ZEF Bonn · West Nile 37.7 37.8 Western 37.6 43.9 South Western 49.6 41.7 Karamoja 53.6 45.0...

1

Impact of Community-Based Health Insurance on Child

Health Outcomes? Evidence on Stunting from Rural

Uganda.

Emmanuel Nshakira Rukundo1, Nicolas Gerber

May 2017

Draft, please do not cite.

Abstract Community-based health insurance has increasingly been integrated into health systems in developing countries. However, there remains limited research on its probable health outcome impacts beyond the conventional health financing and facilitating services access functions. Using a cross-sectional survey, we apply a two- stage residual inclusion instrumental variables method to study the impact of community health insurance on under-five stunting in rural Uganda. Results indicate that in a consistently linear relationship, each year of insurance contributes to about 15.4% less probability of being stunted. Predictive marginal effect reflect that children in households which have had insured for 7 years had a probability of stunting of only 0.329 compared to 0.518 for children in households which never had insurance. The insurance reduction rate of 15.4% per year scaled up to 5 years implies that a child born in a household with insurance would have 77% less probability of stunting at his or her fifth birthday. Our study recommends that developing countries should facilitate the expansion of community health insurance scheme not only for their contribution to health financing but even more for mortality and morbidity aversion. JEL Codes: I130, I150, I100 Keywords: Community-Based Health Insurance; Stunting; Instrumental Variables

1PhD Candidate & Junior Researcher, Centre for Development Research (ZEF) University of Bonn. Walter-Flex.Str. 3, 53113 Bonn, Germany [email protected] or [email protected]

ZEF Bonn

Zentrum für Entwicklungsforschung Center for Development Research Universität Bonn

mailto:[email protected]

mailto:[email protected]

2

1. Introduction

Community-Based Health Insurance (CBHI) schemes are a particular form of health insurance

systems emanating from community social support systems (Criel et al., 2004), often working in the

rural informal sector, operating without profit motivations (Bennett et al., 1998). The schemes evolved

in the 1980s especially in resource-poor developing countries where tax funded and other health

insurance platforms were non-existent (Carrin et al., 2005; Ekman, 2004); and have over the last few

decades evolved as essential buffers for financial protection and enabling health services access,

especially in poor rural communities. They currently form essential building blocks to achieving

universal health coverage in developing countries (Jacobs et al., 2008; Titelman et al., 2015; Wang &

Pielemeier, 2012) and have been adopted across many countries with varying government

involvement.

Studies on community-based health insurance have proliferated in health economics and health policy

literature, often discussing issues on enrolment and drop out (Atinga et al., 2015; De Allegri et al.,

2006; Dror et al., 2016; Mebratie et al., 2015; Panda et al., 2014), facilitating access to health services

(Jütting, 2004; Mebratie et al., 2013; Smith & Sulzbach, 2008; Sood & Wagner, 2016), financial

protection (Nguyen et al., 2011; Nguyen et al., 2012; Sepehri, 2014; Sepehri et al., 2011), and wider

welfare and economy-wide impact on managing shocks and other risk-coping mechanisms during

economic shocks (Asfaw & Braun, 2004; 2004; Landmann & Frölich, 2015; Yilma et al., 2014; Yilma

et al., 2015). However, research evidence remains thin on the impact of community health insurance

on health outcomes such as disease reduction and improved health indicators in children and mothers,

or improved adoption of preventive health behaviours. Several systematic reviews (Adebayo et al.,

2015; Dror et al., 2016; Ekman, 2004; Mebratie et al., 2013; Reshmi et al., 2016; Spaan et al., 2012) do

not report any evidence of effects on health outcomes and only six of nineteen papers reviewed by

Acharya et al (2013) report some health outcomes. The basic question asked by Dror (2014) “What

impact does insurance related improvements in healthcare utilization have on the health of the target

populations?” remains largely unanswered. Nonetheless, there is emerging literature in response to

this research gap. For instance, Hendriks et al (2014, 2016) study the effect on systolic and diastolic

blood pressure in Nigeria while Sood & Wagner (2016) study the effect on post-operative recovery in

India. In terms of child nutrition related health outcomes, we find only one paper, Lu et al (2016) who

study the effect on stunting in Rwanda. Two major differences of our study from that of Lu et al

3

(2016) is that whereas CBHI in Rwanda is not legally compulsory and hence contributed to universal

coverage (Makaka et al., 2012; Nyandekwe et al., 2014), CBHI in Uganda is voluntary. Secondly, Lu

and colleagues use data collected at health facilities which only captures information about children

who visit the health facilities and excludes children whose information was not captured at health

facility level, and hence carries components of self-selection.

Conversely, our research adds to this thin layer of evidence by looking at the impact of health

insurance on stunting among under-5 children in rural south-western Uganda using cross-sectional

data from households and employing instrumental variables overcome endogeneity problems. Our

major contribution to health economics literature is twofold. First, we show that CBHI can influence

other health dimensions beyond its primary goals of resource mobilisation for health systems and

financial protection for households in developing countries. In particular, we show that CBHI

significantly reduces the probability of stunting, a serious problem in developing countries including

Uganda. Secondly, we employ a novel instrumental variables estimation – the Two-Stage Residual

Inclusion (Terza et al., 2008) and following Terza's (2016) Stata implementation code, we combine

both the effect of a treatment – in our case, insurance status and treatment intensity – in our case, the

number of years in insurance, on child stunting. We believe that by asking and answering these

questions, we aim to generate more evidence on the currently weak links between CBHI and health

outcomes.

The rest of this paper is organised as follows. Section two briefly discusses the stunting problem with

a focus on Uganda and section three reviews previous studies on health insurance and health outcomes

with a focus on community health insurance in developing countries. Section four details our

instrumental variable estimation strategy while section five provides our results. A short discussion

and conclusion are presented in section six.

2. The Stunting Problem in Uganda

Low height for age affects an estimated 165 – 170 million children in the world and affects developing

countries disproportionately (Prendergast & Humphrey, 2014; Stevens et al., 2012). The effects are

not only detrimental to a child’s young life but also sustain into adulthood, affecting educational,

health, employment and cognitive abilities later in life (Case & Paxson, 2010; Dewey & Begum, 2011;

Glewwe et al., 2001; Glewwe & Miguel, 2008; Vogl, 2014). Uganda grapples with this problem with

4

over 33% all the children in the country stunted (UBOS & ICF International, 2012). Using the 2014

census (17.7% of 34.8 million under 5 years), this stunting prevalence would translate into 1.8 million

stunted children. It is not simply high but also increasing in some regions such as Central 2 and

Western. The south-western region, where this research took place had the third highest prevalence

rate of 41.7% in 2011. Nutrition practitioners warn that if the current status of stunting is not

improved, stunting could lead to loss of more than half a million lives between 2013 and 2025

(Namugumya et al., 2014). Such a tragedy will disproportionately affect rural households.

Table 1: Stunting in Uganda2

Region Stunting (% of U-5s) 2006 2011

Kampala 22.2 13.5 Central 1 39.2 32.5 Central 2 29.8 36.1 East Central 38.3 33.5 Eastern 36.2 25.3 Northern 40.0 24.7 West Nile 37.7 37.8 Western 37.6 43.9 South Western 49.6 41.7 Karamoja 53.6 45.0 Average 38.1 33.4

Source: Demographic and Health Surveys, 2006 and 2011.

3. Health Insurance and Health Outcomes3

Evidence regarding CBHI effects on health outcomes in developing countries is still scarce and very

recent. For instance, a compilation of research studies about the impacts of health insurance provided

only four studies directly dealing with impacts on health outcomes against 31 and 27 studies dealing

with impacts on the utilisation of care and financial protection respectively (Morsink, 2012). Even

when more recent studies are incorporated on this list, we find only a handful of papers that attempt

to separate health services access from actual health outcomes. This emerging evidence is inconclusive.

Fink et al (2013) analysed the impact of the Nouna scheme in Burkina Faso on mortality and found

2 At the time of writing this paper, preliminary results from the 2016 demographic and health survey indicate stunting prevalence of 29%. However, the due a change in regional reclassification, we have not included those results here. 3 In the literature (Jacobs et al., 2008; Spaan et al., 2012; H. Wang & Pielemeier, 2012), there have been efforts to classify different types of health insurance such community-based health insurance, social health insurance and micro health insurance. However, these classifications also seem to be used interchangeably. For the purposes of this literature review, while focusing on developing countries, we look at these wholesomely from their sole purpose of targeting poor, rural and informal sub-populations.

5

that insurance appeared to have increased overall mortality rather than reduce it. They find that overall

population mortality increased by more than 20% in villages in insurance when compared to overall

average with the highest effect seen in the over 65 age group. However, the study falls short of making

strong conclusions on whether increased mortality was due to insurance. Conversely, Sood et al (2014)

studying the effect of the Vajpayee Arogyashree Scheme (VAS) on mortality in India showed more

positive results from their large sample of over 60,000 households. They found a 0.58 percentage point

difference in mortality from covered conditions in insurance villages compared to non-insurance

villages, which represented a 64% risk reduction. They further looked at the age-specific mortality

caused by caused by insurance-covered conditions and found that the eligible households 52% of the

mortality was in the <60 years population compared to 76% in the same age classification for the

ineligible households implying that overall health services improved in insured villages. Keng & Sheu

(2013) also studied the impact of Taiwan national insurance on mortality of elderly people using a

triple difference method and compared before and after insurance reforms. They found that death

hazard reduced by 16-48% in the least healthy people and by 3-9% in the healthiest people. They,

however, did not find any positive impact on insurance on self-assessed health and functional health.

