importance.33; pos-sibly even posing constraints on firms’ prospects (see Collier and Venables,2007). However, endogeneity problems are inherently present when includingDMA. One basically regresses a measure dominated by a country’s own GDP(DMA) on its GDP per worker (see Head and Mayer (2010) for a (critical) dis-cussion on this issue). Therefore, columns 5–8 in Table 3 show that our resultson FMA, and its two subcomponents (FMA-SSA and FMA-ROW), also comethrough when totally abstracting from DMA. To summarize our main findings:
32. One could argue that our ROW market access measure suffer from too little cross-sectional
variance to find any effect when controlling for country- and year-specific fixed effects. However, the
scatterplots in Figure 3 suggest otherwise. Also, when including SSA and ROW market access separately
we find the same results.
33. Note that this was also already borne out by the estimated effect of DMA in the early years of
our sample. Although significantly positive at the 10 percent level, this effect is very small: a 1 percent
increase in DMA increasing GDP per worker by only 0.004 percent.
Bosker and Garretsen 465
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
access to foreign markets is increasingly important for SSA countries.Moreover, it is access to other SSA countries in particular that is positivelyassociated with income levels.
Additional Robustness Checks
Several issues could still invalidate our main findings. First, even when abstract-ing from domestic market access (DMA), there is still the issue of endogeneity.The assumption under which our baseline results are valid is that, after control-ling for fixed effects and the included control variables, the remaining errorterm is uncorrelated with our measures of foreign market access. One way inwhich this may be violated is when the error term still contains other variablesinfluencing a country’s GDP per worker that are correlated with market access.Another way is reverse causality: if market access not only influences GDP perworker, but GDP per worker in turn also influences market access, the errorterm is by construction correlated with market access.
To control for both possible sources of endogeneity, we employ an instru-mental variable approach,34 using the distance to SSA’s most important exportmarkets in SSA and in the ROW as instruments for our measures of foreignmarket access (i.e. the EU, South Africa and Nigeria; see Figures 2b and B1).The relevance of this approach relies on the arguments put forward to justifythe usefulness of these ‘distance instruments’ (see among others Redding andVenables, 2004 and Hanson, 2006 for more on this). Table 4a shows theresults.
First of all, Table 4a shows that we cannot reject the validity of our instru-ments (see the statistics at the bottom of the Table). The F-statistic for theirjoint significance in the first stage is larger than 10 (Staiger and Stock, 1997),except in column 3 where it is 7.44. Moreover, they always pass the Hansen Jtest for overidentification. The results confirm our baseline findings. Foreignmarket access significantly positively affects income per worker in the latestyears of our sample. When further subdividing this into ROW- and SSAmarket access, we again find that this positive effect holds for access to SSAmarkets only.
A drawback of our IV results is that the distance instruments used are time-invariant. This precludes the use of country-fixed effects. Columns 1–4 ofTable 4b therefore also show results when including each market accessmeasure lagged one period (all control variables are also lagged one period).This to some extent controls for reverse causality, while still allowing for theinclusion of country-fixed effects.35 Reassuringly, all our baseline results againcome through.
34. This also controls for the third way by which endogeneity issues may be raised, that is,
measurement error.
35. Note that this argument breaks down in case of autocorrelation in the residuals. Also including
lagged variables does not solve possible endogeneity resulting from omitted variables or measurement
error.
466 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Finally, our last robustness check extends our empirical model.36 Footnote 5already hinted at the possibility of extending the NEG model that we use (seeAppendix C) to also include an intermediate goods sector. This would (seeRedding and Venables 2004 for the details) add an additional term to equation(5). Besides a country’s market access (i.e. its ease of access of final goodsmarkets), a country’s supplier access (i.e. its ease of access to markets for
TA B L E 4a. IV-Results
dep: ln GDP per worker 1 2 3 4
ln FMA 0.091 2 0.313*** 2
[0.316] 2 [0.008] 2
ln FMA - ROW 2 20.059 2 20.273*2 [0.533] 2 [0.065]
ln FMA - SSA 2 0.064 2 0.176***2 [0.249] 2 [0.001]
Controls see Table 2 (no country FE)nr. obs 268 268 315 315time-period 1993–2000 1993–2000 2001–2009 2001–2009F-stat. instrument 14.02 2 7.44 2
FMA - ROW 2 198.1 2 143.94FMA - SSA 2 17.62 2 18.41p-value over ID-test [0.415] [0.269] [0.501] [0.822]
Notes: P-values, based on robust standard errors, in brackets. ***, **, * denotes significanceat 1, 5, or 10 percent, respectively.
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
TA B L E 4b. Lagged Market Access
Dep: ln GDP Per Worker 1 2 3 4
ln FMA 20.006 2 0.037* 2
[0.695] 2 [0.072] 2
ln FMA - ROW 2 0.005 2 20.012 [0.951] 2 [0.908]
ln FMA - SSA 2 20.007 2 0.042**2 [0.540] 2 [0.024]
Controls see Table 2 (also lagged)nr. obs 203 203 299 299time-period 1993–2000 1993–2000 2001–2009 2001–2009
Notes: P-values in brackets. Bootstrapped p-values on the basis of 200 replications. ***, **, *denotes significance at 1, 5, or 10 percent, respectively.
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
36. We also estimated (5) in first differences as an alternative way to deal with unobserved
country-specific variables that are correlated with market access. Again, we find that SSA market access
is the only component of market access for which we systematically find a significant positive effect on
GDP per worker (results available upon request).
Bosker and Garretsen 467
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E4
c.
Supplier
Acc
ess
dep
:ln
GD
PPer
Work
er1
23
45
6
lnM
A0.0
11*
22
0.0
31**
22
[0.0
96]
22
[0.0
29]
22
lnFM
A2
0.0
10
22
0.0
50**
2
2[0
.646]
22
[0.0
18]
2
lnFM
A-
RO
W2
22
0.0
56
22
20.0
42
22
[0.5
97]
22
[0.5
81]
lnFM
A-
SSA
22
0.0
31*
22
0.0
50**
22
[0.0
80]
22
[0.0
16]
lnD
MA
22
22
22
22
22
22
lnSA
0.0
01
22
0.0
02
22
[0.8
06]
22
[0.7
29]
22
lnFSA
22
0.0
03
22
0.0
07
2
2[0
.851]
22
[0.5
36]
2
lnFSA
-R
OW
22
0.0
91
22
0.0
42
22
[0.2
47]
22
[0.4
51]
lnFSA
-SSA
22
20.0
23**
22
0.0
34***
22
[0.0
44]
22
[0.0
06]
Contr
ols
see
Table
2N
r.obs
268
268
268
315
315
315
Tim
e-per
iod
1993
–2000
1993
–2000
1993
–2000
2001
–2009
2001
–2009
2001
–2009
Note
s:P-v
alu
esin
bra
cket
s.B
oots
trapped
p-v
alu
eson
the
basi
sof
200
replica
tions.
***,**,*
den
ote
ssi
gnifi
cance
at1,5,or
10
per
cent,
resp
ecti
vely
.So
urc
e:A
uth
ors
’analy
sis
base
don
dat
aso
urc
esdis
cuss
edin
the
main
text
or
Appen
dix
A.
468 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
intermediate goods needed in final goods production) would be an importantdeterminant of its income level. A yearly-measure of each SSA country’s sup-plier access can be constructed in a very similar way as its market access (basic-ally replacing m jt with r jt in (4), but again see Redding and Venables 2004 formore details). Table 4c shows the results when also taking account of supplieraccess (SA). Similar to market access, it is very straightforward to decomposeoverall supplier access into domestic supplier access (DSA), and access to sup-pliers in SSA and in the ROW respectively.
For supplier access too, we find that it increased in importance over the lasttwo decades, and that access to SSA suppliers in particular is positively asso-ciated with higher GDP per worker. However, when focusing on foreign sup-plier access, or supplier access as a whole, results are much weaker than thosefor market access. Most importantly for our purposes, all our baseline findingsregarding the importance of market access hold up to also considering coun-tries’ supplier access.
I V. T H E E F F E C T O F D I F F E R E N T P O L I C I E S A I M E D A T I M P R O V I N G
M A R K E T AC C E S S
Our main findings show that improving a country’s market access, and in par-ticular access to other SSA countries, will have significant positive effects on itseconomic development. In this section we use this positive relationshipbetween market access and income levels to gain insight into the relative effectof different policies aimed at improving a country’s market access.37 Forexample, we look at the effect of improving a country’s infrastructure, increas-ing SSA regional integration, or alleviating the burden of landlocked countries.To be able to do so, we change our baseline estimation strategy in one import-ant way (see also Elbadawi and others, 2004 or Redding and Venables 2004).
Our baseline strategy includes importer-year and exporter-year dummieswhen estimating the trade equation (3). This does, however, not allow one “toquantify the effects . . . of particular country characteristics (for example, land-locked or infrastructure), since all such effects are contained in the dummies”(Redding and Venables 2004, p. 75). As such, it becomes impossible to look atthe effect of country-specific policies aimed at lowering trade costs. To over-come this problem we follow Redding and Venables (2004, section 7), and esti-mate the following trade equation instead of equation (3), proxying each
37. We note at this point that our policy experiments do not quantify the full general equilibrium
effects on income levels. We are confining ourselves to the “short-run” effects of improving market
access on income levels. We abstract from any subsequent changes in economic geography induced by
e.g. firms or consumers changing their location decistion as a results of the changes in income levels
induced by the change in market access resulting from one of our policy experiments.
Bosker and Garretsen 469
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
country’s market and supplier access by its GDP (sit � GDPrt
it ; mit � GDPmt
it ):
ln EXijt ¼ rt ln GDPit þ mt ln GDPjt þ b ln Tijt þ 1ijt ð10Þ
In order to conduct our various ‘policy experiments’, we augment our specifi-cation of trade costs (6) by three different exporter- and importer-specificdeterminants of trade costs: being landlocked, being an island, and the state ofa country’s infrastructure, that is,
b ln Tijt ¼ d1 ln Dij þ d2 ln Bij þ d3 ln CLij þ d4 ln CCij þ d5 ln CRij þ d6RFTAijt
þ d7llit þ d8llit þ d9islit þ d10islit þ d11infrait þ d12infra jt
ð11Þ
Furthermore we take note of Martin and others (2008) and control forwhether or not a country is experiencing civil war or civil conflict (seeAppendix A for the full details on all these additional variables included to thetrade equation).
We estimate (10) with (11) substituted for trade costs Tijt for the latest pos-sible year in our sample (which is 2008 instead of 2009 because of missing in-frastructure and conflict data for 2009). We take the latest possible year inorder to make our prediction of the effect of each of our ‘policy experiments’as up-to-date as possible. Table B1 in Appendix B shows the resulting esti-mated coefficients. Again, we allow for a different effect of each trade-costrelated variable on intra-SSA trade. Confining our discussion to the newlyadded country-specific trade cost variables,38 we find a large burden of beinglandlocked, and to a lesser extent of being an island (landlocked countriesexport and import significantly less (84 percent and 125 percent respectively)than coastal nations; and these numbers are 80 percent and 34 percent forisland countries). Moreover, we find that bad infrastructure is an important de-terrent to trade, and exports in particular.
Based on the estimates shown in Table B1, we calculate each SSA country’smarket access and its components [in a similar way as in (9)]. Next, we recal-culate these measures taking into account one of eight different policy experi-ments that we set out in more detail below, and calculate the resulting changein foreign market access (and its two subcomponents). We do not look at do-mestic market access as most of our policy experiments are not that interestingto look at for domestic market access (e.g. no longer being landlocked onlyaffects a country’s trade costs with other countries, and it is hard to thinkabout a country establishing an RFTA with itself ).
38. Given that we no longer include importer- and exporter dummies the results on the magnitude
of the effect of bilateral trade costs on bilateral exports differs from those reported in Table 1.
However, the direction of their effect is never different.
470 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Finally, the effect of the resulting improvement in market access on GDPper worker easily follows from the estimated coefficient(s) on foreign marketaccess in Table 3. In particular, we take our finding on the relative import-ance of access to other SSA markets seriously and use the coefficient onSSA market access, reported in column 8, in combination with the changein SSA market access to get at the overall impact of each policy experimenton economic development.39 Table 5 shows the results of doing this for the
TA B L E 5. Policy Experiments – Increase in Market Access and GDP Perworker
1 2 3 4 5 6
CountryNo LongerLandlocked
NoLongerIsland
Infrastructureþ1 s.d.
AlldistancesHalved
RFTAwithSouthAfrica
SSA -WideFree
TradeZone
Cape Verde
FMA 2 71.89 20.67 63.72 0.34 2.68FMA - SSA 2 51.10 76.57 112.09 2.94 21.11FMA - ROW 2 74.31 10.39 55.26 2 2
GDP per worker 2 2.76 4.14 6.05 0.16 1.14
Botswana
FMA 85.89 2 34.61 75.43 2 0.41FMA - SSA 90.29 2 76.57 112.09 2 1.41FMA – ROW 84.02 2 10.39 55.26 2 2
GDP per worker 4.88 2 4.14 6.05 2 0.08
Central African
Republic
FMA 84.60 2 18.55 61.96 0.06 0.35FMA - SSA 90.29 2 76.57 112.09 0.69 3.83FMA – ROW 84.02 2 10.39 55.26 2 2
GDP per worker 4.88 2 4.14 6.05 0.04 0.21
Ethiopia
FMA 85.01 2 23.81 66.34 2 0.27FMA - SSA 90.29 2 76.57 112.09 2 1.76FMA – ROW 84.02 2 10.39 55.26 2 2
GDP per worker 4.88 2 4.14 6.05 2 0.10
Sudan
FMA 2 2 15.84 59.73 2 0.18FMA - SSA 2 2 76.57 112.09 2 2.99FMA – ROW 2 2 10.39 55.26 2 2
GDP per worker 2 2 4.14 6.05 0.16
Notes: All numbers are percentages. The effect on GDP per worker is calculated by multiply-ing the change in SSA MA by the coefficient on SSA MA reported in column 8 in Table 3.
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
39. It would be straightforward to redo these calculations using any other reported coefficient in
Table 3 or Table 2 for that matter. This would change the absolute effect of each of the different policy
measures on GDP per worker, but it leaves the relative magnitude of each of the different policy
experiments unchanged.
Bosker and Garretsen 471
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
first six of our eight different “policy experiments,” focusing on five differ-ent countries.
Given the way we modelled trade costs, see (11), the first four policy experi-ments affect all countries’ SSA market access and ROW market access similar-ly. This is not true for overall FMA. Since we allowed all trade costs variablesto have a different effect on intra-SSA trade and SSA trade with the ROW re-spectively, changes in FMA depend on the relative importance of SSA marketaccess and ROW market access in overall FMA (explaining why the overallchange in FMA resembles that in ROW market access for Cape Verde andSudan, and that in SSA market access for the other three countries).
Halving distances to all trade partners (a rough proxy for improving SSAcountries’ connectivity through, e.g., cross-border infrastructure projects, ormore effective border procedures) results in the largest improvement in GDPper worker, raising it by about 6 percent. Next comes alleviating a landlockedcountry’s burden of having no direct access to the coast (raising incomes byalmost 5 percent), followed by a 4 percent increase in GDP per worker as aresult of a one standard deviation improvement in a country’s infrastructure(e.g. corresponding to upgrading Ethiopia’s infrastructure to resemble that inBotswana). With a resulting increase of 2.8 percent, alleviating the remotenessof an island country has the smallest effect on GDP per worker.
Finally, columns 5 and 6 show the effects of a newly established RFTA. Theseare also positive but much smaller compared to the other policy experiments.Not surprisingly, the effect on GDP per worker is larger, the larger the numberof new partner countries in the new RFTA.40 (compare columns 5 to 6, or theimpact of the SSA-wide free trade zone on Cape Verde to that on Botswana [aSSA-free trade zone would more than triple the number of RFTA partners forCape Verde whereas ‘only’ doubling it for Botswana]). This much smaller effectof the establishment of an SSA-wide free trade zone compared to those of ourother ‘policy experiments’ should in our view be taken with a pinch of salt. Dueto the plethora of RFTAs officially in existence in SSA in 2008, the average SSAcountry already shared official RFTA membership with 46 percent of its SSAtrade partners. However, the effectiveness of SSA RFTAs in actually implement-ing policies favourable to intra-SSA trade varies widely (compare, e.g., SADC toCENSAD). Our simple dummy variable for the existence of an RFTA is unableto take the varying degrees of effectiveness of each RFTA into account, so thatour findings in Table 5 are most likely an underestimate of the effect on econom-ic development were SSA countries able to establish an SSA-wide trade zone op-erating at the same level effectiveness as e.g. ASEAN or MERCOSUR, let aloneNAFTA or the EU (see also UNCTAD, 2009).
Our last two experiments do not so much concern policy. They are aimed atgiving an idea of the magnitude of spatial spillovers across SSA countries. How
40. Moreover, the impact also depends on the relative importance of a country’s newly added RFTA
partners for its market access compared to that of the countries with which it already shares an RFTA.
472 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
large are the benefits of growth in one particular country for its neighbors as aresult of the increased market access that these neighboring countries enjoy?Figure 4 gives some idea of this. It plots the increase in GDP per worker in allSSA countries resulting from a 10 percent increase in the GDP of one of SSA’seconomic powerhouses, South Africa and Nigeria, against distance to thesecountries.
Given the magnitude of the estimated distance penalty in SSA (see Table 1),we find that the effects of such shocks quickly peters out with distance.Countries located closest to South Africa and Nigeria respectively experiencethe largest spillovers. The overall spillover effect is small compared to some ofour earlier “trade cost experiments” (the nearest neighbors experiencing“spillover-growth” of about 0.2 percent).41 This is due to the fact that for eachcountry, South Africa and Nigeria constitute only one of many trading part-ners, that determine a country’s market access.
V. C O N C L U S I O N S
The role of geography in explaining sub-Saharan Africa’s poor economic per-formance is often confined to its physical geography, focusing on, for example,its hostile disease environment or poor climate. This paper focuses on a
FIGURE 4. Spillovers to Neighbors of Positive 10 Percent GDP Shock inNigeria or South Africa
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
41. This effect is much smaller that the spillover-effects of South African growth on its neighbors
found by Arora and Vamvakidis (2005). Their analysis is a reduced form exercise, making it hard to
compare to our theory-based approach (although our findings could be reconciled with theirs by
arguing that we only capture the trade-induced spillover effect, whereas they capture a composite
spillover effect of South African growth including also other non-trade related spillovers). Moreover,
given their chosen empirical strategy they are unable to include year fixed effects in their panel
estimation so that it is impossible to exclude the possibility that (part of) their findings are driven by an
omitted variable affecting both South African economic growth as well as that of other SSA countries.
Bosker and Garretsen 473
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
different role of geography and establishes the importance of relative or eco-nomic geography for economic development in sub-Saharan Africa (SSA).Using an empirical strategy that is firmly based upon a new economic geog-raphy model, our paper is among the first to test for the importance of marketaccess and thereby of economic geography in explaining the observed differ-ences in economic development between SSA countries.
Building on the framework introduced by Redding and Venables (2004), wefirst construct theory-based measures of market access for manufactures foreach SSA country, relying on bilateral manufacturing trade data to reveal therelative importance of trade costs and market size in determining each coun-try’s market access. In doing so, we explicitly allow for a different impact oftrade costs on intra-SSA trade and SSA trade with the rest of the world(ROW), and subsequently decompose each country’s total market access intomarket access to other SSA countries and into market access to the ROWrespectively.
Using these constructed measures of market access, we estimate the impactof market access for manufactures on GDP per worker. We find that marketaccess positively affects income levels. Economic geography matters for eco-nomic development, also in SSA. Moreover, it has increased in importanceover the last two decades. The relationship between market access and econom-ic development is strongest and most robust during the 2001–2009 period: a 1percent increase in a country’s market access is associated with a 0.03 percentincrease in its GDP per worker. This finding is robust to controlling for othervariables affecting economic development (most notably human capital, insti-tutions and natural resource dependence), to controlling for unobserved hetero-geneity by allowing for country (and year) specific fixed effects, and toinstrumenting market access by distance to major markets.
Arguably even more interesting is our finding that, when decomposing ouroverall market access effect into the respective effects of domestic, SSA-, andROW-market access, access to other SSA markets has the most significant (andalso the most robust) impact on a country’s economic development. Thisfinding becomes less surprising when considering the fact that most SSA coun-tries sell the bulk of their manufacturing exports to other SSA countries(UNCTAD 2009; 2010). Moreover, SSA export growth has been regional inrecent years. Intra SSA-exports have increased faster than SSA-exports to theROW (Easterly and Reshef 2010; UNCTAD 2010). By contrast, SSA exports tothe ROW are to date still dominated by natural resources and agricultural pro-ducts, with only little SSA manufactured goods finding their way to European,US, or Asian markets. Our findings stress the importance of the rest of SSA formost SSA countries’ prospects on developing a more diversified economy witha profitable, exporting, manufacturing sector (one of the backbones of Asia’ssustained growth over the last decades). Improving market access among SSAcountries alleviates the constraint of small domestic market size faced by most
474 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
SSA countries (Collier and Venables 2007), carrying positive effects for eco-nomic development.
Based on our findings, we also show tentative evidence on the impact ofseveral policies specifically aimed at improving SSA countries’ market access.Overall, this lends support to the view that current efforts to improve SSAmarket access by, for example, investing in infrastructure (Sub-Saharan AfricanTransport Policy Program or The Infrastructure Consortium for Africa), allevi-ating the burden of landlocked countries (the Almaty Programme), or byaiming to increase effective intra-SSA integration (African Union), are indeedimportant, although in varying degrees, in further stimulating SSA economicdevelopment.
Above all, see also Henderson, Shalizi and Venables (2001), our results are areminder that distance or relative geography matters for economic develop-ment. Despite room for (policy-induced) improvements in market access, the(economic) remoteness of many SSA countries remains an important burden ontheir economic development prospects.
A P P E N D I X A . D A T A D E F I N I T I O N S A N D S O U R C E S
GDP (also per capita and per worker): Gross Domestic Product (also percapita and per worker) in current US dollars. From World Bank DevelopmentIndicators, 2011, or World Bank Africa Database, 2010.
Distance: Great circle distance between main cities, from CEPII.Internal distance: This often-used specification of Dii reflects the average dis-
tance from the centre of a circular disk with areai to any point on the disk (as-suming these points are uniformly distributed on the disk). It is calculated onthe basis of a country’s area: Dii ¼ 2=3 areai=pð Þ1=2.
Contiguity: Dummy variable indicating if two countries share a commonborder, from CEPII.
Common official language: Dummy variable indicating if two countriesshare a common official language, from CEPII
Common colonizer: Dummy variable indicating if two countries have beencolonized by the same colonizer, from CEPII.
Colony – Colonizer relationship: Dummy variable indicating if two coun-tries have ever had a colony-colonizer relationship, from CEPII.
Landlocked: Dummy variable indicating if a country has no direct access tothe sea.
Island: Dummy variable indicating if a country is an island.Infrastructure index: Following Limao and Venables (2001), the index is
constructed as the unweighted average of four variables (each normalized tohave a mean of 0 and standard deviation 1 over the whole sample period aswell as in each year). As Limao and Venables (2001), we ignore missing values,
Bosker and Garretsen 475
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
making the implicit assumption that the four variables are perfect substitutesto a transport services production function. The four components are:
- Roads: Km road per km2.- Paved roads: Km paved road per km2.- Railways: Km railways per km2.- Telephone main lines: Telephone main lines per 1000 inhabitants.
All four are taken from the World Bank Development Indicators 2011, orthe World Bank Africa Database, 2010.
African regional or free trade agreement: Dummy variable indicating if twocountries in a particular year are both a member of one of the followingAfrican regional or free trade agreements: ECOWAS, ECCAS, COMESA,SADC, UEMOA, CEMAC (or UDEAC), EAC, IGAD, CENSAD, or the recent-ly established AFTZ.
Civil conflict: Dummy variables indicating if a country experienced the useof armed force between two parties, of which at least one is the government ofa state that resulted in at least 25 and at most 999 battle-related deaths, fromthe International Peace Research Institute, Oslo. Source: World DevelopmentIndicators, 2011.
Civil war: Dummy variables indicating if a country experienced the use ofarmed force between two parties, of which at least one is the government of astate that resulted in at least 1000 battle-related deaths, from the InternationalPeace Research Institute, Oslo. Source: World Development Indicators, 2011.
Urbanization rate: Share of the population living in urban areas, from theWorld Bank Development Indicators, 2011, or the World Bank AfricaDatabase, 2010.
Gross primary enrollment: Gross enrollment ratio is the ratio of total enroll-ment, regardless of age, to the population of the age group that officially corre-sponds to the level of education shown. Primary education provides childrenwith basic reading, writing and mathematics skills along with an elementaryunderstanding of such subjects as history, geography, natural science, socialscience, art and music. From the World Bank Development Indicators, 2011,or the World Bank Africa Database, 2010.
Percentage oil rents in total GDP: Oil rents are the difference between thevalue of crude oil production at world prices and total costs of production.From the World Bank Development Indicators, 2011, or the World BankAfrica Database, 2010.
Percentage agriculture in total GDP: Agriculture includes forestry, huntingand fishing as well as cultivation of crops and livestock production. From theWorld Bank Development Indicators, 2011, or the World Bank AfricaDatabase, 2010.
476 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Working population per km2: Data on the working population and a coun-try’s overall area are separately taken from the World Bank DevelopmentIndicators, 2011, or the World Bank Africa Database, 2010.
Polity IV: The "Polity Score" captures a regime’s authority spectrum on a21-point scale ranging from -10 (hereditary monarchy) to þ10 (consolidateddemocracy). From the Polity Project.
Religious similarity: Fraction measuring the probability of two people fromdifferent countries adhering to the same religion. In particular we followHelpman and others (2008) and construct this variable as: (% Protestants incountry i . % Protestants in country j) þ (% Catholics in country i . %Catholics in country j) þ (% Muslims in country i . % Muslims in country j).Information on each country’s religious composition is taken from RobertBarro’s website at Harvard.
Percentage manufacturing exports in merchandise exports:
Manufactures comprise commodities in SITC sections 5 (chemicals), 6 (basicmanufactures), 7 (machinery and transport equipment), and 8 (miscellaneousmanufactured goods), excluding division 68 (non-ferrous metals). From theWorld Bank Development Indicators, 2011, or the World Bank AfricaDatabase, 2010.
APPENDIX B
FIGURE B1. ROW and SSA Market Access, and Distance to the USA andNigeria Resp.
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
Bosker and Garretsen 477
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA B L E B1. Trade Equation Without Importer and Exporter Dummies.Year ¼ 2008
Dep: ln Manuf ExportsCoefficients Coefficients
Variable ROW Extra SSA Variable ROW Extra SSA
ln dist 20.797*** 20.82*** ln GDP exp 1.405*** 0.087[0.00] [0.003] [0.00] [0.472]
contiguity 21.944*** 4.182*** ln GDP imp 0.782*** 20.326***[0.055] [0.00] [0.00] [0.001]
com. lang. 0.528*** 20.378 civil war exp 0.255 2
[0.00] [0.272] [0.207] 2
com. col. 1.20*** 20.258 civil war imp 20.03 2
[0.00] [0.49] [0.888] 2
colonizer 1.611*** 2 civil conflict exp 0.234 2
[0.00] 2 [0.205] 2
RFTA 1.151*** 20.165 civil conflict imp 0.113 2
[0.00] [0.717] [0.498] 2
infra exp 0.416*** 2.647***[0.00] [0.00] p-value Mills’ ratio [0.066]
infra imp 0.192* 0.733 p-value religion [0.000][0.098] [0.233]
ll exp 20.84*** 20.063 nr. obs 14946[0.00] [0.86] censored 11184
ll imp 21.252*** 0.189 uncensored 3762[0.00] [0.56]
isl exp 20.797*** 0.284[0.00] [0.616]
isl imp 20.342** 20.41[0.044] [0.443]
Notes: We report estimated coefficients and not marginal effects. Marginal effects, and theresults for the 1st stage probit are available upon request. The coefficients are used as input in theconstruction of our market access measures. *, **, *** denotes significant at the 5 percent levelin at least 50, 80, or 100 percent of the years.
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
478 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
APPENDIX CIn this Appendix, we briefly set out the new economic geography (NEG)
model that underlies our empirical framework.42 Assume the world consists ofi ¼ 1, . . . ,R countries, each being home to an agricultural43 and a manufactur-ing sector. As in virtually all NEG models, we focus on the manufacturingsector. Moreover, and in line with e.g. Redding and Venables (2004), Breinlich(2006), Knaap (2006) and Head and Mayer (2006), we restrict our attention tothe ‘short-run’ version of the model. This amounts to, as Redding andVenables 2004, p.59) put it, “taking the location of expenditure and produc-tion as given and asking the question what wages can manufacturing firms ineach location afford to pay its workers.”
