Estimating Labor Demand Elasticities from Asymmetric Product Market Shocks †

Elias Einiö*

October 2018

Abstract We identify a plant-level labor demand model from variation in local wage rates induced by product market shocks that are asymmetric across product fields and local labor markets. Our analysis exploits exceptionally rich plant-level panel data and the abrupt decline in ex-ports of specific products in Finland due to the collapse of the Soviet Union. In our setting, the rapid downsizing of the Soviet-dependent industry induces plausibly exogenous varia-tion in the local available supply of labor faced by plants which are not dependent on Soviet import demand. Instrumental variables estimates of short-run own-wage elasticities are small (between –0.1 and –0.2) and substantially overestimated in conventional least squares approaches. The findings suggest that short-term policies reducing labor costs are unlikely to be effective in preventing short-run employment losses during economic downturns. JEL classification: F16; J23; J24; O33. Keywords: Instrumental variables; labor demand functions; natural experiment; plant-level data; structure of labor demand.

† Thanks are due to Manuel Bagues, Peter Fredriksson, Kari Hämäläinen, Steve Machin, Kristiina Huttunen, Andrea

Ichino, Tuomas Pekkarinen, Steve Pischke, John Van Reenen, and seminar participants at the XIII Brucchi Luchino Workshop, EEA-ESEM, HECER, London School of Economics, RES Annual Congress, and 10th Nordic Summer Institute in Labor Economics for their helpful comments and suggestions, and Matti Mitrunen and Jaakko Nelimarkka for the excellent research assistance. The data used in this article are confidential but the authors’ access is not exclu-sive. Funding from the Academy of Finland (grant 134057) is gratefully acknowledged. The authors declare that they have no relevant or material financial interests that relate to the research described in this paper.

* VATT and CEP/LSE. Email: elias.einio@vatt.fi.

1

1 Introduction

How much do firms adjust employment in response to changes in unit labor costs? This

question is of great importance for many labor market policies. Firms’ employment responses

to changes in unit labor costs are especially crucial for temporary policies that aim to improve

employment in the short run by reducing labor costs paid by the employer. Such policies have

gained increasing popularity in many countries during the last recession.1 Labor demand

elasticities are also required to quantify changes in the structure of labor demand (e.g., Katz

and Murphy, 1992; Hamermesh, 1993) and determine optimal tax and labor market policies

(e.g., Lee and Saez, 2012; Kroft et al., 2015; Heathcote, Storesletten, and Violante, 2017).

Despite the tremendous importance of labor demand elasticities in the analysis of labor mar-

kets, surprisingly little work has been undertaken in order to develop causal empirical strate-

gies to identify them.

The main challenge in identifying labor demand elasticities from observed prices and

quantities arises from the simultaneity of labor demand and supply. Although this problem

has been well-acknowledged in the literature (e.g., Frisch, 1933), addressing it has proven to

be difficult because credible instrumental variables that induce exogenous variation in wage

rates are scarce (Acemoglu, Autor, and Lyle, 2004; Lichter, Peichl, and Siegloch, 2015). In

the context of plant- and industry-level labor demand functions, a major concern is that ob-

served variation in wage rates and labor quantities are partly caused by unobserved shocks to

technology. Another key concern is measurement error that tends to attenuate OLS estimates

and which may be a specific problem in disaggregated industry and plant data.

In order to address these challenges, we estimate plant-level labor demand elasticities by

exploiting plausibly exogenous variation in local wage rates induced by unexpected product

demand shocks which are asymmetric across product fields and local labor markets. The idea

is quite simple. If some plants in a local economy face large unexpected shocks to their prod-

uct demand, they will adjust their output and workforce accordingly. For example, a negative

shock to product demand would lead these plants to downsize production and release work-

1 See, e.g., European Commission (2010).

2

force. If labor is not completely mobile and capital stock does not adjust completely in the

short run, the released workers will increase the local available supply of labor. Differences

in the local magnitude of such shocks can induce variation in local wage rates. Such variation

is plausibly exogenous for plants that are not directly affected by the product demand shock

(that is, plants which produce for unaffected product markets). It is important to exclude

plants that are directly affected by the product demand shock from the estimation, because

large product demand shocks may trigger endogenous technology investment and skill up-

grading, as shown by Bustos (2011).

Our empirical application employs the abrupt collapse of exports in Finland due to the col-

lapse of the Soviet Union. The trade between Finland and Soviet Union was based on bilat-

eral trade agreements that defined trade in specific product categories. The abolition of the

trade agreement caused a decline of output among plants that produced in these specific

categories. The fact that the products exported to the Soviet Union covered a limited fraction

of all products manufactured in Finland meant that there were also a significant number of

plants that were not dependent on Soviet import demand. Moreover, because the relative size

of the Soviet-dependent industry varied geographically, the magnitude of the shock due to its

collapse varied across local labor markets. Our identification strategy exploits this geographic

variation in the local shock faced by plants that were not exporting to the Soviet Union. In

this setting, these plants faced differential local labor supply shocks due to the downsizing of

the neighboring Soviet-dependent industry which depended on the historic size of local Sovi-

et-dependent production.

We employ exceptionally rich plant-level panel data on inputs, outputs, and hourly labor

costs by occupation from the census of manufacturing plants in Finland. We observe plant-

level outputs by detailed commodity categories before the collapse of Soviet trade. This

allows us to identify plants that were not exporting to the Soviet Union and construct a meas-

ure of historic local exposure to Soviet trade. We start by showing that the collapse of Soviet

import demand has a large negative effect on output in areas that have historically large

Soviet-dependent industry. We also show that the downsizing of the Soviet-dependent indus-

3

try has only a small impact on local manufacturing employment because it induces significant

re-allocation of labor, within local labor markets, toward plants that are less exposed to Sovi-

et import demand. Moreover, we find a significant divergence in unit labor costs for plants

that are not exporting to the Soviet Union between areas which are historically more and less

exposed to Soviet trade. Hence, the collapse of Soviet trade generates a situation where simi-

lar plants that are not dependent on Soviet import demand face a differential unit labor cost

shock because of the differential historic Soviet specialization of their neighboring industry.

We identify a plant-level labor demand model from this plausibly exogenous variation in unit

labor costs.2

We estimate the model with a variety of IV methods. We provide two-stage least squares

(TSLS) estimates for single-instrument specifications and for over-identified specifications

accounting for the potential variation in the first-stage impacts of the instrument across the

post-collapse years. We also show that our results are robust against a number of potential

sources of bias; our analysis indicates that input supply from our target plants to Soviet-

dependent plants, changes in agglomeration benefits, and shifts in within-group skill distribu-

tions among our target plants are unlikely to drive the results.

We provide two key sets of results. The first provides estimates of plant-level labor de-

mand elasticities. The results suggest that the magnitude of short-run labor demand elastici-

ties is substantially overestimated by conventional OLS approaches. Our preferred IV esti-

mates for plant-level production and non-production labor own-wage elasticities are –0.1 and

–0.2, respectively (the elasticity of substitution between these two worker groups is around

0.3). The magnitude of the corresponding OLS estimates is twice as large. One implication of

the small labor demand responses to changes in wage rates is that short-term policies reduc-

ing labor costs have only a small impact on employment in the short run. The results suggest

that the effectiveness of such policies in reducing short-run employment losses during reces-

sions might be substantially overestimated in conventional OLS approaches. The second set 2 For previous work exploiting spatial variation in historic local industry specialization, see, e.g., Topalova

(2010), who examines the impact of trade liberalization on poverty in India. For previous studies using empiri-cal strategies based on the rise and fall of the Soviet regime and former Eastern Bloc, see, e.g., Glitz (2012), Borjas and Doran (2012), and Falck et al. (2013).

4

of results uses the estimated model to recover within-plant change in the structure of labor

demand. The results indicate a sharp within-plant labor demand shift against production

workers in the 2000s. The results also indicate a considerable divergence in the structure of

labor demand between industries in this period.

We contribute to literature estimating labor demand elasticities and changes in the struc-

ture of labor demand. Our study is closely related to work estimating industry labor demand

models (e.g., Haskel and Slaughter, 2002; Baltagi and Rich, 2005). One of our main contribu-

tions is to advance this literature by proposing an estimation strategy to identify plant-level

labor demand elasticities that exploits local labor market spillovers between industries which

experience abrupt product demand shocks and those which do not. Moreover, we provide (to

our knowledge) the first quasi-experimental estimates of micro-level labor demand elastici-

ties.

Our study is also closely related to a few previous studies that have used quasi-

experimental variation in wage rates in order to estimate state-level labor demand elasticities.

Acemoglu et al. (2004) estimate labor demand elasticities from labor supply shocks stem-

ming from increased labor market participation of women during World War I, while

Hamermesh and Trejo (2000) estimate elasticities of demand for overtime hours from chang-

es in overtime premiums in California in the 1970s and early 1980s. Another important con-

tribution is the study by Ciccone and Peri (2005), who use an IV strategy based on local labor

supply shocks induced by changes in child labor and compulsory school attendance laws to

identify a state-level labor demand model. Using data for the period 1950–1990, they find a

technology-induced demand shift toward more educated workers.3 Our study complements

and extends this strand of research by estimating plant-level labor demand elasticities in a

3 A related strand of research has examined the impacts of labor supply shocks due to migration and immi-

gration (e.g., Angrist, 1996; Borjas, 2003; Glitz, 2012). See also Angrist (1995), who shows that increasing educational attainment reduced skill premiums among Palestinians in the 1980s. For aggregate-level studies, see, e.g., Katz and Murphy (1992), who use time series data to estimate labor demand shifts in an aggregate labor supply and demand framework, Heckman, Lochner, and Taber (1998), who develop an empirically grounded dynamic overlapping-generations general equilibrium model of skill formation that explains rising wage inequality as a consequence of skill-biased technical change, and Heathcote, Storesletten, and Violante (2010), who assess the implications of the rise in U.S. wage inequality for the macroeconomy and for welfare. For a comprehensive discussion and meta-analysis of the literature estimating own-wage elasticities of labor demand, see Lichter, Peichl, and Siegloch (2015).

5

quasi-experimental setting and by providing new evidence on changes in the structure of

labor demand. While our analysis is limited to labor demand shifts within plants, we believe

that our findings are important for understanding recent developments in the labor markets.

Our results suggest that the demand shift against production workers accelerated in the 2000s

when within-plant demand shifts accounted for around half of the overall decline in the pro-

duction labor cost share.

The work is organized as follows. The econometric framework for estimating plant-level

labor demand elasticities and shifts is presented in section 2. Section 3 presents the identifica-

tion strategy based on asymmetric product demand shocks. Section 4 presents the results and

several robustness checks verifying the identification strategy. The final section concludes.

2 Econometric Framework

2.1 Manufacturing Labor Input

To guide our work, we start by documenting key worker groups in manufacturing. Column 1

in table 1 shows wage earnings shares for three major occupational categories in manufactur-

ing: production, professional, and clerical and service workers. Manufacturing labor input is

highly concentrated in the first two categories with production workers accounting for 56.5%

and professional workers for 36.6% of it, whereas clerical and service occupations account

for only 7.1% of it. The rest of the table documents the job task content in these occupations

based on detailed occupational job task data from Acemoglu and Autor (2011). As expected,

production work is highly routine manual task-intensive. Production workers also score

highly in the non-routine manual physical task dimension. On the other hand, work by pro-

fessionals is highly non-routine cognitive task-intensive. The small clerical and service work-

er category scores highly in the routine cognitive task dimension and has relatively low

scores across other tasks.

These observations indicate that manufacturing work is characterized by a clear occupa-

tion structure, with production and professional occupations accounting for around 93% of

labor input. These worker groups are also clearly separated by the key task dimensions, with

6

the former group specialized in routine manual and non-routine physical manual tasks and the

latter group specialized in non-routine cognitive tasks.

2.2 Plant-Level Labor Demand Model

Our econometric model is based on the translog cost function framework.4 We consider a

manufacturing plant producing output by combining production labor, non-production

labor, and capital. We assume that non-production labor input is produced by professional

workers and each unit of it includes a fixed proportion of clerical and service labor.5 The

manufacturer minimizes variable costs given the unit cost of production labor and non-

production labor . With quasi-fixed capital , the plant-level variable cost function is

( , , , , ) = min, + : , ∈ ( , , ′ )

where is the production labor input; is the non-production labor input; and ( , , ′ ) is

the input requirement set where is a vector of productivity factors ( ). Assuming translog

4 For prior related work using the translog framework, see e.g. Machin and Van Reenen (1998), Ciccone and

Peri (2005), and Baltagi and Rich (2005). 5 Formally, this corresponds to a nested model where non-production labor input is produced with a Leontief

technology. The assumption of perfect complementarity between professional and clerical and service labor is motivated by the observation that clerical and service workers are mainly in occupations providing support and assistance to professional workers (appendix table A1). It is unlikely to affect our results because the share of clerical and service workers in manufacturing is low.

Table 1: Manufacturing Occupations and Tasks

Job Task Indices

Routine Non-Routine

Occupation

Wage Earnings

Share, 1995 (%) Manual

Cogni-tive

Cognitive Analytic

Cognitive Interper-

sonal Manual Physical

Manual Interper-

sonal

Production Worker 56.5 1.27 0.20 -0.28 -0.55 1.13 -1.09 Non-Production Worker 43.5 -0.32 0.11 0.68 0.20 -0.40 -0.07 Professional 36.6 -0.35 0.07 0.91 0.34 -0.40 -0.04 Clerical or Service Worker 7.1 -0.18 0.34 -0.54 -0.52 -0.34 -0.23

Notes: The wage earnings share is calculated from the FLEED data. Job task indices are based on data from Acemoglu and Autor (2011). Appendix A.1 provides details of the data construction. Appendix table A1 reports the wage earnings share and job task indices by 2-digit occupation.