Sood & Wagner (2016) also studied the effect VAS insurance on post-operative health and recovery

and found that insured people reported more positive improvements after an operation in all the six

categories of self-assessed health (self-care, usual activities, walking ability, pain, anxiety and overall

health) and statistical significance in three categories – walking ability, pain, and anxiety. Insured

people were also 9.4 percentage points less likely to report an infection after their stay in a hospital,

16.5 percentage points less likely to be re-hospitalized and 5 percentage points more likely to seek

treatment. In Nigeria, Hendrik and colleagues studied the impact of health insurance on hypertension

(Hendriks et al., 2014, 2016). In the first study, Hendriks et al (2014) assessed health outcomes using

changes in blood pressure in two regions in Nigeria. Using the difference in difference methodology,

they found that systolic blood pressure reduced by 5.2 mm Hg more in treatment villages compared

to the control villages Hg while diastolic blood pressure also reduced twice as much. Sustained effects

measured after 5 years indicated that systolic blood pressure maintained a 4.97 mm Hg greater

reduction (Hendriks et al., 2016).

Focusing on health outcomes specifically for children, though the research is scarce, a handful of

papers try to unearth impacts and the general consensus is positive. Two papers evaluating the national

insurance scheme in Rwanda found some positive effects on early childhood stunting and infant

6

mortality. Binagwaho et al (2012) used two rounds of the Rwanda Integrated Households and Living

Standards survey and found height-for-age z-scores (measuring stunting) gains of 0.42cm for children

between 6 and 24 months. They further found that infants (children below one year) were 3-14

percentage points less likely to die in their infancy. A more recent paper on Rwanda also reported that

the probability of stunting for children enrolled in insurance was 14 percentage points lower than

those not enrolled (Lu et al., 2016). A study conducted in the Philippines indicated that insurance

reduced the likelihood of wasting in children by 9-12 percentage points (Quimbo et al., 2011). In

Ghana, Gajate-Garrido & Ahiadeke (2015) find that being insured increased the possibility of children

taking anti-malaria medicine by 25.2% and malaria-related child care seeking increased by 29.5%.

Gajate-Garrido & Ahiadeke view this as an impact of insurance on investing in curative child health

services. Still, in Ghana, Singh et al (2015) found that caregivers with insurance were more likely to

take a child to a health facility even with the slightest of illnesses – an indication of better health

seeking behaviour. In Burkina Faso, (Schoeps et al., 2015) analysis based on data from over 33,000

children found that the risk of mortality was 46% lower for insured children compared to their

uninsured counterparts.

Overall, the question of insurance impacts on health outcomes is more or less still unanswered.

Evidence in this area is still scanty and only emerging especially in areas of foundational importance

such as maternal and child health outcomes. Whilst we do not diminish the importance of other health

outcomes in the broader health spectrum, maternal and child health services remain a priority for

developing countries and learning how insurance could nudge improvements in this area should

remain a priority in both academics and policy-making. For East Africa, a region where CBHI policy

making has gained momentum (Basaza et al., 2013; Abuya et al., 2015) and contributed to almost

universal health coverage in Rwanda (Lu et al., 2012) and where good case studies have existed for a

few decades now, such evidence is required to make informed policies as well as demonstrate impact.

Using our case study from rural south-western Uganda, we hope to make this contribution to literature

– to study the effect of community-based health insurance on child growth indicators, specifically,

child stunting.

7

4. The Data and Estimation Strategy

4.1. Data

Between August 2015 and April 2016, we carried out a household survey in south-western Uganda in

two districts in which the Kisiizi hospital CBHI scheme operates. The Kisiizi CBHI scheme is one of

the 21 and the biggest CBHI scheme in Uganda, providing insurance coverage to over 7,200

households. The scheme is run by a rural hospital and enrolment is based on membership in burial

groups. The enrolment unit is a household. Households pay premiums ranging from an equivalent of

$4 to $8 depending on household size. Membership in insurance is based on group such that for a

household to be insured, it has to be part of group. Most of the groups are pre-existing burial societies.

For a burial society to join insurance, two conditions have to be met. The first is that the group has

to have at least 30 households. For these smaller groups, all the households have to subscribe. The

second condition pertains to larger groups. In these larger groups, some with more than 100

households, at least 50% of households have to subscribe and all members of the subscribing

households are required to enrol. Another important feature of this scheme is the timing of coverage.

In order to further control moral hazard and adverse selection, members are fully covered once they

have been in insurance for more than one year. Once a household member is ill with in only one year

of enrolling, insurance covers only 10% of the cost. These kinds of strategies to limit adverse selection

have been applied in other community insurance scheme. For instance in a Nigerian scheme, enrolled

people wait for about 36 days before they can be covered by insurance (Bonfrer et al., 2015). In the

case of this scheme, a waiting time of one year is certainly more conservative. Typically, insurance

covers basic primary care, maternity care, surgeries, and outpatient and inpatient services. Outpatient

services for chronic illnesses and substance abuse related illnesses and injuries are excluded.

In partnership with this scheme, we used a multi-stage simple random sampling, and surveyed 464

households in fourteen (14) villages in two (2) districts. The survey modules included a household

demographic module which collected information on household occupancy; a child and maternal

health module which collected information on health care seeking behaviour for mothers and children;

a nutrition module which collected information on household food availability and intake. The survey

collected detailed information on household social and economic welfare using durable assets holdings

and other endowments in agriculture, water and sanitation, and housing. The health insurance and

social connectivity modules collected information regarding household insurance status, group

membership and participation, and knowledge of insurance such as premiums and benefit package. In

8

line with emerging tools for understanding enrolment in community insurance in sub-Saharan Africa,

the survey incorporated a detailed module on perceptions about several aspects of health insurance.

Ethical approvals for data collection was conducted by the Mengo Hospital Research and Ethics

Review Committee and an ethical certificated was provided by the Uganda National Council of

Science and Technology (Reference Number SS-39369). Further ethical reviews were undertaken by

the Centre for Development Research, University of Bonn research committee and the local research

partner, Kisiizi Hospital Research Committee. Verbal consent was obtained from the district

administration health teams, local government authorities, village leaders and respondents.

4.1.1. Outcome variable: Child stunting

Our main outcome of interest in this paper is stunting in under-five children. Moderate and severe

stunting affects about 170 million children worldwide and has increased in several sub-Saharan Africa

countries (Stevens et al., 2012). We estimate stunting by means of a dummy of 1 if stunted – in that

the height-for-age z-scores are below -2 standard deviations and 0 otherwise using the World Health

Organisation 2006 reference points (Duggan, 2010; Garza et al., 2006). With large samples, linear

growth improvements can be estimated using a continuous variable of height-for-age z-score (Spears,

2012), however this is sometimes inappropriate in small samples and also becomes harder for policy

relevant explanations (Wang & Chen, 2012). We therefore use a dichotomous classification of stunting

as has been used in other similar studies (Lu et al., 2016).

4.1.2. Independent variables

Exposure: Our treatment exposure was insurance measured by insurance status and the number of

years a household has been in insurance. Insurance in Kisiizi is provided by the Kisiizi community

health insurance scheme, a hospital managed non-profit community scheme. As predetermined by the

scheme history, insurance is provided through groups, mainly pre-existing burial groups. Groups have

governing committees which admit new members and report deregistration of members who drop

out from insurance due to non-payment of premiums or death. First, we constructed a dummy for

insurance status where 1 was if a household was registered as insured and 0 otherwise. We then asked

about the number of years the household has been continuously insured. Our measurement, therefore,

combined both the exposure and the extent of the exposure. These kinds of variables fit well into our

modeling technique of choice – the two-part count data model (Mullahy, 1997, 1998) integrating a

two-stage residual inclusion instrumental variable model (Terza et al., 2008).

9

Other Covariates: Our other covariates can be categorised into four categories, namely; child specific

covariates, household covariates, spatial covariates and village level covariates.

Child-specific covariates: we included the age of the child in months, ranging from 6 to 59 months.

We include the gender dummy that specifies 1 if male and 0 otherwise. Birth weight is an important

determinant of child growth (Wilcox, 2001) and had long term effects such as education and labour

market participation in adulthood (Behrman & Rosenzweig, 2004; Chatterji et al., 2014; Xie et al.,

2016). In our data, we recorded birthweight of 55.4% of the sample – from children who were born

in health facilities. For the remaining 44.6% whom we did not record birthweight, we undertook a

multiple imputation procedure using a linear regression (Rubin, 1987; Schenker & Taylor, 1996). We

impute under an assumption of missing not at random to build a complete dataset for birthweight.4

We also included dummies for exclusive breastfeeding for at least six months, child immunisations

and vitamin A supplementation and child sleeping under a mosquito net.

Household covariates: first, we controlled for household social economic status using a composite

health index that combines agricultural housing and energy, water and sanitation and household

durable asset endowments. We developed a household asset index from household durable assets,

agriculture assets, housing and water and sanitations, using principal components analysis (Filmer &

Pritchett, 1999, 2001; Vyas & Kumaranayake, 2006) Generally, wealth indices provide a robust

measurement of households’ socioeconomic wellbeing especially in the absence of good consumption,

expenditure and income data (Filmer & Pritchett, 1999, 2001; Filmer & Scott, 2012).