FIGURE B2. Share of Manufacturing in Total Merchandise Exports 1993–2009
Notes: The figure shows box-plots for each year in our sample. Each box ranges from the 25th
to the 75th percentile of the distribution of SSA countries’ percentage of manufacturing exports intotal merchandise exports. The horizontal line within each box denotes the median percentage ofmanufacturing exports in total merchandise exports. This median increases from less than 10 %to more than 20 % over our sample period. The lines extending from each box denote the upperand lower adjacent values respectively. These are calculated as 1 . 5 times the interquartile range(IQR), the difference between the 25th and 75th percentile of the overall SSA distribution ofmanufacturing shares in total merchandise trade in each particular year. The countries explicitlyshown in the figure are those for which manufacturing exports constitute a share in overallmerchandise exports that is even higher that the upper adjacent value of the overall SSAdistribution in a particular year.
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
42. See Fujita, Krugman, and Venables (1999), Puga (1999), Head and Mayer (2004) for more
detailed expositions of various NEG models and the derivation of market access and the equilibrium
wage equation in particular.
43. The agricultural sector uses labor and land to produce a freely tradable good under perfect
competition that acts as the numeraire good.
Bosker and Garretsen 479
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA B L E B2. Average Share of Manufacturing in Merchandise Exports
Avg. % Manufacturing in Total Merchandise Exports
Country 1993–2000 2001–2009 1993–2009
Sub-Saharan Africa 29.4 32.4 31.1Djibouti 2 90.7 90.7Lesotho 94.9 88.8 90.0Botswana 89.6 82.9 83.6Cape Verde 81.3 68.2 73.8Mauritius 72.2 66.0 68.9Swaziland 54.4 65.2 63.9South Africa 49.6 54.7 52.3Namibia 55.8 47.1 48.0Central African Republic 49.0 37.4 44.5Senegal 43.1 41.7 42.3Madagascar 28.4 46.5 38.0Eritrea 28.0 40.5 36.3Zimbabwe 31.4 33.5 32.5Togo 12.8 53.1 31.4Kenya 25.4 30.6 28.1Gambia, The 19.0 21.9 20.7Guinea 22.8 18.4 20.4Ghana 14.7 22.3 18.9Tanzania 16.6 19.5 18.6Cote d’Ivoire 14.1 17.2 16.0Comoros 23.6 6.3 14.3Burkina Faso 12.7 11.5 12.1Uganda 6.9 14.5 11.0Malawi 8.4 11.9 10.4Zambia 10.7 10.1 10.4Mali 4.1 13.6 9.9Mozambique 11.6 6.9 8.8Ethiopia 7.5 9.4 8.6Sierra Leone 9.7 7.5 8.6Benin 5.8 8.7 7.0Niger 2.8 9.3 6.6Burundi 2.2 10.2 6.4Rwanda 5.8 5.9 5.9Cameroon 6.8 4.0 5.1Gabon 2.7 7.8 5.0Sao Tome and Principe 2.0 4.3 3.9Congo, Rep. 2.4 2 2.4Seychelles 1.1 3.7 2.4Nigeria 1.6 2.9 2.3Sudan 3.7 0.6 2.1Guinea-Bissau 0.2 1.2 1.0Mauritania 0.2 0.0 0.1
Notes: All numbers denote percentages. Angola, Chad, the Democratic Republic of Congo,Equatorial Guinea, Liberia and Somalia are not shown in the Table due to lack of information.
Source: Authors’ analysis based on data sources discussed in the main text or Appendix A.
480 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
In the manufacturing sector, firms operate under internal increasing returnsto scale, represented by a fixed input requirement ciF and a marginal input re-quirement ci. Each firm produces a different variety of the same good undermonopolistic competition using the same Cobb-Douglas technology combiningtwo different inputs. The first is an internationally immobile factor (e.g. labor),with price wi and input share b, the second is an internationally mobile factorwith price vi and input share g, where g þ b¼ 1.44
Manufacturing firms sell their products to all countries. This involves ship-ping them to foreign markets incurring trade costs in the process. These tradecosts are assumed to be of the iceberg-kind and the same for each variety pro-duced. In order to deliver a quantity xij(z) of variety z produced in country i tocountry j, xij(z)Tij has to be shipped from country i. A proportion (Tij-1) ofoutput ‘is paid’ as trade costs (Tij ¼ 1 if trade is costless). Taking these tradecosts into account gives the following profit function for each firm in country i,
pi ¼XR
j
pijðzÞxijðzÞ=Tij �wbi vgi ci½F þ
XR
j
xijðzÞ� ðC1Þ
where pij(z) is the price of a variety produced in country i.Turning to the demand side, consumers combine each firm’s manufacturing
variety in a CES-type utility function, with s being the elasticity of substitutionbetween each pair of product varieties. Given this CES-assumption, it followsdirectly that in equilibrium all manufacturing varieties produced in country iare demanded by country j in the same quantity (for this reason varieties areno longer explicitly indexed by (z)). Denoting country j’s expenditure on manu-facturing goods as Ej, country j’s demand for each product variety produced incountry i can be shown to be
xij ¼ p�sij EjGðs�1Þ:j ðC2Þ
where Gj is the price index for manufacturing varieties that follows from theassumed CES-structure of consumer demand for manufacturing varieties. It isdefined over the prices, pij, of all goods produced in country i and sold incountry j,
Gj ¼XR
i
nip1�sij
" #1=ð1�sÞ
ðC3Þ
Maximization of profits (C1) combined with demand as specified in (C2) gives
44. Since our main aim is to establish the relevance of market access we, in line with for instance
Breinlich (2006), skip intermediate inputs and thereby ignore supplier access for most of our analysis,
do however see Table 4c for estimation results for supplier access.
Bosker and Garretsen 481
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
the well-known result in the NEG literature that firms in a particular countryset the same f.o.b. price, pi, depending only on the cost of production in loca-tion i, i.e. pi is a constant markup over marginal costs:
pi ¼ wbi vgi cis=ðs� 1Þ ðC4Þ
As a result, price differences between countries in a good produced in country ican only arise from differences in trade costs, i.e. pij ¼ piTij.
Next, free entry and exit drive (maximized) profits to zero, pinpointing equi-librium output per firm at �x ¼ ðs� 1ÞF. Combining equilibrium output withequilibrium price (C4) and equilibrium demand (C2), and noting that in equi-librium the price of the internationally (perfectly) mobile primary factor ofproduction will be the same across countries (vi ¼ v for all i), gives the equilib-rium manufacturing wage:
wi ¼ Ac�1=bi
XR
j
mjTð1�sÞij
zfflfflfflfflffl}|fflfflfflfflffl{MAij
|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}MAi
0BBBB@
1CCCCA
1
bs
ðC5Þ
where A is a constant that contains inter alia the substitution elasticity, s, andthe fixed costs of production, F), and mj denotes country j’s market capacitythat is a combination of its expenditure on manufacturing goods (Ej) and theprice index for manufacturing varieties that it faces (Gj), i.e. mj ¼ Ej Gj
(s21). Inlog-linear form (C5) equates to equation (1) that underlies all our estimationresults in the main text of the paper.
Finally, aggregating demand from consumers in country j for a good pro-duced in country i (see (C2)) over all firms, ni, producing in country i, gives thefollowing aggregate export equation describing the total amount country iexports to county j.
EXij ¼ nip1�si EjG
ðs�1Þj T
ð1�sÞij|fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl}
MAij
¼ si½mjTð1�sÞij �|fflfflfflfflfflffl{zfflfflfflfflfflffl}
MAij
; ðC6Þ
where we make use of the fact that pij ¼ piTij and redefine si ¼ ni pi12s (what
Redding and Venables 2004 refer to as a country’s supplier capacity). Equation(C6) forms the basis of the two-step estimation procedure that we use in ourpaper.
482 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
RE F E R E N C E S
Amiti, M., and B.S. Javorcik. 2008. Trade Costs and Location of Foreign Firms in China. Journal of
Development Economics 85(1–2): 129–149.
Amiti, M., and L. Cameron. 2007. Economic Geography and Wages. Review of Economics and
Statistics 89(1): 15–29.
Amjadi, A., and A.J. Yeats. 1995. Have Transport Costs Contributed to the Relative Decline of
Sub-Saharan African Exports? Some Preliminary Evidence. World Bank Policy Research Paper no.
1559.
Anderson, J.E., and E. van Wincoop. 2004. Trade Costs. Journal of Economics Literature vol. XLII:
691–751.
Arora, V., and A. Vamvakidis. 2005. The implications of South African economic growth for the rest of
Africa. IMF Working Paper WP/05/58.
Bigsten, A., and M. Soderbom. 2006. “What have we learned from a decade of manufacturing enter-
prise surveys in Africa?” The World Bank Research Observer 21: 241–265.
Bosker, E.M., and H. Garretsen. 2010. Trade Costs, Market Access and Economic Geography. Papers
in Regional Science 89 (1): 485–513.
Bosker, E.M., S. Brakman, H. Garretsen, and M. Schramm. 2010. Relaxing Hukou-Increased Labor
Mobility and China’s Economic Geography. CEPR discussion paper 8106.
Breinlich, H. 2006. The spatial income structure in the European Union – what role for Economic
Geography? Journal of Economic Geography 6: 593–617.
Buys, P., U. Deichmann, and D. Wheeler. 2010. Road Network Upgrading and Overland Trade
Expansion in Sub-Saharan Africa. Journal of African Economies 19: 399–402.
Coe, D.T., and A.W. Hoffmaister. 1999. North-South Trade: Is Africa Unusual? Journal of African
Economics 8: 229–256.
Collier, P. 2002. Primary Commodity Dependence and Africa’s Future. Keynote Speech, World Bank’s
Annual Conference on Development Economics.
Collier, P., and J.W. Gunning. 1999. Explaining African Economic Performance. Journal of Economic
Literature 37: 64–111.
Collier, P., and A.J. Venables. 2007. Rethinking trade preferences: how Africa can diversify its exports.
The World Economy 30(8): 1326–1345.
Deichmann, U., S.V. Lall, S.J. Redding, and A.J. Venables. 2008. “Industrial location in developing
countries.” World Bank Research Observer 23(2): 219–246.
Easterly, W., and A. Reshef. 2010. African export successes: surprises, stylized facts, and explanations.
NBER Working Paper no. 16597.
Elbadawi, I., T. Menigstae, and A. Zeufack. 2006. Market access, supplier access, and Africa’s manufac-
tured exports: a firm level analysis. Journal of International Trade & Economic Development 15(4):
493–523.
Fally, T., R. Paillacar, and C. Terra. 2010. Economic geography and wages in Brazil: evidence from
micro-data. Journal of Development Economics 91: 155–168.
Foroutan, F., and L. Pritchett. 1993. Intra-Sub-Saharan African Trade: Is It Too Little?. Journal of
African Economics 2: 74–105.
Frazer, G., and J. van Biesebroeck. 2010. Trade growth under the African growth and opportunity pact.
Review of Economics and Statistics 92(1): 128–147.
Freund, C., and N. Rocha. 2011. What constrains Africa’s exports?. World Bank Economic Review
25(3): 361–386.
Fujita, M., P. Krugman, and A.J. Venables. 1999. The Spatial Economy MIT Press.
Gallup, J.L., J.D. Sachs, and A.D. Mellinger. 1999. Geography and Economic Development,
International Regional Science Review. 22: 179–232.
Bosker and Garretsen 483
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Hanson, G. 2005. Market Potential, Increasing Returns and Geographic Concentration. Journal of
International Economics 67(1): 1–24.
Head, K., and Th. Mayer. 2004. The Empirics of Agglomeration and Trade. in V. Henderson, and J-F.
Thisse (eds.) The Handbook of Regional and Urban Economics, volume IV North Holland 2609–
2665.
———. 2006. Regional wage and employment responses to market potential in the EU. Regional
Science and Urban Economics 36(5): 573–594.
———. 2010. Gravity, Market Potential and Economic Development U British Columbia/Sciences-Po.
Processed.
Helpman, E., M. Melitz, and Y Rubinstein. 2008. Estimating trade flows: trading partners and trading
volumes. Quarterly Journal of Economics Vol. 123, 441–487.
Henderson, J.V., Z. Shalizi, and A.J. Venables. 2001. Geography and Development. Journal of
Economic Geography 1: 81–105.
Hering, L., and S. Poncet. 2010. Market Access and individual wages: evidence from China. Review of
Economics and Statistics 92: 145–159.
Hummels, D. 2001. Toward a Geography of Trade Costs. mimeo, Purdue University
IMF, 2007. Regional Economic Outlook: Sub-Saharan Africa, Washington, D.C.
Johnson, S., J. Ostry, and A. Subramanian. 2006. Levers for growth. Finance and Development 43(1):
28–31.
Jones, B.F, and B.A. Olken. 2008. The anatomy of start-stop growth. Review of Economics and
Statistics 90(3): 582–587.
Knaap, T. 2006. Trade, location, and wages in the United States. Regional Science and Urban
Economics 36(5): 595–612.
Krugman, P. 1991. Increasing Returns and Economic Geography. Journal of Political Economy nr. 3,
483–499
Lall, S.V., Z. Shalizi, and U. Deichmann. 2004. Agglomeration economies and productivity in Indian in-
dustry. Journal of Development Economics 73: 643–673.
Limao, N., and Mayer, Th. 2008. Market Potential and Development. CEPR Discussion Paper no.
6798, CEPR, London.
Ndulu, B. 2007. Challenges to African Growth World Bank.
Nunn, N., and D. Puga. 2011. Ruggedness: the blessing of bad geography in Africa. Review of
Economics and Statistics forthcoming.
Portugal-Perez, A., and J.S. Wilson. 2008. Trade costs in Africa: barriers and opportunities for reform.
Policy research working paper, no.4619. World Bank.
Puga, D. 1999. The rise and fall of regional inequalities. European Economic Review Vol. 43, 303–
334.
Redding, S., and A.J. Venables. 2004. Economic Geography and International Inequality. Journal of
International Economics 62(1): 53–82.
Subramanian, A., and N.T. Tamirisa. 2003. Is Africa Integrated in the Global Economy?. IMF Staff
Papers 50(3): Washington, D.C.
Staiger, D., and J.H. Stock. 1997. Instrumental variables regression with weak instruments.
Econometrica 65: 557–586.
Temple, J., and L. Wossmann. 2006. Dualism and Cross-Country Growth Regressions. Journal of
Economic Growth 11: 187–228.
United Nations Economic Commission for Africa, 2010. Assessing Regional Integration in Africa IV
(ARIA IV) – Enhancing Intra-African Trade. Addis Ababa, Ethiopia.
United Nations Conference on Trade and Development. 2009. Economic Development in Africa Report
2009 – Strengthening Regional Economic Integration for Africa’s Development. New York and
Geneva.
484 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Van Biesebroeck, J. 2005. Exporting raises productivity in Sub-Saharan African manufacturing firms.
Journal of International Economics 67: 373–391.
Wooldridge, J.M. 2003. Introductory Econometrics-A Modern Approach Thomson, USA.
World Bank. 2007. Accelerating Development Outcomes in Africa-Progress and Change in the Africa
Action Plan Africa Region, The World Bank, Washington, D.C.
———. 2008. World Development Report 2009 The World Bank, Washington, D.C.
Bosker and Garretsen 485
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
The Decision to Import Capital Goods in India:Firms’ Financial Factors Matter
Maria Bas and Antoine Berthou
Are financial constraints preventing firms from importing capital goods? Sourcingcapital goods from foreign countries is costly and requires internal or external finan-cial resources. A simple model of foreign technology adoption shows that credit con-straints act as a barrier to importing capital goods under imperfect financial markets.In our study, we investigate this prediction using detailed balance-sheet data fromIndian manufacturing firms having reported information on financial statements andimports by type of good over the period 1997–2006. Our empirical findings shed newlight on the micro determinants of firms’ choices to import capital goods. Baseline es-timation results show that firms with a lower leverage and higher liquidity are morelikely to source their capital goods from foreign countries. Quantitatively, a 10 per-centage point improvement of the leverage or liquidity ratio increases the probabilityof importing capital goods by 11 percent to 13 percent respectively. Different robust-ness tests demonstrate that these results are not driven by omitted variable bias relatedto changes in firm observable characteristics as well as ownership status. These find-ings are also robust to alternative specifications dealing with the potential reversecausality issues. JEL codes: F10, F14, D92
Globalization is characterized by a significant increase in world imports ofcapital goods and intermediate inputs. In developing countries, a number offirms rely on capital goods and inputs from abroad since they are moreadvanced in terms of technology relative to the domestic goods. While the lit-erature on endogenous growth provides theoretical grounds for the role offoreign technology to enhance economic growth, recent firm-level studiesconfirm that firm performance depends critically on the access to inputs used
Maria Bas is an economist at the Centre d’Etudes Prospectives et d’Informations Internationales
(CEPII); her email is [email protected]. Antoine Berthou (corresponding author) is an economist at the
Banque de France and is associate researcher at CEPII; his email is [email protected].
The authors thank Amelie Maingault for her excellent research assistance. We have benefited from
discussions with Jens Arnold, Agnes Benassy-Quere, Matthieu Crozet, Joze Damijan, Joep Konings,
Tibor Besedes, Ben Li, Kalina Manova, Bruno Merlevede, Sandra Poncet, Romain Ranciere and
Elisabeth Sadoulet. We also thank three anonymous referees for their valuable comments to the
manuscript and their constructive suggestions. This research was carried out while Antoine Berthou was
an economist at CEPII, and does not reflect the views of the Banque de France.
THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 486–513 doi:10.1093/wber/lhs002Advance Access Publication February 8, 2012# The Author 2012. Published by Oxford University Press on behalf of the International Bankfor Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions,please e-mail: [email protected]
486
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
in the production of final goods.1 Amiti and Konings (2007) find that input-trade liberalization in Indonesia boosts firm productivity up-to 12 percent,while Khandelwal and Topalova (2010) show that it improved firm productiv-ity by 4.8 percent in India. Goldberg and others (2010) demonstrate that input-tariff cuts in India account on average for 31 percent of the new products intro-duced by domestic firms. Using firm-level data from Argentina, Bas (2011)finds that input tariff reductions are associated with 8 percent increase in theprobability of exporting.2
While the use of foreign technology is expected to increase firm efficiency,foreign technology adoption is conditioned by the access to financial resources.Importing capital goods implies incurring fixed costs associated with gatheringinformation on foreign markets, establishing linkages with foreign suppliers,learning the new technology or adapting the production process, whichrequires external financing.3 In our study, we argue that financial constraintsrepresent an important barrier to firms’ imports of capital goods, thereby limit-ing their opportunities to benefit from technological spillovers of foreigncountries.
First, we present a simple theoretical framework to rationalize the mainmechanisms through which financial access affects firms’ foreign technologychoice. In this framework, using foreign capital goods increases the efficiencyto produce final goods, but requires paying an additional fixed cost. In thepresence of financial constraints wealthier firms have a better access to externalfinance and are more likely to use the foreign technology by importing capitalgoods. Second, we test this relationship between firms’ financial statements andtheir decision to import capital goods using a detailed Indian firm-level dataset,Prowess. This data was collected by the Centre for Monitoring the IndianEconomy (CMIE) for the period 1997–2006.4 During this period, about 75percent of imports of capital goods in India are originated from high incomeOECD countries.5 The Prowess data provides information on financial charac-teristics of firms as well as imports distinguished by type of goods (capitalequipment, intermediate goods, or final goods). This information allows us tocompute the liquidity and leverage ratios that are used throughout the paper tomeasure firms’ financial factors. These balance sheet statements are expected to
1. Ethier (1982), Markusen (1989), Grossman and Helpman (1991), Rivera-Batiz and Romer
(1991) develop theoretical models where foreign technology acts as a driver of economic growth.
2. Kasahara and Rodrigue (2008), Halpern and others (2009), Schor (2004), Kugler and Verhoogen
(2009), Bas and Strauss-Kahn (2011) find empirical evidence that the use of foreign inputs enhances
firms’ total factor productivity, the quality of final goods, and the number of products exported by
firms.
3. See Eaton and Kortum (2001), who quantify that about 25 percent of cross-country productivity
differences can be explained by the relative price of equipment, half of it being due to barriers to trade
in equipment.
4. We focus on the period 1997–2006 in order to maximize the number of firms each year.
5. This number is obtained by using the HS6 product-level bilateral trade BACI dataset from CEPII,
combined with the Broad Economic Product Classification provided by the United Nations.
Bas and Berthou 487
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
be positively related to the borrowing capacity in the presence of financial con-straints. Our empirical strategy disentangles the impact of the liquidity and le-verage ratios of the firm on the decision to invest in foreign capital goods.
Our empirical findings confirm the theoretical prediction that those firmsthat are ex-ante more liquid and less leveraged are more likely to importcapital goods. In our baseline estimations, a 10 percentage point decrease inthe leverage ratio or an equivalent increase in the liquidity ratio for the averagefirm increases the likelihood of importing capital goods by 11 percent and bymore than 13 percent, respectively. These results are robust to changes in firmobservable characteristics such as firm size, capital and skill-intensity. We carryout different tests that demonstrate that our results are not influenced byomitted variable bias related to India’s trade liberalization. Our results remainalso robust to the exclusion of multinational firms, state-owned firms, andlocal business groups.
We provide robustness tests to account for the possibility that using foreigncapital goods may improve financial factors of firms ex-post. First, we focus onthe sample of firms that have started importing foreign capital goods, by con-sidering in the empirical analysis only those firms that did not import capitalgoods in the previous two to four years. As an additional related test, weinclude in the baseline specification the past importer status to take intoaccount previous import experience. Second, we use the measure of externaldependence proposed by Rajan and Zingales (1998) to test whether financialfactors are more important in industries where firms rely more on externalfinance. These results confirm that the leverage (liquidity) of the firm has astrong negative (positive) effect on the probability of importing foreignequipments.
These results complete the existing evidence regarding the determinants offirm performance in the case of the Indian economy. Many of these workshave used the Prowess data over a comparable period of time. Alfaro andChari (2009) show that although the importance of private firms in the Indianeconomy has been growing after the economic reforms in the early 1990s,state-owned firms still represent an important share of total production andassets in some sectors. Khandelwal and Topalova (2010) and Goldberg andothers (2010, 2009) study the micro-economic effects of trade liberalization inIndia. Input-tariffs cuts have contributed significantly to firm productivitygrowth (Khandelwal and Topalova 2010), and also to the ability of firms tointroduce new products (Goldberg and others 2010, 2009). On the same line,Arnold and others (2010) find that product market reforms in services sectorshave an important effect on firm productivity in the manufacturing sector inIndia. Evidence regarding the importance of financial factors in explainingIndian firm performance is more scarce: Topalova (2004) shows that althoughIndian firms improved their financial statements during the period of economicreforms, some firms still face problems servicing their debt obligations.
488 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
This work also contribute to the literature in finance that connects financialfactors and firms’ investment decision. In the presence of information asymmet-ries, uncollateralized external financing becomes more costly than internal fi-nancing, thus introducing a positive relation between a firm’s net worth and itsinvestment decision. This link has been empirically observed for a number ofcountries and surveyed by Hubbard (1998). These studies (Fazzari and others,1988, Whited 1992, Bond and Meghir 1994, Bond and others 2003) use firms’financial indicators such as the cash flow, the debt to assets ratio, or the liquid-ity ratio as proxies for firms’ net worth or collateral. Most of these papers relyon data for OECD economies and show that wealthier firms invest more.Similar evidence is found for Ecuador (Haramillo and others 1996) and Coted’Ivoire (Harrison and McMillan 2003). In a different setting, Gorodnichenkoand Schnitzer (2010) use a survey of firms in Eastern European countries andshow that financial constraints decrease investment in innovation by domesticfirms. Aghion and others (2008) alternatively use measures of firms’ paymentincidents for France to analyze the relation between credit constraints and re-search and development along the business cycle. We build on this literatureand provide new evidence that financial constraints are preventing firmslocated in India to invest in foreign capital equipment.
Previous empirical studies have investigated whether financial constraints in-fluence the export decisions of firms in the United Kingdom (Greenaway andothers 2007) and in several developing economies (Berman and Hericourt2010). The negative effect of financial constraints on export decisions isobserved for the sample of developing countries, but not in the case of the UK.These studies, however, elude the question of financial constraints as a deter-minant of foreign technology adoption through the imports of foreign capitalgoods. This is the focus of our study.
In the next section, we present a simple theoretical framework of import de-cision and credit constraints. Section II describes the data and introduces theestimation strategy. In Section III we present the baseline empirical results.Section IV presents several robustness checks. In the last section, we presentour conclusion.
I . T H E O R E T I C A L M O T I VA T I O N
The aim of this section is to motivate our empirical analysis by introducing asimple model of endogenous adoption of foreign technology. The theory ratio-nalizes the mechanisms through which credit constraints affect firms’ decisionto upgrade foreign technology. The model is based on firm heterogeneity interms of productivity a la Melitz (2003). Firms are also characterized by theirinitial wealth as in Chaney (2005).6 They use this wealth as a collateral to get
6. Previous models of heterogeneous firms and credit constraints have also used this framework to
explain the determinants of export decision. See Manova (2008) and Muuls (2008).
Bas and Berthou 489
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
external finance in the presence of financial constraints. The representativehousehold allocates consumption from among the range of domestic goods ( j)produced using domestic-low technology (Vd) and those produced usingforeign-high technology (Vf ).
7
Production
There is a continuum of firms, which are all different in terms of their initialproductivity (w). This productivity draw is derived from a common distributiondensity g(w), after firms decide to enter the market. Each firm produces its ownvariety in a monopolistic competition market structure. In order to produce thefinal good (y) firms must pay a fixed production cost (F) and they need tocombine two inputs: labor (l) and physical capital (k). There are two types ofcapital equipment goods: domestic (z) and imported (m).8 However, only thosefirms that are productive enough to adopt the foreign technology are able toproduce with imported capital goods. Heterogeneous firms in terms of differentproductivity levels (w) are introduced. Technology is represented by the follow-ing Cobb-Douglas production function that combines labor (l) and capitalgoods (k) to produce output with factor shares h and 1 2 h:
yi ¼ wgi
ki
h
� �h li1� h
� �1�hi ¼ d; ff gð1Þ
The subscript d corresponds to firms producing with domestic technology and fto those producing with foreign technology embodied in imported capitalgoods. The coefficient g represents the efficiency of imported capital goodsrelative to domestic ones. Firms using only domestic capital goods (i ¼ d) haveg ¼ 1 and kd ¼ z. Firms producing with foreign technology (i ¼ f ) combineboth types of capital goods by a Cobb-Douglas function: kf ¼ z
a
� �a m1�a� �1�a
.Firms that decide to adopt foreign technology increase their productivity levelby a factor g . 1. To access imported capital goods firms must pay a fixedforeign technology acquisition cost (FT). The fixed technology costs are asso-ciated with gathering information on foreign markets, learning about theforeign technology and establishing linkages with foreign suppliers of this tech-nology. To keep the model simple, we assume that the fixed cost for domesticcapital goods is included in the fixed production cost. These assumptionsreflect the fact that for a developing country like India, foreign capital goods
7. The standard CES utility function (C) represents the consumer preferences
Cw�1w ¼
Ðj[Vd
Cw�1w
dj djþÐ
j[VfC
w�1w
fj dj. The elasticity of substitution between both types of goods is given
by f . 1. The optimal relative demand functions are: Ci ¼ Pp
i
� �w
C, where P represents the price index,
C the global consumption and pi the price set by a firm.
8. To keep the model simple, we assume that one unit of domestic capital good is produced using
one unit of labor, which is elastically supplied and the wage is normalized to one.
490 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
are more advanced in terms of technology relative to domestic goods, but theyare also associated with ahigher initial investment.9
The first-order condition of monopolistic firms is such that prices reflect a
constant mark-up, r ¼ w�1w
� �, over marginal costs: pi ¼ ci
rw. ci represents the per
unit cost of production, which is different among firms depending on whetheror not they have adopted the foreign imported technology: cd ¼ ph
z and
cf ¼ pzð Þah tmpzð Þ 1�að Þh
g: The price of domestic capital good is pz and the price of
imported capital takes into account transport costs and tariffs (tm): pm ¼ tmpz .
The relative per unit cost is equal tocf
cd¼ t
h 1�að Þm
g. We assume that the efficiency
parameter of imported capital goods (g) is higher than its additional variablecost (tm) relative to domestic ones.10
Combining the demand faced by each firm, qiðwÞ ¼ PpiðwÞ
� �wC; and the price
function, piðwÞ ¼ ci
rAi, revenues are given by riðwÞ ¼ qiðwÞpiðwÞ ¼ P
pi
� �w�1R;
where R ¼ PC is the aggregate revenue of the industry exogenous to the firm.Firm profit is then pi ¼ ri
w� F; where F is the fixed production cost.
Firm’s Decision under Perfect Financial Market Conditions
Only those firms with enough profits to afford the fixed production (F) costwill be able to survive and produce. Profits of the marginal firm are equal to
zero. The zero cutoff condition is given by:rd w
d�ð Þ
w¼ F. The value w�d represents
the productivity cutoff to produce in the domestic market.Once a firm has received its productivity draw, it may also decide to adopt a
foreign technology to reduce its marginal costs on the basis of its profitability.Only a subset of the most productive firms will switch to foreign technologysince the fixed importing cost is higher than the fixed production cost. The con-dition to acquire the foreign technology is given by: pf ðwf�Þ ¼ 0. The value w�f
represents the productivity cutoff to import foreign goods:rf ðwf
�Þw¼ F þ FT .