7

costs and applying Shephard’s lemma yields the input cost share equation:

( ) = = ln( ) + ln( ) + ln( ) + ln( ) + ( ) ln ( ) , (1)

where the production labor cost share is a function of primary input demand variables: the

labor unit costs, capital stock, and output. In this model, the substitution elasticity

between production and non-production labor is = (1 − )⁄ , where = [− + (1 − )]⁄ is the cross-wage elasticity of labor demand. The own-wage

elasticities are = ( + − )⁄ and = (− + (1 − ) − 1 + ) (1 − )⁄ for

production and non-production labor, respectively (Hamermesh, 1993).

The last term on the right-hand side of equation (1) is the joint contribution of the produc-

tivity factors that affect the production labor input share. Rising ( ) reduces the relative

demand for production labor if ( ) < 0 and increases it if ( ) > 0. For example, if ( ) is a

measure of the level of technology, ( ) < 0 implies that technological development is biased

against production workers. may also include other non-primary input demand factors,

such as inputs in foreign affiliates. For example, if ( ) is a measure of international out-

sourcing, ( ) < 0 implies that the relative demand for production labor declines if interna-

tional outsourcing increases.

To empirically implement equation (1), we impose homogeneity of degree one in prices ( + = 0) and constant returns to scale ( + = 0).6 We allow for unobserved heter-

ogeneity across plants and include plant fixed effects to account for it. Moreover, we allow

for ( ) and ( ) to vary across industry and year. The empirical cost share equation for plant

in year and industry is

= + ln / + ln / + ( ) ln ( ) + , (2)

where is the error term. This specification allows for differences in the relative unit labor

cost / across local labor markets . Such differences may arise when local labor

markets are sufficiently isolated so that the impacts on wage rates of asymmetric shocks to 6 These assumptions can be tested in the translog framework. The analysis below indicates that they are valid

in our data.

8

the local available supply of labor are not completely offset by migration.

In equation (2), all demand shifters ( ) are not observed, and even if they were, identify-

ing causal effects of many demand shifters would require valid instruments for each of them.

Instead of trying to estimate each ( ), we account for the demand shift term by setting ∑ ( ) ln ( ) = . Taking first differences from − 1 to to eliminate plant fixed ef-

fects yields the key estimating equation:

∆ = ∆ ln / + ∆ ln / + + ∆ . (3)

Because this regression equation controls for capital stock and is based on short-term varia-

tion in the relative production labor unit cost, an estimate of allows us to recover short-run

labor demand elasticities. = − , represents within-plant changes in the relative

demand for production labor in industry from year − 1 to . Summing this term across

years yields the industry-specific demand shift index ( ) = ∑ for industry in year

with respect to a baseline year . It represents the joint effect of the productivity and other

non-primary labor demand factors on the structure of labor input and provides a measure of

the average within-plant change in the structure of labor demand in industry .

3 Empirical Strategy Based on Asymmetric Product Market Shocks

The main econometric challenge in identifying equation (3) arises from the simultaneity of

the labor cost share and relative production labor unit cost / . A potential

source of confounding variation is a correlated production labor-biased technology shock. For

instance, a skill-biased technology shock would be expected to reduce the relative wage of

production workers, who work typically in occupations with lower skill requirement, and to

simultaneously reduce the production labor cost share, inducing positive bias in the OLS

estimate of . Another potential source of bias is measurement error in hourly wage rates,

which tends to attenuate the OLS estimate of toward zero. Such bias may be more relevant

for our estimations using plant-level wages than for estimations using more aggregate-level

measures of wages (e.g., as in a study by Ciccone and Peri (2005), who use state-level data).

9

can be identified if exogenous variation in unit labor costs faced by plant is available.

We employ exogenous variation in local wage rates induced by product demand shocks that

are asymmetric across product fields and local labor markets. The identification strategy is

based on the following idea. Suppose that some plants or industries in a local economy face

an unexpected shock to their product demand. As a result, they adjust their output and work-

force accordingly. For example, a negative shock to product demand would lead them to

downsize production and release workforce. If labor is not completely mobile and capital

stock does not adjust completely in the short run, the released workers will increase the local

available supply of labor. Differences in the local magnitude of such shocks can induce varia-

tion in local wage rates. Such variation is plausibly exogenous for plants that are not directly

affected by the product demand shock (that is, plants which produce for unaffected product

markets). It is important to exclude plants that are directly affected by the product demand

shock from the estimation, because large product demand shocks may trigger endogenous

technology investment and skill upgrading, as shown by Bustos (2011). We implement this

empirical strategy in an application exploiting asymmetric product market shocks from the

abrupt collapse of the Soviet-dependent industry in Finland.

3.1 Trade Agreement between Finland and Soviet Union

The commodity structure of trade between Finland and Soviet Union was determined by a

bilateral trade agreement. Finnish exports to the Soviet Union were concentrated in relatively

few commodity categories. In 1990, the last year of the trade agreement, 250 commodities at

the 6-digit level of the Harmonized System (HS) classification covered around 91% of ex-

ports from Finland to the Soviet Union. Appendix table B1 displays the top 15 commodities

which covered around 34% of all Finnish exports to the Soviet Union in 1990.

The trade agreement between Finland and Soviet Union was unexpectedly abolished in

December 1990. As a result, the value of Soviet exports fell from 2.52 billion euro in 1990 to

0.90 billion euro in 1991 – a drop corresponding to around 2.7% of manufacturing output

(figure 1). The collapse of Soviet import demand led to a rapid downsizing of production in

10

the Soviet-dependent industry. Because the Soviet-dependent industry was geographically

scattered across the country, the shock from its collapse was asymmetric across local labor

markets. Furthermore, the fact that the products exported to the Soviet Union covered a lim-

ited fraction of all products manufactured in Finland meant that there were also a significant

number of plants that were not directly dependent on Soviet import demand. Our identifica-

tion strategy exploits geographic variation in the unit labor cost shock faced by these plants,

stemming from the differential historic Soviet specialization of their neighboring industry.

3.2 Data Sources and Plant-Level Measure of Soviet Specialization

Our main data source is the Longitudinal Database of Plants in Finnish Manufacturing

(LDPM) provided by Statistics Finland. The LDPM is based on the annual Industrial Struc-

tures survey, which includes all manufacturing plants with at least 20 employees. These

plants account for around 82% of aggregate manufacturing output in our observation period

1980–2008. The LDPM provides detailed information on annual outputs and inputs, includ-

ing value added, capital stock, and labor costs for production and non-production workers.

Importantly, the data provide information on hours worked by these worker groups, which

facilitates the calculation of average plant-level hourly labor costs for production and non-

Figure 1: Aggregate Exports to the Former Soviet Union from Finland, 1987–2002 Notes: Until 1990, the series cover exports to the Soviet Union from Finland. From 1991 onwards, the series cover exports to the same geographic area as in 1990: In 1991, they include exports to the Soviet Union, Es-tonia, Latvia, and Lithuania; from 1992 onwards, they include exports to Armenia, Azerbaijan, Belarus, Es-tonia, Lithuania, Latvia, Russia, Uzbekistan, Tajikistan, Turkmenistan, Kyrgyz Republic, Kazakhstan, and Ukraine. Data on exports are from the OECD ITCS database. Data on manufacturing output are from the Annual National Accounts of Finland provided by Statistics Finland.

0.0%

1.0%

2.0%

3.0%

4.0%

5.0%

6.0%

7.0%

0

1,000

2,000

3,000

4,000

5,000

6,000

7,000

Exports Exports / Manufacturing Output (right axis)

Million euro (2000 prices)

Fenno-Soviet trade agreement abolished in December 1990

11

production workers. It also provides information on the location of the plant at the municipal

level.

The LDPM data are amended with the plant-level 1988 Commodity Statistics survey

(CSS) and OECD International Trade by Commodity Statistics data (ITCS). The CSS covers

around 91% of aggregate LDPM output and provides information on plant-level outputs and

inputs by 6-digit HS commodity. The CSS data are linked to the LDPM with unique plant

codes. To identify plants that are not dependent on Soviet import demand, we construct a

measure of 1990 plant-level Soviet specialization:

, = ∑ ,, , (4)

where , is total national exports of 6-digit HS commodity from Finland to the

Soviet Union in 1990 drawn from the ITCS and is plant ’s share of total national output

of commodity before the abolition of the trade agreement.7 The numerator is the sum of

the plant’s Soviet exports over the commodity categories, predicted by the pre-collapse com-

modity output structure. Scaling by plant-level output in 1990 yields a measure of plant-level

exposure to Soviet trade. In our baseline specification, we define producers which are not

dependent on Soviet import demand as plants with , < 0.001. That is, their output

share of predicted Soviet exports is 0.1% or lower.8 Around 28% of plants in the LDPM fulfil

this criterion in 1990. These plants constitute our baseline estimation sample. Appendix A

provides further details of the data construction and summary statistics for samples relevant

for our analysis.

3.3 Asymmetric Shocks across Local Labor Markets

Because the Soviet-dependent industry was geographically scattered across the country, the

7 At the time of data construction, Statistics Finland was unable to provide access to the 1989 and 1990 CSS

files due to licencing restrictions, and hence we use CSS data for the latest available pre-collapse year 1988. We define = , / , , where , is the output of commodity by plant in 1988 and , is total national output of commodity in 1988.

8 We show below that our results are robust when we choose different thresholds. We prefer the threshold of 0.1% because it sets the predicted share of Soviet exports to a very low level but still maintains a good sample size. For instance, setting the threshold to 0.025% has little impact on the results but it more than halves the number of observations in our baseline sample.

12

effects of the abolition of the trade agreement varied substantially across regions. We measure

1990 local Soviet specialization by

, = ∑ ∑ ,∈ ( )∑ ,∈ ( ) = ∑ , ,∈ ( )∑ ,∈ ( ) , (5)

where ( ) denotes the set of plants in the local labor market . This measure is the sum of

predicted Soviet exports by local plants divided by the sum of their outputs. Figure 2 displays

the spatial variation in 1990 local Soviet specialization by municipality.9 It shows that Soviet-

dependent production was widespread across the country before the collapse of the Soviet

Union. The figure also displays the spatial distribution of plants which were not dependent on

Soviet import demand. Importantly, we observe these plants in areas with both high and low

Soviet specialization.10

Figure 3 compares manufacturing output and employment between municipalities in the

lowest and highest LSS quintiles. Panel A shows that output growth is very similar before the

collapse of Soviet trade in these areas but diverges dramatically as the trade agreement is

abolished with an around 10% drop in output in the most exposed areas, whereas the least

exposed areas are very little affected. Despite the large divergence in output, little differences

emerge in terms of employment (panel B). A potential explanation for this is that the work-

force released by the downsizing Soviet-dependent industry was to a large extent re-

employed by local plants which were not dependent on Soviet import demand. Results in

appendix table B2 provide support for this adjustment mechanism. The table reports estimates

from regressions of annual plant-level employment growth on plant-level Soviet specializa-

tion ( , ) conditioning on municipality fixed effects. The results indicate large and

statistically significant reallocation of labor from more to less exposed plants within local

9 Municipalities have mean acreage of 749 km2. In 1990, there were 460 municipalities. 10 Summary statistics for plants in municipalities above and below the median of 1990 local Soviet Speciali-

zation are provided in appendix table A3.

No199

1

Sm

pe

Soviet Sp

Observa

Ad

otes: Soviet spe90, predicted by

s

n

r

t

t

mallest Municipalitie

rcentile of 1990 out

pecialization (Q

ations in the Es

Norwegian Se

B

dministrative Region

Figure 2: So

ecialization is by the 1988 outp

st

nd

rd

h

h

es (Below the 5th

tput)

Quintiles)

stimation Sam

ea

Baltic Sea

n (‘Maakunta’)

oviet Speciali

based on equatiput structure, to

13

ple

ization by M

ion (5). It is th the municipali

Municipality i

he fraction of aity’s 1990 outpu

in Finland, 1

a municipality’sut.

100km

990

s Soviet exportts in

14

labor markets in 1991, while we detect no significant reallocation of labor before the collapse

of Soviet trade. The table also shows estimates by worker group. The estimates suggest that

the initial relative labor adjustment at the margin is around 0.572 (=0.079/(0.079+0.059)) in

terms of the production worker employment share, while the average production worker

employment share among plants in our estimation sample is around 0.782. This suggests that

the relative local labor supply shock induced by the collapse of the Soviet-dependent industry

was more intensive in non-production labor. Therefore, the shock could be expected to in-

crease the relative wage of production workers (or, analogously, reduce the relative wage of

non-production workers).