Using the PCA method again, we developed a composite perception index derived from 39 questions

that elicited perceptions about health insurance, an issue that has become central to health insurance

studies in developing countries (Abuosi et al., 2016; Jehu-Appiah et al., 2012; Mulupi et al., 2013). We

further included education status of the parents – an important determinant of child health outcomes

(Desai & Alva, 1998; Semba et al., 2008; Vollmer et al., 2016), religion, parental employment,

information access, household dietary diversity and variables for social connectivity.

Spatial covariates: Our research area is generally a highland area for which distance and altitude

present spatial access to social services. In order to control for these, we included altitude of the

4 See Annex 2 for our multiple imputation procedure.

10

household and the differential distance of the household from the insurance hospital relative to other

health facilities of a similar or comparable service level.

Village covariates: Finally, in order to control for geographical invariant determinants, we included

village level covariates. We included a categorical variable for village economy (categorised in forest

villages, trading villages, banana cultivating villages or pastoralist villages) number of insurance groups

in the village, insurance coverage rates in 2010 and dummies for village having a school, road, and a

health unit.

4.2. Identification strategy: Two stage Residual Inclusion estimator

A lot of literature attests to the fact that insurance increases health care utilisation, simply because, for

poor households, it removes financial barriers to access. In fact, in some instances, over consumption

of care might lead to more spending of health (Wagstaff & Lindelow, 2008), hence defeating the

financial protection purpose. Irrespective of the financial protection situation after insurance, we take

that consumption of more health services should, in principle, lead to improved health. But such a

causal relationship is difficult to establish because of endogeneity between health insurance and the

health outcomes of the insured people (Levy & Meltzer, 2004; 2008). Our identification strategy will

therefore utilise an instrumental variable approach to overcome endogeneity problems (Angrist,

Imbens, & Rubin, 1996; Angrist & Imbens, 1995). Instead of the conventional Two-Stage Least

Squares (TSLS) instrumental variable approach, we employed a Two-Stage Residual Inclusion (2SRI)

which is more robust in addressing endogeneity in health economics studies (Cai, Small, & Have, 2011;

Terza, 2017; Terza et al., 2008).

Moreover, our treatment variable measured two related parts; namely the treatment and the treatment

intensity, we used an alternative estimation of the 2SRI which combines these two parts, following

Terza's (2016, p. 34) Stata code implementation guidelines. It is a common occurrence that when

treatment is offered, treatment intensity varies across the treated (Angrist & Imbens, 1995) but also

the proportion of non-compliers is large and non-ignorable (Mullahy, 1998). In a case like this, Mullahy

(1998) proposes a two-part model that assumes that the probability of the treatment happening,

Pr(y>0 |x) is governed by a probability model like a Probit in the first part and that E[ln(y)|y>0, x]

and is a linear function of x, e.g., E[ln(y)|y> 0, x] = xβ (part two).

11

These two steps produce both the probability of the treatment and the predicted treatment intensity.

Residuals are then manually generated and included in the second stage outcome model. Using

Mullahy’s smoking and birthweight data, (Mullahy, 1997, 1998), Terza (2016) employs this version of

a non-linear instrumental variable model to analyse of the effect of maternal smoking on birthweight.

One other convenience of a 2SRI model is the ease with which the tests of endogeneity and

endogeneity control are done. It is important to make sure that the first stage F-statistic for the joint

significance of the instruments meets the threshold level of 10 (Staiger & Stock, 1997; Stock et al.,

2002; Stock & Yogo, 2005). In addition, when the included residuals are significant, it shows that

endogeneity was indeed present and also well controlled for in the model (Gibson et al., 2010; Pizer,

2009; Staub, 2009)

4.2.1. The empirical model specification

Just like the maternal smoking example in Mullahy (1997), a large part of our sample were not enrolled

in insurance (x=0) and in addition, those whose treatment was taken have varying treatment intensities

such that x|x>1 = 1,2… This, therefore, required a two part-part instrumental variable Probit model

(Mullahy, 1998). To model the Two Stage Residual Inclusion (2SRI) model, we follow Terza's (2017)

model execution guidelines with detailed guide on the two-part model provided in Terza (2016). In

the first part of the two-part first stage of the 2SRI, we estimated the probability of being insured (α1)

by regressing the instruments )(Zi and other covariates )(Xi using Probit model.

jijiij ZStatusInsure 210_

Where;

ijstatusInsure_ represents a dummy of whether the household of the under under-5 child i

in village j has insurance,

ijZ is a vector of the instruments used,

jiX stands for a vector of a child, household, spatial and village covariates included in the

model, and

ε = is the error term.

Using this model in this first step of the first stage, we predict and stored the results for the probability

of being insured. Let us call this probability1 .

12

In the second part of the first stage, we fit a Generalised Linear Model (GLM) on the treated (insured)

sub-sample whose dependent variable is treatment intensity measured by the number of years of

continuous insurance.

ijijij XZyearsInsure 210_

From this second step of the first stage, we predict and store the mean years of insurance. We call this

2 .

To estimate the residuals, we manually derive the observed values of treatment status from the product

of predicted treatment status and treatment intensity.

21 * tusInsure_staresidual

The second stage of the 2SRI model, we fitted a GLM model, regressing our outcome indicator on

the extent of the treatment and other covariates.

SRI

ijijij residualXyearsInsureStunted 2

3210 _

In these models, we controlled for individual child-specific variables such as child age, sex, if they took

a vitamin A supplement in the last months, immunisations taken and others. We also controlled for

household-specific variables such as education of the parents, household durable assets, agricultural

assets and use of protected water sources. We further included a perception index developed from a

principle components analysis that takes a single principle component from several dimensions of

perceptions concerning health insurance. We then included several village level controls. Because we

generate the residuals manually through the two step first stage, the standard errors have to be

bootstrapped to make them asymptotically correct (Terza, 2016). We therefore bootstrap up to 10,000

replications for our main model.

4.2.2. A review of the instruments used in health insurance studies

Normally, finding suitable instruments for health insurance is a daunting task and many studies

completely fail in this regard (Chankova, Atim, & Hatt, 2010; Nguyen, 2012; Rajkotia & Frick, 2012;

Sepehri et al., 2011; Wehby, 2013). Moreover, even when instrumental methods have improved in

rigor, instrument validation remains a contestable issue by economists (French & Popovici, 2011;

Rashad & Kaestner, 2004). Nonetheless, due to the difficulties and limitations of randomisation and

the rarity of natural occurrences to facilitate natural experiments (Barrett & Carter, 2010; Sanson-

13

Fisher et al., 2007), the use of instrumental variables in health services research has increased in

prominence (Cawley, 2015).

In health insurance studies, instruments that have been used include variations of employment status

and or firm size (Bhattacharya et al., 2011; Deb et al., 2006; Koç, 2011; Rashad & Markowitz, 2007;

Thornton & Rice, 2008; Wagstaff & Lindelow, 2008; Zheng & Zimmer, 2008); variations on

community or state level enrolment rates or phased rollouts (Binagwaho et al., 2012; Fink et al., 2013;

Galarraga et al., 2010; Jung & Streeter, 2015; Nguyen, 2012; Sosa-Rubí et al., 2009; Strobl, 2016;

Trujillo et al., 2005; Wirtz et al., 2012; Woldemichael et al., 2016); computed probability of enrolling

in insurance (Bhattacharya et al., 2011); and differences in district-level requirements of membership

and subsidies given (Gajate-Garrido & Ahiadeke, 2015; Pan et al., 2015). Other instruments have

included: randomized offer of enrolment (Raza et al., 2016); strict eligibility cut-offs used in

combination with a regression discontinuity design (Palmer et al., 2015); and membership in

microfinance or other social support organisations (Jowett et al., 2004; Woldemichael et al., 2016).

While literature gives several prospective instruments, finding appropriate instruments is hinged on a

clear understanding of the process behind the variable of interest (Angrist & Pischke, 2009, p. 88).

Strong instruments in some studies might not be applicable in other studies mainly because of the

contextualised processes behind each data generation process. One example is the distance to health

facilities (with different variations)– which is a common and credible instrument in studies undertaken

in developed countries with a wide network of health facilities in close geographical proximity, but

which might not work in developing countries with health facilities insufficient and scattered.

4.2.3. Our instruments

The instruments adopted for our study are: (1) ever dropping out of insurance, (2) cluster insurance

demand rate (3) perception of fairness of previously incurred medical costs and (4) delivering in an

insurance providing health facility. We used four instruments because of the model efficiency gains

that more instruments bring (Chao & Swanson, 2005; Hansen et al., 2008; Roodman, 2009). We

expound on each of them here below.

Dropping out of insurance: We expected a negative correlation between insurance and ever

dropping out of insurance. We also expected that ever dropping out of insurance does not have any

direct relationship with child stunting. From a contextual perspective, the decisions on dropping out

of insurance are not individual but rather household and community level decisions. Subscribing to

14

insurance requires enrolment of all household members – locally referred to as the full household

insurance prerequisite. Essentially, the insurance scheme does not allow some members of the

household to be insured and others not. Dropping out, therefore, does not only affect an individual

but rather a household. We, therefore, expected that it would not be in the interest of household heads

to risk the lives of all household members by dropping out. The only reason for dropping out that we

envisaged was a lack of financial resources to pay the required household premiums. This perception

was backed by qualitative data gathered in focus groups held with respondents.