Firm’s Decision under Imperfect Financial Market Conditions
Importing technology embodied in foreign capital goods implies a sunk cost ofinvestment (FT). In the presence of financial constraints, firms cannot use theirfuture expected revenues rf(w) to get external finance ex-ante. In this context,
9. Using product-level imports for India (from the BACI data), we find that about 75 percent of
imports of capital goods in India during the 1997-2006 period are sourced from high income OECD
economies. This confirms that capital goods are mostly imported from countries that are more advanced
in terms of technology.
10. Note that the relative per unit cost is a function of tariffs on capital goods and the efficiency
parameter. A reduction of import tariffs on capital goods reduces the relative per unit costs of foreign
technology. Similar results hold in the case of an increase in the efficiency parameter of foreign
technology (g).
Bas and Berthou 491
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
firms can make use of two sources of cash to finance the extra fixed cost FT.First, firms are able to borrow up to rd(w), which corresponds to the sales ofthe final good for firms using the domestic technology. Financial intermediarieshave perfect information about firms’ profitability in the case where theyproduce with the domestic technology, and will be willing to provide cash inadvance up to rd(w). Second, firms can use their exogenous wealth A as a col-lateral to borrow additional liquidity lA, where l corresponds to the creditmultiplier and is inversely related to the extent of credit constraints in theeconomy, as in Aghion and others (1999).
We assume that the productivity and the exogenous collateral distributionsare independent. The total liquidity that is available to the firm is equal topd(w) þ lA. Importing foreign capital goods relates to the liquidity constraintcondition (LCC) given by
pdðwÞ þ lA � FTð2Þ
We can define the lowest productivity level below which firms with an exogen-ous wealth A, wðAÞ, are liquidity constrained. wðAÞ is given bypdðwðAÞÞ þ lA ¼ FT . Firms that face liquidity constraints have a productivitylevel below wðAÞ: They are not able to import capital goods due to financialconstraints.
Following Chaney (2005), we set w�d ¼ gðFÞ and use the zero cutoff profitconditions and the liquidity constraint condition, equation (2), to define twoproductivity cutoffs11:
w�f ¼F þ FT
F
� � 1w�1 t
h 1�að Þm
g
!w�d;wðAÞ ¼ FT þ F � lA
F
� � 1w�1
w�d
All the firms with a productivity level between maxfw�f ;wðAÞg . w . w�dproduce with domestic technology. Only those firms with a productivityw . maxfw�f ;wðAÞg are able to finance the fixed technological cost of import-ing and thus they use both types of capital goods.
Which are the firms that face credit constraints to import capital goods?There is a subset of firms that are profitable enough to be viable importers, butprevented from accessing foreign capital goods because of liquidity constraints.Firms that have a productivity level w below wðAÞ are liquidity constrained,and are not able to source imported inputs from abroad no matter how profit-able they could be by importing more efficient foreign capital goods. All firmswith a productivity level above w�f could profitably import, if they had sufficientliquidity. Hence, there is a subset of liquidity constrained firms with a
11. For tractability purposes we assume, as in Chaney (2005), that the price index only depends on
local firms’ prices. In the Appendix we define the price index approximation.
492 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
productivity level above wf�, but below wðAÞ: In the appendix we demonstratethe existence of liquidity constrained importers.
Testable Prediction
Firms’ import decision is determined by domestic revenues and by the exogen-ous collateral. These two sources of finance allow firms to afford the fixed tech-nology cost of importing. Using equation (2) we can define the probability thata firm i imports capital goods at time t:
PrðpdþlA�FT . 0Þ ¼ Prðwf�1 1
w
r
cd
� �f�1
RPw�1þlA� F � FTÞ . 0ð3Þ
The probability of importing is directly determined by the two sources offinance. On the one hand, in this monopolistic competition framework withheterogeneous firms, the most productive firms set lower prices and have largerdomestic revenues to finance the fixed importing cost. On the other hand thehigher the exogenous collateral, the greater the financial resources of the firmto afford the fixed foreign technology cost.
Testable prediction: In the presence of financial constraints, wealthier firmsare more likely to import foreign equipment and upgrade foreign technology.
I I . D A T A A N D E M P I R I C A L M E T H O D O L O G Y
In the empirical part of the paper, we present a test of the prediction that isderived from the theoretical model. The empirical strategy is based on the esti-mation of an equation where the import decision of a firm is explained by its fi-nancial factors such as the liquidity or leverage ratios. Estimations areperformed using information for a sample of 3,500 Indian listed companies(Prowess data) over the period 1997–2006.
Data
The Indian firm-level dataset is compiled from the Prowess database by theCentre for Monitoring the Indian Economy (CMIE). This database contains in-formation from the income statements and balance sheets of listed companiescomprising more than 70 percent of the economic activity in the organized in-dustrial sector of India. Collectively, the companies covered in Prowess accountfor 75 percent of all corporate taxes collected by the Government of India. Thedatabase is thus representative of large and medium-sized Indian firms.12
The dataset covers the period 1997–2006 and the information varies byyear. It provides quantitative information on sales, capital stock, income from
12. Since firms are under no legal obligation to report to the data collecting agency, the Prowess
data do not allow properly identifying entry and exit of firms.
Bas and Berthou 493
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
financial and non financial sources, consumption of raw material and energy,compensation to employees and ownership group.
The Prowess database provides detailed information on imports by categoryof goods: finished goods, intermediate goods and capital goods. In our mainempirical specification, we use imports of capital goods (machinery and equip-ment) as a proxy of foreign technology. Although we are not able to test direct-ly for the impact of imported capital goods depending on the countryof origin(e.g developed vs. developing countries), one realistic assumption for the caseof a developing country like India is that most imports of capital goods aresourced from more advanced economies.
The dataset contains also comprehensive information about the financialstatements of firms such as total assets, current assets, total debt and liabilities.We construct two financial variables: (1) the leverage ratio and (2) the liquidityratio. Leverage is the ratio of borrowing over total assets and liquidity ratio ismeasured by the ratio of current assets over total liabilities of the firm.
Summary statistics are provided in the Appendix Table. Our sample containsinformation for 3,500 firms on average each year in organized industrial activ-ities from manufacturing sector for the period 1997–2006. The total numberof observations firm-year pairs is 34,735. In order to keep a constant samplethroughout the paper and to establish the stability of the point estimates, wekeep firms that report information on all the firm and industry level controlvariables. On average 32 percent of firms import capital goods in a year and62 percent of firms import intermediate goods. Firms are categorized by indus-try according to the 4-digit 1998 NIC code (116 industries). Most of the firmsin our sample are private-owned firms (81 percent). 39 percent of firms arelargest firms belonging to local business groups and only 7 percent aremultinational firms. Although our panel of firms is unbalanced, there is nostatistical difference in the average firm characteristics presented in theAppendix Table between the initial year (1997) and the final year (2006) ofour sample.
Two industry-level controls are included in the empirical specifications tocontrol for competitive pressures. Since the period under analysis covers tradeliberalization process started in the early 1990s, we introduce effectivelyapplied output tariffs (collected rates) at the 4-digit NIC code level obtainedfrom the World Bank (WITS).13 In order to capture domestic competition weuse an Herfindhal index computed at the 2-digit NIC industry level. TheHerfindahl index measures the concentration of sales for each industry within2-digit industry categories.
13. Tariffs data provided by WITS are at the industry level ISIC rev 2 4-digit level. We use
correspondence tables to convert tariffs into ISIC rev 3.1. that match almost perfectly with NIC 4-digit
classification.
494 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Empirical Methodology
A unique feature of our database is that firms report imports by type of pro-ducts: finished goods, capital goods and intermediate goods. Keeping in linewith our theoretical framework, the baseline econometric analysis is thereforeperformed on capital goods. The rationale for this is that importing capitalgoods implies incurring fixed costs associated with gathering information onforeign markets, establishing linkages with foreign suppliers, and learningabout the new foreign technology. In the case of a developing economy likeIndia, firms’ importing capital goods decision can be interpreted as foreigntechnology adoption.
We estimate a linear probability model, where the decision of a firm i toimport capital goods from abroad in year t is explained by its financial factorsand additional control variables. Our preferred specification estimates the fol-lowing equation using the following model:
ImporterðisÞðtÞ ¼ b0 þ b1FinanceðiÞðt�1Þ þ b2ZðiÞðt�1Þ þ b3XðsÞðtÞ þ yt þ mi þ nitðIÞ
where Importer(is)(t) is a dummy variable equal to one if the firm i producing in4-digit NIC code industry s, has positive imports of capital goods in year t andzero otherwise. Finance measures firms’ financial statements. The financial vari-ables of interest that we use to proxy the financial factors (the empirical coun-terpart of the exogenous collateral in the model) are the liquidity ratio and theleverage ratio. The liquidity ratio is the share of firms’ current assets over totalliabilities. The liquidity ratio is related to the firm’s ability to pay off its short-terms debts obligations. The leverage ratio indicates the proportion of borrow-ing over total assets of the firm. A higher level of leverage decreases, everythingelse equals, the net worth of the firm. According to the model’s predictions, animprovement of the firm’s wealth (measured by a higher liquidity ratio or alower leverage), increases the access to external finance. Since the access to ex-ternal finance determines the decision to source capital goods from abroad, weexpect a positive coefficient for the liquidity ratio and a negative coefficient forthe leverage ratio.
Unobserved firm characteristics could lead to inconsistent estimates. For thisreason, all estimations include firm-level fixed effects (mi). The introduction offirm fixed effects is important to control for unobservable firm characteristicsthat do not vary over time. Our specification shows how improvements infirms’ financial factors over time affects firms’ decisions to import.
Estimates also include controls for firm and industry characteristics that varyover time. First, we introduce a set of firm level variables (Z(i)(t21)) expressed inlogarithm in year (t-1) that control for observable firm characteristics that
Bas and Berthou 495
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
mightaffect firms’ import choices. We use the value added to measure firms’size (the number of employees is not available in the Prowess data). In alterna-tive specifications, we use firm total factor productivity (TFP) computed usingLevinsohn and Petrin (2003) methodology, by relying on wage bill rather thanlabor.14 Since larger firms tend to be more skill intensive and to pay higherwages, we also control for the wage-bill. As we focus on the import decision ofcapital equipment goods, we also include the past capital intensity of the firmmeasured as total capital stock over the wage-bill. We expect a positive coeffi-cient of capital intensity. The more firms rely on capital goods in the produc-tion process, the more likely they are to import capital goods from abroad.
Second, we introduce a set of industry level variables X(st) that control for ob-servable industry characteristics that might affect firms’ import choices of capitalgoods. Several studies show that competition might enhance firm efficiency andcreate incentives for firms to invest in R&D activities and in foreign technology(Aghion and others 2005). We construct a Herfindahl index at the 2-digit NIC in-dustry level to control for competition in the domestic market. We also control forforeign competition pressures associated with the trade liberalization processexperienced by India at the beginning of the 1990s, by including the average ef-fective applied import tariffs for final goods at the 4-digit NIC industry level.
All explanatory variables are expressed in logarithm and they are lagged by oneperiod. We also introduce year fixed effects to control for macroeconomic shocks(y t). This is an important control since India was affected by the Asian financialcrisis in 1997-1998. The introduction of year fixed effects allows us to control forthe effects of this crisis on both financial statements of firms and their import deci-sions. In the last section we deal explicitly with the potential reversecausalitybetween financial factors and firms’ investment decision in imported capital goods.
I I I . E S T I M A T I O N R E S U L T S
The estimation results of the import decision equation are presented in thisthird section of the article. All estimations are performed using the above men-tioned firm-level data for India (Prowess). The testable prediction from themodel states that “in the presence of financial constraints, wealthier firms aremore likely to import foreign equipment and upgrade foreign technology.”
Baseline Results
Are Financial Factors Related to Firms’ Decision of Sourcing Foreign CapitalGoods? Estimation results of the linear probability model (equation I) are pro-vided in Table (1). The estimation includes firm and year fixed effects. Theeffect of the leverage ratio lagged of one period on firms’ decision to import
14. Because our dataset does not contain the number of employees, we can not rely on the
extension of Olley and Pakes (1996) and Levinsohn and Petrin (2003) developed by Ackerberg and
others (2007) to estimate total factor productivity.
496 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E1
.A
cces
sto
Exte
rnal
Fin
ance
and
Import
of
Capit
al
Goods
Dec
isio
n(1
997
–2006)
Dum
my
equal
one
iffirm
(i)
import
sca
pit
al
goods
int
Dep
enden
tvari
able
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Lev
erage(
i)(t
21)
20.1
56***
20.1
32***
20.1
14***
20.1
14***
(0.0
16)
(0.0
16)
(0.0
16)
(0.0
16)
Liq
uid
ity
rati
o(i
)(t2
1)
0.1
66***
0.1
09***
0.1
26***
0.1
26***
(0.0
26)
(0.0
26)
(0.0
26)
(0.0
26)
Log
valu
eadded
(i)(
t21)
0.0
29***
0.0
16***
0.0
16***
0.0
30***
0.0
15***
0.0
15***
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
Capit
al
inte
nsi
ty(i
)(t2
1)
0.0
36***
0.0
36***
0.0
44***
0.0
44***
(0.0
05)
(0.0
05)
(0.0
05)
(0.0
05)
Log
wage(
i)(t
21)
0.0
55***
0.0
55***
0.0
60***
0.0
60***
(0.0
07)
(0.0
07)
(0.0
06)
(0.0
06)
Outp
ut
tari
ffs(
s)(t
21)
0.0
26
0.0
20
(0.0
36)
(0.0
36)
Her
findahl
index
(s)(
t21)
0.0
01
0.0
00
(0.0
03)
(0.0
03)
Fir
mfixed
effe
cts
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yea
rfixed
effe
cts
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Obse
rvat
ions
34,7
35
34,7
35
34,7
35
34,7
35
34,7
35
34,7
35
34,7
35
34,7
35
R2
0.0
16
0.0
21
0.0
25
0.0
25
0.0
13
0.0
18
0.0
24
0.0
24
Note
s:T
he
table
report
ses
tim
ates
from
linea
rpro
babilit
yes
tim
atio
ns
of
Equat
ion
(I).
The
dep
enden
tvari
able
isa
dum
my
equal
toone
ifth
efirm
iim
port
sca
pit
algoods
int.
All
expla
nat
ory
vari
able
sare
lag
of
one
per
iod.
Fir
ms’
capit
al
inte
nsi
tyis
the
rati
oof
capit
al
ove
rth
ew
age-
bill.
The
financi
al
vari
able
sth
atw
euse
are
leve
rage(
i)and
liquid
ity
rati
o(i
).L
ever
age(
i)is
the
rati
oof
borr
ow
ings
ove
rto
tal
ass
ets
and
liquid
ity
rati
o(i
)is
the
rati
oof
curr
ent
ass
ets
ove
rto
tal
liabil
itie
sof
the
firm
.T
he
outp
ut
tari
ffs
are
atth
e4-d
igit
NIC
indust
ryle
vel
and
the
Her
findahl
index
isat
the
2-d
igit
NIC
indus-
try
leve
l.In
pare
nth
eses
we
report
het
erosk
edast
icit
y-r
obust
standard
ser
rors
.***,*
*,
and
*in
dic
ate
signifi
cance
atth
e1,
5and
10
per
cent
leve
lsre
spec
tive
ly.
Sourc
e:A
uth
ors
’es
tim
atio
ns
usi
ng
Pro
wes
sdat
a.
Bas and Berthou 497
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
capital goods is negative and significant at the 1 percent level: firms having ahigher ratio of borrowing over total assets are less likely to import their capitalgoods from the foreign market (column 1). Next, we include firm level variablesto control for firm characteristics that vary over time and that could be pickingup the effect of firms’ financial factors. As expected, bigger firms are morelikely to import capital goods from abroad (column 2). We next introduce twoadditional firm-level controls: capital intensity and wage-bill in column (3).More capital and skill-intensive firms (paying higher wages) have a higherprobability of upgrading foreign technology. The coefficient of interest remainsrobust and also stable when we control for firm observable characteristics.
Moreover, firms producing in industries growing faster might be less creditconstrained. If this is the case, changes in firms’ financial statements might becapturing the effects of industry characteristics. We address this issue by intro-ducing in column (4) additional controls at the industry level. Concerning thetime variant industry characteristics, both coefficients of output tariff andHerfindahl index are not significant. More importantly, the negative effect ofleverage on firms’ foreign technology decision remains unchanged and robustto the inclusion of this set of industry controls. The estimated coefficientsimply that a 10 percentage point fall in the leverage ratio leads to an 11percent to 15 percent increase in the likelihood of importing capital goods forthe average firm.
We test how a firm’s liquidity ratio affects its probability to upgrade foreigntechnology embodied in imported capital goods in columns (5) to (8). Thelagged liquidity ratio is subsequently introduced. The coefficient of the liquidityratio is positive and significant at the 1 percent level, indicating that moreliquid firms are more likely to import their capital goods from abroad (column5). Results on the liquidity ratio remain robust to the inclusion of firm size,capital and skill-intensity in column (6) and (7) and also to the inclusion ofindustry-level characteristics in column (8). Moreover, the point estimates ofthe liquidity ratio are stable under different specifications. The estimated coeffi-cients imply that a 10 percentage point increase in the liquidity ratio leads to a13 percent to 17 percent increase in the likelihood of importing capital goodsfor the average firm.
Based on these estimations, we use the standard deviation of the leverageand liquidity ratios within firms to have a quantification of their economicimpact on the import decision of capital goods by the average firm. We findthat a one standard deviation reduction of the leverage ratio, corresponding toa decrease of leverage of the average firm by 32 percent, increases the probabil-ity of sourcing capital goods from abroad by 3.6 percent. A one standard devi-ation increase of the liquidity ratio, corresponding to an increase of theliquidity of the average firm by 17 percent, improves the probability of
498 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
importing capital goods by 2.1 percent. These results confirm that firms finan-cial factors are important determinants of the decision to import capitalgoods.15
Are Imports of Intermediate Inputs also Affected by Financial Constraints ?In order to disentangle the mechanisms through which financial access affectfirms’ foreign technology upgrading, we consider separately the special case ofintermediate goods imports. This test allows us to determine whether financialfactors affect differently imports of intermediate goods relative to capitalgoods.
First, we estimate equation (I) for the subsample of firms using foreign inter-mediate goods. Results are reported in Table 2. Once we control for firm andindustry characteristics, the leverage ratio has a negative but not significanteffect on firms’ import decision to use foreign intermediate goods (column (1)).Next, we investigate whether the decision to import intermediate goods is asso-ciated to the decision of upgrading foreign technology embodied in capitalgoods. If such a complementarity exists, the effect of financial factors onimports of intermediates may arise because of its effect through capital goods.To isolate the effect of capital goods decision from foreign input decision, werestrict our sample to firms that have never imported capital goods in theperiod (columns 2). The results are similar to the previous ones. These findingsindicate that credit constraints are not crucial for importing foreign inputs.
In the next columns, we reproduce the same specifications using the liquidityratio. The effect is positive and significant in both cases for the full sample(column 3) and the sample of firms that have never imported capital goods inthe period (column 4). The higher the current assets over total liabilities ratioof the firm, the more likely firms are to import their inputs from abroad.
Given that firms tend to import intermediates on a regular basis, the positiveeffect of the liquidity of the firm may not be specifically related to the decisionto start importing intermediates. Indeed, 62 percent of firms in our samplereport importing intermediates, whereas 32 percent import capital goods(Appendix Table). This evidence suggests that firms find it more difficult toimport capital goods than intermediates due to a larger fixed cost. In order toexplore the effect of financial health of firms on their decision to start sourcinginputs from abroad, we carry out an additional test focusing on firms that havenot imported intermediate goods in the previous two years. These results arereported in columns (5) and (6). As can be seen when we focus on the decisionto start importing intermediate inputs, not only the leverage ratio is not signifi-cant but now the liquidity ratio is no longer significant. Firm size, capital and
15. These results are also robust to alternative econometric specifications, available upon request,
such as Conditional Logit estimations with firm fixed effects. We only report the linear probability
model estimations since the parameters are easier to interpret and their stability easier to establish.
Bas and Berthou 499
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E2
.D
ecis
ion
toim
port
inte
rmed
iate
s
Dum
my
equal
one
iffirm
(i)
import
sin
term
edia
tes
int
(5)
(6)
Dep
enden
tvari
able
(1)
(2)
(3)
(4)
Fir
ms
that
hav
enot
import
edin
puts
inth
epast
2ye
ars
Lev
erage(
i)(t
-1)
20.0
25
20.0
24
20.0
08
(0.0
17)
(0.0
25)
(0.0
21)
Liq
uid
ity
rati
o(i
)(t2
1)
0.0
69***
0.1
24***
0.0
03
(0.0
26)
(0.0
34)
(0.0
34)
Log
valu
eadded
(i)(
t21)
0.0
21***
0.0
18***
0.0
20***
0.0
16***
0.0
15***
0.0
15***
(0.0
03)
(0.0
04)
(0.0
03)
(0.0
04)
(0.0
03)
(0.0
03)
Capit
al
inte
nsi
ty(i
)(t2
1)
0.0
42***
0.0
37***
0.0
45***
0.0
39***
0.0
41***
0.0
41***
(0.0
05)
(0.0
07)
(0.0
05)
(0.0
07)
(0.0
07)
(0.0
07)
Log
wage(
i)(t
21)
0.0
86***
0.0
73***
0.0
87***
0.0
73***
0.0
85***
0.0
86***
(0.0
07)
(0.0
10)
(0.0
07)
(0.0
09)
(0.0
09)
(0.0
09)
Outp
ut
tari
ffs(
s)(t
21)
0.0
39
0.0
95*
0.0
38
0.0
92*
20.0
12
20.0
12
(0.0
42)
(0.0
51)
(0.0
42)
(0.0
51)
(0.0
48)
(0.0
48)
Her
findhal
index
(s)(
t21)
0.0
03
0.0
05
0.0
03
0.0
05
0.0
02
0.0
02
(0.0
02)
(0.0
04)
(0.0
02)
(0.0
04)
(0.0
04)
(0.0
04)
Fir
mfixed
effe
cts
Yes
Yes
Yes
Yes
Yes
Yes
Yea
rfixed
effe
cts
Yes
Yes
Yes
Yes
Yes
Yes
Obse
rvat
ions
34,7
35
18,4
77
34,7
35
18,4
77
16,1
90
16,1
90
R2
0.0
33
0.0
27
0.0
34
0.0
28
0.0
36
0.0
36
Note
s:T
he
table
report
ses
tim
ates
from
linea
rpro
babilit
yes
tim
atio
ns
of
Equat
ion
(I).
The
dep
enden
tvari
able
isa
dum
my
equal
toone
ifth
efirm
iim
port
sca
pit
algoods
int.
All
expla
nat
ory
vari
able
sare
lag
of
one
per
iod.
Fir
ms’
capit
al
inte
nsi
tyis
the
rati
oof
capit
al
ove
rth
ew
age-
bill.
The
financi
al
vari
able
sth
atw
euse
are
leve
rage(
i)and
liquid
ity
rati
o(i
).L
ever
age(
i)is
the
rati
oof
borr
ow
ings
ove
rto
tal
ass
ets
and
liquid
ity
rati
o(i
)is
the
rati
oof
curr
ent
ass
ets
ove
rto
tal
liabil
itie
sof
the
firm
.T
he
outp
ut
tari
ffs
are
atth
e4-d
igit
NIC
indust
ryle
vel
and
the
Her
findahl
index
isat
the
2-d
igit
NIC
indus-
try
leve
l.In
pare
nth
eses
we
report
het
erosk
edast
icit
y-r
obust
standard
ser
rors
.***
,**
and
*in
dic
ate
signifi
cance
atth
e1,
5and
10
per
cent
leve
lsre
spec
tive
ly.
Sourc
e:A
uth
ors
’es
tim
atio
ns
usi
ng
Pro
wes
sdat
a.
500 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
skill intensity are positive, significant and more over their coefficients remainstable relative to the previous estimations.16
These results suggest that credit constraints are more binding for importingcapital goods than intermediate goods. One of the main differences betweenthe decision of importing capital goods and the decision of using foreign inter-mediate inputsis related to the nature of this choice and the way these produc-tion factors enter into the production process. While intermediate goods arevariable inputs that firms have to buy on a regular basis, capital goods arefixed investments (machines and capital equipment) used in the productionprocess that are not renewed every year. As discussed above, the fact that lessfirms import capital goods than intermediates suggests that the barriers to startimporting are larger in the case of capital goods. Moreover, the evidence thatmost foreign capital goods in India are sourced from developed economies rein-forces this idea that importing capital goods is associated with a higher fixedcost. This may be due to a larger up-front cost in the case of foreign technologyupgrading, and also to additional sunk costs related to learning about foreigntechnologies, finding foreign suppliers, and the adaptation period of the pro-duction process. In the theoretical part of the paper, the model shows that thelarger thefixed cost, the more binding are credit constraints. This reasonablyexplains why firms’ financial factors play a higher role for starting importingcapital goods than intermediates.
Financial Constraints Versus Tariffs on Capital Goods. In a context of tradeliberalization, firms could upgrade foreign technology easily thanks to theremoval of import barriers on capital equipment goods. Thereby, the effect ofbetter financial access on foreign technology adoption might just be picking upthe effects of lower tariffs on capital equipment goods.
In the previous specifications, we include tariffs on final goods at the 4 digitindustry level to capture the impact of India’s trade liberalization that tookplace at the beginning of the nineties. We now explore the robustness of ourresults when we take into account tariff reductions on capital goods over theperiod. The average yearly reduction of import tariffs on machinery and equip-ment goods is 2.3 percent during the period. Since trade liberalization in Indiain the early 90s consisted in a unilateral trade reform, we use effectivelyapplied tariff rates at the HS6 product level set by India to the rest of theworld for the period 1997 to 2006. This is possible thanks to the match of thefirm level data with the average import tariff data of products corresponding toHS6 codes between 840000 and 859999 (machinery and mechanical appli-ances) from the World Bank (WITS). The results including the variation of theaverage import tariffs on capital equipment goods are presented in Table 3.
As expected, a reduction of import barriers on capital goods increases thelikelihood of firms to upgrade foreign technology. Most importantly, our
16. The results are similar when we use as an alternative specification, available upon request, the
previous import status of intermediate goods as a control variable relying on the full sample of firms.
Bas and Berthou 501
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E3
.T
rade
liber
aliza
tion
and
import
sof
capit
al
goods
Dum
my
equal
one
iffirm
(i)
import
sof
capit
algoods
Dep
enden
tvari
able
(1)
(2)
(3)
(4)
(5)
Dta
riff
sca
pit
al
goods
20.3
21***
20.2
95***
20.3
09***
20.2
97***
20.3
11***
(0.0
52)
(0.0
51)
(0.0
50)
(0.0
52)
(0.0
51)
Lev
erage(
i)(t
21)
20.1
33***
20.1
33***
(0.0
16)
(0.0
16)
Liq
uid
ity
rati
o(i
)(t2
1)
0.1
62***
0.1
61***
(0.0
26)
(0.0
26)
Log
valu
eadded
(i)(
t21)
0.0
19***
0.0
19***
0.0
18***
0.0
18***
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
Capit
al
inte
nsi
ty(i
)(t2
1)
0.0
42***
0.0
42***
0.0
52***
0.0
52***
(0.0
05)
(0.0
05)
(0.0
05)
(0.0
05)
Log
wage(
i)(t
21)
0.0
46***
0.0
47***
0.0
52***
0.0
53***
(0.0
07)
(0.0
07)
(0.0
07)
(0.0
07)
Outp
ut
tari
ffs(
s)(t
21)
0.0
27
0.0
27
(0.0
29)
(0.0
29)
Her
findahl
index
(s)(
t21)
0.0
01
0.0
01
(0.0
03)
(0.0
03)
Fir
mfixed
effe
cts
Yes
Yes
Yes
Yes
Yes
Obse
rvat
ions
34,7
35
34,7
35
34,7
35
34,7
35
34,7
35
R2
0.0
01
0.0
17
0.0
17
0.0
15
0.0
15
Note
s:T
he
table
report
ses
tim
ates
from
linea
rpro
babilit
yes
tim
atio
ns
of
Equat
ion
(I).
The
dep
enden
tvari
able
isa
dum
my
equal
toone
ifth
efirm
iim
port
sca
pit
algoods
int.
All
expla
nat
ory
vari
able
sare
lag
of
one
per
iod.
Fir
ms’
capit
al
inte
nsi
tyis
the
rati
oof
capit
al
ove
rth
ew
age-
bill.