3.4 Instrument and First Stage

In order to identify , we exploit variation in the shock on the relative wage of production

workers stemming from the collapse of Soviet trade and variation in the historic local size of

the Soviet dependent industry. We implement this strategy by using the 1990 local Soviet

specialization as an instrument for the relative production labor unit cost and estimate equa-

tion (3) with a TSLS procedure based on the following first stage:

6080

100

120

140

160

180

200

Empl

oym

ent (

1990

=100

)

1980 1985 1990 1995 2000Year

Low Soviet-Import Dependence, 1st Quintile

High Soviet-Import Dependence, 5th Quintile

6080

100

120

140

160

180

200

Rea

l Out

put (

1990

=100

)

1980 1985 1990 1995 2000Year

Low Soviet-Import Dependence, 1st Quintile

High Soviet-Import Dependence, 5th Quintile

A. Real Output B. Employment

High Soviet Specialization, 5th Quintile Low Soviet Specialization, 1st Quintile

Figure 3: Output and Employment in Municipalities with Low and High Soviet Specialization

Notes: This figure compares manufacturing output and employment between municipalities in the lowest and highest 1990 local Soviet specialization quintiles. The 20th and 80th percentiles of 1990 local Soviet specialization are 0.54% and 6.14%, respectively. Real output and employment are calculated from the LDPM.

15

∆ ln / = , + ∆ ln / + + ∆ , (6)

where , is a function of the 1990 local Soviet specialization and is a year in the

post-collapse period 1991–1995. Our most simple specification uses a single instrument,

, = ln , . This specification assumes that the impact of the instrument

is equivalent across all post-collapse years. We can allow for differential impacts of the in-

strument across years by interacting it with year dummies. This might be important because

annual changes in local relative wage rates induced by the instrument are unlikely to be

constant over time. Fully interacting the log of 1990 local Soviet specialization with year

dummies yields five instruments. In the most general first-stage specification, we also allow

for the potential non-linear effects of the log of 1990 local Soviet specialization:

, = ln , ∙ ( = )+ ln , ∙ ( = ) . (7)

Here ( = ) is an indicator function equal to one in the post-collapse year and zero

elsewhere.

The over-identified models use many instruments. This may raise concerns about weak in-

strument bias. Fortunately, it turns out that the just- and over-identified estimators recover

very similar point estimates. This indicates that weak instruments are not a major source of

bias in the over-identified models. The over-identified models turn out to have additional

value in our analysis because they provide higher precision of the estimation due to the better

fit of the first-stage. We also follow a standard procedure for over-identified models and

compare the TSLS estimates to the limited information maximum likelihood (LIML) and

bias-corrected IV estimates (e.g., Stock, Wright, and Yogo, 2002; Hahn and Hausman, 2003;

Angrist and Pischke, 2009; Kolesár et al., 2015). Finding point estimates of a similar magni-

tude for the conventional TSLS and these alternative estimators provide further credence to

the validity of the over-identified models.11

11 We also test for weak identification (Kolesár, 2012; Kolesár et al., 2015).

16

4 Results

This section first presents the estimates of the parameters of the plant-level labor demand

model. We then address a number of potential robustness concerns. The last part of the sec-

tion presents within-plant changes in the relative demand for production labor implied by the

estimated model. Standard errors are corrected for clustering at the plant level in all plant-

level estimations.12

4.1 Parameters of the Labor Demand Model and Labor Demand Elasticities

Table 2 presents the main estimates of the parameters of the labor cost share equation (3). The

first column displays the OLS estimates. The OLS coefficient on the relative production labor

unit cost ( ) is 0.088 and implies a plant-level short-run elasticity of substitution between

production and non-production labor of 0.60 at the sample mean of the production labor cost

share (0.68). The corresponding own-wage elasticities are –0.28 and –0.47 for non-

production and production workers, respectively.

As discussed above, the OLS estimates may be confounded by several potential sources of

bias, including unobserved correlated technology shocks and measurement error in plant-

level hourly wage rates. To account for them, we estimate the TSLS model given by equa-

tions (3) and (6). Column 2 displays the TSLS and first-stage estimates for the just-identified

model using local Soviet specialization in 1990 as an instrument for the relative production

labor unit cost and data for the period 1991–1995. The first-stage coefficient on the instru-

ment is 0.007 and significant at the 5% risk level. This first-stage effect indicates that the

relative labor supply shock induced by the instrument was, on average, more intensive in

non-production labor, which is consistent with the results reported in section 3.3.

Despite the strong first-stage, the coefficient on the relative production labor unit cost in

the cost share equation is not significant at the conventional confidence levels (0.143 with a

standard error of 0.120). One potential explanation for this is that the single-instrument speci-

12 We also experimented with clustering standard errors by municipality and administrative region (maakun-

ta), which gave similar and in many cases slightly smaller standard errors compared to clustering at the plant level.

17

Tabl

e 2:

Par

amet

er E

stim

ates

of t

he L

abor

Cos

t Sha

re E

quat

ion

Estim

ator

: Co

st Sh

are

Equa

tion

Spec

ifica

tion:

(1)

OLS

Tr

anslo

g

(2)

TSLS

Tr

anslo

g

(3)

TSLS

Tr

anslo

g

(4)

TSLS

Tr

anslo

g

(5)

TSLS

Tr

anslo

g

(6)

TSLS

Tr

anslo

g

(7)

TSLS

Tr

anslo

g

(8)

LIM

L Tr

anslo

g

(9)

BTSL

S Tr

anslo

g

(10)

M

BTSL

S Tr

anslo

g

(11)

TS

LS

CES

0.08

8***

0.

143

0.16

7*

0.18

1*

0.21

5**

0.15

6**

0.14

6**

0.16

6**

0.18

5*

0.21

4 0.

288

(0.0

08)

(0.1

20)

(0.0

95)

(0.0

95)

(0.0

85)

(0.0

62)

(0.0

58)

(0.0

80)

(0.1

02)

(0.1

39)

(0.3

07)

-0

.008

***

-0.0

08**

* -0

.007

***

-0.0

08**

* -0

.008

***

-0.0

08**

* -0

.019

**

-0.0

19*

-0.0

19*

-0.0

18

-0.1

33**

(0

.002

) (0

.002

) (0

.002

) (0

.002

) (0

.002

) (0

.002

) (0

.009

) (0

.010

) (0

.011

) (0

.012

) (0

.060

)

1st S

tage

Lo

g

–

0.

007*

* 0.

010*

* 0.

010*

* S

ee a

ppen

dix

tabl

e B3

(0.0

03)

(0.0

03)

(0.0

03)

Lo

g

*199

4 Ye

ar D

umm

y

–

-0

.019

**

(0.0

09)

Obs

erva

tions

2,

881

2,

881

2,33

0 2,

881

2,88

1 2,

881

2,81

3 2,

813

2,81

3 2,

813

2,81

3

Addi

tiona

l Exc

lude

d In

strum

ents:

Log

*

Full

Year

Dum

mie

s N

o

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Log

*

Full

Year

Dum

mie

s N

o

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Log

Capi

tal I

nten

sity

in t-

2 N

o

No

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Angr

ist-P

ischk

e Fi

rst-S

tage

F-S

tatis

tics:

∆ln(/)

–

6.91

9.

01

4.50

3.

19

2.71

2.

53

2.53

2.

53

2.53

2.

42

∆ln( /)

–

– –

– –

– 6.

63

6.63

6.

63

6.63

5.

72

Te

st fo

r the

Val

idity

of I

nstru

men

ts

P-

valu

e –

–

– –

– –

– –

0.97

0.

97

–

Not

es: O

LS a

nd IV

esti

mat

es fo

r equ

atio

n (3

). Th

e sa

mpl

e in

clud

es p

lant

s tha

t wer

e no

t dep

ende

nt o

n So

viet

impo

rt de

man

d in

199

0. T

he o

utco

me

is th

e ch

ange

in th

e pr

oduc

tion

labo

r cos

t sha

re, e

xcep

t in

the

CES

spec

ifica

tion

in c

olum

n 11

, whe

re th

e ou

tcom

e is

the

log-

chan

ge in

the

ratio

of p

rodu

ctio

n an

d no

n-pr

oduc

tion

wor

ker e

mpl

oym

ent.

For a

ll sp

ecifi

catio

ns,

and

a

re c

oeffi

cien

ts on

∆ln(/

) and ∆ln( /

) , respect

ivel

y. L

ocal

Sov

iet s

peci

aliz

atio

n (

) is

the

fract

ion

of a

mun

icip

ality

’s So

viet

exp

orts

in 1

990,

pre

dict

ed b

y th

e 19

88 o

utpu

t stru

ctur

e, to

the

mun

icip

ality

’s 19

90 o

utpu

t. A

ll sp

ecifi

catio

ns in

clud

e in

dustr

y×ye

ar d

umm

ies.

All

spec

ifica

tions

use

dat

a fo

r th

e pe

riod

1991

–199

5. ∆ln(

/) is

treat

ed a

s en

doge

nous

in c

olum

ns 2

–11.

∆ln(/) is

treat

ed a

s en

doge

nous

in

colu

mns

7–1

1. C

olum

n 8

show

s th

e LI

ML

estim

ates

. Col

umns

9 a

nd 1

0 sh

ow re

sults

for t

he b

ias-

corre

cted

and

mod

ified

bia

s-co

rrect

ed T

SLS

estim

ates

(Kol

esár

et a

l., 2

015)

. The

p-v

alue

s fo

r the

adj

uste

d te

st fo

r the

val

idity

of i

nstru

men

t by

Kol

esár

(201

2) (s

ee a

lso K

oles

ár e

t al.,

201

5) a

re d

ispla

yed

for t

he B

TSLS

and

MBT

SLS

estim

ates

. Pan

el-ro

bust

stand

ard

erro

rs a

re in

par

enth

eses

. The

90%

, 95%

, and

99

% c

onfid

ence

leve

ls ar

e de

note

d by

*, *

*, a

nd *

**, r

espe

ctiv

ely.

18

fication assumes that the impact of the instrument is equivalent across the post-collapse years.

However, it is likely that the adjustment in relative wage rates is not constant across years.

For example, it could be expected to be larger in the first years following the collapse.

In order to examine the potentially different impacts of the instrument across the post-

collapse years, we estimate the first-stage for our most general specification in equation (7)

using the second order polynomial of the 1990 local Soviet specialization interacted with year

dummies as instruments. The resulting first-stage coefficients on the instruments and the

corresponding marginal effects are displayed in columns 2 and 3 in appendix table B3. For

this specification, the point estimates of the marginal effects are positive for all other years

except for 1994. This motivates two additional specifications. Column 3 in table 2 shows

results for a just-identified model excluding the year 1994. The TSLS coefficient for the

production labor unit cost is 0.167 and significant at the 10% risk level, and it has a smaller

standard error compared to column 2, despite the lower number of observations. Column 4

shows results for an over-identified model where the instruments are the 1990 local Soviet

specialization and its interaction with the 1994 year dummy. This specification allows for

different first-stage impacts of the instrument in 1994. The estimate for the relative produc-

tion labor unit cost is 0.181 and significant at the 10% risk level.

A more general IV specification allows the impact of the 1990 local Soviet specialization

to vary freely across the post-collapse years. To implement this, we estimate models based on

the first-stage in equation (7) and the corresponding equation using linear rather than quadrat-

ic function of the instrument (all first-stage coefficients are provided in appendix table B3).

Column 5 shows estimates for the latter model, where a linear term for the 1990 local Soviet

specialization is fully interacted with year dummies. The TSLS estimate for the relative

production labor unit cost from this specification is 0.215 and it has a smaller standard error

than the estimates in columns 2–4. The improvement is due to the better fit of the first-stage

regression, which leads to more precise identification of the second-stage parameters. In

column 6, the specification based on the first-stage equation (7) provides an estimate of 0.156

with a standard error of 0.062. This is very close to the corresponding estimate from the just-

19

identified model in column 3 with improved precision.

Overall, the TSLS estimates for the relative production labor unit cost are of a larger

magnitude than the corresponding OLS estimate and suggest that the OLS estimate is nega-

tively biased and biased toward zero. The bias toward zero is consistent with attenuation due

to measurement error in plant-level hourly wage rates. On the other hand, the direction of the

bias is not consistent with confounding skill-biased technology shocks because such shocks

would be expected to induce positive bias in the coefficient on the relative production labor

unit cost, as discussed above. Nevertheless, because the direction of the potential bias in-

duced by skill-biased technology shocks and measurement error is different, the results

should not be taken as evidence of the absence of the former source of bias. However, the

results do suggest that the attenuation bias dominates the potential bias from skill-biased

technology shocks in our setting. This conclusion should not be generalized to other settings

at the more aggregate level where measurement error is likely to be less important.

The specifications in columns 1–6 treat capital intensity as exogenous. To account for the

potential reverse causality between the production labor cost share and capital intensity, the

specification in column 7 uses the log of capital intensity in − 2 as an instrument for its

concurrent change. This approach is based on the assumption that current shocks to the labor

input mix do not affect capital intensity two years earlier. Allowing for endogenous capital

intensity has very little impact on the estimate of . The coefficient on capital intensity ( )

is –0.019 and indicates statistically significant complementarity between capital and non-

production labor. The coefficient on the relative production labor unit cost is 0.146 and im-

plies own-wage labor demand elasticities of –0.11 for production and –0.23 for non-

production labor at the sample mean of the production labor cost share. The magnitude of

these elasticities is small and indicates that employers adjust employment only a little in

response to changes in unit labor costs in the short run. The corresponding plant-level short-

run elasticity of substitution between production and non-production labor is also small

(around 0.33). This elasticity is substantially smaller than the IV estimate of the state-level

long-run substitution elasticity between more and less educated workers of 1.5 estimated by

20

Ciccone and Peri (2005). This is expected because the plant-level short-run elasticity does not

incorporate labor substitution within or between plants due to capital adjustment in the long

run. It is also consistent with the view that wage elasticities are typically smaller at the micro-

level.13

The finding that the just-identified IV models provide very similar results compared to the

over-identified IV models indicates that weak instruments are not a major source of bias in

the over-identified models. To further check the validity of the over-identified models, we

follow Kolesár et al. (2015) and estimate the model with the LIML and bias-corrected TSLS

(BTSLS) estimators, as well as with the modification of the latter estimator (MBTSLS). The

MBTSLS estimator is less efficient compared to the BTSLS estimator, but it has the ad-

vantage of being consistent under both many-instruments and many-covariates asymptotics.