“I have been a member for 5 years and it has been really helpful because usually, I am operated when I am

giving birth and the health bill is sometimes 600,000shs which I would not manage to pay by myself. And I

would only pay like 50,000shs. Like recently, I paid only 50,000 on a bill of 408,200shs; which I am really

grateful for. But this year I did not enrol because of not having the money ready, though I am preparing and

saving the little money I get such that I can be able to join next year.”

Another respondent said:

“We have left (insurance) because of lack of money to invest for insurance. The money to invest for insurance

was too much for me I thus chose to step aside…I was asked to pay 70,000 for my family (Me, my husband

and 4 children) I didn’t get the money and thus didn’t enrol.”

Female respondents in from a focus group

Moreover, insurance groups operate as self-help village social support groups in which dropping out

is highly discouraged and in some instances, punishable. Most of the groups were founded on kin

relationships in which trust and social support are highly valued and actions the contravene the norm

could invite punishment (Katabarwa et al., 1999; Katabarwa et al., 2000; Twikirize, 2009). In addition,

these groups operate savings and loans facilities, similar to health savings accounts (Dupas &

Robinson, 2013), which assist in raising premiums. In the event of the group’s finances not being

enough for all members, members who fall short (due to the size of their households) have an

opportunity of taking credit to meet premium requirements. This kind of underlying social support

and risk-sharing has been observed in other community insurance schemes in India (Ravallion &

Chaudhuri, 1997; Townsend, 1994) and in Rwanda (Binagwaho et al., 2012). In instances where

members would not able to raise premiums and the group was not able to support, the household

15

would end up dropping out, largely for these financial reasons. Once members were financially

capable, they could re-join insurance by paying premiums of the subsequent year.

Due to this operative nature of groups in this scheme, dropping out was significantly lower in other

schemes in Uganda without these stringent group requirements (Kyomugisha et al., 2009) and even

other CBHI in other developing countries (Dong et al., 2009; Mebratie et al., 2015). In our sample,

13.6% of the respondents had ever dropped out; and 7.4% of those insured at the time of the survey

had ever dropped out. There could be some concern that since household decisions have a significant

group or community bearing, members might drop out for reasons not of household nature but rather

of a group nature. For instance, corruption, misuse of group’s health fund collections and

disagreements between group leaders and members could influence household membership status.

But we do not believe this to be the case. We asked respondents who reported to have ever dropped

out their main reason for dropping out and in confirmation of our theory, close to 80% reported

inability to raise premiums.

Figure 1: Reasons for dropping out of insurance

Source: Survey data

Even when group coherence and social support are central features of this scheme, financial reasons

were the major determinants of dropping out. In order block reverse causality through income and

wealth, we include a wealth index in the model to control for household socioeconomic status.

Another possible reason could be information and education, and we include these in our model too.

02

04

06

08

0

Pe

rcen

t

Lim

ited

info

rmat

ion

High

prem

ium

s

Sta

rted

own

fam

ily

Con

flicts with

gro

up le

ader

Oth

er

Reasons for drop out

16

In addition, we assessed simple correlations between the endogenous variable (insurance status and

insurance intensity), the instruments and the outcome variable. As shown in Table 2 below, the

correlation coefficient was -0.159, and -0.105 between dropping out and insurance status years in

insurance respectively and insignificant with regard to stunting.

Cluster insurance demand rate: Another instrument we used was the cluster insurance demand rate.

We asked the non-insured households if they would like to join any of the available community groups

including health insurance –registered groups. Cluster insurance demand rate was therefore computed

as the ratio of uninsured households who wish to join an insurance-registered group in each village.

householdsinsuredTotal

insurancejointowishinghouseholdsUninsuredratedemandinsuranceCluster

__

________

The demand rate ranged from 0 to 75 percent. We tested the correlational association and the

predictive OLS regression association. We found that cluster insurance rate had a correlation of -0.120

with insurance, -0.109 with years in insurance and only 0.038 with stunting. In addition, OLS

regression indicated that cluster insurance demand was also strongly associated with insurance status

and years in insurance while not associated with stunting5.

Perception of fairness in hospital costs: The third instrument employed in our analysis was a

perception of how fair the hospital costs were at the most recent hospital visit. Health services are

delivered at government health centres with no official charges after the removal of the formal health

care user fees in 2001 (Deininger & Mpuga, 2005; Xu et al., 2006) while health private not for profit

services charge fees (Amone et al., 2005) and receive some subsidisation from the government

(Nabyonga Orem et al., 2011; Okwero et al., 2010; Pariyo et al., 2009). Nonetheless, unofficial

payments are highly reported at both public and private facilities (Hunt, 2010) and cost-related barriers

still impede almost 50% of the women from accessing health services (UBOS & ICF International,

2012).

Insurance services are only available at private health facilities. Moreover, public health facilities are

plagued by inconsistent services, drug stock-outs and overcrowding, high rates of bribery and basic

infrastructure scarcity; a situation that has given more space to private health providers (Amone et al.,

2005; Konde-Lule et al., 2010; Pariyo et al., 2009) and increased the attractiveness of micro-insurance

5 More results on the instruments are presented in Annex 1

17

in places it is available (Twikirize & O’Brien, 2012) . Given this choice between free but inconsistent

and unpredictable public health services and expensive private health services, we reasoned that people

use private services and hence buy insurance if they believe that they receive better value for money.

To measure this perception of fairness, we asked a question “do you think the cost you paid for your

health care at your most recent visit to a health facility was fair?” We coded response to this question

as a dummy indicating 1 is the respondent felt the costs were fair and 0 otherwise. We expected that

the relationship between fairness of health costs and insurance should be complimentary, such that a

positive attribution of fairness would be more likely to join insurance.

In Table 2 below, we show the correlation coefficient for the perception of fairness of costs and

insurance to be positively strong. We also show that this perception was not correlated with stunting.

Table 2: Correlation between the instruments, insurance, and stunting

Insurance status Years in insurance Stunting

Ever dropped out of insurance -0.1593 -0.1053 0.0559

Cluster insurance demand rate -0.1202 -0.1090 0.0382

Perceptions of fairness of costs 0.119 0.1276 -0.0347

Delivery from an insurance providing facility 0.2269 0.1543 -0.0372

Delivering in an insurance-providing health facility: Our final instrument was a dummy that

indicated 1 if the mother delivered from an insurance providing health facility and 0 otherwise.

Mothers had a choice to deliver in a private health facility that provided insurance or any other public

or private facility that did not provide insurance services. As with the other instruments, we expected

that delivering in an insurance-providing private facility would be strongly correlated with insurance

status and not with stunting. Results supporting this thinking are shown, also in Table 2 above.

One other way of testing the validity of the instruments was to carry out standard two –sample t-tests

to observe the significance of the mean differences across the instruments by the distribution of the

outcome variable.

Table 3: T-tests for mean differences of the outcome variable (stunting) across the instruments Overall Stunted Not stunted mean diff t-stat

Dropping out of insurance 13.60 15.82 11.94 -3.88 (-1.20) Cluster insurance demand rate 48.25 49.07 47.65 -1.42 (-0.82) Perception of fairness of costs 39.22 37.25 40.67 3.43 (0.75) Delivery in insurance facility 42.46 40.31 44.03 3.72 (0.80)

N 464 196 268 464

18

t statistics in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

Table 3 shows these results, which indicates that none of the instruments presents any significant

differences in the mean distribution of our outcome – stunting. Both Tables 2 and 3 indicate a similar

message; that the only way in which the instruments influence the outcome variable is through

stunting.

One other approach we used in qualifying these instruments was carrying out OLS regressions to

reveal the association between the instruments, the endogenous variable, and the outcome variable.

We present these results in the Annex. The overall message is that our instruments were strongly

associated with both the treatment (insurance status) and the treatment intensity (the number of years

in insurance) and not associated with stunting prevalence.

5. Results

5.1. Descriptive Results

Out of our sample of 464 households, 43.8% were enrolled in insurance at the time of the survey.

This was found to be substantially higher than previously reported insurance coverage of 15% in the

same area (Dekker & Wilms, 2010) or 30% reported by Twikirize & O’Brien (2012) indicating

significant improvements in coverage. In 30.4% of the households, at least one of the parents had

attended at least secondary level education. The average age of mothers interviewed was 30.2 years

while the average age of the children whose data was recorded was 30.2 months.

48.1% of the children included in the survey were male, and only 29.1% of the children had received

all child immunisations. These included BCG, Polio DPT, and Pneumococcal Pneumonia vaccine.

Our data collection process did not capture the actual dates of administration of the vaccines so this

measurement is not an indication of timely full immunisation as desired in the national schedule for

childhood immunizations. Some 63.4% of the mothers had exclusively breastfed their youngest child

for up to 6 months. Only 16.8% of the households had all under-five children sleeping under a long-

lasting mosquito net (LLIN). Furthermore, about 55.4% of the children had been born in a health

facility. This was significantly higher than the regional average of 40.3% of children born in health

facilities as measured in the 2011 DHS (UBOS & ICF International, 2012).

For our outcome variable, 42.2% of the children whose anthropometric data was recorded were

stunted – with a height-for-age Z-score (HAZ) of below -2 standard deviations. Stunting prevalence

19

was similar to the regional average of 41.7% for south-western Uganda and much higher than the

national average of 33.1% (UBOS & ICF International, 2012).

In terms of distribution of respondents by village economies, some 18.1% of our respondents were

located in forestry villages while 26.1% were in banana cultivation villages. Around 36.6% of the

respondents were located in trading villages while the remaining 19.2% were from villages whose

pastoralism was the main economic activity. More descriptive statistics are presented in Tables 4

below.