The
financi
al
vari
able
sth
atw
euse
are
leve
rage(
i)and
liquid
ity
rati
o(i
).L
ever
age(
i)is
the
rati
oof
borr
ow
ings
ove
rto
tal
ass
ets
and
liquid
ity
rati
o(i
)is
the
rati
oof
curr
ent
ass
ets
ove
rto
tal
liabil
itie
sof
the
firm
.T
he
outp
ut
tari
ffs
are
atth
e4-d
igit
NIC
indust
ryle
vel
and
the
Her
findahl
index
isat
the
2-d
igit
NIC
indus-
try
leve
l.In
pare
nth
eses
we
report
het
erosk
edast
icit
y-r
obust
standard
ser
rors
.***
,**
and
*in
dic
ate
signifi
cance
atth
e1,
5and
10
per
cent
leve
lsre
spec
tive
ly.
Sourc
e:A
uth
ors
’es
tim
atio
ns
usi
ng
Pro
wes
sdat
a.
502 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
results remain unaffected by the introduction of import tariffs on capitalgoods. Once we take into account directly the effects of trade reform, a reduc-tion of the leverage ratio and an increase in the liquidity of the firm have botha positive impact on the probability of adopting a foreign technology.Comparing the point estimates of the coefficients of the leverage and liquidityratios with those reported in baseline estimations (Table 1) reveals that thecoefficients on the variables of interest remain stable. Estimation results inTable 3 show that a 10 percentage point reduction of the leverage ratioincreases the probability of upgrading foreign technology by 13 percent.Similarly, a 10 percentage point increase in the liquidity ratio is associatedwith an increase in the likelihood of importing capital goods of 16 percent inthis specification. These effects are very comparable to the baselinespecification.
Additional firm characteristics. This section presents alternative sensitivitytests related to other firm-characteristics that might be driving our results.
In the previous estimations, firm controls include the wage bill, value addedand capital intensity. Although value added is positively correlated with firmproductivity, it is a raw measure for efficiency gains. As an additional test weuse firm total factor productivity measured by Levinsohn and Petrin (2003)methodology, using wage bill as a proxy for firms’ labor utilization. Our find-ings show that most productive firms have a higher probability of upgradingforeign technology (columns (1) and (2) of Table (4)). The inclusion of firms’total factor productivity does not modify the results as compared with thebaseline specification. The point estimates suggest that a reduction of leverageratio of 10 percentage points or an analogous increase in the liquidity ratioincreases the probability of sourcing capital goods from abroad by 14 and 15percent, respectively.
Next we address whether firms’ ownership is driving our results. Previousstudies on multinational firms show that foreign firms in developing countriestend to use more advanced technologies and be more productive relative to do-mestic firms (Javorcik 2004). One reason may be that foreign multinationalshave a better access to finance, and are more likely to source capital goodsfrom abroad. Javorcik and Spatareanu (2009) also show that the suppliers ofmultinationals are less credit constrained.17 In general, the fact that foreigncompanies are wealthier firms and use more advanced technology could poten-tially explain our results. In order to address this issue, we exclude from oursample multinational firms in columns (3) and (4) of Table (4). Our coefficientsof interest on financial variables remain robust and stable when we restrict thesample to domestic firms, implying that financial factors matter when consider-ing the sample of domestic firms.18
17. Manova and others (2009) also show that in the case of China, multinationals have a better
propensity to export in sectors where firms are typically more financially vulnerable.
18. We thank anonymous referees for having pointed out this channel.
Bas and Berthou 503
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E4
.A
ddit
ional
firm
chara
cter
isti
cs
dum
my¼
1if
firm
import
sca
pit
al
goods i
t=
1
Contr
oll
ing
for
TFP
Subsa
mple
Non-M
NF
Subsa
mple
Pri
vat
efirm
sSubsa
mple
Non-B
usi
nes
sgro
ups
Dep
enden
tvari
able
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Lev
erage(
i)(t
-1)
20.1
41***
20.1
19***
20.1
18***
20.1
20***
(0.0
18)
(0.0
17)
(0.0
17)
(0.0
20)
Liq
uid
ity
rati
o(i
)(t2
1)
0.1
48***
0.1
32***
0.1
19***
0.0
94***
(0.0
28)
(0.0
27)
(0.0
29)
(0.0
30)
TFP(i
)(t2
1)
0.0
26***
0.0
28***
(0.0
09)
(0.0
09)
Log
valu
eadded
(i)(
t21)
0.0
16***
0.0
15***
0.0
15***
0.0
15***
0.0
11***
0.0
11***
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
Capit
al
inte
nsi
ty(i
)(t2
1)
0.0
38***
0.0
48***
0.0
35***
0.0
43***
0.0
34***
0.0
42***
0.0
32***
0.0
40***
(0.0
05)
(0.0
05)
(0.0
05)
(0.0
05)
(0.0
05)
(0.0
05)
(0.0
06)
(0.0
06)
Log
wage(
i)(t
21)
0.0
82***
0.0
87***
0.0
52***
0.0
57***
0.0
51***
0.0
56***
0.0
58***
0.0
63***
(0.0
07)
(0.0
07)
(0.0
07)
(0.0
07)
(0.0
07)
(0.0
07)
(0.0
08)
(0.0
08)
Outp
ut
tari
ffs(
s)(t
21)
0.0
25
0.0
18
0.0
42
0.0
37
0.0
12
0.0
06
20.0
06
20.0
11
(0.0
38)
(0.0
38)
(0.0
37)
(0.0
38)
(0.0
40)
(0.0
40)
(0.0
44)
(0.0
44)
Her
findahl
index
(s)(
t21)
0.0
02
0.0
02
0.0
01
0.0
01
0.0
02
0.0
01
0.0
02
0.0
01
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
Fir
mfixed
effe
cts
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yea
rfixed
effe
cts
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Obse
rvat
ions
32,9
36
32,9
36
32,2
75
32,2
75
28,3
01
28,3
01
21,3
21
21,3
21
R2
0.0
26
0.0
24
0.0
25
0.0
23
0.0
23
0.0
21
0.0
24
0.0
21
Note
s:T
he
table
report
ses
tim
ates
from
linea
rpro
babilit
yes
tim
atio
ns
of
Equat
ion
(I).
The
dep
enden
tvari
able
isa
dum
my
equal
toone
ifth
efirm
iim
port
sca
pit
algoods
int.
All
expla
nat
ory
vari
able
sare
lag
of
one
per
iod.
Fir
ms’
capit
al
inte
nsi
tyis
the
rati
oof
capit
al
ove
rth
ew
age-
bill.
The
financi
al
vari
able
sth
atw
euse
are
leve
rage(
i)and
liquid
ity
rati
o(i
).L
ever
age(
i)is
the
rati
oof
borr
ow
ings
ove
rto
tal
ass
ets
and
liquid
ity
rati
o(i
)is
the
rati
oof
curr
ent
ass
ets
ove
rto
tal
liabil
itie
sof
the
firm
.T
he
outp
ut
tari
ffs
are
atth
e4-d
igit
NIC
indust
ryle
vel
and
the
Her
findahl
index
isat
the
2-d
igit
NIC
indus-
try
leve
l.In
pare
nth
eses
we
report
het
erosk
edast
icit
y-r
obust
standard
ser
rors
.***
,**
,and
*in
dic
ate
signifi
cance
atth
e1,
5and
10
per
cent
leve
lsre
spec
tive
ly.
Sourc
e:A
uth
ors
’es
tim
atio
ns
usi
ng
Pro
wes
sdat
a.
504 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Besides, previous works using the same firm-level dataset have emphasizedthe role of state-owned firms in India (Topalova 2004, Alfaro and Chari 2009).State-owned firms might benefit from special access to credit from State-owedbanks. In order to address this issue, we restrict the sample to private firms(columns 5 and 6). The point estimates of leverage and liquidity ratio remainrobust and stable. Finally, the largest domestic firms belonging to the Indianbusiness groups in India could also benefit from a better access to finance. Ourresults are robust to the exclusion of these groups from the estimation sample(columns 7 and 8). These tests confirm that different firm ownership character-istics are not picking up our results.
Robustness Analysis
One of the challenges when investigating the relationship between the access toexternal finance and firms’ technology adoption decisions is the potentialreverse causality. In the medium or long run, importing foreign capital goods isexpected to increase the profitability of the firm and therefore its financial state-ments (reduce the leverage or increase the liquidity ratio). This mechanismwould result in a positive bias in the relation between imports and financialfactors of the firm. In the short run, the cost associated with the imports of anew technology is expected to increase the leverage of the firm, or decrease itsliquidity. This mechanism would result in a negative bias. We perform two ro-bustness checks to address this potential reverse causality issue.19
Decision to Start Importing Capital Goods. We explore the robustness ofour baseline specification when we restrict our sample to firms that have notimported capital goods in the previous years. We investigate whether an in-crease in the access to external finance is associated with the decision to startsourcing capital goods from abroad. By focusing on firms that have notimported capital goods in the previous period, this specification deals with thepossible endogeneity issues between financial access and foreign technologyadoption that the previous specifications might suffer.
The estimates from linear probability estimations of equation (I) with firmand year fixed effects for the restricted sample of firms that have not importedcapital goods in the last two years are reported in columns (1) to (2) ofTable 5. In this case, the coefficients of the financial variables are smaller com-pared to the baseline specification due to the reduction of the sample size from34,735 observations to almost 21,000. The point estimates indicate that a 10percentage point reduction of the leverage ratio increases the probability to
19. In estimations available upon request, we also carry out a two-stage least square (2SLS) linear
probability model where the liquidity ratio, the leverage ratio, and the capital intensity are instrumented
with lagged values (three to four years) and the mean capital intensity of the industry. The results
remain robust for the leverage ratio under the instrumental variable specification, while the liquidity
ratio is no longer significant. However, this last result should be interpretedwith caution, given the
reduction of the sample size due to the use of lagged instruments. This restriction leaves us with half of
the sample under the IV relative to the baseline estimations.
Bas and Berthou 505
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
start importing capital goods by 6 percent (column (1)). Similarly, a 10 per-centage point increase in the amount of liquidity increases the probability toupgrade foreign technology for the first time by 5 percent (column (2)). Whenwe restrict our sample to firms that have not imported capital goods in the lastfour years, the effect of the leverage ratio is still negative, significant and stable,while the liquidity ratio is no longer significant (columns (3) and (4)).
As an alternative test we include the past import experience in the baselineestimations. In this case, we keep the full sample of firms and include thelagged importer status of the firm measured by a dummy variable that is equalto one if the firm has been an importer of capital goods in the previous years.This specification allows us to take into account the past experience of import-ing capital goods that can reduce the fixed costs in the present.20 These resultsare reported in Table 6. As expected the previous import status has a positiveeffect on the decision of importing capital goods in year t. The point estimates
TA B L E 5. Decision to start importing capital goods
dummy ¼ 1 if firm imports capital goodsit ¼ 1
Firms that do not import capital goods in the previous
Two years Four years
Dependent variable (1) (2) (3) (4)
Leverage(i)(t21) 20.058*** 20.052***(0.012) (0.012)
Liquidity ratio (i)(t21) 0.050** 0.022(0.021) (0.020)
Log value added(i)(t21) 0.004* 0.004* 0.000 0.001(0.002) (0.002) (0.002) (0.002)
Capital intensity(i)(t21) 0.012*** 0.016*** 0.013*** 0.017***(0.004) (0.004) (0.004) (0.003)
Log wage(i)(t21) 0.029*** 0.032*** 0.038*** 0.040***(0.006) (0.005) (0.005) (0.005)
Output tariffs(s)(t21) 20.055* 20.058** 20.069*** 20.072***(0.028) (0.029) (0.027) (0.027)
Herfindahl index(s)(t21) 20.002 20.003 20.001 20.001(0.002) (0.002) (0.002) (0.002)
Firm fixed effects Yes Yes Yes YesYear fixed effects Yes Yes Yes YesObservations 20,993 20,993 18,600 18,600R2 0.014 0.013 0.024 0.022
Notes: The dependent variable is a dummy equal to one if the firm i imports capital goods in tand have not imported in the previous two years (columns 1 and 2) or four years (columns 3 and4). We use the same control variables as in Table 1. ***, ** , and * indicate significance at the 1,5 and 10 percent levels respectively.
Source: Authors’ estimations using Prowess data.
20. We thank an anonymous referee for having pointed out this channel.
506 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
of leverage and liquidity ratio remain almost unchanged relative to the onespresented in the baseline specifications in Table 1. These findings confirm theimportance of financial access to start sourcing capital goods from abroad.
Dependence with Respect to External Finance. As a final exercise, we usethe measure of firms’ dependence on external finance (“external dependence”),proposed by Rajan and Zingales (1998) and updated by Braun (2002) andBraun and Larrain (2005), to identify an exogenous effect of financial con-straints on capital goods imports across different industries. In the presence offinancial constraints, the borrowing capacity of a firm is closely related to its fi-nancial statement. Financial constraints are therefore expected to affect morethe investment decision in sectors where firms rely more on the use of externalfinance.
TA B L E 6. Controlling for past import experience
dummy ¼ 1 if firm imports capital goodsit ¼ 1
Dependent variable (1) (2) (3) (4)
Importer status(i)(t21) 0.121*** 0.110*** 0.124*** 0.111***(0.009) (0.009) (0.009) (0.009)
Leverage(i)(t21) 20.143*** 20.109***(0.015) (0.015)
Liquidity ratio (i)(t21) 0.167*** 0.132***(0.024) (0.024)
Log value added(i)(t21) 0.014*** 0.013***(0.003) (0.003)
Capital intensity(i)(t21) 0.029*** 0.036***(0.005) (0.005)
Log wage(i)(t21) 0.044*** 0.049***(0.006) (0.006)
Output tariffs(s)(t21) 0.020 0.014(0.034) (0.035)
Herfindahl index(s)(t21) 0.000 0.000(0.003) (0.003)
Firm fixed effects Yes Yes Yes YesYear fixed effects Yes Yes Yes YesObservations 34,735 34,735 34,735 34,735R2 0.031 0.037 0.029 0.036
Notes: The table reports estimates from linear probability estimations of Equation (I). The de-pendent variable is a dummy equal to one if the firm i imports capital goods in t. All explanatoryvariables are lag of one period. Firms’ capital intensity is the ratio of capital over the wage-bill.The financial variables that we use are leverage(i) and liquidity ratio(i). Leverage(i) is the ratio ofborrowings over total assets and liquidity ratio(i) is the ratio of current assets over total liabilitiesof the firm. The output tariffs are at the 4-digit NIC industry level and the Herfindahl index is atthe 2-digit NIC industry level. In parentheses we report heteroskedasticity-robust standardserrors. ***, ** and * indicate significance at the 1, 5 and 10 percent levels respectively.
Source: Authors’ estimations using Prowess data.
Bas and Berthou 507
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E7
.Im
port
sof
capit
al
goods
-dep
enden
ceon
exte
rnal
finance
inth
ein
dust
ry
dum
my¼
1if
firm
import
sca
pit
al
goods i
t¼
1
Dep
enden
tvari
able
(1)
(2)
(3)
(4)
Lev
erage(
i)(t
21)
20.0
77***
20.0
98*
(0.0
24)
(0.0
51)
Lev
erage(
i)(t
21)�
Ext.
Dep
.(s)
20.1
24*
20.1
21*
(0.0
69)
(0.0
70)
Lev
erage(
i)(t
21)�
Cap.I
nt(
s)0.2
58
(0.5
37)
Liq
uid
ity
rati
o(i
)(t2
1)
0.0
48
20.0
33
(0.0
36)
(0.0
66)
Liq
uid
ity(i
)(t2
1)�
Ext.
Dep
.(s)
0.2
41***
0.2
55***
(0.0
87)
(0.0
89)
Liq
uid
ity(i
)(t2
1)�
Cap.I
nt.
(s)
1.0
16
(0.6
50)
Log
valu
eadded
(i)(
t21)
0.0
16***
0.0
16***
0.0
15***
0.0
15***
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
Capit
al
inte
nsi
ty(i
)(t2
1)
0.0
36***
0.0
36***
0.0
44***
0.0
44***
(0.0
05)
(0.0
05)
(0.0
05)
(0.0
05)
Log
wage(
i)(t
21)
0.0
55***
0.0
55***
0.0
60***
0.0
60***
(0.0
07)
(0.0
07)
(0.0
07)
(0.0
07)
Her
findahl
index
(s)(
t21)
0.0
01
0.0
01
0.0
01
0.0
01
(0.0
03)
(0.0
03)
(0.0
03)
(0.0
03)
Outp
ut
tari
ffs(
s)(t
21)
0.0
18
0.0
19
0.0
14
0.0
17
(0.0
37)
(0.0
37)
(0.0
37)
(0.0
37)
Fir
mfixed
effe
cts
Yes
Yes
yes
Yes
Yea
rfixed
effe
cts
Yes
Yes
Yes
Yes
Obse
rvat
ions
33,7
73
33,7
73
33,7
73
33,7
73
R2
0.0
26
0.0
26
0.0
24
0.0
25
Note
s:***
,**
and
*in
dic
ate
signifi
cance
atth
e1,
5and
10
per
cent
leve
lsre
spec
tive
ly.
Lev
erage(
i)(t
-1)�
Ext.
Dep
.and
Liq
uid
ity(i
)(t-
1)�
Ext.
Dep
.are
inte
ract
ion
vari
able
bet
wee
nth
eL
ever
age
rati
oand
the
vari
able
exte
rnal
dep
enden
cepro
vid
edby
Bra
un
(2002)
and
Bra
un
and
Larr
ain
(2005).
Lev
erage(
i)(t
-1)�
Cap.I
nt.
isth
ein
tera
ctio
nof
the
leve
rage
rati
ow
ith
the
capit
al
inte
nsi
tyof
the
indust
ry,
als
opro
vid
edby
Bra
un
(2002).
Exte
rnal
de-
pen
den
ceand
capit
al
inte
nsi
tyare
sect
or-
spec
ific,
wit
hIS
ICre
v.
23-d
igit
scl
ass
ifica
tion.
Sourc
e:auth
ors
’es
tim
atio
ns
usi
ng
Pro
wes
sdat
a.
508 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
The empirical strategy proposed by Rajan and Zingales (1998) is adapted tothe context of our study.21 The measure of external dependence at the 2-digitindustry level updated by Braun (2002) and Braun and Larrain (2005) is inter-acted with our measures of firms’ financial statements. The baseline empiricalspecification is then augmented with the Leverage(i)(t-1)� Ext. Dep.(s) andLiquidity ratio (i)(t-1)� Ext. Dep.(s) variables. The coefficient on the inter-action variable between the leverage of the firm and the external dependence ofthe industry is expected to be negative, and the coefficient on the interactionbetween the liquidity ratio of the firm and the degree of external dependence isexpected to be positive: in the presence of financial constraints, the liquidityratio and leverage of the firm are expected to be more closely related to theimports of foreign capital goods for firms that rely more on the use of externalfinance.
Estimation results are reported in Table 7. The leverage ratio is interactedwith the external dependence variable in column (1). The coefficient on theinteraction variable reports a negative sign, confirming that the negativeimpact of the leverage of the firm, on its probability to import foreign capitalgoods, is higher in sectors where firms require more external finance. This es-timation is replicated in column (2), including as well an interaction betweenthe leverage of the firm and the capital intensity of the industry. The capitalintensity of the industry is provided by Braun (2002), and is sector-specific.This new variable allows to control for the possibility that importing capitalgoods is more likely to affect firms’ financial factors in sectors where firmsare typically more capital intensive. Since the external dependence of the firmand the capital intensity are positively correlated, reverse causality would biasthe coefficient on the Leverage(i)(t-1)� Ext. Dep.(s) variable. Estimationresults in column (2), though, confirm that the point estimate of the inter-action term, Leverage(i)(t-1)� Ext. Dep.(s), is robust and stable under thisspecification.
A similar analysis where the liquidity ratio is interacted with the sectoral ex-ternal dependence and the sectoral capital intensity is provided in columns (3)and (4). The coefficient on the Liquidity ratio(i)(t-1)� Ext. Dep.(s) is positiveand significant (column 3) and it remains robust and stable when we introducethe interaction term between the liquidity of the firm and the capital intensityof the industry in column (4). These sensitivity tests therefore provide addition-al evidence confirming that the liquidity of the firm affects the import decisionof capital goods.
21. Rajan and Zingales (1998) propose to identify the effect of financial development on economic
growth, using an interaction term between the country’s financial development and the industry level of
external dependence. The degree of dependence on external finance is a technology parameter
(measured using Compustat data for the United States), and is independent of countries’ characteristics.
The coefficient on the interaction term is therefore expected to be unrelated to countries’ characteristics,
and unaffected by future economic growth.
Bas and Berthou 509
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
I V. C O N C L U D I N G R E M A R K S
Adopting foreign technology is costly and requires using internal and externalfinancial resources. This paper investigates the influence of firms’ financialfactors on their decision to source foreign capital goods. We test whetherfirms that experience an improvement in their financial statements have ahigher probability to upgrade foreign technology embodied in importedcapital goods. We find strong evidence that this is the case in India. Firmswith a lower leverage and a higher liquidity have a higher probability of up-grading foreign technology. Different sensitivity tests demonstrate that theseresults are not driven by omitted variable bias related to changes in firm ob-servable characteristics (size, capital, and skill intensity) as well as ownershipstatus (multinational, state-owned firms and local Indian business groups).Finally, these findings are also robust to alternative specifications dealing withthe potential reverse causality issues between financial factors and foreigntechnology adoption.
Our findings suggest that financial market imperfections have a negativeeffect on purchases of foreign technology. This is an important issue for aggre-gate productivity growth in developing countries, like India, that rely heavilyon foreign technology in their production process. One important policy impli-cation of our findings is that the success of trade reforms is closely related tothe capacity of the financial intermediaries to provide funding to domesticfirms.
A . T H E O R E T I C A L A P P E N D I X
A.1. Price index approximation
Following Chaney (2005), we assume that the price index only depends onlocal firms’prices and that foreign firms do not face any liquidity constraints.The price index approximation is
P �ðw�w
d�
pdðwÞ1�wLdFwðwÞ ! 1
1�w
We define a function g(.) in the following way:
gð:Þ : w�f�1 ¼ w
m
ðw�w�
wf�1dFwðwÞ� �
� F , w� ¼ gðFÞ
510 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
A.2. Credit constrained firms
A sufficient condition for the existence of liquidity constraints importers iscf
cd, 1: This is the assumption that we introduce concerning the relative per
unit cost is then equal tocf
cd¼ t
h 1�að Þm
g, 1: This condition implies that the effi-
ciency parameter of imported capital goods is higher than its additional vari-able cost relative to domestic ones ðg . t
h 1�að Þm Þ:
Proposition 1: Under the assumption that x , 1, there is a subset of firms(denoted F) subject to liquidity constraints with a productivity level betweenw�f , w , wðAÞ:
B . E M P I R I C A L A P P E N D I X
Proof In order to prove that F is not empty we investigate whether wð0Þ . wf�:
F þ FT � lA
F
� � 1w�1
wd� .F þ FT
F
� � 1w�1 cf
cd
� �wd�
Appendix Table : Descriptive statistics of Indian manufacturing firms(1997–2006)
Mean Std. Dev.
Number of firms
Average number of firms per year 3,473Importers of capital goods (%) 32Importers of intermediate goods (%) 62Private firms (%) 81Local business groups (%) 39Foreign firms 7 percentFinancial variables
Liquidity ratio 0.50 0.20Leverage ratio 0.38 0.31Firm level characteristics
Value added 50 214Wage bill 7.65 46Capital stock 113 534Industry level controls
Effectively applied output tariffs (NIC 4 digit) 0.30 0.13Herfindahl index (NIC 2 digit) 0.94 0.78
Notes: Mean values and standard errors in parentheses are reported. Leverage(i) is the ratio ofborrowings over total assets and liquidity ratio(i) is the ratio of current assets over total liabilitiesof the firm.
Source: authors calculations based on Prowess data.
Bas and Berthou 511
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
RE F E R E N C E S
Ackerberg, D., K. Caves, and G. Frazer. 2007. “Structural Identification of Production Functions.”
Working paper, UCLA.
Aghion, P., P. Askenazy, N. Berman, G. Cette, and L. Eymard. 2008. “Credit Constraints and the
Cyclicality of R&D Investment: Evidence from France.” PSE Working Papers 2008-26.
Aghion, P., A. Banerjee, and T. Piketty. 1999. “Dualism and Macroeconomic volatility.” Quarterly
Journal of Economics 114: 1359–1397.
Aghion, P., N. Bloom, Blundell, R. Griffith, and P. Howitt. 2005. “Competition and Innovation: An
Inverted U Relationship.” Quarterly Journal of Economics 120: 701–728.
Alfaro, L., and A. Chari. 2009. “India Transformed? Insights from the Firm Level 1988-2005.” NBER
Working Papers 15448.
Amiti, M., and J. Konings. 2007. “Trade liberalization, Intermediate Inputs, and Productivity: Evidence
from Indonesia.” American Economic Review 97(5): 1611–1638.
Arnold, J., B. Javorcik, M. Lipscomb, and A. Mattoo. 2010. “Services Reform and Manufacturing
Performance: Evidence from India.” CEPR Discussion Paper 8011.
Bas, M. 2011. “Input-trade Liberalization and Firm Export Decisions: Evidence from Argentina.”
Journal of Development Economics 97 (2012) 481–493.
Bas, M., and V. Strauss-Kahn. 2011. “Does Importing more Inputs Raise Exports? Firm Level Evidence
from France.” Working Paper CEPII.
Berman, N., and J. Hericourt. 2010. “Financial Factors and the Margins of Trade : Evidence from
Cross-country Firm-level Data.” Journal of Development Economics 93 (2): 206–217.
Bond, S., J. Elston, J. Mairesse, and B. Mulkay. 2003. “Financial Factors and Investment in Belgium,
France, German and the UK: A comparison using Company Panel Data.” Review of Economics and
Statistics 85: 153–165.
Bond, S., and C. Meghir. 1994. “Dynamic Investment Models and the Firm’s Financial Policy.” Review
of Economic Studies 61(2): 197–222.
Braun, M. 2002. “Financial Contractibility and Asset Hardness.” Mimeo.
Braun, M., and B. Larrain. 2005. “Finance and the Business Cycle: International, Inter-industry
Evidence.” Journal of Finance 60(3): 1097–1128.
Chaney, T. 2005. “Liquidity Constrained Exporters.” Chicago University. Processed.
Eaton, J., and S. Kortum. 2001. “Trade in Capital Goods.” European Economic Review 45(7):
1195–1235.
Ethier, W. 1982. “National and International Returns to Scale in the Modern Theory of International
Trade.” American Economic Review 72(3): 389–405.
Fazzari, S. M., R. G. Hubbard, and B. C. Petersen. 1988. “Financing Constraints and Corporate
Investment.” Brookings Papers on Economic Activity 19(1988-1): 141–206.
Goldberg, P., A. Khandelwal, N. Pavcnik, and P. Topalova. 2009. “Trade Liberalization and New
Imported Inputs.” American Economic Review Papers and Proceedings 99(2): 494–500.
Goldberg, P. K., A. K. Khandelwal, N. Pavcnik, and P. Topalova. 2010. “Imported intermediate inputs
and Domestic Product Growth: Evidence from India.” The Quarterly Journal of Economics 125(4):
1727–1767.
Gorodnichenko, Y., and M. Schnitzer. 2010. “Financial Constraints and Innovation: Why Poor
Countries Don’t Catch Up.” NBER Working Paper No. 15792.
Greenaway, D., A. Guariglia, and R. Kneller. 2007. “Financial Factors and Exporting Decisions.”
Journal of International Economics 73(2): 377–395.
Grossman, G., and E. Helpman. 1991. “Innovation and Growth in the Global Economy.” Cambridge,
MA: MIT.
Halpern, , Koren, and Szeidl. 2009. “Imported Inputs and Productivity.” CEPR N. 5139.
512 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Haramillo, F., F. Schiantarelli, and A. Weiss. 1996. “Capital Market Imperfections before and after
Financial Liberalization: An Euler Equation Approach to Panel Data for Ecuadorian Firms.” Journal
of Development Economics 51(2): 367–386.
Harrison, A. E., and M. S. McMillan. 2003. “Does Direct Foreign Investment Affect Domestic Credit
Constraints?” Journal of International Economics 61(1): 73–100.
Hubbard, R. G. 1998. “Capital-market Imperfections and Investment.” Journal of Economic
Literatture 36(1): 193–225.
Javorcik, B. S. 2004. “Does Foreign Direct Investment Increase the Productivity of Domestic Firms? in
Search of Spillovers through Backward Linkages.” American Economic Review 94(3): 605–627.
Javorcik, B. S., and M. Spatareanu. 2009. “Liquidity Constraints and Firms’ Linkages with
Multinationals.” World Bank Economic Review 23(2): 323–346.
Kasahara, H., and J. Rodrigue. 2008. “Does the Use of Imported Intermediates Increase Productivity?
plant-Level Evidence.” Journal of Development Economics 87: 106–118.
Khandelwal, A., and P. Topalova. 2010. “Trade Liberalization and Firm Productivity: The Case of
India.” The Review of Economics and Statistics Forthcoming.