Column 8 shows the LIML estimates, while columns 9–10 provide estimates for the bias-

corrected estimators. Reassuringly, the LIML and BTSLS estimates are of a similar magni-

tude as the TSLS estimates. As expected, the MBTSLS estimates in column 10 are less pre-

cise compared to the estimates in columns 8–9, but they are of a similar magnitude as the

other IV estimates. Finally, for the BTSLS and MBTSLS specifications, we tested the validity

of the instruments with the adjusted test by Kolesár (2012) (see also Kolesár et al., 2015). We

did not reject the null hypothesis that the instruments are valid (that is, there is no evidence of

direct effects of the instruments on the outcome).14 These findings provide further reassur-

ance that weak instruments do not bias the parameter estimates.

CES Specification. In the last column of table 2, we show TSLS estimates for the constant-

elasticity-of-substitution (CES) specification, which is similar to the translog equation (3),

but uses the log-change in the ratio of production and non-production labor employment as

the outcome. In this specification, the coefficient on the relative production labor unit cost is

equivalent to the substitution elasticity between production and non-production labor. The

point estimate is 0.288 and it is fairly close to the corresponding translog elasticity (0.333),

13 See e.g. Lichter, Peichl, and Siegloch (2015), who argue that firm-level estimates of wage elasticities are

typically smaller than industry-level estimates. 14 P-values were 0.973 for both specifications.

21

although it is not significant at conventional significance levels. We prefer the translog speci-

fication because it provides a better fit and allows us to test some of the key assumptions of

the econometric model.

Testing the Price Homogeneity and CRS Assumptions. For the specification in column 6

(treating capital intensity as exogenous) we tested the assumption of homogeneity of prices

by allowing for unrestricted coefficients on the production ( ) and non-production ( ) la-

bor unit cost variables and by testing the null hypotheses + = 0, which we did not

reject (p-value was 0.503). We also tested for the constant returns to scale (CRS) assumption

by allowing for unrestricted coefficients on the capital stock ( ) and value added ( ) vari-

ables and by testing the null hypotheses + = 0. We ran three different tests for this null

hypothesis. The first was similar to the specification in column 6, treating these variables as

exogenous, but allowed for unrestricted and . The second was similar to the specifica-

tion in column 7, treating these variables as endogenous, but allowed for unrestricted and

and added the log of value added in − 2 as an instrument. The third test allowed addi-

tionally for unrestricted and and performed a joint test of the homogeneity of prices

and CRS assumptions. None of these tests rejected the null hypotheses (p-values were 0.471,

0.661, and 0.554, respectively).

4.2 Robustness Checks

In table 3, we further examine the robustness of our estimates against a number of potential

threats for identification. The baseline estimates in column 1 correspond to the TSLS esti-

mates in column 7 of table 2, which we prefer because they have high precision and they treat

both the relative production labor unit cost and capital intensity as endogenous.

Industry competition. We start by examining whether the results are robust when we ex-

clude from the analysis plants which face less competition. To do this, we calculate the Her-

findahl index for 2-digit industries using plant-level value added data. Column 2 excludes

plants in industries with the index above 0.1, which is generally considered to be a fairly

good level of competition. Column 3 excludes plants in industries with the index above 0.25,

22

Tabl

e 3:

Rob

ustn

ess A

naly

sis

(1)

Base

line

(2)

H-in

dex

belo

w

0.10

(3)

H-in

dex

belo

w

0.25

(4)

Base

line

+ 19

90

Ener

gy

Inte

nsity

(5)

+ Re

gion

D

umm

ies

(6)

+ 19

89

Wag

es

(7)

Inpu

t Su

pply

to

Sovi

et-

Dep

ende

nt

Indu

stry

Belo

w 1

%

of O

utpu

t

(8)

Inpu

t Su

pply

to

Sovi

et-

Dep

ende

nt

Indu

stry

Belo

w 2

0%

of O

utpu

t

0.14

6**

0.

158*

* 0.

147*

* 0.

148*

* 0.

144*

* 0.

154*

* 0.

159*

* 0.

162*

*

(0.0

58)

(0

.058

) (0

.058

) (0

.057

) (0

.059

) (0

.056

) (0

.058

) (0

.055

)

-0

.019

**

-0

.017

* -0

.019

**

-0.0

20*

-0.0

20*

-0.0

23*

-0.0

17*

-0.0

22*

(0

.009

)

(0.0

10)

(0.0

09)

(0.0

11)

(0.0

11)

(0.0

12)

(0.0

09)

(0.0

12)

O

bser

vatio

ns

2,81

3 2,

549

2,79

3 2,

760

2,76

0 2,

719

2,73

9 2,

788

Not

es:

The

spec

ifica

tions

are

bas

ed o

n th

e sp

ecifi

catio

n in

col

umn

7 of

tab

le 2

. a

nd

are

coe

ffici

ents

on ∆ln(

/) and

∆ln( /

) , respect

ivel

y. C

olum

n 1

repl

icat

es th

e es

timat

es in

col

umn

7 of

tabl

e 2.

The

spe

cific

atio

ns in

col

umns

2-3

sho

w re

sults

for

plan

ts in

2-d

igit

indu

strie

s fo

r whi

ch th

e H

erfin

dahl

inde

x is

belo

w 0

.1 a

nd 0

.25,

resp

ectiv

ely.

The

spe

cific

atio

n in

col

umn

4 ad

ds th

e lo

g of

ene

rgy

inte

nsity

in 1

990

inte

ract

ed w

ith y

ear

dum

mie

s as

con

trol v

aria

bles

to th

e ba

selin

e sp

ecifi

catio

n. T

he s

peci

ficat

ion

in

colu

mn

5 ad

ds a

dmin

istra

tive

regi

on (

maa

kunt

a) d

umm

ies.

The

spec

ifica

tion

in c

olum

n 6

adds

log

s of

198

9 pr

oduc

tion

and

non-

prod

uctio

n w

orke

r wag

es. T

he s

peci

ficat

ion

in c

olum

n 7

(8) c

orre

spon

ds to

the

spec

ifica

tion

in c

olum

n 1

but e

xclu

des

plan

ts w

ith th

e pr

edic

ted

inpu

t sup

ply

to th

e lo

cal S

ovie

t-dep

ende

nt in

dustr

y m

ore

than

1%

(20%

) of o

utpu

t in

1990

. Pan

el-ro

bust

stand

ard

erro

rs a

re

in p

aren

thes

es. T

he 9

0%, 9

5%, a

nd 9

9% c

onfid

ence

leve

ls ar

e de

note

d by

*, *

*, a

nd *

**, r

espe

ctiv

ely.

23

which are often considered to be uncompetitive. The results are very little affected by these

restrictions, suggesting that plants facing little competition do not drive our results.

Soviet energy imports. The abolition of the trade agreement did not only affect Soviet de-

mand for Finnish products, but it also resulted in a collapse in Finnish imports from the Sovi-

et Union, the bulk of which were energy inputs.15 Gorodnichenko et al. (2012) argue that the

Finnish manufacturing sector was adversely affected by the reduction in the supply of cheap

Soviet energy inputs and the resulting increase in energy prices, which affected plants with

high energy intensity the most. To examine the robustness of the results against the energy

price shock, we control for energy intensity in 1990 (the costs of energy inputs divided by

value added) in column 4. This has virtually no impact on the point estimates, suggesting that

shocks to energy prices are unlikely to drive our results.

Unobserved local trends. While plant fixed effects account for unobserved heterogeneity

across local labor markets, a potential threat for identification are unobserved local trends

correlated with relative wage rates and production labor cost share. We examine the robust-

ness of our results against them by including dummies for administrative regions (maakunta,

see figure 2). In the differenced equation (3), these dummies control for region-specific

trends. Reassuringly, including region dummies (column 5) has little impact on the estimates,

which suggests that local trends are unlikely to be a major source of bias.

Selection by local wage rates. Column 6 adds controls for plant-level production and non-

production labor unit costs in 1989. These variables control for the potential confounding

selection of plants by pre-collapse local wage rates. Adding these controls has little impact on

the estimates. We also experimented with a specification adding within-plant changes in the

relative production labor unit cost from 1989 to 1990 and the size of the neighboring industry

in the same municipality as controls. These specifications gave also very similar results.

Skill composition of the local labor supply shock. Another concern is that the skill compo-

sition of the workforce released by the Soviet-dependent industry may be different compared

to the skill distribution in our target plants. For instance, if production workers released by 15 In the period 1986–1990, fuels and crude oil accounted for around 62% of imports of manufacturing in-

puts from the Soviet Union to Finland (Foreign Trade 1990, Vol. 2, The Finnish Board of Customs).

24

the Soviet-dependent industry are less skilled than production workers in our target plants,

spurious correlation due to shifts in the within-group skill distribution within the target plants

may bias our analysis. However, these plants are very similar in terms of average production

labor shares and hourly wages of both production and non-production labor (appendix table

A2). To further assess whether the population of workers which moved from the Soviet-

dependent industry was significantly different compared to workers in the target plants, we

test for differences in the distribution of annual wage earnings between workers who worked

at one of our target plants in 1990 ( , < 0.001) and workers who worked at a plant

with predicted Soviet exports more than 10% of output in 1990 ( , > 0.1) and left it

in 1991. Appendix figure B1 shows the distribution of annual wage earnings in these two

groups in the period 1992–1995, separately for production and non-production occupations.16

Reassuringly, the distributions are very similar for both occupational categories and provide

no evidence of significant within-group differences between workers in the target plants and

workers who left a Soviet-dependent plant. We also test for differences in the means of the

distributions in a regression controlling for municipality-year fixed effects (appendix table

B4). We are unable to reject the null hypothesis that the mean of annual wage earnings is the

same for the two worker groups for either of the occupational categories. These findings

suggest that variation in wage rates induced by the Soviet trade shock is unlikely to be driven

by changes in the within-group skill composition in our target plants.

Local input supply and agglomeration. One concern is that the collapse of the Soviet-

dependent industry affects our target plants through changes in local agglomeration bene-

fits.17 There are three potential mechanisms associated with agglomeration. First, the shock

may affect our target plants through local input supply. In order to assess whether this could

16 The data are constructed by linking the measure of plant-level Soviet specialization to FLEED annual

wage earnings by a unique plant identifier. We do not observe whether a person worked full-time. Hence, we include worker-year observations for which annual wage earnings are positive and the individual did not receive unemployment benefits. We exclude the year 1991 because of the concern that some workers who left a Soviet-dependent plant in that year may have worked fewer hours due to the transition, which would reduce their annual wage earnings, on average. Alternatively, wage rates could be used as a proxy for worker skill, but individual-level data on hourly wages are unavailable for the early 1990s.

17 For a detailed discussion of agglomeration mechanisms, see, e.g., Duranton and Puga (2004), Glaeser and Gottlieb (2009), and Moretti (2011).

25

affect our results, we use CSS data on 1988 plant-level inputs and outputs by 6-digit HS

commodity to calculate the predicted plant-level output share of products which were used as

inputs in the neighboring Soviet-dependent industry.18 Columns 7 and 8 display the results for

specifications where we exclude plants with the predicted supply of inputs to the local Soviet-

dependent industry more than 1% and 20% of their annual output. These restrictions exclude

only a few observations and have little impact on the estimates.19 This suggests that the diffu-

sion of the shock through local input supply channels is unlikely to drive our results. Second,

the downsizing of the Soviet-dependent industry may affect our target plants through changes

in productivity spillovers. Previous research has shown that large plants with high productivi-

ty may induce substantial local spillovers (Greenstone, Hornbeck, and Moretti, 2010). Such

spillovers are unlikely to be significant in our setting because plants in the Soviet-dependent

industry are not technologically more advanced or significantly larger compared to our target

plants. Overall, Soviet-dependent plants and our target plants are very similar on average (see

appendix table A2). This suggests that changes in local productivity spillovers due to the

decline of the Soviet-dependent industry are unlikely to be significant. Third, the negative

shock in high-exposure areas may affect local wage rates through changes in the thickness of

the local labor markets. However, the fact that manufacturing employment does not diverge

much between the high- and low-exposure areas before and after the collapse of Soviet trade

(panel B in figure 3) suggests that labor market thickness was unlikely to vary much between

these areas.

Representativeness of the estimation sample. Finally, appendix table A4 shows that the es-

timation sample includes plants from all major 2-digit industries in the Finnish manufactur-

ing.

Overall, our findings suggest that the IV estimates of the parameters of the labor demand 18 To calculate this measure, we first approximate the amount of input used in Soviet-dependent produc-

tion by plant : = , , where is the plant’s usage of input in 1988 and , is the plant’s predicted output share of Soviet exports in 1990. Then the predicted usage of input in Soviet-dependent production in location is = ∑ ∈ ( ) and the predicted output share of inputs supplied to local Soviet production by plant in location is = ∑ / , where is the plant’s output share of commodity in location .