5.2. Empirical Results

5.2.1. Main model results

In Table 5 below, we present instrumental variable results with all covariates detailing the average

treatment effect of insurance on stunting. Our model reveals that controlling for all child-specific,

household, spatial and village level covariates, the probability of stunting reduces by 15.4% for each

year a household is insured. With 400 replications of bootstrapped standard errors, these results are

significant at 5%. Our main results further indicate that the probability of stunting increased by 8.5%

for each extra additional month of a child’s life span. However, this relationship is nonlinear in such

a way that the probability of stunting increases by 8.5% per year only up to 32 months. This is indicated

by the negative but statistically significant square term of the child’s age.

Our model results further indicate that one extra kilogram of birthweight reduces the probability of

stunting by 31.6%. Moreover, children with one or both parents having attended at least some

secondary education had a 31.9% less probability of stunting while children in households with four

or fewer people had a 33.8% less probability of stunting. In addition, once we control for insurance

and other variables, the probability of stunting reduces across richer wealth quintiles compared to the

poorest. However, minimal statistical significance is observed only with respect fourth quintile.

Nonetheless, the richer households seem to benefit more as we observe larger coefficients of reduced

probability of stunting.

20

Table 4: Summary Statistics VARIABLES Type & Description of the variable mean min max SD

Stunting Dichotomous: 1 if the child’s HAZ score was below -2 standard deviations and 0 otherwise. 0.422 0 1 0.494 Insurance status Dichotomous: 1 if the household was in insurance and 0 otherwise 0.438 0 1 0.497 Years in Insurance Count: Ranging from 0 to 11 years (only for the currently insured) 5.07 0 11 3.108 Age of child in months Continuous: Ranging from 6 to 59 months 30.137 5.55 60.58 15.217 Child is male Dichotomous: 1 if child is male 0.481 0 1 0.500

Birthweight6 Continuous: observed of imputed birthweight 3.18 1 5.6 0.521

Vitamin A supplement Dichotomous: 1 if the child had Vitamin A supplement in six months prior to the survey 0.772 0 1 0.420

Taken all child immunisations Dichotomous: 1 if child had all immunizations; namely Measles, DPT, BCG, Polio, and Pneumococcal vaccine

0.291 0 1 0.455

Exclusive breastfeeding Dichotomous: 1 if child was exclusively breastfed for up to six months 0.634 0 1 0.482 All U5 slept under LLIN Dichotomous: 1 if all the U5 children in the household slept under an LLIN the previous night 0.168 0 1 0.374 Religion: Catholic

Protestant Other religion

Categorical: Catholic Protestant Other religion

0.5043 0.3858 0.1099

0 0 0

1 1 1

0.501 0.487 0.313

Mother’s age Continuous: age of the mother recorded in years 30.20 14 56.5 7.164 Parental education (at least one) Dichotomous: 1 if at least one parent has attended at least secondary education 0.304 0 1 0.460 Household >=4 Dichotomous: 1 if household has more than four members 0.407 0 1 0.492 Wealth Index Continuous: Index developed by principle components analysis for agriculture endowments -3.47e-09 -1.799 8.323 1.348 Household diet diversity score Continuous: Number of food groups consumed by a household in the last 24 hours. 4.080 0 8 1.280 Insured neighbour Dichotomous: 1 if the respondent knows that the neighbour is in insurance 0.690 0 1 0.463 Casual labourer Dichotomous: 1 if both parents are casual labourer 0.066 0 1 0.486

Perception index Continuous: Index developed from PCA of 7 perceptions with 40 5-point Likert scale questions about health insurance

-1.24E-09 -3.3306 3.59997 1.741

Listening to radio Dichotomous: 1 if the respondent listened to radio daily 0.569 0 1 0.496 Aggregate group membership Continuous: known number of groups the respondent and head of household belonged to 1.829 0 5 1.003

Differential distance Dichotomous: 1 if the difference of the distance from household to insurance hospital and distance from household to alternative government health centre is less than average

0.565 0 1 0.496

Household altitude Continuous: recorded household altitude in metres above sea level 1720.235 1482.38 2099.9 119.634 No. of insurance groups in village Continuous: Number of insurance groups in the village 2.778 1 6 1.832 Village has a health centre Dichotomous: 1 if village has a health centre 0.401 0 1 0.491 Village has a road Dichotomous: 1 if the village has a road 0.916 0 1 0.278 Village has school Dichotomous: 1 if village has a school 0.634 0 1 0.482 Village economy: Pastoral villages Banana villages Forest villages Trade villages

Categorical: 1 if pastoral village 2 if banana village 3 if forest village 4 if trade village

0.192 0.261 0.181 0.366

0 0 0 0

1 1 1 1

0.394 0.440 0.385 0.482

Cost fairness perception Dichotomous: 1 if the respondent thought that the cost of care at the previous hospitalisation was fair 0.392 0 1 0.489 Insurance dropout Dichotomous: 1 if a respondent has ever dropped out of insurance 0.136 0 1 0.343 Cluster demand Continuous: % of uninsured respondents who would like to join insurance in each village 0.482 0 0.75 0.183 Deliver in insurance facility Dichotomous: 1 if delivered in a health facility providing insurance 0.425 0 1 0.495

N 464

6 56% of the sample had an observed birthweight and the remaining 44% was inputted. See the annex for detailed supplementary information about the imputation procedure

21

Table 5: Instrumental variable results for the impact of insurance on stunting

(1) VARIABLES stunted

Years in insurance -0.154** (0.0709) Age of the child (in months) 0.0846*** (0.0232) Age squared -0.00132*** (0.000355) Gender of the child (1=male) 0.0962 (0.146) Vitamin A supplementation -0.294 (0.214) All child immunisations -0.150 (0.361) Vitamin A ## Child immunisation 0.369 (0.396) All U-5 sleep in LLIN 0.149 (0.197) Birthweight -0.316** (0.153) Exclusive breastfeeding 0.417*** (0.154) Religion (base= Catholic) Protestant 0.233 (0.187) Others -0.0624 (0.275) Parental secondary education -0.319* (0.181) Household <=4 -0.338** (0.160) Wealth quintile (base= quintile 1) Quintile 2 -0.318 (0.230) Quintile 3 -0.167 (0.254) Quintile 4 -0.369 (0.247) Quintile 5 -0.335 (0.280) Household Diet Diversity Score -0.0235 (0.0642) Have an insured neighbour 0.127 (0.199) Casual employment 0.0531 (0.296) Composite perception index -0.00900 (0.0512) Composite satisfaction index -0.00345

22

(0.0275) Perceptions# satisfaction indices 0.0106 (0.0160) Waiting time -0.000698 (0.000723) Number of groups 0.491** (0.244) Number of groups squared -0.0831 (0.0541) Differential distance -0.365 (0.245) Household altitude 0.000480 (0.00137) Village has a school 0.253 (0.354) Village has a health centre -0.0842 (0.290) Number of insurance groups in village -0.00872 (0.110) Insurance coverage in 2010 1.430 (1.049) Forest village 0.260 (0.423) Banana village 0.245 (0.296) Residuals 0.205*** (0.0796) Constant -1.439 (2.452) First stage F-statistic 26.56*** AIC 1.325 Observations 464

Bootstrapped standard errors with 10000 replications in parentheses *** p<0.01, ** p<0.05, * p<0.1

We further explore how the behaviour of the probability of being stunted for each additional year of

insurance by the household. In order to know this, we employ the analysis of marginal effects across

the years of insurance by re-running the outcome equation, this time with categorical conditioning of

the years of insurance. The results (in Table 7 in Annex 2) indicate that the probability of stunting

reduces consistently and approaches zero for children in households that have been in insurance for

20 years.

Figure 2: Predictive margins for Insurance Years

23

The marginal plot of the predicted probability of stunting across years of insurance measures two

things, that is; (1) the shift in number of years a household has been in insurance in reference to

households with no insurance and, (2) the effect of this shift on the probability of being stunted. The

results presented in the graph above and in Table 8 in annex indicate that for children in households

with no insurance (zero years), the probability of stunting was 0.518. This probability reduced as

households became and remained insured, reaching 0.329 for households that were insured up to

seven years. The probability of stunting continue to decline but is no longer strongly significant after

seven years. There are two interpretations that can be made from this result. The first one is that as

households shift further on the right, thus increasing on the years they have been in insurance, the

comparisons with the base category of those without insurance is no longer significant because the

behaviour and practices that influence stunting through insurance have been well entrenched in the

group with more years of insurance compared to those without. This interpretation is supported by

the results in Table 8 in annex which indicate that children in households that had insurance for six

years were almost 1.7 times less likely to be stunted compared to children in households without

insurance. The other interpretation is that as sample reduces as years of insurance increase. This

implies that our results after about 7 years are not strong because of limited predictive power of the

small sample.