Kugler, M., and E. Verhoogen. 2009. “Plants and Imported Inputs: New Facts and an Interpretation.”
American Economic Review Papers and Proceedings 99(2): 501–507.
Levinsohn, J., and A. Petrin. 2003. “Estimating Production Functions Using Inputs to Control for
Unobservables.” Review of Economic Studies 70(2): 317–341.
Manova, K. 2008. “Credit Constraints, Heterogeneous Firms and International Trade.” NBER Working
Paper 14531.
Manova, K., S.-J. Wei, and Z. Zhang. 2009. “Firm Exports and Multinational Activity under Credit
Constraints.” Stanford University. Processed.
Markusen, J. 1989. “Trade in Producer Services and in other Specialized Intermediate Inputs.”
American Economic Review 79(1): 85–95.
Melitz, M. 2003. “The Impact of Trade on Intra-industry Reallocations and Aggregate Industry
Productivity.” Econometrica 71: 1695–1725.
Muuls, M. 2008. “Exporters and Credit Constraints. A Firm-Level Approach.” National Bank of
Belgium Working Paper.
Olley, G., and A. Pakes. 1996. “The Dynamics of Productivity in the Telecommunications Equipment
Industry.” Econometrica 64(6): 1263–1297.
Rajan, R., and L. Zingales. 1998. “Financial Dependence and Growth.” American Economic Review 3:
559–586.
Rivera-Batiz, L., and P. Romer. 1991. “International Trade with Endogenous Technological Change.”
European Economic Review 35(4): 971–1001.
Schor, A. 2004. “Heterogeneous Productivity Response to Tariff Reduction: Evidence from Brazilian
Manufacturing Firms.” Journal of Development Economics 75: 373–396.
Topalova, P. 2004. “Overview of the Indian Corporate Sector: 1989–2002.” IMF Working Papers
04/64.
Whited, T. M. 1992. “Debt, Liquidity Constraints, and Corporate Investment: Evidence from Panel
Data.” Journal of Finance 47(4): 1425–1460.
Bas and Berthou 513
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Coffee Market Liberalisation and the Implicationsfor Producers in Brazil, Guatemala and India*
Bill Russell, Sushil Mohan, and Anindya Banerjee
The standard approach to modelling the relationship between world and producerprices of coffee does not incorporate the effects of changing government policies andmarket structures. These changes have led to large structural breaks in the relationshipbetween the prices implying the standard estimates are biased. We model coffee pricesin Brazil, Guatemala and India allowing for the structural breaks and show that theliberalisation of coffee markets has benefited producers substantially both in terms ofa higher share of the world price of coffee and higher real prices. This suggests thatcalls to re-regulate coffee markets may be misplaced. JEL Classification: Q11, Q17,Q18, C32, C52, F13, F14
Before the 1990s, unilateral and multilateral interventions in coffee marketswere common. The governments of most coffee-producing countries consideredregulation of coffee marketing and pricing necessary because of coffee’s im-portance as a source of foreign exchange and government revenue.1 For majorcoffee-producing countries such as Brazil and Colombia the main objective forregulation was to raise world coffee prices. Countries also used regulation tomaintain fixed producer prices so as to shield coffee producers (hereafter
* Bill Russell (corresponding author), Economic Studies, School of Business, University of Dundee,
Dundee DD1 4HN, United Kingdom. þ44 1382 385165 (work phone), þ44 1382 384691 (fax), email
[email protected]. Sushil Mohan, Economic Studies, School of Business, University of Dundee, United
Kingdom. Anindya Banerjee, Department of Economics, Birmingham Business School, University of
Birmingham, United Kingdom. We would like to thank David Hendry and Hassan Molana for their
helpful advice, Ivan Carvalho, Denis Seudieu and Martin Wattam from the International Coffee
Organization for help in providing the data, and Tom Doan and Pierre Perron for graciously making
available the programmes for estimating structural breaks. A Supplementary Appendix to this article is
available at http://wber.oxfordjournals.org and http://billrussell.info. The data are available at http://
billrussell.info.
1. Unless specified otherwise, ‘coffee’ means green (raw or un-roasted) beans and coffee prices imply
prices of green beans. Coffee producers include growers and/or semi-processors, who sell their coffee as
cherries, parchment or green beans. If producers sell their coffee as cherry or parchment, the prices are
converted to green beans by using a ‘green bean equivalent’.
THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 514–538 doi:10.1093/wber/lhr055Advance Access Publication January 10, 2012# The Author 2012. Published by Oxford University Press on behalf of the International Bankfor Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions,please e-mail: [email protected]
514
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
referred to as producers) from price fluctuations and assured them a minimumprice.2
At the multilateral level the interventions generally took the form of regula-tion of export supply and prices. The first International Coffee Agreement(ICA) was signed in 1962 by the major coffee producing (exporting) and con-suming (importing) countries. According to its regulatory provisions, basicexport quotas were allocated to each of the exporting member countries forexport to importing member countries and they were tightened if internationalcoffee prices fell below a particular level and loosened when they rose abovethat level. The monitoring requirements of the ICA supported the domesticregulation of coffee markets in coffee-producing countries to ensure compliancewith quota restrictions.3 The ICA was undermined by some member countriesdistributing their exports at lower prices through non-member countries, the in-ability to agree on quotas and the continuing fragmentation of the geographyof production. Quotas were operational between October 1963 to December1973, October 1980 to March 1986, and November 1987 to July 1989 (ICO,1989). The agreement was suspended in 1989.
The suspension of the International Coffee Agreement (ICA) in 1989 andbroader economic reforms including exchange-rate reforms in most developingcountries in the late 1980s resulted in most coffee-producing countries liberalis-ing their coffee sector by replacing state-controlled marketing systems withmarkets run by private agents. The pace and scope of liberalisation has variedacross countries but has led to a more competitive international coffee marketin which producer prices reflect more accurately the domestic and internationalmarket conditions.
In terms of their impact on producer welfare the interventions are generallyregarded as unsuccessful. The cost of reduced volatility seemed too high giventhat the administered prices were usually far below the certainty equivalentthat would be accepted by producers.4 Jarvis (2005) and Mehta and Chavas(2008) find that the interventions resulted in high levels of rent seeking by arange of beneficiaries including bureaucrats, intermediaries such as coffee mar-keting boards and foreign importers, but not by producers. The accruing ofthese economic rents to intermediaries in the coffee supply chain creates alarger margin between the international and producer prices of coffee resultingin a lower share of the international price of coffee going to producers.
However, some commentators do not accept these assertions and raise con-cerns on the effects of these reforms on the real price of coffee received by pro-ducers. They feel that the principal beneficiaries of liberalisation have been
2. See Akiyama (2001).
3. For a brief description of coffee market interventions see Supplementary Appendix S1 available
at http://wber.oxfordjournals.org and http://www.billrussell.info. See also Raffaeli (1995), Gilbert
(1996), Bates (1997), McIntire and Varangis (1999), Jerome and Ogunkola (2000), Akiyama (2001),
Varangis et al. (2002) and Winter-Nelson and Temu (2002).
4. See Akiyama et al. (2001), Krivonos (2004) and Anderson et al. (2008).
Bill Russell, Sushil Mohan and Anindya Banerjee 515
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
coffee roasters and international traders who are able to capture all the monop-oly rents and dictate the prices they are willing to pay to producers.5 Latelythere have been calls for a return to some form of coffee market interventionsor regulations on the grounds that the liberalisation of coffee markets has notimproved the plight of producers.6
To evaluate if producers were better or worse off due to the regulation andinterventions in coffee markets requires us to compare the actual returns theyreceived with some counter-factual return that they would have received in un-regulated markets. The terminal (or international) price of coffee has not beendirectly regulated or administered over the last four decades.7 Consequently, ifthe ‘law of one price’ between the terminal and producer prices of coffee holdsthen we can use the terminal price to calculate the corresponding un-regulatedproducer price of coffee. This allows us to compare the actual returns to produ-cers with the return that producers would have received in an un-regulatedmarket.
Unfortunately, the recent extensive literature does not provide unambiguousempirical evidence in support of the ‘law of one price’. The results depend onthe level of aggregation of the data and the methods employed to ‘test’ the rela-tionship between prices in different markets. For example, the work on com-modity markets by Ardeni (1989) and Baffes (1991) provide mixed evidenceon accepting the ‘law’ while Mundlak and Larson (1992), Michael et al.(1994), Vataja (2000) and Batista and da Silveira (2010) provide evidence thatsupports the ‘law’. Moreover, this literature mostly focuses on the ‘law’ interms of exchange rate movements, distance between markets, shipping costsand price discrimination.
We argue instead that the relationship between the prices of a commodity inany two markets is also highly dependent on the prevailing international andnational policies. This is especially important when long samples of data areexamined and policy interventions by governments are extensive and changingas in the case of coffee markets. What is important, therefore, is how we in-corporate the effects of changing coffee market policies into our model ofcoffee prices.
We use the term ‘policies’ to include all measures to implement the inter-national and domestic agreements such as administering the producer priceand controlling the production and marketing of coffee as well as local taxes,export levies and subsidies, value added taxes, the setting of coffee gradingstandards, exchange rate and foreign exchange regulations and the provision ofcredit. Many of these policies change frequently and are difficult to document.However, they redistribute income either towards or away from producers and
5. See for example Fitter and Kaplinsky (2001), Ponte (2002), Oxfam (2002), Calfat and Flores
(2002), Shepherd (2004), Talbot (2004) and Daviron and Ponte (2005).
6. For example see ActionAid (2008) and South Centre (2008).
7. The terminal price of coffee is the spot price of coffee as traded in international markets and the
producer price of coffee is the cash price received at the ‘gate’ by producers.
516 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
other participants in the coffee supply chain and may introduce a shift, orbreak, in the mean of the share of the terminal coffee price going toproducers.8
The changes in policies at both the domestic and international levels and theeventual liberalisation of coffee markets raise a number of issues when model-ling coffee prices. For example, consider the terminal and producer prices forBrazilian, Guatemalan and Indian Arabica coffee measured in US cents perpound.9 There are three striking features of the data. First, the difference, orgap, between the terminal and producer prices of coffee varies considerablyover time in all three countries. This gap is an indirect measure of the costs oftransferring coffee from the producer to the terminal markets. BetweenJanuary 1973 and December 1989 the average transfer costs were around 85(141 per cent of the average producer price), 45 (54 per cent), and 58 (72 percent) US cents per pound of coffee for Brazil, Guatemala and India respective-ly. Since January 1990, following liberalisation, there has been a markeddecline in these costs to around 18 (28 per cent), 32 (46 per cent) and 30 (42per cent) US cents per pound of coffee respectively. It is unlikely that suchlarge reductions in transfer costs can be explained by changes in freight, hand-ling and related costs alone. It is more likely the reductions are due to changesin the economic rents received by intermediaries and governments in the trans-fer process arising from the greater degree of vertical integration in coffeemarkets.10
The variation in the gap between the terminal and producer coffee prices foreach of these countries appears to provide prima facie evidence that the ‘law ofone price’ does not hold. The ‘law’ suggests that the prices of two identicalgoods in two separate markets will differ in equilibrium by the cost of transfer-ring the goods between the markets. Importantly this implies that the twoprices will evolve together and that the gap between the two prices in equilib-rium will be constant. This ‘law’ is predicated on an unchanging economic en-vironment which in the case of coffee markets is difficult to sustain. Changesin coffee market policies at the domestic and international levels will alter theeconomic environment and may lead to discrete changes in the gap betweenthe two prices. Consequently, the changes that we observe in the gap betweenthe two coffee prices may be due to either a change in the economic environ-ment (i.e. changes in policy) or because the ‘law’ does not hold. Therefore, toexamine empirically the dynamics of coffee prices with a model that incorpo-rates the ‘law’ we need to control for the effects of changes in coffee marketpolicies.
8. For example, if the administered producer price of coffee is not changed in line with the terminal
price of coffee it will lead to a change in the share of the terminal price received by producers.
9. Prices are in nominal terms unless stated otherwise. See Supplementary Appendix S2 for details
of the data and S3 for a graph of the coffee price data.
10. See Mohan (2007).
Bill Russell, Sushil Mohan and Anindya Banerjee 517
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
The second feature follows from the first. Large changes in transfer costsassociated with changing government policies, regulations and market struc-ture may cause structural breaks in the mean of the producers’ share of theterminal price of coffee. The share is the ratio of the producer to the ter-minal price of coffee and is referred to in the paper as the coffee priceratio. These breaks are at times quite sudden and persistent as demon-strated in Figure 1 of the coffee price ratios for the three countries.11
Third, the nominal price of coffee received by producers at the beginningof the 21st century is much the same as it was in the early 1970s suggest-ing a large fall in real terms. This fall in the real price of coffee is demon-strated in Table 1.
All of these features lead us to argue that if the terminal and producer pricesof coffee are closely related then any modelling of the two prices must take thestructural breaks in the coffee price ratio into account. Otherwise estimatingthe model will result in biased and poor estimates that may lead to incorrectinferences. We therefore model the relationship between the terminal and pro-ducer prices of coffee in Brazil, Guatemala and India allowing for the breaks inthe mean of the coffee price ratio. These three countries are chosen due to thevariation in their coffee policies and market structures over time. We demon-strate that the modelling approach we adopt is successful in dealing with thisvariation.
We develop a two-step model of coffee prices. In the first step we employthe Bai and Perron (1998) (Bai-Perron) technique to identify multiple breaks inthe mean of the natural logarithm of the coffee price ratio. In the second stepwe estimate a vector autoregressive error correction model (VAR-ECM) ofcoffee prices conditioned on the identified breaks. A major advantage of thisapproach is that if the law of one price holds then the estimated error correc-tion term is equivalent to the log of the coffee price ratio and the structuralbreaks in the coffee price ratio identified in the first step are simultaneously thebreaks in the mean of the error correction term. A further advantage of this ap-proach is that we can examine directly the empirical relevance of the ‘law ofone price’ after accounting for the influence of changing policies on coffeeprices.
The next section sets out the standard approach to modelling coffee pricesbased on the ‘law of one price’ and explains the biases from not allowing forthe breaks in the coffee price ratio. Section II reports the results of the esti-mated models allowing for these breaks. We find that once we account for thebreaks in the coffee price ratio the ‘law of one price’ is strongly supported bythe data. This allows us to show that the producers’ share of the terminal priceof coffee in equilibrium has increased in all three countries since liberalisationto around 0.85 in Brazil and India and 0.79 in Guatemala. Assuming that
11. In the empirical analysis that follows including estimating the breaks in the coffee price ratio,
unit root tests and the estimation of the models, the variables are in natural logarithms.
518 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
liberalised markets led to these higher producer shares we demonstrate inSection III that the loss of revenue to producers from coffee market regulationsand interventions over the years is substantial. Our analysis shows that produ-cers have benefited since liberalisation from an increase in real prices, outputand higher share in the terminal price of coffee. This indicates that calls for a
FIGURE 1. THE COFFEE PRICE RATIO
Note: Thick line is the coffee price ratio (not logged). The horizontal thin lines are the meanof the coffee price ratio for the ‘regimes’ estimated by the Bai-Perron technique as described inSupplementary Appendix S4.
Bill Russell, Sushil Mohan and Anindya Banerjee 519
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E1
.N
om
inal
and
Rea
lC
off
eePri
ces
Bra
zil
Nom
inal
Valu
esR
eal
Valu
e
Pro
duce
rP
rice
Ter
min
alP
rice
Tra
nsf
erC
ost
sP
roduce
rP
rice
Ter
min
alP
rice
Tra
nsf
erC
ost
s1973
31.0
269.2
038.1
91
11
1990
55.5
882.9
729.3
90.5
832
0.4
048
0.2
599
2007
98.2
3111.6
613.4
40.9
273
0.4
725
0.1
030
Per
centa
ge
Change
1973
–1990
72.7
(3.3
)19.9
(1.1
)2
23.0
(21.5
)2
41.7
(23.1
)2
59.5
(25.2
)2
74.0
(27.6
)1990
–2007
83.3
(3.6
)34.6
(1.8
)2
54.3
(24.5
)59.0
(2.8
)16.7
(0.9
)2
60.7
(25.3
)1973
–2007
216.7
(3.4
)61.4
(1.4
)2
64.8
(23.0
)2
7.3
(20.2
)2
52.8
(22.2
)2
89.7
(26.5
)
Guat
emala
Nom
inal
Valu
esR
eal
Valu
e
Pro
duce
rP
rice
Ter
min
alP
rice
Tra
nsf
erC
ost
sP
roduce
rP
rice
Ter
min
alP
rice
Tra
nsf
erC
ost
s1973
46.0
362.3
016.2
71
11
1990
54.5
889.4
634.8
70.4
004
0.4
848
0.7
235
2007
98.0
9123.5
525.4
60.6
240
0.5
807
0.4
581
Per
centa
ge
Change
1973
–1990
18.6
(1.0
)43.6
(2.2
)114.3
(4.6
)2
60.0
(25.2
)2
51.5
(24.2
)2
27.6
(21.9
)1990
–2007
79.7
(3.5
)38.1
(1.9
)2
27.0
(21.8
)55.9
(2.6
)19.8
(1.1
)2
36.7
(22.7
)1973
–2007
113.1
(2.3
)98.3
(2.0
)56.5
(1.3
)2
37.6
(21.4
)2
41.9
(21.6
)2
54.2
(22.3
)
520 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
India
Nom
inal
Valu
esR
eal
Valu
e
Pro
duce
rP
rice
Ter
min
alP
rice
Tra
nsf
erC
ost
sP
roduce
rP
rice
Ter
min
alP
rice
Tra
nsf
erC
ost
s1973
45.3
762.3
016.9
31
11
1990
66.7
589.4
622.7
00.4
968
0.4
848
0.4
527
2007
108.3
4123.5
515.2
10.6
992
0.5
807
0.2
630
Per
centa
ge
Change
1973
–1990
47.1
(2.3
)43.6
(2.2
)34.1
(1.7
)2
50.3
(24.0
)2
51.5
(24.2
)2
54.7
(24.6
)1990
–2007
62.3
(2.9
)38.1
(1.9
)2
33.0
(22.3
)40.8
(2.0
)19.8
(1.1
)2
41.9
(23.1
)1973
–2007
138.8
(2.6
)98.3
(2.0
)2
10.2
(20.3
)2
30.1
(21.0
)2
41.9
(21.6
)2
73.7
(23.9
)
Note
s:Fig
ure
sin
bra
cket
sare
the
com
pounded
annualise
dper
centa
ge
change.
Rea
lvalu
esare
calc
ula
ted
wit
hre
fere
nce
toth
eU
Nin
dex
of
unit
valu
esof
export
s.Sim
ilar
resu
lts
are
obta
ined
ifth
ere
al
pri
ceof
coff
eeis
mea
sure
din
term
sof
the
Unit
edSta
tes
consu
mer
pri
cein
dex
(CPI)
.A
nnual
valu
esare
aver
ages
of
the
month
lyvalu
es.
Nom
inal
valu
esare
inU
Sce
nts
per
pound.
Tra
nsf
erco
sts
are
the
dif
fere
nce
bet
wee
nth
ete
rmin
al
and
pro
duce
rpri
ce.
The
table
on
the
right
show
sth
eper
centa
ge
changes
inth
eU
Nin
dex
of
unit
valu
esof
export
sand
the
Unit
edSta
tes
CPI.
Unit
Val
ues
of
Export
sU
SC
PI
1973
–1990
196.2
(6.6
)194.1
(6.6
)1990
–2007
15.3
(0.8
)58.7
(2.8
)1973
–2007
241.5
(3.7
)366.6
(4.6
)
Bill Russell, Sushil Mohan and Anindya Banerjee 521
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
return to coffee market interventions of the kind that existed prior to liberalisa-tion are not justified.
I . M O D E L L I N G C O F F E E P R I C E S
The standard approach to modelling coffee prices can be motivated with refer-ence to the Enke (1951), Samuelson (1952), Takayama and Judge (1971)spatial models of prices and the ‘law of one price’.12 These models argue thatwhen two markets attain equilibrium the prices differ only by the costs oftransferring the goods between the markets. The transfer costs include the ship-ping and storage costs associated with moving produce between markets alongwith the economic rents of intermediate agents in the supply chain. In thismodel, demand and supply shocks are fully transmitted between the twomarkets in equilibrium.
The standard approach incorporates Samuelson’s (1957) spatial competitivemarket where prices of the same good in two markets diverge subject to theconstraint that:
� TC12 � P1 � P2 � TC21 ð1Þ
where P1 and P2 are prices in markets one and two respectively and TC12 arethe costs of transferring the goods from market 1 to market 2 and TC21 is thereverse. Samuelson points out that with positive transport costs both equalitiescannot hold simultaneously and that if both inequalities hold then there is notrade between the two markets. However, if one inequality and one equalityholds then there may be a uni-directional trade from the lower priced marketto the higher priced market. In our case we assume the trade flow is from theproducer to the terminal market for coffee such that arbitrage in a competitivemarket leads to:
PT ¼ PP þ TCPT ð2Þ
where PT . PP and PT and PP are the terminal and producer prices of coffee re-spectively and TCPT are the transfer costs associated with transferring coffeebetween the producer and terminal markets. The equilibrium in coffee pricesbetween the terminal and producer markets can then be represented in thesespatial models as:
PP; t ¼ PT; t�Ue ð3Þ
where Ue is the constant ratio that coffee prices in the two markets attain inequilibrium and the ‘t’ subscript indicates the time period of the data. Notethat even though it is assumed that the physical trade in coffee is uni-
12. The model can be thought of as ‘spatial’ in terms of the geographic locations of producer and
terminal coffee markets.
522 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
directional the causation between the two prices in equation (3) is conceptuallybi-directional.
With Ue , 1 in equation (3) we can identify transfer costs, TCt, associatedwith the movement of coffee between the two markets such thatTCt ¼ PT; t � PP; t which are also measured in price per unit of coffee. In theshort run the equilibrium relationship (3) need not apply due to the incompletetransfer of information between markets and other rigidities. For example, pro-ducer prices may not immediately respond to changes in the terminal price ofcoffee due to the fragmented nature of the supply chain and the need for allprices and costs in that chain to adjust. One might also assume that there aresmall ‘menu costs’ associated with adjusting costs and prices within the supplychain implying the producer price may diverge from its equilibrium level.However, in the long run we expect all prices and costs to respond to competi-tive pressures and the equilibrium relationship of equation (3) will hold if in-formation is shared efficiently between markets and all agents make normalprofits.
An assumption of this model that is not often highlighted in the literature isthat the transfer costs in equilibrium are a fixed ratio of both the terminal andproducer prices of coffee. For example, transfer costs as a ratio to, or share of,the terminal price of coffee in equilibrium is:
TCt
PT; t¼ PT; t � PP; t
PT; t¼ 1 � Ue ð4Þ
Importantly, this implies that if the equilibrium in coffee prices is adequatelydescribed by this model, the statistical process of the transfer costs is the sameas that of the two coffee prices in equilibrium. If this were not the case, whenequilibrium coffee prices are attained in the two markets, transfer costs willnot have returned to its equilibrium ratio in terms of the respective coffeeprices. This implication is important for our modelling of coffee prices below.
The equilibrium relationship in equation (3) above has a straightforwardtime series interpretation in terms of a vector autoregressive-error correctionmodel of coffee prices. Consider the following error correction representationof a two variable VAR of order k:
Dxt ¼ d þ ab0 xt�1 þXk�1
i¼1
PiD xt�i þ 1t ð5Þ
where, xt ¼pP
pT
� �t
, and a ¼ a11 a12
a21 a22
� �is a matrix of equilibrium speed of
adjustment coefficients, b is a matrix containing the equilibrium vectors
1 b1 2
1 b2 2
� �, Pi is a matrix of short-run coefficients, lower case variables are in
Bill Russell, Sushil Mohan and Anindya Banerjee 523
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
natural logarithms and D represents the change in the variable such thatDxt ¼ xt � xt�1.
The error correction representation argues that the producer and terminalprices of coffee move together and coffee prices converge on the equilibrium re-lationship. Changes in coffee prices depend on the deviation from the equilib-rium relationship and the speed of adjustment (i.e. the error correctionmechanism) along with lagged changes in the prices of coffee.
The standard approach to estimating the VAR-ECM of coffee prices pro-ceeds within a cointegration framework.13 If coffee prices, pP and pT are inte-grated processes and cointegrate, the matrices a and b must be of reduced rankand in our case would be equal to one. No cointegration implies a rank ofzero.
While it is common to model price data as integrated processes, in realityprices cannot be ‘truly’ integrated as they have a lower boundary of zero. It ismore likely, therefore, that prices are trend stationary processes with breaksand it is these breaks that make the data appear integrated.14 For coffee pricesthe trend is very small or non-existent and the natural logarithm of coffeeprices and the coffee price ratio appear to be stationary processes over the pastthirty four years. This can be verified by the ADF and KPSS univariate unitroot tests reported in columns 1 and 2 of Table 2. Note however that eventhough the trend is difficult to identify in the coffee price data the ADF testindicates that the coffee price ratio is a trend stationary process for all threecountries. Consequently, the standard cointegration approach to modellingcoffee prices is not appropriate and the modelling framework is a VAR-ECMof stationary price variables with possible breaks in the error correction term.
The equilibrium relationship in equation (5) can be written:
b0 xt ¼ pP; t þ bpT; t ¼ zt ð6Þ
where zt is a stationary process. Furthermore, if b ¼ 21 then the equilibriumvector can be interpreted as the equilibrium relationship of equation (3) where:
pP; t � pT; t ¼ ut ð7Þ
and the error correction mechanism is now equivalent to the natural logarithmof the coffee price ratio, ut.
Note that when b ¼ 21, the standard model in equations (3) and (6) dis-plays long-run homogeneity which is equivalent to accepting the ‘law of oneprice’. This means that a 1 per cent increase in one coffee price leads to a 1 percent increase in the other price in equilibrium so that transfer costs as a propor-tion of either the terminal or the producer price are unchanged. This also
13. For example see Rapsomanikas et al. (2004), Fortenbery and Zapta (2004), Krivonos (2004)
and Alizadeh and Nomikos (2005).
14. See Perron (1989).
524 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
implies that transfer costs have increased by 1 per cent in the long run. Withtrending price and cost variables we need b ¼ 21 so that a persistent changein the level of coffee prices does not lead to a persistent change in the gapbetween the two prices in equation (5). In other words, b ¼ 21 implies that achange in the level of prices leaves the coffee price ratio unchanged in equilib-rium and transfer costs increase in line with prices.
Two important assumptions concerning the equilibrium relationship (7) andthe VAR-ECM analysis are that the coffee price ratio, ut, converges on a con-stant value, ue, in equilibrium and that ut is a stationary process with a con-stant mean equal to ue. Looking closely at Figure 1 we can however seediscrete shifts in the mean of the coffee price ratio. Thus, the finding of
TA B L E 2. Univariate Unit Root Tests
Series Original Data De-meaned Data
Brazil ADF KPSS ADF KPSSLn Producer Price 23.48 0.0688*10% 24.33 0.043710%
[0.0089] [0.0005]Ln Terminal Price 22.91 0.1195*5% 24.02 0.042710%
[0.0454] [0.0014]Ln Ratio of Coffee Prices 23.33* # 0.1975*1% 27.57# 0.019710%
[0.0634] [0.0000]Guatemala
Ln Producer Price 23.01 0.165810% 24.33 0.046810%
[0.0350] [0.0005]Ln Terminal Price 22.91 0.1252*5% 24.02 0.043410%
[0.0454] [0.0014]Ln Ratio of Coffee Prices 26.10* 0.1363*5% 211.61 0.023610%
[0.0000] [0.0000]India
Ln Producer Price 23.00 0.108110% 23.84# 0.125210%
[0.0361] [0.0027]Ln Terminal Price 22.91 0.1252*5% 24.62 0.123210%
[0.0454] [0.0001]Ln Ratio of Coffee Prices 25.71* 0.1154*10% 210.00 0.020310%
[0.0000] [0.0000]
Notes: The data are in natural logarithms as indicated by ‘Ln’. ‘Ln Ratio of Coffee Prices’, ut,is measured as pP,t 2 pT,t. ADF is the augmented Dickey-Fuller t-statistics that assumes a null hy-pothesis of a unit root in the data. Associated probability values are shown as [ ]. ADF 5 per centcritical values with a constant is – 2.87 and with a trend and constant -3.43. A lag length of onein the ADF test was chosen on the basis of SIC in all cases except when zero as indicated by #.
KPSS is the Kwiatkowski-Phillips-Schmidt-Shin LM test statistic that assumes a null hypothesisthat the data are stationary. KPSS 1, 5 and 10 per cent critical values with a constant are 0.7390,0.4630 and 0.3470 respectively and with a trend and constant 0.2160, 0.1460 and 0.1190. Thepercentage shown with the KPSS test statistic is the significance level that the null hypothesis isaccepted at.