19 The estimates in appendix table B5 show that setting the input supply threshold to 0.1% or 5% has little impact on the results.

26

model are robust against numerous potential sources of bias. We next use the identified model

to simulate within-plant changes in the structure of labor demand.

4.3 Within-Plant Changes in the Structure of Labor Demand

This section presents changes in the relative demand for production labor implied by the

estimated plant-level labor demand model. Taking expectations on both sides of equation (3)

conditional on industry and year yields

= ∆ | , − [∆ ln / | , ]− [∆ ln( / ) | , ]. (8)

Our measure of the average within-plant shift in the structure of labor demand in industry

from the year 1980 to is then

= . (9)

are calculated from equation (8) by using estimates of and and replacing the expec-

tation terms with the corresponding industry-by-year sample means. We use the TSLS esti-

mates in column 7 of table 2, where both coefficients have high precision and both variables

are treated as endogenous, and sample means from the full LDPM sample.20 Figure 4 displays

the results by industry over the period 1980–2008. It also shows a series for all industries

which is the weighted average of 2-digit industry indices with annual industry labor cost as a

weight. The series for all industries shows a 7.1 percentage point within-plant decline in the

relative demand for production labor between 1980 and 2008. This corresponds to a decadal

reduction of around 2.5 percentage points and explains around 42% of the overall decline in

the production labor cost share in this period, which was 16.8 percentage points. In the 1980s,

the decadal reduction is around 2.8 percentage points, while it is only 1.2 percentage points in

the 1990s. The shift in the structure of labor demand accelerates again in the 2000s when the

decadal reduction is around 2.9 percentage points. In this period, within-plant shifts in the

structure of labor demand account for around half of the overall decline in the production 20 The findings are robust in a wide range of IV estimates in table 2. See appendix A for details of the LDPM

sample.

labor co

The

shift di

decadal

in 1980

tries div

21 Ap

share by

-.2-.1

5-.1

-.05

0.0

5R

elat

ive D

eman

d Fo

r Blu

e-C

olla

r Lab

orR

elat

ive

Dem

and

for P

rodu

ctio

n La

bor (

1980

=0)

Notes: classifi(NACEindustrby-yeathen avto cons

ost share.

industry-sp

ffers betwe

l rate of the

0–1989 to 2.

verge at an

ppendix table industry and s

1980 1982 1

F

This figure disication is the 2-E). The series ary from equatioar means of ln(veraged by 2-chstruct the series

pecific serie

een industri

industry-sp

.0 in 1989–

n increasing

B6 shows chasubperiod.

984 1986 198

FoodWoodChemMineMachTrans

Figure 4: Ch

splays estimated-character levelare constructed ons (8) and (9), ( / ) and lnharacter industrys for all industri

es in figure

ies over the

pecific relat

1999, and r

gly rapid pa

anges in the r

88 1990 1992

ddmicalsral Prod.

hinerysport Eq.

hanges in the

d changes in thl of Statistical Cby first calculausing the TSLSn( / ) from thy with annual 2es.

27

4 suggest t

e observatio

tive product

reaches 3.0 i

ace. In the

relative deman

1994 1996 1Year

TextilePaperPlastiMetal ElectrAll Ind

Relative De

he relative demClassification ofating relative pS estimates of he 1980–2008 2-digit industry

that the tim

on period. T

tion labor de

in 1999–20

2000s, the

nd for produc

998 2000 200

esrc and Rubber PProd.

rical & Optical Edustries

emand for Pro

and for productof Economic Acproduction labo

and in coLDPM samplelabor cost as a

me pattern o

The standar

emand shift

08.21 This in

largest dec

tion labor and

2 2004 2006

Prod.

Eq.

oduction Lab

tion labor by intivities in the Er demand indicolumn 7 of tabl. These 2-digit weight. A simi

of the labor

rd deviation

t increases f

ndicates tha

cline in the

d production

2008

Textiles

Electrical & OTransport Eq

Mineral Produ

bor

ndustry. The indEuropean Commces by 2-digit Nle 2 and the indt industry indicilar procedure i

demand

n of the

from 1.0

at indus-

relative

labor cost

Optical Eq. q.

ucts

dustry munity NACE dustry-es are s used

28

demand for production labor occurs in the production of electrical and optical equipment,

transport equipment, and textiles, with a decadal reduction of 8.7, 6.1, and 6.0 percentage

points, respectively.

4.4 The Impacts of Offshoring and ICT

What are the sources of the relative production labor demand shifts unrecovered by our anal-

ysis? In this section, we examine this question by focusing on changes in offshoring and ICT.

These two major labor demand shifters have been shown to affect the structure of employ-

ment and wages;22 hence, we should be able to find a relationship between measures of them

and our relative production labor demand index. In order to examine this, we estimate the

following industry-level regression:

= + ln , + ln , + + + , + . (10)

Here is the index of the relative production labor demand in 2-digit industry in year .

To account for the potential time gap before the effects of offshoring and ICT are fully real-

ized, we lag them by two years. The model includes industry fixed effects and industry-

specific trends. We also control for lagged R&D intensity, , . ICT is measured by com-

puter and programming expenses. Offshoring is measured by imports of industrial intermedi-

ate inputs, as in, e.g., Feenstra and Hanson (1996).23 All industry regressions are weighted by

industry labor cost.

Table 4 shows the results. Odd columns show estimates for specifications without the year

dummies as control variables while even columns show estimates for specifications including

them. Columns 1 and 2 show the OLS estimates. The coefficient on offshoring is negative

22 See, e.g., Berman et al., (1994), Machin and Van Reenen (1998), Autor et al. (2003), Autor et al. (2008),

Michaels et al. (2014), and Goos et al. (2014). 23 Computer and programming expenses include costs of equipment and programming; consulting related to

automatic data processing; design and programming of software; activities related to computer operations and data processing; database hosting; repair and maintenance of office equipment and computers; other data pro-cessing services, e.g. software engineering services; and IT software maintenance and consulting. The data are available by 2-digit industry in the Industrial Statistics database maintained by Statistics Finland. Offshoring is calculated from 2-digit industry input-output tables maintained by Eurostat, and available for the period 1995-2007. Data after 2007 uses a considerably coarser industry classification and hence cannot be used to extend the 2-digit industry data.

29

and highly significant for both specifications, whereas the coefficient on ICT is negative and

significant at the 10% risk level for the specification excluding the year dummies. Including

the year dummies reduces the degrees of freedom and the precision of the estimation. How-

ever, it has little impact on the point estimate.

To account for the potential correlation between offshoring, ICT, and unobserved con-

founding industry shocks within the domestic manufacturing sector, we use lagged U.S.

industry imports of intermediate inputs from the same industry and the U.S. industry use of

computer services, both measured in − 3, as instruments.24 Columns 3 and 4 (5 and 6) show

result for specifications treating offshoring (ICT) as endogenous, whereas columns 7 and 8

treat both variables as endogenous. The IV coefficients on offshoring are of a similar magni-

tude or larger compared to the corresponding OLS estimates across specifications. The IV

24 U.S. computer services are inputs from NAICS industry 5415 (computer systems design and related ser-

vices) drawn from the U.S. Bureau of Labor Statistics nominal use tables. U.S. offshoring is imported intermedi-ate inputs from own industry drawn from the U.S. Bureau of Economic Analysis import matrixes.

Table 4: The Effects of Offshoring and ICT on the Relative Demand for Production Labor

(1) (2) (3) (4) (5) (6) (7) (8) OLS OLS IV IV IV IV IV IV

Offshoring -0.020*** -0.017*** -0.022 -0.029 -0.020*** -0.018** -0.027** -0.032 (0.007) (0.006) (0.017) (0.023) (0.006) (0.007) (0.011) (0.033) ICT -0.009* -0.009 -0.009* -0.009 -0.024** -0.012 -0.019** -0.002 (0.005) (0.006) (0.005) (0.005) (0.011) (0.019) (0.009) (0.022)

Industry fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Industry trends Yes Yes Yes Yes Yes Yes Yes Yes Year dummies No Yes No Yes No Yes No Yes Endogenous variables [1st stage F-statistics]

- - Offs. [16.47]

Offs. [7.312]

ICT [4.523]

ICT [3.269]

Offs. [19.398],

ICT [5.838]

Offs. [5.057],

ICT [2.247]

Notes: N=171. 2-digit industry data for the period 1999–2007. The table shows OLS and IV estimates from regressions of the estimated relative production labor demand index at the 2-digit industry level on logs of industry offshoring and ICT lagged two years. All regressions are weighted by industry labor cost. Columns 3–4 use the log of U.S. imported intermediate inputs from own industry lagged three years as an instrument. Columns 5–6 use the log of U.S. input usage from NAICS industry 5415 (computer systems design and related services) lagged three years as an instrument. Columns 7–8 use both instruments. All specifications control for industry R&D intensity. First-stage estimates for the IV specifications are shown in appendix table B8. For specifications excluding R&D intensity, see appendix table B9. Panel-robust standard errors are in parentheses. The 90%, 95%, and 99% confidence levels are denoted by *, **, and ***, respectively.

30

estimate in column 7 is –0.027 and significant at the 5% risk level. The precision of the esti-

mation reduces considerably when year dummies are added (column 8) due to the decline in

the degrees of freedom, but this has little impact on the point estimate. Most of the IV coeffi-

cients on ICT are also of a similar magnitude or larger compared to the corresponding OLS

estimates. The point estimate in column 6, treating ICT as endogenous and including year

dummies, is –0.012 but insignificant. Estimates in columns 5 and 7 are both larger and signif-

icant at the 5% risk level. The coefficient in column 8 is smaller, but it has low precision due

to the lower degrees of freedom and limited first-stage variation because the specification

includes year dummies and treats both ICT and offshoring as endogenous.25

To put the size of these estimates into perspective, we calculated how much predicted val-

ues from the model explain the estimated overall labor demand shift. The smallest coefficient

on offshoring of –0.017 (p<0.01) implies that offshoring explains around one-third of the

decline in the relative demand for production labor between 2000 and 2008. Some caution is

in order when interpreting the results for ICT due to the small IV estimate in column 8. Nev-

ertheless, the OLS coefficient of –0.009 (p<0.01) in column 1, which is smaller than the IV

estimates in columns 5 and 6 treating ICT as endogenous, implies that ICT explains also

around one-third of the relative production labor demand shift. Overall, these findings sug-

gest that offshoring accounts for a substantial fraction of variation in our industry-specific

relative production labor demand indices. The results for ICT are slightly weaker, but the

preferred estimates indicate effects of a similar magnitude.

5 Summary and Conclusions

In this paper, we provided quasi-experimental estimates of plant-level labor demand elastici-

ties and calculated within-plant changes in the structure of labor demand implied by them.

Our IV strategy exploited the abrupt decline in exports of specific products in Finland due to

the collapse of the Soviet Union. The shock induced by the collapse of Soviet trade was

25 We also provide results for specifications using one year-lagged and concurrent ICT and offshoring as co-

variates. These specifications give mainly less precise and smaller estimates (appendix tables B7–B8). Exclud-ing R&D intensity had little impact on the results (appendix table B9).

31

asymmetric across product fields and local labor markets. As a result, it induced plausibly

exogenous variation in the local available supply of labor faced by plants which were not

dependent on Soviet import demand. We used these shocks as a source of variation in unit

labor costs to identify a plant-level labor demand model.

Our results suggest that the magnitude of the plant-level, short-run own-wage labor de-

mand elasticity is small for both production and non-production workers. The analysis sug-

gests that plant-level labor demand responses to changes in wage rates can be substantially

overestimated in conventional OLS approaches. Our results are in line with the view that

measurement error in plant-level labor unit cost data is the primary source of estimation bias.

We emphasize that this conclusion should not be generalized to more aggregate-level settings,

such as studies using state-level data (e.g., Ciccone and Peri, 2005), where measurement error

is likely to be much less important.

The results indicate that employers adjust employment only a little in response to changes

in unit labor costs in the short run. One implication of this finding is that temporary labor

market interventions lowering employers’ labor costs, which have become popular in many

countries during the last recession, are unlikely to be effective in preventing short-run em-

ployment losses during economic downturns.

A simulation of the model uncovers a sharp within-plant shift in the structure of labor de-

mand against production workers in the 2000s. The within-plant shift accounts for around

half of the reduction in the production labor share in this period. Moreover, our findings

suggest that the pace of the shift in the structure of labor demand vary substantially across

industries in the 2000s. The rising industry heterogeneity points to the possibility that the

structure of labor demand between labor markets with different industry specialization may

have diverged recently.

Our research design, based on asymmetric shocks across product fields and local labor

markets, can be also implemented with spatial industry data. Because asymmetric product

market shocks are common (due to changes in trade policies, for instance), we believe that it

can be useful for future studies seeking to identify labor demand models for other countries

32

and time periods.

33

[Not for publication unless otherwise requested]

Appendix A: Data Sources and Summary Statistics

A.1 Data on the Wage Earnings Share and Job Task Indices by Occupation

The research-use sample of the Finnish Linked Employer-Employee Data (FLEED) contains

individual-level information on taxable wage earnings and occupation (2-digit level of the

ISCO-88 classification) for individuals aged 15–64. We use this information to calculate the

wage earnings share in 1995 by occupation. For job task indices by occupation, we use the

Acemoglu and Autor (2011) job task data which are based on the 4-digit SOC-2000 classifi-

cation. We use the 4-digit correspondence table from the U.S. National Crosswalk Service

Center to obtain indices for the 4-digit ISCO-88 classification.26 After this, the 4-digit job

task indices were aggregated to the 2-digit level by using employment weights in the Ace-

moglu-Autor data (at the time of the analysis, 4-digit ISCO codes were unavailable in

FLEED and therefore we were unable to use Finnish employment by occupation as a weight

in this stage of the aggregation).