0.2

.4.6

.8

Pre

dic

ted M

ean

Pro

ba

bili

ty o

f S

tuntin

g

0 1 2 3 4 5 6 7 8 9 10 13 20Number of Years in Insurance

Predictive Margins of Insurance Years with 95% CIs

24

5.2.2. Some non-linear and heterogeneous treatment effects

In Model 1 below, we studied the effect of insurance on stunting for children of two groups of

mothers: those below the average age of 30.2 years and those above this average age. The results

indicate that while each year of insurance reduces the probability of stunting by about 18.4%, the

probability of stunting in children whose mothers were older than the average mother’s age of 30.2

years, could increase by 9.9%. This implies that for older mothers, the net reduction in the probability

of stunting was 8.15% while for younger mothers, the probability of reduction in stunting for their

children was more than twice as much at 18.4%

Table 6: Effect of insurance on stunting in selected sub-groups

(1) (2) VARIABLES Effect with

mother’s age Effect with child age

groups

Years in insurance -0.184** -0.262** (0.0849) (0.110) Mother age > 30.2 years 0.0454 (0.209) Mother age > 30.2 years #Years in insurance 0.0999** (0.0493) Age categories (base: 6-11 months)

12-23 months 0.119 (0.309)

24-35 months 0.498 (0.303)

36-47 months 0.380 (0.325)

48-60 months -0.370 (0.353) Age category# Insurance 12-23 months # Number of years in insurance 0.143 (0.0980) 24-35 months # Number of years in insurance 0.0336 (0.0912) 36-47 months # Number of years in insurance 0.121 (0.0941) 48-60 months # Number of years in insurance 0.0838 (0.0986) Residuals 0.153** 0.239*** (0.0769) (0.0833) Constant -1.602 -0.135 (2.454) (2.543) First stage F-Statistic 26.77*** 26.88*** AIC 1.398 1.398 Observations 464 464


25

In Model 2 of Table 6, we carried out interactions of the treatment with various child age groups in

order to segregate between effects on different age groups. Our results indicate that with this

heterogeneous analysis, the probability of stunting reduces be 26.2% however, we do not observe a

significant effect in any age group.

Imbens & Angrist (1994) suggest a more precise measure of treatment effects, the local average

treatment effect, which is the average treatment effect on the treated. To estimate this, we carried out

reduced sample regressions on the subsample whom treatment was taken, that is, those who were

insured. In addition, we examined the linearity of the treatment effects to understand if the insurance

does not have an effect after a certain time. To estimate this, we included a square term of years on

insurance in the regression and present the result in Table 7.

Table 7: Local average treatment effects and linearity of treatment effects

(1) (2) VARIABLES LATE Years squared

Years in insurance -0.236 -0.133 (0.229) (0.170) Years squared -0.00196 (0.0144) Residuals 0.270 0.185 (0.237) (0.188) Residuals (Years squared) 0.00188 (0.0152) Constant -1.743 -1.421 (5.885) (2.462) First stage F-statistic 26.56*** 26.56*** AIC 1.434 1.334 Observations 203 464


Results presented in Model 1 of Table 8 above indicate negative but insignificant coefficient relating

to the local average treatment effect, implying that though the probability of stunting reduces for the

insured, it is not significant, possibly due to a small sub-sample on which the analysis is undertaken.

In Model 2 of Table 7, we include the square term on the number of years in insurance to determine

the linear behaviour of years in insurance. In essence, we would like to know at what number of years

in insurance the gains start diminishing. Our results indicate no change in the sign of the coefficient

on the square term for number of years in insurance but results are insignificant. This implies that the

observed relationship in Table 6 with regards to number of years in insurance is linear in all through

26

the years and does not change.

5.2.3. A comment on peculiar results regarding the effect of exclusive breastfeeding

Finally, our main results present a strange finding concerning exclusive breastfeeding. Results indicate

that controlling for all other things, exclusive breastfeeding was associated with a 14% increase in the

probability of stunting. This finding is worrying, especially in view of the fact that exclusive

breastfeeding is an integral component of child nutrition and anthropometric improvement

(Kamudoni et al., 2015; Kamudoni et al., 2007; Kuchenbecker et al., 2015; Marques et al., 2015).

However, our findings deviate from this position. Our sample indicated that 63.4% of the mothers

exclusively breastfed up to six months, similar to the 2012 demographic and health survey report of

63.2% (UBOS & ICF International, 2012). There could be a measurement error in our formulation of

the breast feeding question, that is driving this finding.

Public health and maternal health experts have found that data on exclusive breastfeeding is highly

contentious and could sometimes be unreliable (Aarts et al., 2000; Cupul-Uicab et al., 2009; Gillespie

et al., 2006; Greiner, 2014). Some suggest that the question ‘exclusive breastfeeding at six months’

should not be used at all (Debra, 2011). The challenges with the reliability of breastfeeding data could

emanate from the way in which the questions are framed. However, we also believe that reliability of

this question’s outcome could be at the risk of recall bias. One dimension of recall bias is that after

several years, mothers do not exactly remember for how long they breastfed their children. Several

studies in developed countries have found that mothers tend to overestimate the duration of exclusive

breastfeeding (Aarts et al., 2000; Gillespie et al., 2006). We imagine the situation would be even worse

in developing countries. For instance, Cupul-Uicab et al (2009) found that measurement of exclusive

breastfeeding was unreliable especially when a mother has four or more children. We imagine that the

same challenges might be true for a country like Uganda, and more so in our sample where only 18.3%

of the mothers surveyed had attended secondary education.

The other recall bias dimension is that sometimes respondents provide answers which they imagine

to be favourable with the researcher even when they are not correct in reality, by either underreporting

or over overestimating (Martinelli & Parker, 2009). This type of bias has been suspected in a malaria

control study in Uganda which included a site close to our area of study. Kamya et al (2015) studied

the parasitic prevalence and malaria incidence in households where children under 5 years had been

given LLINs. At two study sites, reported LLIN usage rates were between 51% and 78.5% but

27

episodes of malaria ranged from 182 to 1,546 over a 2 year period. Most intriguingly, the study site

that reported high LLIN usage of 78.5% also had the highest episodes and the highest malaria

incidence rates of 2.8 incidences per person per year. For such a paradox, Meshnick (2015) suggested

that recall bias could not be ruled out. We think similar scenarios might be at play with exclusive

breastfeeding. In view of this scenario, we conducted further analysis in which exclusive breastfeeding

was left out of the model. Table 8 below presents the effect of insurance based on analysis with and

without considering exclusive breastfeeding.

Table 8: Results with and without exclusive breastfeeding

(1) (2) VARIABLES With Excl.

Breastfeeding Without Excl. Breastfeeding

Years in insurance -0.154** -0.168** (0.0709) (0.0724) Exclusive breastfeeding 0.417*** (0.154) Residuals 0.205*** 0.221*** (0.0796) (0.0815) Constant -1.439 -1.648 (2.452) (2.452) First stage Wald test 26.56*** 26.25*** AIC 1.325 1.336 Observations 464 464


Comparing the two models; main model with exclusive breastfeeding (Model 1) and without exclusive

breastfeeding (Model 2) in Table 8, we find that once we do not include exclusive breastfeeding, the

results were biased upwards by about 6.5% compared to Model 1. This bias does not worry the

relevance of our results. However, leaving out an important variable such as exclusive breastfeeding

in the measure of child health outcomes would imply a loss of very important information. Future

studies should look into more precise ways of measuring exclusive breastfeeding to reduce

measurement errors and their related biases that could influence the overall inference.

5.2.4. Robustness checks

Even when insurance uptake has increased over the years, only 43.8% of our sample were insured.

The implication of this is that we have a large number of zeros in our data and hence there is over

28

dispersion in our data. For robustness checks we therefore elicit to use alternative models which are

consistent with count data with large number of zeros. Normally, count data is analysed with Poisson

models, however due to substantial number of zeros in the data, zero inflated models such as Zero-

inflated Poisson have been developed to deal with such problems (Cameron & Trivedi, 2009, Chapter

17; Mullahy, 1986). We hence use two versions of the instrumental variable estimation. The first one

is an IV estimation where the first stage is a zero-inflated Poisson model and a second stage Probit

model, specifying first stage probit prediction of excess zeros. Since we have a dummy outcome we

also implement a standard IV-Probit model. In the table below we report the results of these models

against our main 2SRI model.

Table 9: Robustness checks with alternative specifications

Main model Comparison models

VARIABLES IV-2SRI IV with ZIP IV-PROBIT Years in insurance -0.154** -0.149** -0.165* (0.0709) (0.0628) (0.0925) Residuals 0.205*** 0.198*** (0.0796) (0.0689) Constant -1.439 -1.162 -1.281 (2.452) (2.142) (1.931) First stage Wald test 26.56*** 57.17*** 18.65*** AIC 1.325 1.323 Wald test for exogeneity 2.72* Observations 464 464 464

2SRI Model standard errors bootstrapped with 10000 replications and other models, robust standard errors in parentheses. Reported significance levels *** p<0.01, ** p<0.05, * p<0.1

The results indicate that an IV model with first stage Zero-inflated Poisson model produces results

not very different from our main model. The IV-Probit results are also stable in the same ranges as

our main model though only weakly significant. We present these full results in the annex tables.

6. Discussion and Conclusion

This study reveals that on average, CBHI reduces the probability of stunting by up to 15.4% for each

year of insurance. This relationship is linear throughout. More intuitively, it implies that if a child was

born in 2015 at the time of the survey, and the household subscribes to insurance and remains so for

a period of at least five years, the child would have a 77% less probability of being stunted. Conversely,

if a child was born in 2015 and was not in insurance at all until the age of five, he or she would have

29

their probability of stunting elevated by 77%.

Another interpretation of our results is to intuitively scale up this reduction proportion in out sample

over a period of five years. Given that in our sample of 464 children, 42.2% were stunted, we know

that 196 children were stunted and have a linear reduction of 15.4% per year. Over a 5-year period,

an additional 90 under-5 children would be protected from stunting due to insurance.