De-meaned data adjusts the original data for the shifts in mean as identified by the Bai-Perrontechnique for each country (see Supplementary Appendix S4). * indicates a significant trend inthe unit root test. In all other cases the trend is insignificant and excluded prior to inference.
Bill Russell, Sushil Mohan and Anindya Banerjee 525
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
stationarity of the coffee price ratio notwithstanding, it is worth investigatingwhether breaks are present in this series. These visual shifts in the ratio can betested formally by applying the Bai-Perron technique, which allows for breaksin otherwise stationary processes, for identifying multiple structural breaks inthe mean of the coffee price ratio, ut.
15 A justification for the likely presence ofsuch breaks is provided in the paragraph below. The Bai-Perron technique iden-tifies 11, 8 and 10 discrete shifts in the mean of the coffee price ratio forBrazil, Guatemala and India and the shifting mean coffee price ratio are shownin Figure 1 by horizontal thin lines.16
Our reasons for proceeding with our analysis with the maintained as-sumption that the identified breaks in the coffee price ratio are valid aretwo-fold. First, incorporating the breaks improves the estimated model’s de-scription of the coffee price data considerably. Second, the converse assump-tion, namely that the coffee price ratio is stationary with a constant mean, isdifficult to sustain. For example, if the converse were true then all thechanges in coffee market policies over the past thirty five years have had nopersistent impact on the share of the terminal price of coffee that goes toproducers. More importantly, the vociferous arguments surrounding themerits of either regulating coffee markets or de-regulating coffee markets arelargely irrelevant as neither policy would have had any persistent impact onthe coffee price ratio. Therefore, both these implications of assuming a sta-tionary process with constant mean are hard to defend empirically and at apolicy level.
The number of breaks that we find in the mean of the coffee price ratio mayseem large relative to that usually reported in the applied structural breaks lit-erature. However, this result is consistent with producer prices being largelyadministered during the first seventeen years of the data and the coffee marketbeing de-regulated in irregular steps over the next seventeen years.Consequently, over the entire thirty four years examined in the empirical ana-lysis, the coffee price ratio can be characterised as being subject to discrete andirregular shocks. Finding around ten breaks over a thirty-four year periodimplies that a break occurred only once every thirty six months on averagewhich is broadly consistent with the rate of structural change and interventionsin coffee markets over the period studied. We return to the issue of the validityof the identified structural breaks in Section II following the estimation of themodels below.
It appears, therefore, that the two important assumptions mentioned above,namely that the coffee price ratio is a stationary process with a constant meanand converges on a constant value in equilibrium, are not valid. This suggeststhe mean of the coffee price ratio may be discretely time-varying and the
15. See Supplementary Appendix S4 for details of the Bai-Perron estimates of the structural breaks.
16. If the breaks in the coffee price ratio are valid this implies Type 1 and Type 2 errors in the ADF
and KPSS tests respectively in Table 2.
526 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
equilibrium coffee price ratio could be written in a more general form that con-tains a trend, t, and n shift dummies Di:
pP; t þ b1 pT; t þ b2 t þXn
i¼1diDi ¼ vt ð8Þ
The inclusion of the trend is intended to capture a systematic divergencebetween the terminal and producer prices in equilibrium. This would occur iftransfer costs evolved differently to coffee prices over the longer term. Transfercosts may be driven by factors such as the wage rate, productivity, shippingrates, insurance, technological advances incorporated in the supply chain, realinterest rates, energy costs and inventory control which may not affect to thesame extent the prices of homogeneous agricultural products like coffee. Theshift dummies represent discrete structural breaks in the coffee price ratio inresponse to changing domestic and international coffee market policies.
If estimation of the model proceeds assuming that the mean ratio of coffeeprices is time invariant as implicitly (or explicitly) assumed in the standard lit-erature then the estimates of the equilibrium coefficients, b, and the adjustmentcoefficients in equation (5) will be poor and biased if the coffee price ratio, ut,is in fact non-stationary (i.e as described by (8) above). The direction of thebias in b depends on whether the coffee price ratio is on average increasing ordecreasing over the period. The adjustment coefficients will be biased down-wards if the shifts in the mean coffee price ratio are not accounted for. Fromthe perspective of the estimated model it will appear that the coffee price ratiois taking a long time (i.e. the speed of adjustment is low) to return to the equi-librium coffee price ratio following a change in the mean coffee price ratio inthe data. Consequently, how we model the coffee price ratio may affect ourestimates in important ways.
I I . E S T I M A T I N G A VA R - E C M O F C O F F E E P R I C E S
The models are estimated in natural logarithms using the monthly averageInternational Coffee Organization (ICO) Indicator Price for Arabica coffee(‘Brazilian Natural’ Arabica for Brazil and ‘Other Mild’ Arabica for Guatemalaand India) as a measure of terminal prices and the monthly average producerprice for Arabica coffee in Brazil, Guatemala and India for the period January1973 to October 2007. All price data are in nominal terms and measure amonthly average price in US cents per pound. Further details concerning thedata are provided in Supplementary Appendix S2.
Modelling coffee prices without structural breaks
Estimates of the standard VAR-ECM model that do not account for the struc-tural breaks in the mean of the coffee price ratio are reported in Table 3. Whilethere is some variation in the estimates there is a strong pattern in the results.A trend is significant for all three countries suggesting that there has been a
Bill Russell, Sushil Mohan and Anindya Banerjee 527
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E3
.V
AR
Err
or
Corr
ecti
on
Model
of
Coff
eePri
ces:
Model
1–
Sta
ndard
Model
wit
hout
Bre
aks
BR
AZ
IL
Equilib
rium
Coef
fici
ents
Adju
stm
ent
Coef
fici
ents
pP
pT
Tre
nd
Dp
PD
pT
Unre
stri
cted
1.0
000
20.7
010
(24.7
)2
0.1
828
(23.1
)2
0.0
630
(23.0
)2
0.0
044
(20.3
)R
estr
icte
d1.0
000
21.0
000
20.0
048
(24.3
)2
0.0
351
(21.6
)0.0
199
(1.2
)T
LR
R¼
0.2
875,
TT¼
0.1
513,
LM
1¼
0.2
688,
LM
2¼
0.4
058.
Norm
ality¼
0.0
000.
Sta
tionari
ty:
pP¼
0.0
883,
pT¼
0.0
121.
Excl
usi
on:
pP¼
0.0
121,
pT¼
0.0
883,
Tre
nd¼
0.0
830.
Exogen
eity
:P
P¼
0.0
494,
pT¼
0.8
594.
GU
AT
EM
AL
A
Equilib
rium
Coef
fici
ents
Adju
stm
ent
Coef
fici
ents
pP
pT
Tre
nd
Dp
PD
pT
Unre
stri
cted
1.0
000
20.9
571
(213.9
)2
0.0
495
(22.2
)2
0.2
022
(26.0
)0.0
218
(1.0
)R
estr
icte
d1.0
000
21.0
000
20.0
528
(22.3
)2
0.1
935
(25.9
)0.0
268
(1.3
)T
LR
R¼
0.6
281,
TT¼
0.0
564,
LM
1¼
0.
1204,
LM
2¼
0.
0683.
Norm
ality¼
0.0
000.
Sta
tionari
ty:
pP¼
0.0
000,
pT¼
0.0
000.
Excl
usi
on:
pP¼
0.0
000,
pT¼
0.0
000,
Tre
nd¼
0.0
381.
Exogen
eity
:p
P¼
0.0
000,
pT¼
0.3
724.
IND
IA
Equilib
rium
Coef
fici
ents
Adju
stm
ent
Coef
fici
ents
pP
pT
Tre
nd
Dp
PD
pT
Unre
stri
cted
1.0
000
20.7
783
(212.0
)2
0.0
514
(22.4
)2
0.0
913
(24.0
)0.0
864
(3.7
)R
estr
icte
d1.0
000
21.0
000
20.3
474
(22.4
)2
0.0
508
(22.8
)0.0
788
(4.3
)T
LR
R¼
0.0
316,
TT¼
0.0
364,
LM
1¼
0.3
603,
LM
2¼
0.5
283.
Norm
ality¼
0.0
000.
Sta
tionari
ty:
pP¼
0.0
000,
pT¼
0.0
000.
Excl
usi
on:
pP¼
0.0
000,
pT¼
0.0
000,
Tre
nd¼
0.0
226.
Exogen
eity
:p
P¼
0.0
003,
pT¼
0.0
009.
Note
s:R
eport
edas
()
are
t-st
atis
tics
.T
he
tren
dis
mult
iplied
by
100.
The
model
sand
stat
isti
csare
esti
mat
edw
ith
two
lags
of
the
core
vari
able
sand
an
effe
ctiv
esa
mple
of
416
month
lyobse
rvat
ions
for
the
per
iod
January
1973
toO
ctober
2007.
The
num
ber
of
lags
was
chose
nby
alikel
ihood
rati
ote
stfo
rla
gre
duct
ion.
TL
RR
and
TT
are
the
finit
esa
mple
Bart
lett
corr
ecte
dpro
babilit
yvalu
esof
the
test
of
the
equilib
rium
rest
rict
ion
that
b¼
21
and
the
likel
ihood
rati
oex
clusi
on
test
of
the
esti
mat
edtr
end
resp
ecti
vely
.L
M1
and
LM
2are
the
pro
babilit
yvalu
esof
the
Lagra
nge
Mult
iplier
test
sof
no
seri
al
corr
elat
ion
inth
eer
rors
of
lags
1and
2re
spec
tive
ly.
Norm
ality
isth
epro
babilit
yvalu
eof
the
Doorn
ik-H
anse
nte
stfo
rnorm
al
erro
rs.
Sta
tionari
ty,
Excl
usi
on
and
Exogen
eity
are
the
pro
babil
ity
valu
esof
the
likel
ihood
rati
ote
sts
that
pP
and
pT
(and
tren
dif
applica
ble
)are
stat
ionary
,ca
nbe
excl
uded
from
the
equil
ibri
um
rela
tionsh
ipand/o
rw
eakly
exogen
ous
resp
ecti
vely
.E
stim
ated
wit
hC
AT
S2.0
.
528 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
trend increase in the producer price of coffee relative to terminal prices overthe period examined. The estimated models for Brazil and India appear to berelatively poor representations of the dynamics of the coffee prices with the ad-justment coefficients largely insignificant for Brazil and small for India. Theresults for Guatemala are the best behaved where we accept the long-runhomogeneity restriction that b ¼ 21 with the adjustment coefficients moder-ately large and significant. Consequently, with the possible exception ofGuatemala, we might conclude that a VAR-ECM is a relatively poor represen-tation of the dynamics of coffee prices. The inconclusive nature of these resultsis consistent with the view that structural breaks are an important feature ofthe price series being modelled and therefore need to be taken into account inorder to develop a more accurate description of the system generating coffeeprices.
Modelling coffee prices with structural breaks
Breaks in the mean of the coffee price ratio are due to breaks in the componentcoffee price series. However, simultaneous breaks of equal magnitude will notaffect the ratio. That is, the coffee price ratio will break if either the breaks inthe price series occur at different points of time or when occurring simultan-eously they are of different magnitudes. While there may be numerous eventsthat impact simultaneously on both price series the discussion above focuses onchanges to policy and market structure that impact on the coffee price ratio.Therefore, our preferred approach is to identify breaks in the mean of thecoffee price ratio directly by applying the Bai and Perron (1998) algorithm.The identified breaks are those shown in Figure 1 and reported inSupplementary Appendix S4. Once we allow for these breaks in the coffeeprice ratio the univariate unit root tests reported in columns 3 and 4 of Table 2lead to the unambiguous conclusion that the natural logarithm of the prices ofcoffee and their ratio are stationary.
Estimates of the VAR-ECM incorporating the breaks leads to the resultsreported in Table 4. The breaks are introduced as level shifts in the equilibriumrelationship which account for the level shifts in the mean coffee price ratioevident in Figure 1. The inclusion of the shift dummies now leads to the rejec-tion of the trend in the equilibrium relationship at the 5 per cent level for allthree countries. This implies the shift dummies dominate the trend as a proxyfor the shifts in the equilibrium coffee price ratio over the period.17 The shiftdummies are highly significant which reinforces our conclusion that theBai-Perron identified breaks are valid. The restriction that b ¼ 21 can now beeasily accepted for all three countries at the 5 per cent level implying we
17. Over the full sample the mean coffee price ratio increases and this results in a positive estimated
trend in the standard model (see Figure 1 and Table 3). However, the structural breaks in Model 2
reported in Table 4 represent the evolution of the mean coffee price ratio better, resulting in the trend
becoming insignificant.
Bill Russell, Sushil Mohan and Anindya Banerjee 529
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA B L E 4. VAR Error Correction Model of Coffee Prices: Model 2–Bai-PerronBreak Adjusted
BRAZIL
Equilibrium Coefficients Adjustment Coefficients
pP pT DpP DpT
Unrestricted 1.0000 20.9058 (217.2) 20.2652 (26.7) 20.0338 (21.1)Restricted 1.0000 21.0000 20.2420 (25.9) 20.0036 (20.1)
TLRR ¼ 0.2085, LM1 ¼ 0.5148, LM2 ¼ 0.0577. Normality ¼ 0.0000. Stationarity: pP ¼ 0.0000,pT ¼ 0.0000. Exclusion: pP ¼ 0.0000, pT ¼ 0.0000. Exogeneity: pP ¼ 0.0000, pT ¼ 0.3981.Dummies: December 1974, 20.2449 (22.8); March 1977, 0.3939 (4.8); August 1979, 0.2806(3.3); August 1981, 20.1225 (21.5); December 1984, 20.5702 (27.6); April 1987, 0.4842(5.7); April 1989, 20.5049 (26.1); December 1991, 20.2544 (23.8); August 1996, 20.0645(21.1); November 2000, 0.1965 (2.6); and December 2002, 20.2014 (22.8).
GUATEMALA
Equilibrium Coefficients Adjustment Coefficients
pP pT DpP DpT
Unrestricted 1.0000 20.9189 (223.8) 20.4547 (210.2) 20.0578 (21.9)Restricted 1.0000 21.0000 20.4117 (29.4) 0.0852 (2.9)
TLRR ¼ 0.1309, LM1 ¼ 0.0001, LM2 ¼ 0.6317. Normality ¼ 0.0000. Stationarity: pP ¼ 0.0000,pT ¼ 0.0000. Exclusion: pP ¼ 0.0000, pT ¼ 0.0000. Exogeneity: pP ¼ 0.0000, pT ¼ 0.1319.Dummies: June 1975, 0.2924 (5.4); December 1979, 20.1562 (23.3); October 1983, 20.3316(26.1); May 1986, 0.3101 (4.8); May 1988, 0.0607 (1.1); April 1993, 0.1043 (1.8); April 1995,20.1062 (21.7); and February 1998, 20.1693 (22.8).
INDIA
Equilibrium Coefficients Adjustment Coefficients
pP pT DpP DpT
Unrestricted 1.0000 21.0391 (224.1) 20.1699 (25.8) 0.2250 (7.8)Restricted 1.0000 21.0000 20.1813 (26.0) 0.2243 (7.5)
TLRR ¼ 0.4704, LM1 ¼ 0.0083, LM2 ¼ 0.0777. Normality ¼ 0.0000. Normality ¼ 0.0000.Stationarity: pP ¼ 0.0000, pT ¼ 0.0000. Exclusion: pP ¼ 0.0000, pT ¼ 0.0000. Exogeneity:pP ¼ 0.0000, pT ¼ 0.0000.Dummies: March 1976, 0.5463 (10.5); June 1978, 20.2231 (23.9); June 1980, 20.1491(22.6); September 1982, 0.2442 (5.0); December 1986, 20.2019 (24.3); June 1989, 20.1842(23.6); April 1992, 20.1102 (22.0); April 1994, 0.4594 (8.1); September 1996, 20.2669(26.2); and November 2004, 20.2021 (25.1).
Notes: The shift dummies are those estimated in the restricted model. For further details see thenotes to Table 3.
530 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
simultaneously accept the ‘law of one price’ in relation to terminal and produ-cer coffee prices. The error correction mechanisms are also strongly significantwith large adjustment coefficients.
Therefore our results for the VAR-ECM with structural breaks are in accord-ance with the theoretical spatial model of prices. For example, the terminaland producer prices of coffee move very closely together as demonstrated byaccepting the restriction that b ¼ 21 and the existence of a powerful error cor-rection mechanism in all three countries. Consequently, these results stronglyfavour the ‘law of one price’ once we adjust for the breaks in the mean of thecoffee price ratio which we characterise as due to changing coffee market pol-icies. The large adjustment coefficients reported in Table 4 imply that devia-tions from the equilibrium coffee price ratio are relatively small and transitory.Following a shock to the coffee price ratio, half of the adjustment to equilib-rium occurs in around 2 1
2 months for Brazil and 1 month for Guatemala andIndia.18 This can be seen in Figure 1 where the estimated equilibrium coffeeprice ratios are equivalent to the horizontal thin lines and deviations in, ut,from the equilibrium ratio are mostly small and short lived.
Finally, it appears that modelling the discrete shifts in the coffee price ratioas structural breaks plays an important role in explaining the behaviour ofcoffee prices and the associated changing shares in the terminal prices going toproducers. Extending the standard analysis to include the breaks improves ourunderstanding of the behaviour of coffee prices significantly.
Are the identified structural breaks valid?
There are two ways to approach this question. The first is to identify particularevents in the regulation of coffee markets that coincide roughly with thebreaks in the coffee price ratio used in the empirical analysis. To this endSupplementary Appendix S1 provides a brief description of the regulations inthe coffee markets of each of the three countries and relates the changing regu-lations with the breaks in the coffee price ratio reported in SupplementaryAppendix S4. Of the twenty nine breaks identified in the data of the threecountries by the Bai-Perron technique, twenty roughly coincide with historicalevents described in Supplementary Appendix S1.
The second approach focuses on the empirical results themselves and asksare the results robust to the number of breaks? In particular, is the largenumber of identified breaks causing the high speed of adjustment back to theequilibrium coffee price ratio and the acceptance of the law of one price? Toexamine whether our results are robust to the number of breaks we repeat theanalysis by imposing the number of breaks to be half those identified in themodels reported in Table 4 and to allow the Bai-Perron technique to choose
18. The adjustment speeds are calculated from simulations based on the estimates in Table 4.
Bill Russell, Sushil Mohan and Anindya Banerjee 531
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
the breaks optimally subject to this constraint.19 The estimates imposing halfthe number of breaks are found to be very similar qualitatively and statisticallyto those reported in Table 4 that incorporate the full number of breaks.20 Forexample, for all three countries: (i) the appropriate modelling framework is aVAR-ECM of stationary price variables with possible breaks in the error cor-rection term; (ii) the trend remains insignificant; (iii) the long-run restriction ofb ¼ 21 can be accepted suggesting the law of one price continues to hold;(iv) the adjustment coefficients are slightly smaller but similar in magnitudeand significance to those reported in the ‘full’ model reported in Table 4; (v)the shift dummies are highly significant; and (vi) the residuals are largely freeof serial correlation as in the full model.
Therefore our estimated model is robust to halving the number of breaks.Consequently, if we reject the number of breaks that we find using statisticaltests and instead impose a smaller number of breaks so as to conform with ourprior views of the data then our general empirical results along with the asso-ciated policy conclusions and economic implications are largely unaffected.
I I I . T H E LO S S I N R E V E N U E T O P R O D U C E R S F R O M
C O F F E E M A R K E T R E G U L A T I O N
The equilibrium coffee price ratio is particularly useful for examining theimpact of government policies and changing market structures on the coffeemarket as it abstracts from short-run variations and shocks to this ratio.Following the liberalisation of coffee markets the equilibrium coffee price ratioin levels has increased to around 0.85 for Brazil and India and 0.79 forGuatemala in the most recent period. Given that both the producer and termin-al prices of coffee are largely market determined in the most recent period fol-lowing liberalisation we might infer the coffee price ratios are theirun-regulated market values. Furthermore, the empirical evidence presented inSection II suggests the ‘law of one price’ is strongly supported by the data.Consequently, as the terminal price of coffee has been largely market deter-mined over the entire period we can calculate the counter-factual unregulatedproducer price of coffee by multiplying the terminal price of coffee in eachperiod by the un-regulated market value for the coffee price ratio in 2007. Thisallows us to estimate the loss of revenue to producers due to regulation byexamining the extent to which the actual producer price of coffee deviatesfrom its counterfactual unregulated value.
The thin line in Figure 2 is the nominal loss in revenue to producers due tothe regulation of coffee markets calculated as Qt † PT; t† Ue
2007 � Uet
� �where
Qt is the production of coffee and Ue2007 and Ut
e are the equilibrium coffee
19. The ‘half-breaks’ models imposes 6, 4 and 5 breaks for Brazil, Guatemala and India
respectively.
20. The results are available in Supplementary Appendix S5.
532 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
price ratios in 2007 and in each year for each country respectively. This calcu-lation assumes that production and the terminal price of coffee are independentof the market based coffee price ratio. Note that the law of one price meansthat Ue is ‘unit free’ and a real number. This allows the unit of the loss to pro-ducers to be measured in US cents. The thick line in Figure 2 shows the realloss to producers measured in 2007 prices.
FIGURE 2. THE NOMINAL AND REAL LOSSES OF COFFEE PRODUCERS
Note: The thin and thick lines are the nominal and real values (in 2007 prices) of the loss toproducers. Real values are in terms of the UN index of unit values of exports.
Bill Russell, Sushil Mohan and Anindya Banerjee 533
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
We observe that producers suffered from a loss of revenue in all three coun-tries during the 1970s and early 1980s. In 2007 prices, the real losses to produ-cers peak at around US$7 billion in the early 1980s for Brazil and aroundUS$0.45 billion for Guatemala and India in the mid-1970s. The losses trail offto negligible levels in the most recent periods following liberalisation in allthree countries indicating that market interventions over the years have notbeen in the overall interest of producers. Note that these measures of loss arein respect to the producers alone and not necessarily to the country as a whole.The loss, or a part of the loss, may simply represent a transfer from the produ-cer to either the government or intermediaries in the transfer process.
I V. A R E C O F F E E P R O D U C E R S B E T T E R O F F D U E T O L I B E R A L I S A T I O N ?
To answer this question we consider the impact that liberalisation has had onthe revenue and input costs of producers. The impact on revenues can bethought of in terms of three interrelated issues. First, what has happened to theproducer price of coffee? Second, what has happened to the quantity of coffeeproduced? And third, what has happened to the share of the terminal price ofcoffee that goes to the producer?
The common expectation from coffee market interventions was that theywould result in an increase in the terminal and producer prices of coffee. If thiswere the case then the producer price of coffee should have fared better duringthe period of regulation than after liberalisation. This is not supported byTable 1. In the seventeen years prior to liberalisation, producer coffee price in-flation is only half that of the ‘world’ inflation rate of 6.6 per cent as measuredby the UN index of unit values of exports or the United States CPI. Producerprices increased by 3.3, 1.0 and 2.3 per cent per annum in Brazil, Guatemalaand India but in real terms they fell over the period prior to liberalisation by42, 60 and 50 per cent respectively.
In contrast, during the period 1990 to 2007 when coffee markets witnessedthe phase of liberalisation, producer prices increased by 3.6, 3.5, and 2.9 percent per annum in Brazil, Guatemala and India compared with ‘world’ inflationof less than 1 per cent per annum as measured by the UN index of unit valuesof exports. This means that the producer price of coffee increased by around60, 55 and 40 per cent in real terms after liberalisation.21 Coffee prices there-fore did not match the growth in ‘world’ prices during the years of coffeemarket regulation up to 1990 but increased by more than ‘world’ prices afterthe liberalisation of coffee markets.
The data also shows there has been a large increase in coffee production fol-lowing liberalisation in all three countries. Average coffee production in the
21. In the period 1990 to 2007 United States CPI deviates from the UN index of unit values of
exports (see Table 1). However, if we use United States CPI as a measure of world inflation we find that
producer prices still increase in real terms by around 15, 13 and 2 per cent respectively.
534 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
seventeen years after 1990 is 1.3, 1.3 and 1.8 times the coffee production in theseventeen years prior to 1990 in Brazil, Guatemala and India respectively (ICO,2007). There may be many reasons for this increase in production and wecannot say that the increase is solely due to liberalisation. But to the extent thatliberalisation caused an increase in producer prices we can expect its effect onproduction was beneficial. We can also expect some increase in productionbecause of the removal of restrictions that were imposed on the production ofcoffee prior to liberalisation. Akiyama (2001) argues that the removal of con-straints of the international quota system meant that exports increased alongwith production. In addition, the liberalisation meant that higher quality coffeeis no longer mixed with lower grades prior to export and the opportunity toexport and obtain a premium on higher quality coffee has expanded. There wastherefore an added incentive after liberalisation for producers to supply betterquality coffee, which in turn has helped the growth in the consumption of coffeeworldwide. In any case, as far as producers are concerned they have gained bothin terms of prices and production following liberalisation. The empirical analysisabove also demonstrates that the equilibrium coffee price ratio has increased sys-tematically since liberalisation of coffee markets in all three countries. This in-crease means that liberalisation has improved the returns to producers byreducing the net transfer costs throughout the coffee supply chain.
Therefore, the answer to the question of whether producers are better off interms of the revenue that they receive since liberalisation appears to be unam-biguously yes. Producers have benefited from a higher real price of coffee,higher coffee production and a higher share of the terminal price of coffee.
The impact of liberalisation on the input costs of producers is obscured bythe lack of data at the producer level.22 However, there are studies that esti-mate the relative input shares for the production of coffee. For example, ICO(1996/97) reports that the annual cost shares for small holder Arabica produc-tion in India are 54, 26, 7 and 14 per cent for labour, materials (fertiliser,pesticide, bags, seedlings), processing (machinery, energy) and overheads re-spectively. Consequently, if the prices of these inputs are largely determined atthe economy wide level and mostly responsive to macroeconomic policies andthe exchange rate then the evolution of input costs are independent of liberal-isation. It would then be reasonable to argue that there are no input cost impli-cations to liberalisation and that the benefits to producers in terms of revenueoutlined above correspond to the total benefits to producers.23
22. The lack of input cost data may be overcome in the near future. In April 2010 the Executive
Director of the International Coffee Council asked all members of the organisation to provide input cost
data for crop years 2000/01 to 2009/10. See ICO (2010).
23. We have not come across any studies showing that liberalisation has increased input costs. It is
possible that input costs after liberalisation increased faster than revenues because of factors other than
liberalisation. However, producers are still better off after liberalisation (due to the higher revenues)
because without liberalisation they would have been even worse off financially.
Bill Russell, Sushil Mohan and Anindya Banerjee 535
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
V. C O N C L U S I O N A N D P O L I C Y I M P L I C A T I O N S
This paper argues that to understand the dynamics of coffee prices we needto allow for structural breaks in the equilibrium coffee price ratio due tochanges in government policies. To simply undertake the standard analysisignoring the breaks is inadequate. In contrast, estimating the model allow-ing for the shifts in the coffee price ratio as reported in Table 4 providesestimates that are consistent with our understanding of coffee markets.First, terminal and producer prices move closely together in equilibrium.Second, shocks to the relationship between the two coffee prices are eradi-cated very quickly. Third, liberalisation of the coffee markets has coincidedwith increases in the coffee price ratio which is equivalent to the producers’share of the terminal price of coffee. In Brazil, the equilibrium share hasrisen from 0.6254 in the late 1980s to 0.8461 in the most recent period upto October 2007. Over the same period the equilibrium share to producershas increased from 0.6325 to 0.7896 for Guatemala and from 0.5485 to0.8494 for India.
The systematic increase in the producers’ equilibrium share of terminalcoffee prices over the last seventeen years has greatly benefited the producers inthese three coffee producing countries. The combination of the increase in theequilibrium producers share and the increase in the real price of coffee afterliberalisation beginning in 1990 suggest that the benefit to producers in termsof revenue in 2007 is around 2.15, 0.15 and 0.22 $US billion for Brazil,Guatemala and India.24 This can be compared with the actual payments tocoffee producers of 4.63, 0.47 and 0.65 $US billion in 2007.
It should be stressed that the analysis does not suggest that liberalisationposes no risk for producers. Instead it gives rise to new problems by expos-ing producers to the vagaries of the market. In particular, concerns havebeen raised that liberalisation has exposed producers to the full volatility ofcoffee prices.25 The international community and policy planners could domore for producers by helping to develop missing credit and insurancemarkets so as to handle this increase in volatility. However, leaving asidethe issue of what more could be done for producers, this paper emphasisesthat responding to calls for returning to the kind of interventions that
24. The total benefit to producers in 2007 is calculated as:
Q2007†PT;2007† Ue2007
� �� Q2007†
PT;2007
1þ dLð Þ17 † Ue1989
h iwhere 1þ dLð Þ17 is the compounded annual real
increase in terminal prices during the 17 years of liberalisation. The first component of the calculation is
what the producer actually received in 2007. The second component is what the producer would receive
if the real terminal price was lower,PT;2007
1þ dLð Þ17 ; and they only received the 1989 equilibrium producers
share, U1989e . The difference between the two components is the benefit to producers from higher real
coffee prices and a higher equilibrium share of terminal prices. This measure of the benefit assumes that
the quantity of coffee produced has been unaffected by the liberalisation of the coffee markets.