The wage earnings share and job task indices by 2-digit occupation are displayed in table

A1. In order to generate the job task indices in table 1, we aggregated the 2-digit indices with

Finnish employment by 2-digit occupation calculated from FLEED as a weight.

A.2 Longitudinal Database of Plants in Finnish Manufacturing

The main plant-level data source of this study is the Longitudinal Database of Plants in Finn-

ish Manufacturing (LDPM) provided by Statistics Finland. The LDPM is based on the annual

Industrial Structures survey (IS). For the years 1980–1994, the IS covers all plants with at

least 5 employees, and for the years 1995–2008 it covers plants whose parent company had at

least 20 employees. Therefore, all plants with at least 20 employees are covered over the

period 1980–2008. We restrict our analysis to these plants. They cover around 82% of nation-

26 webdata.xwalkcenter.org/ftp/DOWNLOAD/xwalks/SOC2000xISCO88.zip

34

al manufacturing output during the observation period.

The LDPM provides information on value added, capital stock, and labor costs and work

hours for both production and non-production workers.

Labor costs include wage bill and employer contributions such as compulsory insurance

payments. The category “production worker” includes all persons directly engaged in produc-

tion or the related activities of the establishment. These include packers, service staff,

maintenance staff, construction staff, machinists, and stokers, for example. Low-level super-

visors involved in the actual production are also included in this category. The category “non-

production worker” refers to all other employees not directly engaged in production tasks.

These are typically employees engaged in supervision, technical services, administration, and

sales. The LDPM also provides information on the location of the plant at the municipal

level.

The LDPM capital stock variable is calculated from plant-level investments with the per-

petual inventory method using a depreciation rate of 0.10. For old plants for which initial

investment in entry year is not observed fire insurance values are used as the initial value for

the capital stock. LDPM also provides measures of real output and capital stock in 2000

prices, which are calculated by deflating the nominal variables with industry price indices

drawn from the national accounts.

A.3 Commodity Statistics Survey and OECD ITCS Data

The Commodity Statistics survey (CSS) provides information on outputs and inputs used by a

plant at the level of a 6-digit commodity. We use the 1988 CSS file to estimate each plant’s

share of national output of a 6-digit commodity before the abolition of the bilateral trade

agreement. The CSS sampling frame corresponds closely to the LDPM sampling frame and

the data cover around 91% of the aggregate LDPM output in 1988. The CSS data are linked

to the LDPM with unique plant codes.27

Data on Finnish exports to the former Soviet Union in 1990 are drawn from the OECD

27 The CSS data are based on an annual survey targeting all manufacturing plants with a parent company of

at least 10 employees.

35

ITCS database. We use exports by 6-digit commodity. Both the 1988 CSS and 1990 ITCS

data are based on the same 1988 Harmonized System (HS) commodity classification. There-

fore, calculating measures of plant- and local-level Soviet specialization in equations (4) and

(5) is straightforward.

A.4 Summary Statistics for Relevant Samples

Table A2 presents summary statistics for relevant samples used in the analysis. The first

column displays means and standard deviations for the 1990–1995 sample. The sample is

constructed from the LDPM by excluding plant-year observations falling into the first or last

year of a plant’s existence in the LDPM panel to avoid observations for years in which plants

may have entered or exited the market in the middle of the year. We also exclude plants in the

smallest municipalities falling below the 5th output percentile, for which measures of local

Soviet-specialization are based on only a few plants.28 The second column displays summary

statistics for plants that are not dependent on Soviet import demand ( , < 0.001).

This sample is used to estimate equation (3) and it covers around 28% of plants in the 1990–

1995 sample. The third column contains summary statistics for the 1980–2008 sample.

Means from this sample by industry and year are used to construct the measures of the rela-

tive production labor demand shift (equations (8) and (9)). Table A3 tabulates 1989 summary

statistics for plants which were not dependent on Soviet import demand, separately for plants

located in municipalities with 1990 local Soviet specialization below and above the median

of 0.030. Table A4 displays the industry distribution for the full sample and for the sample of

plants that are not dependent on Soviet import demand.

28 This excludes 19 smallest municipalities, which account for around 0.1% of the aggregate LDPM output.

36

Table A1: Wage Earnings Share and Job Task Indices by 2-Digit Occupation

Job Task Indices Routine Non-Routine

Occupation

Wage Earn-ings Share 1995 (%) Manual

Cogni-tive

Cognitive Analytic

Cognitive Interper-

sonal Manual Physical

Manual Inter-

personal A. Production Worker Machine operator or assembler 12.9 1.96 0.56 -0.44 -0.59 0.92 -1.28 Stationary plant or related operator 8.6 1.68 0.33 -0.07 -0.42 0.94 -1.13 Precision or related trades worker 2.5 1.32 0.65 -0.24 -0.95 0.43 -1.05 Skilled agricultural or fishery worker 0.5 1.30 -1.31 -0.84 -0.65 1.17 -1.29 Driver or related water traffic operator 2.3 1.22 0.35 -0.70 -0.91 2.17 -0.46 Other craft or related trades worker 3.7 1.03 0.15 -0.54 -0.56 0.35 -1.10 Laborer in manufacturing or construction 3.8 0.89 0.16 -0.73 -0.45 1.00 -1.15 Metal, machinery, or related trades worker 19.8 0.82 -0.03 -0.08 -0.53 1.46 -1.04 Extraction or building trades worker 2.4 0.74 -0.49 -0.18 -0.33 1.38 -0.89 B. Professional Worker Engineering science associate professional 10.4 0.46 0.41 0.58 -0.26 0.17 -0.57 Life science and health associate professional 1.4 -0.03 0.57 0.92 1.07 -0.01 1.17 Life science and health professional 0.3 -0.09 0.53 1.23 1.30 -0.01 1.32 Corporate manager 6.3 -0.62 -0.66 0.90 1.57 -0.56 0.57 Engineering science professional 8.5 -0.63 0.26 1.56 0.08 -0.71 -0.69 Manager of small enterprises 0.4 -0.73 -1.30 0.87 1.29 -0.25 0.61 Other associate professional 5.9 -0.76 0.17 0.43 0.03 -0.65 0.34 Other professional 3.2 -1.03 -0.36 1.16 0.51 -0.88 0.69 Teaching professional 0.2 -1.05 -1.01 1.15 1.36 -1.05 1.57 C. Service or Clerical Worker Customer services clerk 0.3 0.34 1.35 -0.77 -0.34 -0.35 0.24 Sales and services elementary occupation 1.3 0.20 -0.77 -1.49 -1.12 0.10 -0.86 Personal or protective services worker 0.8 0.11 -0.42 -0.74 -0.23 0.18 0.21 Office clerk 3.9 -0.28 0.90 -0.32 -0.54 -0.57 -0.25 Salesperson or demonstrator 0.8 -0.69 -0.18 0.19 0.13 -0.44 0.27 Notes: Data from the FLEED and Acemoglu and Autor (2011).

37

Table A2: Summary Statistics 1990–1995 Sample 1980–2008 Sample

(1)

All

(2) Plants Not

Dependent on Soviet Import

Demand

(3)

All

Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.

Labor Cost, Total (‘000) 3,247 6,244 3,103 6,177 3,221 7,880Labor Cost, Non-production labor (‘000) 1,204 2,843 1,140 2,940 1,205 4,740Labor Cost, Production labor (‘000) 2,042 3,811 1,963 3,595 2,016 4,227Employment, Total 128 220 122 228 127 235Employment, Non-production labor 38 84 35 85 36 102Employment, Production labor 91 149 87 152 91 155Hours, Total (‘000) 205 344 193 342 210 387Hours, Non-production labor (‘000) 63 141 58 143 61 175Hours, Production labor (‘000) 142 224 135 216 148 249Production Labor Cost Share 0.676 0.161 0.701 0.171 0.691 0.168Production Labor Employment Share 0.746 0.145 0.762 0.153 0.757 0.153Production Labor Hour Share 0.738 0.147 0.755 0.157 0.753 0.154Real Capital Stock (2000 Prices) (‘000) 9,496 30,644 10,248 30,020 7,992 30,333Real Output (2000 Prices) (‘000) 21,464 73,049 22,323 55,242 22,456 101,492Real Value Added (2000 Prices) (‘000) 7,012 19,023 6,675 15,030 7,184 34,339Capital Intensity 2.242 58.64 2.062 8.172 2.429 170.5Production Labor Unit Cost 13.1 3.21 13.41 3.31 12.79 6.680Non-Production Labor Unit Cost 18.4 4.8 18.53 5.3 17.92 9.110Prod. Labor Unit Cost, 1989 10.63 2.560 10.78 2.640 10.65 2.592Non-Prod. Labor Unit Cost, 1989 15.60 4.214 15.44 4.508 15.57 4.184Relative Production Labor Unit Cost 0.743 0.214 0.764 0.225 0.738 0.227Energy Intensity 0.114 1.657 0.141 0.593 0.322 42.30Energy Intensity, 1990 0.080 0.139 0.106 0.176 0.084 0.195Soviet Specialization in 1990: Plant 0.052 0.174 0 0 0.054 0.170 Municipality 0.048 0.075 0.040 0.060 0.048 0.096 Municipality, Log Levels -3.458 0.978 -3.712 1.110 -3.452 0.969Observations 7,479a 2,881b 70,806c Notes: The number of observations for 1989 production labor unit cost, 1989 non-production labor unit cost, 1990 energy intensity, and plant-level Soviet specialization is, respectively: a – 7453, 7434, 7362, and 7479; b – 2851, 2838, 2808, and 2881; c – 57543, 57210, 56209, and 52857.

38

Tabl

e A3:

Sum

mar

y St

atist

ics b

y Lo

cal S

ovie

t Spe

cial

izat

ion,

198

9

Plan

ts Lo

cate

d in

Mun

icip

aliti

es

with

Loc

al S

ovie

t Spe

cial

izat

ion

Belo

w

the

Med

ian

Plan

ts Lo

cate

d in

Mun

icip

aliti

es

with

Loc

al S

ovie

t Spe

cial

izat

ion

Abo

ve

the

Med

ian

Mea

n St

d. D

ev.

Mea

n St

d. D

ev.

La

bor C

ost,

Tota

l (‘0

00)

2,57

1 5,

492

2,43

2 4,

591

Labo

r Cos

t, N

on-p

rodu

ctio

n la

bor (

‘000

) 87

7 2,

283

881

2,31

4 La

bor C

ost,

Prod

uctio

n la

bor (

‘000

) 1,

694

3,44

6 1,

551

2,54

4 Em

ploy

men

t, To

tal

118

216

114

205

Empl

oym

ent,

Non

-pro

duct

ion

labo

r 31

77

31

74

Em

ploy

men

t, Pr

oduc

tion

labo

r 87

14

6 83

14

3 H

ours

, Tot

al (‘

000)

19

9 36

9 18

4 28

8 H

ours

, Non

-pro

duct

ion

labo

r (‘0

00)

54

141

52

124

Hou

rs, P

rodu

ctio

n la

bor (

‘000

) 14

4 24

0 13

2 18

5 Pr

oduc

tion

Labo

r Cos

t Sha

re

0.72

5 0.

156

0.70

9 0.

176

Prod

uctio

n La

bor E

mpl

oym

ent S

hare

0.

788

0.13

6 0.

774

0.15

5 Pr

oduc

tion

Labo

r Hou

r Sha

re

0.78

4 0.

140

0.76

8 0.

161

Real

Cap

ital S

tock

(200

0 Pr

ices

) (‘0

00)

9,74

3 30

,841

6,

989

21,1

11

Real

Out

put (

2000

Pric

es) (

‘000

) 21

,462

51

,714

15

,553

36

,049

Re

al V

alue

Add

ed (2

000

Pric

es) (

‘000

) 6,

704

17,0

55

5,63

8 15

,867

Ca

pita

l Int

ensit

y

1.81

3 3.

962

1.59

2 2.

939

Nom

inal

Pro

duct

ion

Labo

r Uni

t Cos

t 10

.51

2.72

10

.84

2.73

N

omin

al N

on-P

rodu

ctio

n La

bor U

nit C

ost

15.0

3 4.

40

15.6

0 4.

71

Rela

tive

Prod

uctio

n La

bor U

nit C

ost

0.73

8 0.

219

0.73

0 0.

192

Ener

gy In

tens

ity

0.13

6 0.

329

0.09

0 0.

198

Ener

gy In

tens

ity, 1

990

0.12

3 0.

156

0.11

7 0.

559

Sovi

et sp

ecia

lizat

ion

in 1

990:

Pla

nt

0.00

0 0.

000

0.00

0 0.

000

M

unic

ipal

ity

0.01

6 0.

009

0.06

5 0.

070

M

unic

ipal

ity, L

og L

evel

s -4

.446

0.

976

-2.8

97

0.45

7 O

bser

vatio

ns

366a

357b

Not

es: D

ata

for p

lant

s whi

ch w

ere

not d

epen

dent

on

Sovi

et im

port

dem

and

in 1

990.

a –

Num

ber o

f obs

erva

tions

for 1

990

ener

gy in

tens

ity is

364

. b –

Num

ber

of o

bser

vatio

ns fo

r 199

0 en

ergy

inte

nsity

is 3

50.