Table 10: Some intuition on would-be stunting reduction

Number stunted Reduction rate Lives protected

Year 0 196 0.154 Year 1 165.816 0.154 30.184

Year 2 140.2803 0.154 25.53566

Year 3 118.6772 0.154 21.60317

Year 4 100.4009 0.154 18.27628

Total lives protected 95.59912

CBHI, therefore, makes a huge difference in improving nutritional indicators of rural children. For a

country where about one-third of children under the age of five are stunted, such an intervention

could have broader benefits for human growth, economic development and human capital formation.

Our results are in line with findings from Rwanda (Lu et al., 2016) where the authors found that

enrolment in CBHI significantly reduced the odds of being stunted by between 39% and 44%.

One of the channels of impact is health service utilisation. Very many studies have found that health

insurance coverage greatly facilitates utilisation of health services for mothers (Wang, Temsah, &

Mallick, 2014) and for children (Gajate-Garrido & Ahiadeke, 2015; Singh et al., 2015). Our data and

findings are in line with this literature, thus creating a significant pathway of impact. For instance, our

data shows that insured mothers were more likely to undertake four or more antenatal care (ANC)

visits and also receive the essential ANC services compared to the uninsured while insured mothers

were also more likely to undertake a postnatal care visit. Moreover, when asked if a child had had all

of cough, fever and diarrhoea in fourteen days prior to the survey, only 3.5% of children in insured

households reported the affirmative compared to 6.1% of the children in uninsured households.

Table 11: Differences in services utilisation as pathways of impact

Overall Mean

Mean Insured

Mean uninsured

Mean difference

t-statistic

30

Attended at least 4 ANC visits 71.55 89.16 57.85 -0.313*** (-7.88) Attended at least 1 postnatal visit 82.76 84.73 81.24 -0.0350 (-0.99) Received essential ANC services 40.73 50.25 33.33 -0.169*** (-3.73) Reported the 3 child illnesses 04.96 03.45 06.13 0.0268 (1.32)

N 464 203 261 464 Significance of the mean difference reported for * p < 0.05, ** p < 0.01, *** p < 0.001

These differences are not causal, nonetheless, they give a compelling correlation through which health

services utilisation can translate into health improvements.

Furthermore, these findings also highlight the effect of other factors that reduce stunting such as

parental education (Vollmer et al., 2016) and birthweight (Behrman & Rosenzweig, 2004; Wilcox,

2001) for child health outcomes. Our study however highlights once again on the challenges of

accurately estimating exclusive breastfeeding in household and nutrition surveys. These challenges

breed measurement errors that might grossly influence the inferences made in child nutrition studies.

We recommend that future surveys develop more rigorous and multiple quality assurance methods of

generating precise responses for questions such as exclusive breastfeeding.

In conclusion, this study contributes to the thin layer of literature on community-based health

insurance and health outcomes in developing countries – specifically by interrogating the situation

using a rural Ugandan case study. Our findings indicate that enrolling and remaining enrolled in

community-based health insurance reduces the probability of child stunting by up 77% over a five-

year period. For a country like Uganda, grappling with child stunting problem, promoting and

facilitating the scale-up and long-term survival of community-based insurance schemes would be one

of the pathways to avert millions of mortality and morbidity.

31

7. Bibliography

Aarts, C., Kylberg, E., Hornell, A., Hofvander, Y., Gebre-Medhin, M., & Greiner, T. (2000). How exclusive is exclusive breastfeeding? A comparison of data since birth with current status data. International Journal of Epidemiology, 29(6), 1041–1046.

Abuosi, A. A., Domfeh, K. A., Abor, J. Y., Nketiah-Amponsah, E., Blanchet, N., Fink, G., … Edgman-Levitan, S. (2016). Health insurance and quality of care: Comparing perceptions of quality between insured and uninsured patients in Ghana’s hospitals. International Journal for Equity in Health, 15(1), 76.

Abuya, T., Maina, T., & Chuma, J. (2015). Historical account of the national health insurance formulation in Kenya: experiences from the past decade. BMC Health Services Research, 15, 56.

Acharya, A., Vellakkal, S., Taylor, F., Masset, E., Satija, A., Burke, M., & Ebrahim, S. (2013). The Impact of Health Insurance Schemes for the Informal Sector in Low-and Middle-Income Countries: A Systematic Review. The World Bank Research Observer, 28(2), 236–266.

Adebayo, E. F., Uthman, O. A., Wiysonge, C. S., Stern, E. A., Lamont, K. T., & Ataguba, J. E. (2015). A systematic review of factors that affect uptake of community-based health insurance in low-income and middle-income countries. BMC Health Services Research, 15(1), 543.

Amone, J., Asio, S., Cattaneo, A., Kweyatulira, A. K., Macaluso, A., Maciocco, G., … Santini, S. (2005). User fees in private non-for-profit hospitals in Uganda: a survey and intervention for equity. International Journal for Equity in Health, 4(1), 6.

Angrist, J. D. ., Imbens, G. W. ., & Rubin, D. B. . (1996). Identification of Causal Effects Using Instrumental Variables. Journal of the American Statistical Association, 91(434), 444–455.

Angrist, J. D., & Imbens, G. W. (1995). Two-Stage Least Squares Estimation of Average Causal Effects in Models With Variable Treatment Intensity. Journal of the American Statistical Association, 90(430), 431–442.

Angrist, J. D., & Pischke, J.-S. (2009). Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton University Press.

Angrist, J. D., & Pischke, J.-S. (2015). Mastering Metrics: The Path from Cause to Effect. Princeton, New Jersey: Princeton University Press.

Asfaw, A., & Braun, J. Von. (2004). Can community health insurance schemes shield the poor against the downside health effects of economic reforms? The case of rural Ethiopia. Health Policy, 70(1), 97–108.

Asfaw, A., & von Braun, J. (2004). Is Consumption Insured against Illness ? Evidence on Vulnerability of Households to Health Shocks in Rural Ethiopia. Economic Development and Cultural Change, 53(1), 115–129.

Atinga, R. A., Abiiro, G. A., & Kuganab-Lem, R. B. (2015). Factors influencing the decision to drop out of health insurance enrolment among urban slum dwellers in Ghana. Tropical Medicine and International Health, 20(3), 312–321.

Barrett, C. B., & Carter, M. R. (2010). The power and pitfalls of experiments in development economics: Some non-random reflections. Applied Economic Perspectives and Policy, 32(4), 515–548.

Basaza, R. K., O’Connell, T. S., & Chapčáková, I. (2013). Players and processes behind the national health insurance scheme: a case study of Uganda. BMC Health Services Research, 13, 357.

Behrman, J. R., & Rosenzweig, M. R. (2004). Returns to Birthweight. Review of Economics and Statistics, 86(2), 586–601.

Bennett, S., Creese, A., & Monasch, R. (1998). Health Insurance Schemes for People Outside Formal Sector Employment. Current Concerns. Geneva: World Health Organisation.

Bhattacharya, J., Bundorf, M. K., Pace, N., & Sood, N. (2011). Does Health Insurance Make You

32

Fat? In M. Grossman & N. H. Mocan (Eds.), Economic Aspects of Obesity (pp. 35–64). Universty of Chicago Press.

Binagwaho, A., Hartwig, R., Ingeri, D., & Makaka, A. (2012). Mutual Health Insurance and the Contribution to Improvements in Child Health in Rwanda (Passauer Diskussionspapiere: Volkswirtschaftliche Reihe No. V-66-12) (Vol. 31).

Bonfrer, I., Poel, E. Van De, Gustafsson-Wright, E., & van Doorslaer, E. (2015). Effects of a subsidized voluntary health insurance on insured and uninsured in Nigeria (BMG Working Paper No. 1).

Cai, B., Small, D. S., & Have, T. R. T. (2011). Two-stage instrumental variable methods for estimating the causal odds ratio: Analysis of bias. Statistics in Medicine, 30(15), 1809–1824.

Cameron, A. C., & Trivedi, P. K. (2009). Microeconometrics using Stata. Stata Press. Carpenter, J., & Kenward, M. G. (2013). Multiple Imputation and its Application (1st ed.). John Wiley &

Sons. Carrin, G., Waelkens, M.-P., & Criel, B. (2005). Community-based health insurance in developing

countries: a study of its contribution to the performance of health financing systems. Tropical

Medicine & International Health : TM & IH, 10(8), 799–811. Case, A., & Paxson, C. (2010). Causes and Consequences of Early-Life Health. Demography, 47(0),

S65–S85. Cawley, J. (2015). A selective review of the first 20 years of instrumental variables models in health-

services research and medicine. Journal of Medical Economics, 18(9), 721–34. Chankova, S., Atim, C., & Hatt, L. (2010). Ghana’s National Health Insurance Scheme. In M.-L.

Escobar, C. C. Griffin, & R. P. Shaw (Eds.), The Impact of Health Insurance in Low- and Middle-Income Countries (pp. 58–88). Washington DC: The Brookings Institution.

Chao, J. C., & Swanson, N. R. (2005). Consistent Estimation with a Large Number of Weak Instruments. Econometrica, 73(5), 1673–1692.

Chatterji, P., Kim, D., & Lahiri, K. (2014). Birth Weight and Academic Achievement in Childhood. Health Economics, 23, 1013–1035.

Criel, B., Atim, C., Basaza, R., Blaise, P., & Waelkens, M. P. (2004). Editorial: Community health insurance ( CHI ) in sub-Saharan Africa: researching the context. Tropical Medicine and International Health, 9(10), 1041–1043.