Alternatively, if coffee output has increased since liberalisation then our estimate understates the ‘true’
benefit to producers.
25. For example see Gemech et al. (2011).
536 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
existed prior to the liberalisation of coffee markets is not in the interest ofproducers.
RE F E R E N C E S
ActionAid, 2008. Commodity Dependence and Development: Suggestions to Tackle the Commodities
Problem. Report prepared for the ActionAid and the South Centre, April, Johannesburg and Geneva.
Akiyama, T., 2001. Coffee Market Liberalization since 1990. In: T. Akiyama, J. Baffles, D.F. Larson, and P.
Varangis Eds., Commodity Market Reforms: Lessons of Two Decades, World Bank, Washington DC.
Akiyama, T., J. Baffes, D. Larson, and P. Varangis, 2001. Market Reforms: Lessons from Country and
Commodity Experiences. In: T. Akiyama, J. Baffles, D.F. Larson, and P. Varangis Eds., Commodity
Market Reforms: Lessons of Two Decades, World Bank, Washington DC.
Alizadeh, A., and N. Nomikos, 2005. Agricultural Reforms and the Use of Market Mechanisms for
Risk Management. Report prepared for the Futures and Options Association, London.
Anderson, K., M. Kurzweil, W. Martin, D. Sandri, and E. Valenzuela, 2008. Measuring Distortions to
Agricultural Incentives Revisited, World Trade Review, vol. 74, pp. 675–704.
Ardeni, P.G., 1989. Does the Law of One Price Really Hold for Commodity Prices? American Journal
of Agricultural Economics, vol. 71, pp. 303–28.
Baffes, J., 1991. Some Further Evidence on the Law of One Price, American Journal of Agricultural
Economics, vol. 4, pp. 21–37.
Bai, J., and P. Perron, 1998. Estimating and Testing Linear Models with Multiple Structural Changes.
Econometrica, vol. 66, pp. 47–78.
Bates, R.F., 1997. Open Economy Politics: the Political Economy of the World Coffee Trade, Princeton
University Press, Princeton.
Batista, J.C., and G.B da Silveira, 2010. Trade Costs and Deviations from the Law of One Price,
American Journal of Agricultural Economics, vol. 92, pp. 1011–23.
Calfat, G., and R. Flores, 2002. Government Actions to Support Coffee Producers — an Investigation of
Possible Measures from the European Union Side. Economics working paper, vol. 448. Getulio
Vargas Foundations, Brazil.
Daviron, B., and S. Ponte, 2005. The Coffee Paradox: Global Markets, Commodity Trade and the
Elusive Promise of Development. Zed Books, London.
Enke, S., 1951. Equilibrium among Spatially Separated Markets: Solution by Electrical Analogue,
Econometrica, vol. 191, pp. 40–47.
Fitter, R., and R. Kaplinsky, 2001. Who Gains from Product Rents as the Coffee Market becomes more
Differentiated? A Value Chain Analysis, IDS Bulletin Paper, May 2001, University of Sussex, UK.
Fortenbery, R., and H. Zapta, 2004. Developed Speculation and Underdeveloped Markets – the Role of
Futures Trading on Export Prices in Less Developed Countries’, European Review of Agricultural
Economics, vol. 314, pp. 451–71.
Gemech, F., S. Mohan, A. Reeves, and J. Struthers, 2011. Market-based Price-Risk Management:
Welfare Gains for Coffee Producers from Efficient Allocation of Resources. Oxford Development
Studies, vol. 39, no. 1, pp. 49–68.
Gilbert, C., 1996. International Commodity Agreements: An Obituary Notice, World Development,
vol. 241, pp. 1–19.
ICO, 1989. Agreement: Renegotiation of the International Coffee Agreement, Outline Joint Approach for
Implementing the Universal Quota Concept, ICC 51/5 E 4/6, International Coffee Organization, London.
ICO, 1996/1997. Coffee Market Studies: India, International Coffee Organization, London.
———, 2007. Coffee Statistics, International Coffee Organization, London.
Bill Russell, Sushil Mohan and Anindya Banerjee 537
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
———, 2010. Production Costs, ICC 105-6, International Coffee Council, 105th Session 22 – 24
September 2010, International Coffee Organization, London.
Jarvis, L.S., 2005, The Rise and Decline of Rent-Seeking Activity in the Brazilian Coffee Sector: Lessons
from the Imposition and Removal of Coffee Export Quotas, Working Paper No. 04-004,
Department of Agricultural and Resource Economics and the Giannini Foundation, University of
California, California, USA.
Jerome, A., and O. Ogunkola, 2000. Characteristics and Behavior of African Commodity/Product
Markets and Market Institutions and their Consequences for Economic Growth. CID Working
Papers 35, Center for International Development, Harvard University, USA.
Krivonos, E., 2004. The Impact of Coffee Market Reforms on Producer Prices and Price Transmission,
Policy Research Working Paper No. 3358, July, World Bank, Washington DC.
McIntire, J., and P. Varangis, 1999. Reforming Cote d’Ivoire’s Cocoa Marketing and Pricing System,
Policy Research Working Paper No. 2081, March, World Bank, Washington DC.
Mehta, A., and J-P. Chavas, 2008. Responding to the Coffee Crisis: What can we Learn from Price
Dynamics? Journal of Development Economics, vol. 85, pp. 282–311.
Michael, P., A.R. Nobay, and D. Peel, 1994. Purchasing Power Parity Yet Again: Evidence from Spatially
Separated Commodity Markets, Journal of International Money and Finance, vol. 136, pp. 637–57.
Mohan, S., 2007. Market-based Price-risk Management for Coffee Producers, Development Policy
Review, vol. 253, pp. 333–54.
Mundlak, Y., and D.F. Larson, 1992. On the Transmission of World Agricultural Prices, The World
Bank Economic Review, vol. 6, pp. 399–422.
Oxfam, 2002 Mugged: Poverty in your Coffee Cup, Oxfam, Oxford.
Perron, P., 1989. The Great Crash, the Oil Price Shock, and the Unit Root Hypothesis, Econometrica,
vol. 57, no. 6, November, pp. 1361–401.
Ponte, S., 2002. The ‘Latte Revolution’? Regulation, Markets and Consumption in the Global Coffee
Chain. World Development, vol. 30, pp. 1099–122.
Raffaeli, M., 1995. Rise and Demise of Commodity Agreements: An Investigation into the Breakdown
of International Commodity Agreements, Woodhead Publishing, Cambridge, UK.
Rapsomanikis, G., D. Hallam, and P. Conforti, 2004. Market Integration and Price Transmission in
Selected Food and Cash Crop Markets of Developing Countries: Review and Applications,
Commodity Market Review 2003-2004, Food and Agricultural Organisation, Rome.
Samuelson, P.A., 1952. Spatial Price Equilibrium and Linear Programming, American Economic
Review, vol. 42, pp. 560–80.
———, 1957. Intertemporal price equilibrium. Welwirtschaftliches Archiv, Band, 79, pp. 181–221.
Shepherd, B., 2004. Market Power in International Commodity Processing Chains: Preliminary Results
from the Coffee Market. GEM, Sciences-Po, Paris.
South Centre, 2008. From Declaration to Action on Commodities: Making the Turning Point at
UNCTAD X11, Policy Brief 14, South Centre, Geneva.
Takayama, T., and G.G. Judge, 1971. Spatial and Temporal Price Allocation Models, Amsterdam,
North Holland.
Talbot, J.M., 2004. Grounds for Agreement: the Political-Economy of the Coffee Commodity Chain.
Rowman and Littlefield, Oxford.
Varangis, P., D. Larson, and J.R. Anderson, 2002. Agricultural Markets and Risks - Management of the
Latter, not the Former, Policy Research Paper No. 2793, World Bank, Washington DC.
Vataja, J., 2000. Should the Law of One Price be Pushed Away? Evidence from International
Commodity Markets, Open Economies Review, vol. 11, pp. 399–415.
Winter-Nelson, A., and A. Temu, 2002. Institutional Adjustment and Transaction Costs: Product and
Input Markets in the Tanzanian Coffee System, World Development, vol. 304, pp. 561–74.
538 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Implications of COMTRADE CompilationPractices for Trade Barrier Analyses and
Negotiations
Alexander J. Yeats
U.N. Commodity Trade (COMTRADE) statistics have major shortcomings for manyanalyses relating to tariffs and other trade barriers. The use of a cost-insurance-freightvaluation base for these data results in an upward (sometimes severe) bias in theimplied dutiable value of imports for countries that utilize free-on-board tariffs. Thisproblem can be greatly exacerbated by the “general” trade system procedure used tocompile the U.N. statistics, as opposed to the “special” trade practice used for theWorld Trade Organization Integrated Database. U.S. International Trade Commissionstatistics show that the combined effects of these biases can reach magnitudes that pre-clude the legitimate use of COMTRADE for many tariff-trade simulations or relatedtrade negotiations. JEL codes: F13, F17
To facilitate meaningful cross-country comparisons, the United Nations StatisticalOffice (UNSO) justifiably requests that members report import statistics to theCommodity Trade (COMTRADE) Statistics Division on a commoncost-insurance-freight (c.i.f.) basis, even in the case of countries like the UnitedStates, Canada, Australia, and New Zealand, where free-on-board (f.o.b.) valu-ation tariffs are used. As a result, COMTRADE overstates the actual dutiablevalue of these countries’ imports. A relevant question, which appears to have beengiven little consideration, is whether this valuation procedure invalidates the use ofCOMTRADE for analysis of these countries’ tariffs and similar trade restrictions.
A second important point relates to the “reporting system” used for thecompilation of COMTRADE statistics. Most countries employ the “general”system, which includes imports for direct consumption as well as importsunder customs bond or into officially designated foreign trade zones (FTZs).The latter are not subject to national tariffs unless they eventually clear
Alexander Yeats is a former member of the “trade team” in the World Bank’s Development
Economics group. His email address is [email protected]. Major parts of this paper were written while
the author was a consultant to the World Bank’s Africa Region Department. The author wishes to
thank Francis Ng, members of the USITC “trade data web” support staff, and an anonymous referee for
comments and suggestions.
THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 539–555 doi:10.1093/wber/lhr053Advance Access Publication October 28, 2011# The Author 2011. Published by Oxford University Press on behalf of the International Bankfor Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions,please e-mail: [email protected]
539
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
customs controls. Statistics compiled by the United States International TradeCommission (USITC) can be used to determine if inclusion of FTZ transactionsin COMTRADE causes the U.N. data to significantly further misstate the duti-able value of imports. These questions are of importance because recent effortshave attempted to utilize COMTRADE for tariff change simulations andrelated issues in multilateral trade negotiations.
Before proceeding, one important qualification should be noted. In this study,the term “COMTRADE bias” is sometimes used. This is in no way intended asa criticism of the methodology or procedures used in constructing the databasewhich are fully consistent with the appropriate and intended applications of theU.N. statistics. Rather, the term is directed at users who appear to be unawareof important basic characteristics of COMTRADE and have attempted toemploy the data in ways that are incorrect and inappropriate.
I . C H A R A C T E R I S T I C S O F T R A D E P R O J E C T I O N M O D E L S
During the Uruguay Round, the World Bank and UNCTAD (1987) developeda partial equilibrium projection model (named SMART, for Software forMarket Analysis and Restrictions to Trade) to simulate the effects of proposedtariff cuts in the negotiations. The projections employed tariff line level importstatistics contained in the WTO Integrated Data Base since it was acknowl-edged that the COMTRADE data then available were too aggregate, weresometimes tabulated in inappropriate values, and were compiled using a meth-odology not appropriate for tariff simulations. The fact that far more detailedsix-digit Harmonized System (HS) statistics have now become generally avail-able has generated renewed interest in the possible use of COMTRADE fortariff analyses and simulations (see http://go.worldbank.org/IJIR5D0T80).
In partial equilibrium trade models, the projected trade creation for producti (TCi) resulting from a tariff cut is normally derived from
TCi ¼Mi � ed � Dti=ð1þ tiÞ � ð1� ðed=esÞÞ ð1Þ
where Mi is the initial value of imports of the product, ed and es are the importelasticities of demand and supply for the item, and ti is the import tariff (seeWorld Bank and UNCTAD 1987 or UNCTAD 1986 for previous applications).Two important points are evident from this equation. First, if the initial valueof imports is overstated by a given amount (say 20 percent), the projectedvalue for trade creation will be upward biased by this same percentage.Second, in cases where the percentage overstatement in the trade base is greaterthan the applied nominal tariff, the trade creation projection error will exceedthe actual value of this parameter. Similar issues arise if tariff changes are ana-lyzed in a computable general equilibrium (CGE) framework.
Although different procedures have been used for estimating trade diversion,valuation biases in COMTRADE will generate similar errors for these
540 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
projections. For example, Baldwin and Murray (1977) simulate trade diversionusing a methodology which incorporates the trade creation projection estimateas an explanatory factor. In this and other empirical approaches (see WorldBank and UNCTAD 1987), the higher the COMTRADE valuation bias, thelarger will be the resulting error in trade diversion projections.
I I . T H E M A G N I T U D E O F C R O S S - P R O D U C T VA L U A T I O N B I A S E S
The United States International Trade Commission (USITC) provides onlinepublic access to national import statistics down to the level of individual ten-digitHTS products (see http://USITC.gov). Aside from the f.o.b. value of imports fordirect consumption, this system (named “trade data web”) also provides infor-mation on international transport and insurance costs incurred in bringing theproduct to the first U.S. port of entry. As requested by the United NationsStatistical Office, these c.i.f. import values are reported for inclusion in U.N.COMTRADE records. As a result, COMTRADE overstates the dutiable value ofimports for countries like Australia, Canada, New Zealand, and the UnitedStates that utilize f.o.b. import tariffs. A key question relating to tariff analysesand projections concerns the magnitude of this bias.
Since the USITC “trade data web” provides both free-on-board importvalues and international transport costs for all imported products it can beemployed to quantify the tariff valuation bias for countries trading with theUnited States. Specifically, the nominal bias for imports from a given countrycan be derived by taking the ratio of all transport and insurance costs for agiven product to its landed f.o.b. value.
For an empirical assessment of the bias, the following procedure wasemployed. First, a selection of 45 “test” countries was chosen with an effortmade to achieve as much geographic and economic diversity as possible. Theselection included countries in Asia, Africa, Europe, and South America, islandcountries like the Maldives, Fiji, and Sri Lanka, as well as several that werelandlocked (Nepal and Paraguay). Next, available statistics were drawn fromthe USITC trade data web to compute nominal transportation costs for all six-digit HS level products imported from these countries by the United States.These freight factors were then ranked in ascending order, that is, from thelowest nominal transport cost for each product to the highest. The percentilevalues for each country’s distribution of biases are given in Table 1.
As an illustration, the table shows that the United States imported 132 indi-vidual 6-digit HS products from the Cote d’Ivoire with a total 2007 value of$639 million. (All dollar amounts are current U.S. dollars.) The mediannominal freight factor (i.e., the valuation bias) was 12.7 percent, while three-tenths of all shipments had a freight factor of 25 percent or more. A similarshare of imports from Ghana, Nepal, Togo, and the Maldives incorporatedbiases exceeding 20 percent, while the biases for one-tenth of all imports fromBelize, Senegal, Honduras, Nepal and other countries exceeded 35 percent. Ten
Yeats 541
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA B L E 1. The Distribution of Valuation Biases for All Six-Digit HS ProductsExported to the United States by Forty-Five Partner Countries
2007 United States Imports
Percentile Values for the
Distribution of Biases (%)
Exporter No. of 6-digit HS Products Value ($ million) Median 70th 90th 95th
Maldives 18 4 9.8 30.4 95.6 119.7
Tonga 21 8 40.0 51.0 78.9 90.0
Cote d’Ivoire 132 639 12.7 25.1 59.4 79.6
Ghana 264 212 13.6 22.8 49.5 78.0
Senegal 108 21 8.9 15.0 38.2 74.6
Fiji 164 191 8.0 13.9 34.3 65.5
Guinea 61 139 7.1 24.4 56.0 64.9
Belize 114 113 6.1 11.6 47.4 62.5
Honduras 589 4,101 7.5 14.9 36.8 61.5
Uruguay 474 524 8.3* 13.5 31.0 58.5
Egypt, Arab Rep. 718 2,545 8.1 13.7 31.5 54.7
Togo 49 6 12.9 22.8 36.7 54.0
Nepal 350 97 12.1 20.0 39.1 53.4
Guyana 150 146 8.7 18.1 36.4 52.1
Ecuador 808 6,540 8.8 14.7 34.1 51.2
Peru 1,162 5,489 7.9 11.8 30.1 50.4
Bolivia 362 377 7.7 12.1 28.5 49.4
Bangladesh 435 3,635 7.6* 11.7 26.2 46.9
Argentina 1,556 4,820 7.9 12.6 28.5 44.5
Guatemala 779 3,269 8.1 12.9 29.5 44.2
Philippines 1,550 9,813 7.0 11.6 24.8 44.1
Panama 532 391 6.7 12.3 26.4 43.4
Chile 1,028 9,784 7.6 12.5 28.8 43.1
Greece 879 1,297 6.3 11.0 25.7 41.0
Sri Lanka 604 2,178 6.6* 10.2 22.1 39.4
St. Lucia 74 36 7.5 11.7 26.6 39.0
Paraguay 134 80 9.6 15.4 28.2 38.0
New Zealand 1,480 3,316 5.3 9.2 20.8 37.9
Costa Rica 973 4,209 5.8 10.0 22.8 36.9
Vietnam 1,489 11,425 7.7* 12.6 25.5 36.7
Sierra Leone 174 60 4.2 8.1 21.4 35.9
Australia 2,436 8,971 5.2 9.0 23.2 35.0
Pakistan 1,060 3,831 8.2* 11.5 23.0 34.9
Turkey 1,869 4,897 6.8 10.6 22.8 34.8
Thailand 2,355 23,793 6.1 10.6 21.0 34.2
Indonesia 1,816 15,208 6.7 11.5 22.6 33.7
Colombia 1,486 10,034 6.2 10.0 20.8 29.9
Brazil 2,738 27,193 6.1 9.8 19.2 29.5
Poland 1,780 2,350 5.4 8.5 18.6 28.9
Tunisia 453 478 4.3 6.9 18.9 28.2
Morocco 601 664 8.1 9.6 19.9 27.5
Austria 2,249 10,893 4.4 7.3 17.5 27.4
Cyprus 155 18 4.5 7.3 18.7 26.8
India 3,342 25,113 7.0 9.9 17.7 24.6
Taiwan 3,198 39,853 5.6 8.1 15.3 23.5
MEDIAN VALUE 602 2,264 7.6 11.7 26.3 42.1
* The median U.S. tariff exceeds the median nominal transport cost ratio for this country.
Note: As an illustration, the table shows 30 percent of the valuation biases for the Coted’Ivoire exceed 25.1 percent. Thirty percent of all biases for Guyana exceed 18.1 percent, while10 percent of Senegal’s biases exceeded 38.2 percent.
542 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
percent of the combined U.S. imports from all countries listed in thetable incurred nominal freight costs exceeding 26 percent.
Overall, for the 45 countries, only five (Uruguay, Bangladesh, Sri Lanka,Vietnam, and Pakistan) faced median Unites States tariffs that exceeded theirmedian ad valorem transport costs. As a result, the percentage error in tradecreation estimates for most countries will exceed the actual percentage changein this parameter, often by very high margins. The implications are thatCOMTRADE statistics can generate highly inaccurate estimates of the leveland composition of the trade response to tariff changes.
While Table 1 examined the distribution of individual countries’ valuationbiases across all exports, an important related question is whether significantdifferences exist in the biases between product groups. If so, one would havean interest in identifying those sectors where COMTRADE most seriouslyoverstates the dutiable value of imports. Second, if the valuation biases withinspecific product groups fall in a relatively narrow range this might suggest thepossibility of employing a standard f.o.b.–c.i.f. correction factor to offset, orreduce, the bias. This approach would be similar to the ten percent factoremployed by the IMF to account for differences in export and import statistics.
For relevant information, U.S. trade statistics for the countries in Table 1were combined into six different regional groups (i.e., Southern Cone SouthAmerica, Other South America and the Caribbean, Oceania, sub-SaharanAfrica, Southeast Asia, and South Asia. Next, the COMTRADE bias was com-puted for U.S. imports of several broad categories of goods from these countrygroups. For four of the six country groups the valuation bias for textiles andfootwear products exceeds the median bias for all goods, while meat, fish andvegetable imports frequently incur higher than average nominal transportationcosts. Almost one-third of all meat exports from Oceania had nominal transportcosts exceeding 21 percent, while the corresponding freight rate for sub-SaharanAfrican vegetable products was about 31 percent. One possible explanation forthese results is that due to their perishable nature, food products rely moreheavily on relatively expensive air transport to access U.S. markets.
Other products sectors with relatively high valuation biases include hidesand leather goods, as well as articles of plastic and glassware products.Conversely, nominal freight costs for vehicles and machinery, optical goods,and scientific instruments are generally among the lowest in the table. Theoverall diversity of the biases across groups provided little evidence that theCOMTRADE valuation problem could be corrected by a standard f.o.b.–c.i.f.adjustment factor.1
1. While this investigation focused on COMTRADE valuation biases for the United States, several
published transport cost studies indicate the conclusions concerning similar unacceptably high biases
can be generalized to countries like Australia, Canada, and New Zealand that also employ
free-on-board import tariffs. See Curtis and Chen (2003), Conlon (1982), Pomfret and Sourdin (2010),
Hummels (2007), or Lloyd (1976) among others.
Yeats 543
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
I I I . E F F E C T S O F G E N E R A L T R A D E R E P O R T I N G P R A C T I C E S
While utilization of an inappropriate cost-insurance-freight valuation base cansignificantly overstate the dutiable value of U.S. imports, there is an additionalproblem that may produce even greater biases. Specifically, two methodologiesare used for compiling import statistics; namely, the general and special record-ing systems. Special trade statistics tabulate the value of goods imported direct-ly for final consumption. This exchange encounters any existing tariffs andrelated trade control measures so special trade data are submitted by the U.S.to the WTO Integrated Data Base. In contrast, general trade statistics recordthe value of merchandise imports, either for direct immediate consumption orinto bonded warehouses and foreign trade zones (FTZs) under customscustody. Imports under the general trade regime destined for FTZs are exemptfrom tariffs unless they are redirected toward domestic markets for consump-tion. Due to these special import provisions, general trade statistics, which areemployed in the COMTRADE database, have major shortcomings for analysesof trade restrictions.2
As a result of their compilation procedures, general statistics may seriously mis-state dutiable import values, and may also fail to correctly identify the goodsfacing trade restrictions. A hypothetical example can illustrate this point. Assumethe U.S. imports $10 billion of crude petroleum (HTS 270900) into a foreigntrade zone. In the zone the shipment is further processed (refined) into distillatefuel oils (HTS 271019) and this product is then transferred to the domesticmarket for consumption—at which point applicable U.S. tariffs are assessed.Under normal zone procedures, importers generally have the option of payingduties on the original materials imported into the zone or on the finished fabri-cated product. Since the nominal equivalent of specific tariffs on distillate fuels isrelatively lower than those on crude oil, the former would be reported oncustoms forms. The statistical records of these transactions would be as follows:
(i) COMTRADE would record statistics on the actual value ($10 billion) ofcrude oil imports. In contrast, the WTO Integrated Database would notreport any imports of crude oil, because this specific product did notcross the customs frontier for domestic consumption.
(ii) The IDB would report statistics on the distillate fuel oil imports, sincethese are the shipments upon which relevant United States import dutiesare assessed. In contrast, U.N. COMTRADE would not report import
2. Recent surveys indicate that more than 250 general purpose United States FTZs have been
established. These zones are considered to be outside of U.S. Customs territory for the purpose of
import duty liability. Therefore, goods destined for FTZs are not subject to customs tariffs unless they
formally enter into U.S. Customs territory—at which point they will be reported to the IDB.
Merchandise shipped to foreign countries from FTZs is exempt from duty payments. This provision is
especially important for firms that import components to manufacture finished products for export.
Various activities can be conducted in a zone, including assembly, packaging, storing, cleaning,
repacking, sorting, grading, testing, labeling, repairing, combining foreign or domestic components, or
further processing. See MacLeod (2000) for a useful discussion of activities in U.S. foreign trade zones.
544 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
information for these goods due to their significant physical transform-ation within U.S. geographic territory, that is, the foreign trade zone.
Table 2 provides examples of the magnitude of the general system bias usingstatistics on selected six-digit HS U.S. imports from Korea. The table shows thedutiable f.o.b. value for each item (column 3), which is reported to the WTOIntegrated Data Base. These numbers represent the value of goods imported fordirect consumption and, as such, are subject to existing tariffs. In addition,column 5 shows the corresponding c.i.f. general import value reported toCOMTRADE. Differences between these values indicate the magnitude of biasassociated with the use of the U.N. data for tariff analyses and/or projections.
Column (4) has been added to help assess the relative size of the f.o.b.–c.i.f.valuation bias in COMTRADE data as opposed to biases originating fromtabulating data using the general trade system. As an example, the c.i.f. valueof dutiable imports of gear boxes (HTS 870840) is $1.5 million higher thanthe dutiable customs value, while the difference between the latter and the c.i.f.general import value is $319 million. These comparisons indicate that thegeneral trade compilation practices account for almost all of the differences inimport values reported in the IDB and U.N. COMTRADE.
The clear impression from Table 2 is that COMTRADE biases are of a mag-nitude that invalidates use of these statistics for tariff simulations or negotia-tions. Biases incorporated in the U.N. data may seriously misdirect nationalpriorities for a liberalization across products, and may also significantly over-state overall potential trade gains. Specifically,
– COMTRADE overstates the dutiable value of the first item (HTS252329 — Portland Cement) by about $57 million, or 56 percent. Sincethe customs c.i.f. and general import values are equal ($158.9 million),the bias is entirely attributable to the inappropriate (for tariff analysis)valuation base employed for the U.N. statistics.
– Differences of just under 1 billion dollars (approximately 67 percent)occur in the import value for petroleum oils (HTS 271019) reported tothe IDB for domestic consumption and the general import total ($2.4billion) recorded in COMTRADE. Only about $63 million, or 7 percentof the difference, is attributable to the alternative valuation bases (seecolumns 3 and 4).
– COMTRADE-IDB differences exceeding several thousand percent occurfor rare earth imports (HTS 284610). The table indicates that about 95percent of all imports of this product were destined for the processingzones. As such, the IDB correctly does not record most shipments of thisitem. Similarly, the import value reported in COMTRADE for organicsolvents (HTS 381400) is more than ten times greater than the dutiablevalue of imports for domestic consumption.
– COMTRADE overstates the dutiable value of electrical or nonelectricaltapes (HTS 391990) by about $44 million, or approximately 300
Yeats 545
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E2
.E
xam
ple
sof
the
Magnit
ude
of
Tra
de
Sys
tem
and
Valu
atio
nB
iase
sin
CO
MT
RA
DE
Dat
aon
U.S
.Im
port
sfr
om
Kore
a
2006
Import
Valu
e($
000)
CO
MT
RA
DE
Bia
s
HT
SN
o.