39

Table A4: Industry Composition 1990–1995 Sample

(1)

All

(2) Plants Not Dependent

on Soviet Import Demand

NACE Industry Code Observations Share (%) Observations Share (%) 15 1,649 15.9 967 33.6 16 20 0.19 20 0.69 17 316 3.05 92 3.19 18 292 2.82 9 0.31 19 181 1.75 51 1.77 20 911 8.79 377 13.1 21 480 4.63 113 3.92 22 815 7.87 330 11.5 23 32 0.31 - - 24 481 4.64 191 6.63 25 514 4.96 24 0.83 26 657 6.34 200 6.94 27 237 2.29 108 3.75 28 811 7.83 39 1.35 29 1,273 12.3 79 2.74 30 24 0.23 - - 31 304 2.93 7 0.24 32 223 2.15 77 2.67 33 133 1.28 39 1.35 34 212 2.05 32 1.11 35 153 1.48 64 2.22 36 642 6.20 62 2.15

Appen

Notes: Thindividualworked inand includployment specializa

ndix B: A

his figure showls who worked n a plant not dede worker-year benefits. Data

ation is linked to

Additional

s distributions in a Soviet-dep

ependent on Soobservations foon individual-l

o the FLEED da

Tables an

Figure Bof annual wagependent plant inviet import dem

for which annualevel annual waata by a unique

Panel A

Panel

40

nd Figure

1: Annual We earnings for in 1990 ( ,mand in 1990 (al wage earningage earnings ar plant identifier

A. Production

B. Non-produ

es

Wage Earningproduction and> 0.1) and( , < 0.gs are positive are from the FLEr.

occupations

uction occupat

s d non-productiod left it in 1991001). Data covand the individuEED. The mea

tions

on workers, sep; and for indivver the period

dual did not receasure of plant-le

parately for iduals who 1992–1995 eive unem-evel Soviet

41

Ta

ble

B1: T

op 1

5 So

viet

Exp

ort C

omm

oditi

es in

199

0

Com

mod

it y

Expo

rts, U

SD

% o

f Sov

iet

Expo

rts

% o

f All

Expo

rts

% o

f Man

ufac

-tu

ring

Out

put

Te

leph

onic

or t

ele g

raph

ic sw

itchi

ng a

ppar

atus

19

3,32

2,28

5 5.

7 0.

8 0.

4 Pr

efab

ricat

ed b

uild

ings

10

7,95

6,40

1 3.

2 0.

5 0.

2 Ra

ilwa y

car

s n.e

.s., o

pen,

with

side

s > 6

0 cm

hig

h 10

5,42

0,52

0 3.

1 0.

5 0.

2 Fl

oatin

g, su

bmer

sible

dril

ling

or p

rodu

ctio

n pl

atfo

rm

89,0

01,1

75

2.6

0.4

0.2

Pape

r, fin

e, w

oodf

ree,

40

- 150

g/m

2, u

ncoa

ted

88,5

40,2

80

2.6

0.4

0.2

Chem

ical

woo

d pu

lp, d

issol

ving

gra

des

77,2

87,4

66

2.3

0.3

0.2

App

arat

us, f

or c

arrie

r-cur

rent

line

s yste

ms,

n.e.

s. 68

,536

,926

2.

0 0.

3 0.

1 Ta

nker

s 65

,452

,989

1.

9 0.

3 0.

1 Pa

per,

fine,

woo

d-co

ntai

nin g

, unc

oate

d, n

.e.s.

62

,427

,258

1.

9 0.

3 0.

1 C y

clic

am

ides

, der

ivat

ives

, n.e

.s., s

alts

ther

eof

54,5

85,4

07

1.6

0.2

0.1

Infa

nt fo

ods o

f cer

eals,

flou

r, sta

rch

or m

ilk

53,0

68,6

93

1.6

0.2

0.1

Pape

r, m

ulti-

ply,

cla

y co

ated

, n.e

.s.

50,4

86,4

37

1.5

0.2

0.1

Boot

s, so

le ru

bber

or p

lasti

c up

per l

eath

er, n

.e.s.

45

,988

,073

1.

4 0.

2 0.

1 Ra

ilwa y

tank

car

s 45

,646

,409

1.

4 0.

2 0.

1 W

arsh

ips,

lifeb

oats,

hos

pita

l shi

ps, v

esse

ls n.

e.s.

44,6

75,8

70

1.3

0.2

0.1

Not

es: D

ata

from

OEC

D IT

CS d

atab

ase

and

Ann

ual N

atio

nal A

ccou

nts f

or F

inla

nd. C

olum

n 1

show

s tot

al e

xpor

ts fro

m F

inla

nd to

the

Sovi

et U

nion

in

199

0 fo

r 6-d

igit

HS

com

mod

ity c

ateg

orie

s. “N

.e.s.

” sta

nds f

or “

not e

spec

ially

spec

ified

.”

42

Table B2: Plant-Level Soviet Specialization and Employment Growth, 1989–1995

(1) (2) (3) (4)

Outcome:

Employment, All Workers

Outcome:Employment,

Production Workers

Outcome: Employment,

Non-Production Workers

Observations

1989-1990 -0.079 -0.042 -0.036 3,003 (0.057) (0.026) (0.034) 1990-1991 -0.138** -0.079** -0.059** 2,965 (0.054) (0.029) (0.029) 1991-1992 -0.052 -0.038 -0.014 2,678 (0.035) (0.025) (0.012) 1992-1993 0.001 0.003 -0.002 2,371 (0.020) (0.014) (0.008) 1993-1994 -0.059 -0.043 -0.015 2,153 (0.053) (0.039) (0.016) 1994-1995 -0.008 -0.001 -0.007 1,845

(0.024) (0.019) (0.012) Notes: Coefficients in each row in columns 1–3 are from separate plant-level regressions of the annual change in employment of a worker category indicated by the column title on plant-level 1990 Soviet specialization in percentage points (PSS in equation (4) multiplied by 100). All specifications control for municipality fixed effects. Standard errors clustered by municipality are in parentheses. The 90%, 95%, and 99% confidence levels are denoted by *, **, and ***, respectively.

43

Table B3: Additional First-Stage Estimates for the Instruments

(1)

(2) (3)

Specification

Corresponding to Column 5 in

Table 2

Specification Corresponding to Column 6

in Table 2

Marginal Effects for Column 2

ln( ) ∗ Dummy(Year=1991) 0.020**

(0.007) 0.084** (0.038)

0.019 (0.013)

ln( ) ∗ Dummy(Year=1992) -0.009 (0.006)

0.008 (0.029)

0.007 (0.010)

ln( ) ∗ Dummy(Year=1993) 0.021** (0.011)

0.007 (0.030)

0.017 (0.011)

ln( ) ∗ Dummy(Year=1994) -0.003 (0.009)

-0.026 (0.031)

-0.013 (0.011)

ln( ) ∗ Dummy(Year=1995) 0.006 (0.007)

-0.002 (0.047)

0.015 (0.016)

ln( ) ∗ Dummy(Year=1991) 0.010** (0.004)

ln( ) ∗ Dummy(Year=1992) 0.000 (0.003)

ln( ) ∗ Dummy(Year=1993) -0.002 (0.003)

ln( ) ∗ Dummy(Year=1994) -0.002 (0.003)

ln( ) ∗ Dummy(Year=1995) -0.003 (0.005)

N 2,881 2,881 2,881

Notes: First-stage coefficients on the instruments. The dependent variable is the annual log-change in the relative production labor unit cost from year − 1 to . The year dummy label in each row corresponds to year . The sample includes plants which were not dependent on Soviet import demand. Data for the years 1991-1995. Local Soviet specialization ( ) is the fraction of a municipality’s Soviet exports in 1990, predicted by the 1988 output structure, to the municipality’s 1990 output (equation (5)). All regressions include controls for log capital intensity and industry×year dummies. Column 3 shows marginal effects for the specification in column 2, evaluated at the sample mean of local Soviet specialization. Panel-robust standard errors are in parentheses. The 90%, 95%, and 99% confidence levels are denoted by *, **, and ***, respectively.

44

Table B4: Testing For Differences in Annual Wage Earnings

AllProduction

Occupations

Non-Production

Occupations

Left Soviet-dependent plant in 1991 (0,1) -346.4

(353.5) -486.8 (395.7) 359.8

(307.9)

Observations 327,932 211,476 116,436

Notes: This table shows results from individual-level regressions of annual wage earnings on a dummy variable equal to one if an individual worked for a Soviet-dependent plant in 1990 ( , > 0.1) and left it in 1991; and equal to zero if he/she worked in a plant not dependent on Soviet import demand in 1990 ( , < 0.001). Data are from FLEED, cover the period 1992–1995, and include worker-year observations for which annual wage earnings are positive and the individual did not receive unemployment benefits. Wage earnings are in euro. All specifications control for municipality×year fixed effects. Standard errors clustered by municipality are in parentheses.

45

Table B5: Additional Robustness Analysis Results

Alternative Thresh-olds for the Predicted

Plant-Level Soviet Export Share (PSS)

Alternative Thresholds for

the Predicted Share of Input Supply to the Neighboring Soviet-Dependent Industry

(1) (2) (3) (4) (5)

Baseline 0.025% 0.25% 0.1% 5%

0.146** 0.146** 0.146** 0.161** 0.157**

(0.058) (0.068) (0.058) (0.059) (0.056)

-0.019** -0.031** -0.019** -0.017* -0.017*

(0.009) (0.015) (0.009) (0.009) (0.009)

Observations 2,813 1,402 3,669 2,694 2,750 Notes: The specifications are based on the specification in column 7 of table 2. and are coeffi-cients on ∆ ln( / ) and ∆ ln( / ), respectively. Column 1 replicates the estimates in column 7 of table 2. The specification in column 2 (3) corresponds to the specification in column 1 but excludes plants with the predicted plant-level Soviet export share more than 0.025% (0.25%) of output in 1990. The specification in column 4 (5) corresponds to the specification in column 1 but excludes plants with the predicted input supply to the local Soviet-dependent industry more than 0.1% (5%) of output in 1990. Panel-robust standard errors are in parentheses. The 90%, 95%, and 99% confidence levels are denoted by *, **, and ***, respectively.

46

Table B6: Estimated Changes in the Relative Demand for Production Labor (%)

1980–2008 1980–1989 1989–1999 1999–2008

Demand Cost Share Demand Cost

Share Demand Cost Share

Demand Cost Share

All Industries -2.5 -6.0 -2.8 -6.1 -1.2 -4.3 -2.9 -6.0

Textiles -4.9 -6.5 -3.4 -3.1 -4.1 -6.8 -6.0 -7.5

Electrical and Optical Equipment -4.7 -13.3 -4.4 -13.6 0.0 -8.9 -8.7 -13.8

Transport Equipment -4.2 -1.2 -4.1 -5.0 -1.5 5.3 -6.1 -4.1

Wood -2.5 -2.9 -3.9 -5.7 -1.0 2.2 -1.9 -4.9

Rubber and Plastic Products -2.2 -1.5 -2.6 -3.3 -0.8 2.0 -2.6 -3.1

Chemicals -2.1 -3.9 -1.2 -4.7 -4.7 -2.7 0.5 -3.4

Machinery and Equipment -2.1 -5.4 -3.2 -6.5 0.7 -4.6 -3.3 -3.7

Paper, Publishing, and Printing -1.9 -3.3 -1.7 -5.1 -1.8 -0.6 -1.7 -3.6

Food, Beverages, and Tobacco -1.5 -1.3 -3.4 -2.9 -2.9 -1.0 2.3 0.5

Metal Products -1.3 -1.4 -2.6 -3.6 -0.1 0.4 -1.1 -0.6

Non-Metallic Mineral Products -0.2 -1.2 -2.0 -0.7 1.6 -0.2 -0.4 -2.4

Notes: This table displays estimated changes in the relative demand for production labor (corresponding to figure 4) and changes in the production labor cost share by industry and time period. All changes are converted to decadal rates.

47

Table B7: The Effects of Offshoring and ICT on the Relative Demand for Production Labor, OLS Estimates

(1) (2) (3) (4) (5) (6) Covariate lag (years): 0 1 2 0 1 2

Offshoring 0.010 -0.000 -0.020*** -0.000 -0.009 -0.017***

(0.010) (0.007) (0.007) (0.010) (0.007) (0.006) ICT -0.007 -0.006 -0.009* -0.005 -0.002 -0.009

(0.007) (0.006) (0.005) (0.006) (0.006) (0.006) Observations 171 171 171 171 171 171

Notes: OLS estimates weighted by industry labor cost. 2-digit industry data for the years 1999–2007. The table shows OLS estimates from regressions of the estimated relative production labor demand index at the 2-digit industry level on logs of industry offshoring and ICT. Columns 1–3 include industry fixed effects and industry time trends, while columns 4–6 add year dummies. All specifications control for industry R&D intensity. Panel-robust standard errors are in parentheses. The 90%, 95%, and 99% confidence levels are denoted by *, **, and ***, respectively.