Des
crip
tion
Duti
able
Cust
om
s(f
.o.b
.)*
Rec
ord
edC
ust
om
s(c
.i.f
.)G
ener
al
Import
s(c
.i.f
.)**
Per
cent
(%)
Valu
e($
000)
252329
Port
land
cem
ent
101,9
85
158,8
60
158,8
60
55.8
56,8
75
271019
Oth
erpet
role
um
oil
s1,4
37,7
21
1,5
00,4
82
2,4
05,2
61
67.3
967,5
39
284610
Rare
eart
hco
mpounds
27
42
713
2,5
67.4
686
381400
Org
anic
solv
ents
130
140
1,5
94
1,1
28.8
1,4
65
382490
Oth
ercu
lture
dcr
ysta
ls11,2
76
11,7
83
15,1
35
34.2
3,8
59
390319
Oth
erpoly
mer
s13,8
41
14,8
75
20,7
20
49.7
6,8
80
391910
Sel
fadhes
ive
pla
tes
4,1
37
4,4
32
6,0
86
47.1
1,9
49
391990
Ele
ctri
cal
or
non-e
lect
rica
lta
pes
14,4
77
15,2
22
58,0
48
301.0
43,5
71
392630
Fit
tings
for
furn
iture
1,5
90
1,7
04
2,6
76
68.3
1,0
86
400912
Pip
esw
ith
fitt
ings
284
302
636
124.2
352
401039
Conve
yer
bel
ts2,5
28
2,6
52
3,2
38
28.1
710
540753
Wove
nsy
nth
etic
fabri
c472
511
1,6
32
245.8
1,1
60
551612
Dye
dw
ove
nfa
bri
cs105
112
305
191.0
200
620690
Wom
en’s
blo
use
s185
194
467
152.1
282
721633
Iron
shapes
15,6
88
17,1
93
20,3
53
29.7
4,6
65
730820
Iron
stru
cture
s17,9
63
26,0
32
26,0
32
44.9
8,0
69
731814
Sel
fta
ppin
gsc
rew
s2,2
48
2,5
39
2,9
14
29.7
666
731824
Cott
erpin
s295
311
639
116.9
345
830170
Lock
key
s514
528
3,6
79
616.1
3,1
65
840734
Inte
rnal
com
bust
ion
engin
es19,6
25
20,1
73
46,1
73
135.3
26,5
48
840991
Part
sfo
rair
craf
ten
gin
es83,1
58
86,1
94
159,7
11
92.1
76,5
53
842131
Oil
or
fuel
filt
ers
15,8
66
17,5
34
20,3
53
28.3
4,4
86
848130
Copper
valv
es7,3
74
7,6
04
10,8
55
47.2
3,4
81
848330
Tra
nsm
issi
on
bea
rings
4,3
37
4,4
67
15,9
50
267.8
11,6
13
848410
Met
al
gask
ets
1,3
75
1,4
39
2,6
66
93.9
1,2
91
850110
Ele
ctri
cal
moto
rs29,9
40
31,2
74
37,5
21
25.3
7,5
81
546 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
850140
Oth
eralt
ernat
ing
curr
ent
moto
rs18,7
19
19,5
70
24,4
46
30.6
5,7
27
850511
Ele
ctro
magnet
sof
met
al
2,1
02
2,1
45
5,3
34
153.8
3,2
32
851110
Spark
plu
gs
2,3
73
2,5
12
4,0
19
69.4
1,6
46
851120
Ignit
ion
magnet
s1,1
81
1,1
84
2,0
50
73.6
869
851130
Ignit
ion
coil
s3,5
62
3,8
43
8,7
56
145.8
5,1
94
851140
Sta
rter
moto
rs25,4
03
26,0
06
33,3
58
31.3
7,9
55
851150
Oth
ergen
erat
ors
14,8
40
15,1
55
28,7
79
93.9
13,9
39
851890
Part
sof
tele
phone
hea
dse
ts10,4
06
10,6
36
13,1
38
26.3
2,7
32
852692
Radio
rem
ote
contr
ol
appara
tus
3,3
88
3,6
12
7,2
60
114.2
3,8
71
852812
Monit
ors
and
pro
ject
ors
266,4
92
272,1
65
640,9
37
140.5
374,4
45
853650
Ele
ctro
nic
AC
swit
ches
40,5
88
42,1
91
53,2
88
31.3
12,7
01
853690
Oth
erel
ectr
onic
swit
ches
12,0
98
12,5
49
18,5
74
53.5
6,4
76
854430
Ignit
ion
wir
ing
sets
6,7
05
7,0
92
12,2
87
83.3
5,5
82
870829
Oth
erpart
sof
auto
mobil
es212,0
32
232,8
15
258,0
40
21.7
46,0
07
870839
Part
sof
bra
kes
175,6
42
184,3
37
215,4
21
22.6
39,7
79
870840
Gea
rboxes
and
part
s63,0
23
64,5
23
382,8
67
507.5
319,8
44
870880
Susp
ensi
on
syst
ems
and
part
s13,5
26
14,2
58
17,7
05
30.9
4,1
79
900220
Len
ses
and
filt
ers
266
271
12,9
21
4,7
54.6
12,6
55
950639
Oth
ergym
nast
iceq
uip
men
t6,1
36
6,5
06
15,0
24
144.8
8,8
88
TO
TA
LO
FA
BO
VE
2,6
65,6
21
2,8
47,9
72
4,7
76,4
21
79.2
2,1
10,8
00
*T
he
Unit
edSta
tes
report
sth
est
atis
tics
inth
isco
lum
nto
the
WT
OID
B.
**
The
Unit
edSta
tes
report
sth
est
atis
tics
inth
isco
lum
nto
U.N
.C
OM
TR
AD
E.
The
UN
SO
may
recl
ass
ify
som
eU
SIT
Cdat
ato
confo
rmw
ith
the
gen
eral
trea
tmen
tof
reim
port
edpro
duct
s.
Yeats 547
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
percent. Compilation practices for general imports were a major cause ofthe overall difference between these statistics and the data reported tothe IDB.
– If COMTRADE were used as a trade base for simulations of the effectsof import tariff changes for woven synthetic fabrics (HTS 540753) theupward bias in the projection error would be 246 percent. Similar pro-blems occur for dyed woven fabrics (HTS 551612), where COMTRADEoverstates the dutiable value of imports by approximately 190 percent.General imports account for most of the difference betweenCOMTRADE data and the relevant numbers in the IDB.
– Similarly, COMTRADE would generate an upward bias in projectionsof the effects of tariff changes for lock keys (HTS 830170) exceeding600 percent. The corresponding bias for transmission bearings (HTS848330) would be over $11.6 million, or almost 270 percent. Thegeneral reporting system accounts for almost all of the differencesbetween the WTO and U.N. statistics.
– Differences of approximately $700 million occur between the true duti-able values for the combined imports of radio remote control apparatus(HTS 852692) plus gear boxes and parts (HTS 870840) and the generalimport totals ($1.0 billion) recorded for these items in COMTRADE.The U.N. statistics overstate the dutiable value of gear boxes by over500 percent.
– The value reported in COMTRADE for imports of photographic lensesand filters (HTS 900220) is almost 50 times greater than the dutiablevalue of imports for consumption reported in the Integrated Data Base.Other examples in the table also reflect similar, very high, COMTRADEbiases.
Overall, for the products listed in Table 2 the combined import values reportedto COMTRADE are approximately $2.1 billion higher than the true dutiablevalue of these goods. These comparisons indicate the U.N. statistics overstatethe dutiable value of imports (and potential trade creation gains from a tariffliberalization) by about 80 percent—see the column totals. In general, majorstatistical discrepancies occur between COMTRADE and IDB data for rawmaterials, semi-finished goods, and components imported into FTZs forfurther processing.3
3. There are several reasons to believe that the biases associated with general trade statistics may be
even greater for some other countries. Assemble operations for parts and components are often a major
activity in foreign trade zones and high wage countries like the United States typically do not have an
extensive competitive advantage in these operations. Ng and Yeats (1999) construct multicountry
“revealed” comparative advantage indices for the assembly of parts and components. Their results
suggest that these types of operations, and potential biases from general statistics, may be much greater
in low wage, relatively high skill countries like Thailand, Malaysia, Indonesia, and the Philippines.
548 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Even stronger negative conclusions concerning the magnitude ofCOMTRADE biases emerge from statistics on U.S. imports from Austria. Forexample, Table 3 shows general imports of cars with cylinders under 3,000cc(HTS 870323) from Austria are reported as $3.0 billion in COMTRADE,while the dutiable value of these items imports reported to the IDB is roughly$2.1 billion (or 218 percent) lower. General imports of $333 million arereported for spark ignition engines (HTS 840732) which is almost fifty timesthe dutiable value of these shipments. Overall, the reported general imports ofthe 18 items listed in Table 4 are approximately $3 billion higher than the duti-able customs value of imports. Differences for several products, like articles ofmagnesium, actually exceed one thousand percent.4
Two key points emerge from these statistics.
– First, COMTRADE data have the capacity to significantly misdirect na-tional priorities for a tariff liberalization across products. This point isreflected in the fact that rankings of general import product values maydiffer substantially from those based on actual dutiable values of imports. Asan example, the value of U.S. general imports of spark ignition engines($333 million) is the third highest in the table even though the actual duti-able import value ($7 million) was exceeded by 8 other items—often by veryhigh margins.– Second, the COMTRADE general statistics may also provide very inaccur-
ate and unreliable information concerning the magnitude of potential overallgains resulting from a trade liberalization. This point is reflected in the factthat the actual dutiable value of U.S. imports from Austria is approximately40 percent lower than totals reflected in the general trade statistics (see thememo item in the table). This figure represents the potential overall error intrade creation projections that utilize COMTRADE statistics.
A further major defect of COMTRADE is that it may indicate no, or relativelylimited, domestic consumption of a good occurred when there was in fact sig-nificant dutiable trade. This situation is the converse of that reflected inTables 2 and 3 where COMTRADE overstated the customs value of imports.The underreporting problem occurs when imported raw materials, semi-finished goods, or components experience significant transformation in aforeign trade zone before shipment to domestic markets. As a result of the pro-cessing activity the originally imported good may be classified under a different
4. As noted, one cause of the COMTRADE-IDB statistical discrepancies is that a product
experienced significant transformation in an EPZ and cleared customs under a different HS classification
than that recorded in COMTRADE. In addition, the finished product may never have entered the
domestic market. For example, the automotive products listed in 3 may only have had modifications to
comply with domestic standards required in their final destinations in (say) Central or South America.
These products would be reported in COMTRADE because they entered United States geographic
territory, but would not be recorded in the IDB because they never cleared the U.S. customs frontier in
any form.
Yeats 549
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E3
.E
xam
ple
sof
the
Magnit
ude
of
Tra
de
Sys
tem
and
Valu
atio
nB
iase
sin
U.N
.C
OM
TR
AD
ED
ata
on
Unit
edSta
tes
Import
sfr
om
Aust
ria
2007
Import
Valu
e($
000)
CO
MT
RA
DE
Bia
s
HT
SN
o.
Des
crip
tion
Duti
able
Cust
om
s(f
.o.b
.)*
Rec
ord
edC
ust
om
s(c
.i.f
.)G
ener
al
Import
s(c
.i.f
.)**
Per
cent
(%)
Valu
e($
000)
870323
Car
cyli
nder
sunder
3000cc
955,8
78
963,8
89
3,0
38,8
58
217.9
2,0
82,9
80
870324
Car
cyli
nder
sove
r3000cc
93,7
32
94,3
16
645,0
34
588.2
551,3
02
840734
Spark
ignit
ion
engin
es7,0
48
7,1
63
332,8
70
4,6
22.9
325,8
23
870840
Gea
rboxes
and
part
s7,2
09
7,3
97
12,1
10
68.0
4,9
01
840991
Part
sfo
rair
craf
ten
gin
es14,8
10
15,2
77
18,5
97
25.6
3,7
87
870829
Oth
erpart
sof
auto
mobil
es36,7
57
38,4
41
40,2
59
9.5
3,5
02
392330
Bott
les
and
flask
sgoods
22,0
30
23,5
14
24,7
98
12.6
2,7
68
848180
Copper
taps,
cock
sand
valv
es18,9
29
19,3
39
20,7
71
9.7
1,8
42
392390
Oth
erco
nta
iner
s13,0
77
14,1
37
14,6
14
11.8
1,5
37
848190
Hand
oper
ated
appli
ance
part
s13,8
51
14,6
82
15,2
61
10.2
1,4
10
220860
Vodka
inco
nta
iner
s1,1
31
1,3
06
2,3
12
104.4
1,1
81
870899
Part
sof
trac
tors
23,6
30
24,1
52
24,3
87
3.2
757
810411
Art
icle
sof
magnes
ium
51
52
738
1,3
58.7
687
854430
Ignit
ion
wir
ing
sets
4,4
74
4,5
27
5,0
75
13.4
601
848220
Taper
edro
ller
bea
rings
4,6
76
4,9
06
5,2
00
11.2
524
40690
Oth
erch
eese
s4,9
68
5,1
71
5,4
49
9.7
481
711719
Oth
erro
pes
,ca
ble
sand
chain
s4,3
37
4,3
81
4,6
67
7.6
331
848250
Cyli
ndri
cal
roll
erbea
rings
3,2
66
3,2
84
3,5
07
7.4
241
TO
TA
LO
FA
BO
VE
1,2
29,8
53
1,2
45,9
33
4,2
14,5
06
242.7
2,9
84,6
53
Mem
oIt
emIm
port
sof
All
Goods
7,7
36,0
54
7,9
41,6
42
10,8
93,3
25
40.8
3,1
57,2
71
*T
he
Unit
edSta
tes
report
sth
est
atis
tics
inth
isco
lum
nto
the
WT
OID
B.
**
The
Unit
edSta
tes
report
sth
est
atis
tics
inth
isco
lum
nto
U.N
.C
OM
TR
AD
E.
The
UN
SO
may
recl
ass
ify
som
eU
SIT
Cdat
ato
confo
rmw
ith
the
gen
eral
trea
tmen
tof
reim
port
edpro
duct
s.
Sourc
e:A
llst
atis
tics
com
pute
dfr
om
USIT
CT
rade
Dat
aW
eb.
550 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
TA
BL
E4
.E
xam
ple
sof
CO
MT
RA
DE
Ove
r-and
Under
-rep
ort
ing
Duti
able
Unit
edSta
tes
Ener
gy
Import
s
2008
Import
Valu
es($
million)
CO
MT
RA
DE
Bia
s
HT
SN
o.
Des
crip
tion
Duti
able
Cust
om
s(f
.o.b
.)*
Rec
ord
edC
ust
om
s(c
.i.f
.)G
ener
al
Import
s(c
.i.f
.)**
Per
cent
(%)
Valu
e($
million)
270900
Cru
de
pet
role
um
oil
s274,9
50.2
281,8
25.3
363,3
91.1
32.2
88,4
41.0
271112
Pro
pane
6,6
09.1
6,8
29.0
3,5
74.7
245.9
23,0
34.4
271113
Buta
ne
2,6
65.8
2,7
38.4
1,3
16.6
250.6
21,3
49.1
271114
Eth
yle
ne
5,3
85.9
5,5
73.8
1,0
64.8
280.2
24,3
21.1
271119
Oth
erpet
role
um
gas
598.1
620.3
173.0
271.1
2425.1
271129
Oth
erpro
pane
10,1
89.6
10,5
38.7
84.1
299.2
210,1
05.5
271311
Pet
role
um
coke
14,9
06.6
15,4
51.9
174.8
298.8
214,7
31.8
271320
Pet
role
um
coke
bit
um
en4,5
07.2
4,6
48.3
752.0
283.3
23,7
55.2
271390
Oth
erPet
role
um
resi
due
670.5
699.1
68.3
289.8
2602.2
AB
OV
ER
EFIN
ED
PR
OD
UC
TS
45,5
32.7
47,0
99.6
7,2
08.2
284.2
238,3
24.5
*T
he
Unit
edSta
tes
report
sth
est
atis
tics
inth
isco
lum
nto
the
WT
OID
B.
**
The
Unit
edSta
tes
report
sth
est
atis
tics
inth
isco
lum
nto
UN
CO
MT
RA
DE
.T
he
UN
SO
may
recl
ass
ify
som
eU
SIT
Cdat
ato
confo
rmw
ith
the
gen
eral
trea
tmen
tof
reim
port
edpro
duct
s.
Sourc
e:A
llst
atis
tics
com
pute
dfr
om
USIT
CT
rade
Dat
aW
eb.
Yeats 551
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
COMTRADE HTS heading than the final product. In these cases, statistics onthe end products for domestic consumption are reported to the IDB, but notthe components that originally entered the FTZ. COMTRADE would notrecord statistics on the final product actually consumed, but would reportimports of the raw materials.
While it can be difficult to precisely identify the goods fabricated in FTZs,the situation is less problematic for energy imports since refinery operationsgenerally create a range of products that are clearly petroleum or natural gasderivatives. For example, Table 4 shows the IDB is reporting total U.S. crudepetroleum (HTS 270900) imports whose value is about $88 billion less thanCOMTRADE. The overstated COMTRADE statistics could cause exporters toplace a higher than warranted priority on liberalization of existing U.S. specifictariffs on crude oil imports. However, this strategy would be misdirected sincemuch of the discrepancy can be accounted for by differences in reportedIDB-COMTRADE imports of refined energy products. This transformation isthe reason why reported IDB imports of petroleum coke and bitumen are ap-proximately $18 billion higher than the corresponding figures in U.N.COMTRADE.5
Aside from energy products, COMTRADE may significantly underreportdutiable imports of a broad range of products that experience transformationin FTZs. Statistics on foodstuffs, beverages and tobacco, machinery, electron-ics, and transport equipment are among the sectors often affected by thisproblem. In extreme cases, COMTRADE fails to report any imports of itemsfor which the IDB shows dutiable trade exceeding millions of dollars occurred.Examples include U.S. imports of unwrought zinc (HTS 790111) and marinepropulsion engines (HTS 840721) from Korea. In other cases, COMTRADEsignificantly under-reported dutiable imports by over $400 million in the caseof motorized transport equipment (HTS 870323) from Korea, and by over $2billion for cellular telephones (HTS 851712) from China.
I V. C O N C L U S I O N S
This study examined characteristics of COMTRADE statistics to assess theirutility for tariff analysis and related applications in multilateral trade negotia-tions. This issue is of major importance since recent attempts have been madeto use COMTRADE for tabulating the value of imports subject to tariff andnontariff restrictions and simulating the trade response to negotiated tariffchanges. Accurate and reliable information on these points are key
5. An important point is that discrepancies between dutiable customs and general statistics often
become sharply smaller at higher levels of aggregation. This occurs when individual six-digit HS
products differentiate between unassembled components and the assembled form of a good while these
items are combined in a single category at (say) the four-digit level. Statistics compiled at very high
levels of aggregation (like two-digit data) may completely conceal the magnitude of the differences
occurring in the underlying, more detailed statistics.
552 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
requirements for the formulation of national trade strategies, or to supportmultilateral trade negotiations. For several reasons, negative conclusions werereached regarding the utility of unadjusted U.N. statistics for such efforts.
First, a serious problem exists concerning the valuation base employed forCOMTRADE. These statistics overstate the dutiable value of all United Statesimports, often very significantly, since they are expressed incost-insurance-freight values although the U.S. employs free-on-board importtariffs.6 As a result, the error in COMTRADE-based trade creation projectionscould seriously misdirect national priorities in multilateral negotiations. Thisprojection error extends across all regional groups of countries, as well asmajor product categories. Evidence was cited that indicates these problems alsooccur for other countries like Australia, Canada, and New Zealand thatemploy free-on-board import tariffs.
Second, the general trade compilation procedure used for COMTRADE maygreatly amplify the detrimental effect of the valuation bias. In some cases,general import statistics overstated the dutiable value of individual six-digit HSproducts by several hundred percent, or by billions of dollars. Automotiveequipment, machinery, electronics and energy products were often prone tothis statistical bias. Third, COMTRADE may incorrectly identify specific itemswhich are, or are not, subject to tariffs and other trade restrictions. This is dueto the fact that the U.N. records tabulate information on products entering acountry’s geographic territory, but may fail to record relevant information onthe nature and value of the goods actually clearing customs. This problemoccurs when imports experience significant transformation in foreign tradezones and then clear customs under a different HTS code than that recorded inCOMTRADE.7 Another possible cause is that the processed products were for-warded to final destinations in third countries and did not clear U.S. customsin any form. As a result, COMTRADE may both seriously overstate and under-state dutiable import values.
6. A further problem is that COMTRADE is often compiled at too high a level of aggregation to be
accurately used for tariff analyses and/or projections. Some six digit HS products (the lowest level of
detail available in COMTRADE) may contain multiple tariff lines having widely divergent import
duties. As an example, the six-digit HS product 610439 (women’s suits) has two tariff lines with duties
of 0.0 and 24.0 percent. The average of these duties (12.0 percent) would not accurately reflect the level
of protection afforded either product. Similarly, HS product 640199 (waterproof footwear) contains
four line items with duties ranging from 0 to 39.5 percent. These are not extreme outliers as HS code
210690 (other edible food preparations) incorporates 42 individual tariff line products.
7. The relative magnitude of tariffs on production inputs and the processed product provides a
useful indicator of where the largest COMTRADE-IDB data discrepancies may occur. As noted, firms
operating in FTZs have the option of declaring imports of either the production inputs or the final good
on customs vouchers when the item is transferred to the domestic market for consumption. In situations
where tariffs are relatively high on the inputs, an incentive would exist to declare imports of the
fabricated product to customs (which would be reported to the IDB), while COMTRADE would record
statistics on the unprocessed components initially imported into the zone.
Yeats 553
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Each of these factors is of major importance by itself. However, taken to-gether, the combined biases can reach magnitudes that clearly preclude the le-gitimate use of unadjusted COMTRADE data for trade projections andnegotiations. For example, recent USITC statistics report FTZ imports of $285billion that were exempt from tariffs. To put this value in perspective, theseimports were only slightly lower than the total combined customs value of U.S.imports from Japan, Germany, and the United Kingdom. In addition, transportand insurance charges on all U.S. imports were $65.8 billion—a figure thatreflects the further overall COMTRADE bias associated with tabulation ofcost-insurance-freight trade values.
Third, statutory regulations exempt certain imports from tariffs. These includeall U.S. government imports, imports for the treatment of specific medical pro-blems, and imports by overseas territories—all of which were about $24 billion.Altogether, these combined biases totaled $366 billion, or almost 20 percent ofthe customs value of all United States imports. However, as previously noted, therelative importance of the general trade bias may be larger in other countries(particularly those in East Asia) where international production sharing is prac-ticed more extensively than in the United States. These biases would normally beincorporated in COMTRADE data given the U.N. recommendation that tradestatistics be tabulated using general reporting practices:
“The general trade system provides a more comprehensive recording of theexternal trade flows than does the special system. It is recommended, therefore,that countries use the general system for compilation of their international mer-chandise trade statistics” (U.N., Department of Economic and Social Affairs(ESA/STAT/AC.137.5, p. 30).
The key point that follows is that analyses of tariffs and other trade barriersshould ideally utilize tariff line level import statistics compiled on the same valu-ation base employed for assessing these duties. Furthermore, the data must accur-ately account for specific exemptions like those normally afforded foreign tradezones or government entities, as well as for country specific exemptions asso-ciated with preferences or the withholding of “most-favored-nation” trade status.Brenton and others (2009) presents a useful illustration of the nature of theserequired adjustments within the context of formulating national structural adjust-ment policies. The United Nations (2007) provides comprehensive informationon the trade compilation practices of about 40 countries, which should also beuseful for identifying required adjustments. As this study shows, a failure to prop-erly account for these factors may adversely influence a country’s strategies inmultilateral negotiations or the formulation of national trade policies.
RE F E R E N C E S
Baldwin, Robert, and Tracy Murray. 1977. “MFN Tariff Reductions and Developing Country Benefits
Under the GSP.” The Economic Journal 87(March): 30–46.
554 T H E W O R L D B A N K E C O N O M I C R E V I E W
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
Brenton, Paul, et al. 2009. “Assessing the Adjustment Implications of Trade Policy Changes Using
TRIST.” World Bank. http://siteresources.worldbank.org.
Curtis, John, and Shenjie Chen. 2003. “Transport Costs and Changes in Canada’s Trade Pattern.”
World Economy 26(July): 975–91.
Conlon, Richard. 1982. “Transport Costs and Tariff Protection of Australian Manufacturing.”
Economic Record, 58(March): 73–81.
Laird, Sam, and Alexander Yeats. 1986. “The UNCTAD Trade Policy Simulation Model.” United
Nations Conference on Trade and Development. Geneva.
Lloyd, Peter. 1976. “Transport Costs on Indonesia’s Exports: The Australian Case.” Bulletin of
Indonesian Economic Studies 12(November): 116–20.
Hummels, David. 2007. “Transportation Costs and International Trade in the Second Era of
Globalization.” Journal of Economic Perspectives 21(3): 131–54.
MacLeod, Ian. 2000. "Foreign Trade Zones." International Trade Administration. http://ia.ita.doc.gov.
Ng, Francis, and Alexander Yeats. 1999. “Production Sharing in East Asia: Who Does What for Whom
and Why?” World Bank Policy Research Paper 2197. Washington, D.C.
Pomfret, Richard, and Patricia Sourdin. 2010. “Why Do Trade Costs Vary?” Review of World
Economics 146(4): 709–30.
United Nations Department of Economic and Social Affairs. 2007. ESA/STAT/AC.137.5. “International
Merchandise Trade Statistics: Concepts and Definitions.” New York.
United Nations Statistical Office. “International Merchandise Trade Statistics: National Compilation
and Reporting Practices.” http://UN.org/UNSD/tradeport/introduction_mm.asp.
World Bank and UNCTAD. 1987. “SMART – Software for Market Analysis and Restrictions to
Trade.” Washington, D.C.
Yeats 555
at International Monetary Fund on January 30, 2013
http://wber.oxfordjournals.org/
Dow
nloaded from
THE WORLD BANKECONOMIC REVIEW
editorsAlain de Janvry and Elisabeth Sadoulet, University of California at Berkeley
assistant to the editor Marja Kuiper
editorial boardHarold H. Alderman, World Bank (retired)Chong-En Bai, Tsinghua University, ChinaPranab K. Bardhan, University of California,
BerkeleyThorsten Beck, Tilburg University,
Netherlands Johannes van Biesebroeck, K.U. Leuven,
Belgium Maureen Cropper, University of Maryland,
USAAsli Demirgüç-Kunt, World Bank Jean-Jacques Dethier, World BankQuy-Toan Do, World BankFrédéric Docquier, Catholic University of
Louvain, BelgiumEliana La Ferrara, Università Bocconi, ItalyFrancisco H. G. Ferreira, World BankAugustin Kwasi Fosu, United Nations
University, WIDER, Finland
Caroline Freund, World BankPaul Glewwe, University of Minnesota,
USAPhilip E. Keefer, World BankNorman V. Loayza, World BankWilliam F. Maloney, World BankDavid J. McKenzie, World BankJaime de Melo, University of GenevaUgo Panizza, UNCTADNina Pavcnik, Dartmouth College, USAVijayendra Rao, World BankMartin Ravallion, World BankJaime Saavedra-Chanduvi, World BankClaudia Paz Sepúlveda, World BankJonathan Temple, University of Bristol,
UKDominique Van De Walle, World BankChristopher M. Woodruff, University of
California, San Diego
The World Bank Economic Review is a professional journal used for the dissemination of research indevelopment economics broadly relevant to the development profession and to the World Bank inpursuing its development mandate. It is directed to an international readership among economists andsocial scientists in government, business, international agencies, universities, and development researchinstitutions. The Review seeks to provide the most current and best research in the field of quantita-tive development policy analysis, emphasizing policy relevance and operational aspects of economics,rather than primarily theoretical and methodological issues. Consistency with World Bank policy playsno role in the selection of articles.
The Review is managed by one or two independent editors selected for their academic excellence inthe field of development economics and policy.The editors are assisted by an editorial board composedin equal parts of scholars internal and external to the World Bank. World Bank staff and outsideresearchers are equally invited to submit their research papers to the Review.
For more information, please visit the Web sites of the Economic Review at Oxford University Pressat www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals.
Instructions for authors wishing to submit articles are available online at www.wber.oxfordjournals.org.Please direct all editorial correspondence to the Editor at [email protected].
Forthcoming papers in
• The Impact of the Global Food Crisis on Self-Assessed Food SecurityDerek D. Headey
• How Is the Liberalization of Food Markets Progressing? MarketIntegration and Transaction Costs in Subsistence EconomiesWouter Zant
• Decomposing the Labor Market Earnings Inequality:The Public and Private Sectors in Vietnam, 1993–2006Clément Imbert
• Chinese Trade Reforms, Market Access and Foreign Competition:the Patterns of French ExportersMaria Bas and Pamela Bombarda
• Firms Operating under Electricity Constraints in Developing CountriesPhilippe Alby, Jean-Jacques Dethier & Stéphane Straub
• Antidumping, Retaliation Threats, and Export PricesVeysel Avsar
• Information and Participation in Social ProgramsDavid Coady, César Martinelli, and Susan W. Parker
THE WORLD BANKECONOMIC REVIEW
THE WORLD BANKECONOMIC REVIEW
Is There a Metropolitan Bias? The relationship between povertyand city size in a selection of developing countries
Céline Ferré, Francisco H.G. Ferreira, and Peter Lanjouw
Impact of SMS-Based Agricultural Information on Indian Farmers Marcel Fafchamps and Bart Minten
Crises, Food Prices, and the Income Elasticity ofMicronutrients:Estimates from Indonesia
Emmanuel Skoufias, Sailesh Tiwari, and Hassan Zaman
Economic Geography and Economic Development in Sub-Saharan Africa
Maarten Bosker and Harry Garretsen
The Decision to Import Capital Goods in India: Firms’ FinancialFactors Matter
Maria Bas and Antoine Berthou
Coffee Market Liberalisation and the Implications for Producers in Brazil, Guatemala and India
Bill Russell, Sushil Mohan, and Anindya Banerjee
Implications of COMTRADE Compilation Practices for Trade Barrier Analyses and Negotiations
Alexander J. Yeats
Volume 26 • 2012 • Number 3
www.wber.oxfordjournals.org
THE WORLD BANK1818 H Street, NWWashington, DC 20433, USAWorld Wide Web: http://www.worldbank.org/E-mail: [email protected]
2V
olume 26 • N
umber 3 • 2012
TH
E W
OR
LD
BA
NK
EC
ON
OM
IC R
EV
IEW
Pages 351–555
ISSN 0258-6770 (PRINT) ISSN 1564-698X (ONLINE)