48

Table B8: The Effects of Offshoring and ICT on the Relative Demand for Production Labor, IV Estimates

(1) (2) (3) (4) (5) (6)

Covariate lag (years): 0 1 2 0 1 2

A. Endogenous Variable: Offshoring

Offshoring 0.010 0.009 -0.022 -0.021 -0.025 -0.029 (0.019) (0.013) (0.017) (0.027) (0.030) (0.023)ICT -0.007 -0.006 -0.009* -0.004 -0.001 -0.009 (0.007) (0.007) (0.005) (0.007) (0.004) (0.005)1st Stage:

US Offshoring, t-3 0.422*** 0.422*** 0.422*** 0.255** 0.255** 0.255** (0.104) (0.104) (0.104) (0.094) (0.094) (0.094) B. Endogenous Variable: ICT

Offshoring 0.010 -0.001 -0.020*** -0.003 -0.014 -0.018**

(0.009) (0.006) (0.006) (0.008) (0.009) (0.007)ICT -0.001 0.005 -0.024** 0.006 0.021 -0.012 (0.007) (0.012) (0.011) (0.023) (0.021) (0.019)1st Stage: US Computer Services, t-3 0.370** 0.370** 0.370** 0.343* 0.343* 0.343*

(0.174) (0.174) (0.174) (0.181) (0.181) (0.181)

C. Endogenous Variables: Offshoring and ICT

Offshoring 0.012 0.010 -0.027** -0.030 -0.037 -0.032 (0.016) (0.010) (0.011) (0.108) (0.060) (0.033)ICT -0.003 -0.003 -0.019** 0.021 0.033 -0.002

(0.010) (0.010) (0.009) (0.077) (0.050) (0.022)1st Stage for Offshoring: US Offshoring, t-3 0.371*** 0.371*** 0.371*** 0.216** 0.216** 0.216** (0.086) (0.086) (0.086) (0.098) (0.098) (0.098)US Computer Services, t-3 0.183** 0.183** 0.183** 0.161* 0.161* 0.161* (0.068) (0.068) (0.068) (0.090) (0.090) (0.090)1st Stage for ICT: US Offshoring, t-3 -0.279* -0.279* -0.279* -0.011 -0.011 -0.011 (0.141) (0.141) (0.141) (0.258) (0.258) (0.258) US Computer Services, t-3 0.355** 0.355** 0.355** 0.356** 0.356** 0.356** (0.168) (0.168) (0.168) (0.159) (0.159) (0.159) Observations 171 171 171 171 171 171Notes: 2-digit industry data for the period 1999–2007. The table shows IV estimates from regressions of the estimat-ed relative production labor demand index at the 2-digit industry level on logs of industry offshoring and ICT. All regressions are weighted by industry labor cost. Panel A uses the log of U.S. imported intermediate inputs from own industry lagged three years as an instrument. Panel B uses the log of U.S. input usage from NAICS industry 5415 (computer systems design and related services) lagged three years as an instrument. Panel C uses both instruments. Columns 1–3 include industry fixed effects and industry time trends while columns 4–6 add year dummies. All specifications control for industry R&D intensity. Panel-robust standard errors are in parentheses. The 90%, 95%, and 99% confidence levels are denoted by *, **, and ***, respectively. Angrist-Pischke first-stage F-Statistics: Panel A: 16.47 (columns 1–3) and 7.312 (columns 4–6). Panel B: 4.523 (columns 1–3) and 3.569 (columns 4–6). Panel C: 19.398 for offshoring and 5.838 for ICT (columns 1–3) and 5.057 for offshoring and 2.247 for ICT (columns 4–6).

49

Table B9: The Effects of Offshoring and ICT on the Relative Demand for Production Labor, Alternative IV Estimates Excluding R&D Intensity (1) (2) (3) (4) (5) (6)

Covariate lag (years): 0 1 2 0 1 2 A. Endogenous Variable: Offshoring

Offshoring 0.010 0.010 -0.023 -0.020 -0.018 -0.021 (0.019) (0.014) (0.017) (0.023) (0.043) (0.031) ICT -0.008 -0.008 -0.009** -0.004 -0.003 -0.009 (0.006) (0.007) (0.004) (0.006) (0.004) (0.005) 1st Stage: US Offshoring, t-3 0.414*** 0.414*** 0.414*** 0.248** 0.248** 0.248**

(0.109) (0.109) (0.109) (0.095) (0.095) (0.095) B. Endogenous Variable: ICT

Offshoring 0.009 -0.003 -0.018** -0.004 -0.017* -0.016* (0.008) (0.005) (0.007) (0.008) (0.010) (0.008) ICT -0.001 0.007 -0.025** 0.008 0.025 -0.014 (0.006) (0.014) (0.011) (0.024) (0.021) (0.022) 1st Stage: US Computer Services, t-3 0.348* 0.348* 0.348* 0.305* 0.305* 0.305*

(0.193) (0.193) (0.193) (0.154) (0.154) (0.154)

C. Endogenous Variables: Offshoring and ICT

Offshoring 0.013 0.012 -0.029*** -0.028 -0.027 -0.028 (0.015) (0.011) (0.010) (0.034) (0.058) (0.035) ICT -0.003 -0.004 -0.018* 0.020 0.030 -0.005

(0.010) (0.010) (0.009) (0.044) (0.041) (0.020) 1st Stage for Offshoring: US Offshoring, t-3 0.355*** 0.355*** 0.355*** 0.207* 0.207* 0.207* (0.089) (0.089) (0.089) (0.099) (0.099) (0.099) US Computer Services, t-3 0.184** 0.184** 0.184** 0.157 0.157 0.157 (0.076) (0.076) (0.076) (0.093) (0.093) (0.093) 1st Stage for ICT: US Offshoring, t-3 -0.329** -0.329** -0.329** -0.060 -0.060 -0.060 (0.133) (0.133) (0.133) (0.259) (0.259) (0.259) US Computer Services, t-3 0.358* 0.358* 0.358* 0.336** 0.336** 0.336** (0.195) (0.195) (0.195) (0.141) (0.141) (0.141) Observations 171 171 171 171 171 171Notes: Specifications correspond to table B8 but exclude R&D intensity as a control variable. The 90%, 95%, and 99% confidence levels are denoted by *, **, and ***, respectively. Angrist-Pischke first-stage F-statistics: Panel A: 14.37 (columns 1–3) and 6.775 (columns 4–6). Panel B: 3.251 (columns 1–3) and 3.947 (columns 4–6). Panel C: 16.158 for offshoring and 5.992 for ICT (columns 1–3) and 5.195 for offshoring and 3.362 for ICT (columns 4–6).

50

References

Acemoglu, D., and Autor, D. (2011). Skills, tasks and technologies: implications for employ-ment and earnings. In Handbook of Labor Economics, Vol. 4, ed. Ashenfelter, O., and Card, D. Amsterdam: Elsevier.

Acemoglu, D., Autor, D. H., and Lyle, D. (2004). Women, war, and wages: the effect of female labor supply on the wage structure at midcentury. Journal of Political Economy, 112(3), 497-551.

Angrist, J. D. (1995). The economic returns to schooling in the West Bank and Gaza Strip. American Economic Review, 85(5), 1065-87.

Angrist, J. D. (1996). Short-run demand for Palestinian labor. Journal of Labor Economics, 14(3), 425-53.

Angrist, J. D., and Pischke, J. (2009). Mostly Harmless Econometrics: An Empiricist's Com-panion. Princeton: Princeton University Press.

Autor, D., Katz, L. F., and Kearney, M. S. (2008). Trends in US wage inequality: revising the revisionists. Review of Economics and Statistics, 90(2), 300-23.

Autor, D., Levy, F., and Murnane, R. J. (2003). The skill content of recent technological change: an empirical exploration. Quarterly Journal of Economics, 118(4), 1279-333.

Baltagi, B. H., and Rich, D. P. (2005). Skill-biased technical change in US manufacturing: a general index approach. Journal of Econometrics, 126(2), 549-70.

Berman, E., Bound, J., and Griliches, Z. (1994). Changes in the demand for skilled labor within U.S. manufacturing: evidence from the Annual Survey of Manufacturers. Quarterly Journal of Economics, 109(2), 367-97.

Borjas, G. J. (2003). The labor demand curve is downward sloping: reexamining the impact of immigration on the labor market. Quarterly Journal of Economics, 118(4), 1335-74.

Borjas, G. J., and Doran, K. B. (2012). The collapse of the Soviet Union and the productivity of American mathematicians. Quarterly Journal of Economics, 127(3), 1143-203.

Bustos, P. (2011). Trade liberalization, exports, and technology upgrading: evidence on the impact of MERCOSUR on Argentinian firms. American Economic Review, 101(1), 304-40.

Ciccone, A., and Peri, G. (2005). Long-run substitutability between more and less educated workers: evidence from US states, 1950–1990. Review of Economics and Statistics, 87(4), 652-63.

Duranton, G., and Puga, D. (2004). Micro-foundations of urban agglomeration economies. In Handbook of Urban and Regional Economics, vol. 4, ed. Henderson, J. V., and Thisse, J. Amsterdam: North-Holland.

51

European Commission (2010). Employment in Europe 2010. Luxembourg: Publications Office of the European Union.

Falck, O., Guenther, C., Heblich, S., and Kerr, W. R. (2013). From Russia with love: the impact of relocated firms on incumbent survival. Journal of Economic Geography, 13(3), 419-49.

Feenstra, R. C., and Hanson, G. H. (1996). Globalization, outsourcing, and wage inequality. American Economic Review, 86(2), 240-45.

Frisch, R. (1933). Pitfalls in the Statistical Construction of Demand and Supply Curves. Leipzig: Hans Buske, Verlag.

Glaeser, E. L., and Gottlieb, J. D. (2009). The wealth of cities: agglomeration economies and spatial equilibrium in the United States. Journal of Economic Literature, 47, 983–1028.

Glitz, A. (2012). The labor market impact of immigration: a quasi-experiment exploiting immigrant location rules in Germany. Journal of Labor Economics, 30(1), 175-213.

Goos, M., Manning, A., and Salomons, A. (2014). Explaining job polarization: routine-biased technological change and offshoring. American Economic Review, 104(8), 2509-26.

Gorodnichenko, Y., Mendoza, E. G., and Tesar, L. L. (2012). The Finnish Great Depression: from Russia with love. American Economic Review, 102(4), 1619-43.

Greenstone, M., Hornbeck, R., and Moretti, E. (2010). Identifying agglomeration spillovers: evidence from winners and losers of large plant openings. Journal of Political Economy, 118(3), 536-598.

Hamermesh, D. S. (1993). Labor Demand. Princeton: Princeton University Press. Hamermesh, D. S., and Trejo, S. J. (2000). The demand for hours of labor: direct evidence

from California. Review of Economics and Statistics, 82(1), 38-47. Hahn, J., and Hausman, J. (2003). Weak instruments: diagnosis and cures in empirical econ-

ometrics. American Economic Review, 93(2), 118-25. Haskel, J. E., and Slaughter, M. J. (2002). Does the sector bias of skill-biased technical

change explain changing skill premia? European Economic Review, 46(10), 1757-83. Heathcote, J., Storesletten, K., and Violante, G. L. (2010). The macroeconomic implications

of rising wage inequality in the United States. Journal of Political Economy, 118(4), 681-722.

Heathcote, J., Storesletten, K., and Violante, G. L. (2017). Optimal tax progressivity: An analytical framework. Quarterly Journal of Economics, 132(4), 1693-754.

Heckman, J. J., Lochner, L., and Taber, C. (1998). Explaining rising wage inequality: explo-rations with a dynamic general equilibrium model of labor earnings with heterogeneous agents. Review of Economic Dynamics, 1(1), 1-58.

52

Katz, L. F., and Murphy, K. M. (1992). Changes in relative wages, 1963–1987: supply and demand factors. Quarterly Journal of Economics, 107(1), 35-78.

Kolesár, M. (2012). Minimum distance approach to inference with many instruments. Un-published mimeo, Princeton University.

Kolesár, M., Chetty, R., Friedman, J. N., Glaeser, E. and Imbens, G. W. (2015). Identification and inference with many invalid instruments, Journal of Business and Economic Statistics, 33, 474-84.

Kroft, K., Kucko, K. J., Lehmann, E., and Schmieder, J. F. (2015). Optimal income taxation with unemployment and wage responses: a sufficient statistics approach. NBER Working Papers: No. 21757. National Bureau of Economic Research.

Lee, D., and Saez, E. (2012). Optimal minimum wage policy in competitive labor markets. Journal of Public Economics, 96(9-10), 739-49.

Lichter, A., Peichl, A., and Siegloch, S. (2015). The own-wage elasticity of labor demand: a meta-regression analysis. European Economic Review, 80, 94-119.

Machin, S., and Van Reenen, J. (1998). Technology and changes in the skill structure: evi-dence from seven OECD countries. Quarterly Journal of Economics, 113(4), 1215-44.

Michaels, G., Natraj, A., and Van Reenen, J. (2014). Has ICT polarized skill demand? evi-dence from eleven countries over twenty-five years. Review of Economics and Statistics, 96(1), 60-77.

Moretti, E. (2011). Local labor markets. In Handbook of Labor Economics. Vol. 4, ed. Ash-enfelter, O., and Card, D. Amsterdam: Elsevier.

Stock, J. H., Wright, J. H., and Yogo, M. (2002). A survey of weak instruments and weak identification in generalized method of moments. Journal of Business and Economic Sta-tistics, 20(4), 518-29.

Topalova, P. (2010). Factor immobility and regional impacts of trade liberalization: evidence on poverty from India. American Economic Journal: Applied Economics, 2(4), 1-41.