Cupul-Uicab, L. A., Gladen, B. C., Hernández-Ávila, M., & Longnecker, M. P. (2009). Reliability of reported breastfeeding duration among reproductive-aged women from Mexico. Maternal and Child Nutrition, 5(2), 125–137.

De Allegri, M., Kouyaté, B., Becher, H., Gbangou, A., Pokhrel, S., & Sanon, M. (2006).

Understanding enrolment in community health insurance in sub-Saharan Africa : a population-based case – control study in rural Burkina Faso. Bulletin of the World Health Organization, 84(11), 852–858.

Deb, P., Li, C., Trivedi, P. K., & Zimmer, D. M. (2006). The effect of managed care on use of health care services: Results from two contemporaneous household surveys. Health Economics, 15(7), 743–760.

Debra, H. J. (2011). Complexities and subtleties in the measurement and reporting of breastfeeding practices. International Breastfeeding Journal, 6(1), 5.

Deininger, K., & Mpuga, P. (2005). Economic and welfare impact of the abolition of health user fees: Evidence from Uganda. Journal of African Economies, 14(1), 55–91.

Dekker, M., & Wilms, A. (2010). Health Insurance and Other Risk-Coping Strategies in Uganda: The Case of Microcare Insurance Ltd. World Development, 38(3), 369–378.

Desai, S., & Alva, S. (1998). Maternal education and child health: is there a strong causal relationship? Demography, 35(1), 71–81.

33

Dewey, K. G., & Begum, K. (2011). Long-term consequences of stunting in early life. Maternal and Child Nutrition, 7(SUPPL. 3), 5–18.

Dong, H., De Allegri, M., Gnawali, D., Souares, A., & Sauerborn, R. (2009). Drop-out analysis of community-based health insurance membership at Nouna, Burkina Faso. Health Policy, 92(2–3), 174–179.

Dror, D. M. (2014). Health Microinsurance Programs in Developing Countries. In A. J. Culyer (Ed.), Encyclopedia of Health Economics (Vol. 1, pp. 412–421). Elsevier Ltd.

Dror, D. M., Shahed Hossain, S. A., Majumdar, A., Koehlmoos, T. L. P., John, D., & Panda, P. K. (2016). What factors affect voluntary uptake of community-based health insurance schemes in low- and middle-income countries? A systematic review and meta-analysis. PLoS ONE, 11(8), 1–31.

Duggan, M. B. (2010). Anthropometry as a tool for measuring malnutrition: impact of the new WHO growth standards and reference. Annals of Tropical Paediatrics, 30(1), 1–17.

Dupas, P., & Robinson, J. (2013). Why Don’t the Poor Save More? Evidence from Heath Savings Experiments. American Economic Review, 103(4), 1138–1171.

Eddings, W., & Marchenko, Y. V. (2012). Diagnostics for multiple imputation in Stata. Stata Journal, 12(3), 353–367.

Ekman, B. (2004). Community-based health insurance in low-income countries: A systematic review of the evidence. Health Policy and Planning, 19(5), 249–270.

Enders, C. K. (2010). Applied Missing Data Analysis. New York; London: The Guilford Press. Filmer, D., & Pritchett, L. H. (1999). The Effect of Household Wealth on Educational Attainment:

Evidence from 35 Countries. Population and Development Review. Filmer, D., & Pritchett, L. H. (2001). Estimating wealth effects without expenditure data--or tears:

an application to educational enrollments in states of India. Demography, 38(1), 115–32. Filmer, D., & Scott, K. (2012). Assessing asset indices. Demography, 49(1), 359–392. Fink, G., Robyn, P. J., Sié, A., & Sauerborn, R. (2013). Does health insurance improve health?

Evidence from a randomized community-based insurance rollout in rural Burkina Faso. Journal of Health Economics, 32(6), 1043–1056.

French, M. T., & Popovici, I. (2011). That Instruments is Lousy! In Search of Agreement When Using Instrumental Variables Estimation in Substance Abuse Research. Health Economics, 20(2), 127–146.

Gajate-Garrido, G., & Ahiadeke, C. (2015). The effect of insurance enrollment on maternal and child health care utilization: The case of Ghana (No. 1495). Washington DC.

Galarraga, O., Sosa-Rubi, S. G., Salinas-Rodriguez, A., & Sesma-Vazquez, S. (2010). Health insurance for the poor: Impact on catastrophic and out-of-pocket health expenditures in Mexico. European Journal of Health Economics, 11(5), 437–447.

Garza, C., de Onis, M., Martorell, R., Onyango, A. W., Victora, C. G., Lartey, A., … Borghi, E. (2006). Assessment of differences in linear growth among populations in the WHO Multicentre Growth Reference Study. Acta Paediatrica, Supplement, 56–65.

Gibson, T. B., Song, X., Alemayehu, B., Wang, S. S., Waddell, J. L., Bouchard, J. J. R., & Forma, F. (2010). Cost sharing, adherence, and health outcomes in patients with diabetes. The American Journal of Managed Care, 16(8), 589–600.

Gillespie, B., D’Arcy, H., Schwartz, K., Bobo, J., & Foxman, B. (2006). Recall of age of weaning and other breastfeeding variables. International Breastfeeding Journal, 1(1), 4.

Glewwe, P., Jacoby, H. G., & King, E. M. (2001). Early childhood nutrition and academic achievement: A longitudinal analysis. Journal of Public Economics, 81(3), 345–368.

Glewwe, P., & Miguel, E. A. (2008). The Impact of Child Health and Nutrition on Education in Less Developed Countries. In P. T. Schultz & J. A. Strauss (Eds.), Handbook of Development

34

Economics (Vol. 4, pp. 3561–3606). Elsevier B.V. Greiner, T. (2014). Exclusive breastfeeding: measurement and indicators. International Breastfeeding

Journal, 9(1), 18. Hansen, C., Hausman, J., & Newey, W. (2008). Estimation With Many Instrumental Variables.

Journal of Business & Economic Statistics, 26(4), 398–422. Hendriks, M. E., Rosendaal, N. T. A., Wit, F. W. N. M., Bolarinwa, O. A., Kramer, B., Brals, D., …

Schultsz, C. (2016). Sustained effect of health insurance and facility quality improvement on blood pressure in adults with hypertension in Nigeria: A population-based study. International Journal of Cardiology, 202, 477–484.

Hendriks, M. E., Wit, F. W. N. M., Akande, T. M., Kramer, B., Osagbemi, G. K., Tanovic, Z., … Schultsz, C. (2014). Effect of Health Insurance and Facility Quality Improvement on Blood Pressure in Adults With Hypertension in Nigeria: A Population-Based Study. JAMA Internal Medicine, 24(4), 555–563.

Horton, N. J., & Kleinman, K. P. (2007). Much Ado About Nothing: A Comparison of Missing Data Methods and Software to Fit Incomplete Data Regression Models. The American Statistician, 61(1), 79–90.

Horton, N. J., & Lipsitz, S. R. (2001). Multiple Imputation in Practice: Comparison of Software Packages for Regression Models with Missing Variables. The American Statistician, 55(3), 244–254.

Hunt, J. (2010). Bribery in health care in Uganda. Journal of Health Economics, 29(5), 699–707. Imbens, G. W., & Angrist, J. D. (1994). Identification and Estimation of Local Average Treatment

Effects. Econometrica, 62(2), 467–475. Jacobs, B., Bigdeli, M., Pelt, M. Van, Ir, P., Salze, C., & Criel, B. (2008). Bridging community-based

health insurance and social protection for health care - A step in the direction of universal coverage? Tropical Medicine and International Health, 13(2), 140–143.

Jehu-Appiah, C., Aryeetey, G., Agyepong, I., Spaan, E., & Baltussen, R. (2012). Household perceptions and their implications for enrolment in the National Health Insurance Scheme in Ghana. Health Policy and Planning, 27(3), 222–233.

Jowett, M., Deolalikar, A., & Martinsson, P. (2004). Health insurance and treatment seeking behaviour: Evidence from a low-income country. Health Economics, 13(9), 845–857.

Jung, J., & Streeter, L. J. (2015). Does health insurance decrease health expenditure risk in developing countries? The case of China. Southern Economic Journal, 82(2), 361–384.

Jütting, J. P. (2004). Do community-based health insurance schemes improve poor people’s access to health care? Evidence from rural senegal. World Development, 32(2), 273–288.

Kamudoni, P., Maleta, K., Shi, Z., & Holmboe-Ottesen, G. (2007). Infant feeding practices in the first 6 months and associated factors in a rural and semiurban community in Mangochi District, Malawi. Journal of Human Lactation, 23(4), 325–332.

Kamudoni, P., Maleta, K., Shi, Z., & Holmboe-Ottesen, G. (2015). Exclusive breastfeeding duration during the first 6 months of life is positively associated with length-for-age among infants 6-12 months old, in Mangochi district, Malawi. European Journal of Clinical Nutrition, 69(1), 96–101.

Kamya, M. R., Arinaitwe, E., Wanzira, H., Katureebe, A., Barusya, C., Kigozi, S. P., … Dorsey, G. (2015). Malaria transmission, infection, and disease at three sites with varied transmission intensity in Uganda: Implications for malaria control. American Journal of Tropical Medicine and Hygiene, 92(5), 903–912.

Katabarwa, M., Mutabazi, D., & Richards, F. O. (1999). Monetary incentives and community-directed health programmes in some less-developed countries. Lancet, 354(9193), 1909.

Katabarwa, N. M., Richards Jr, F. O., & Ndyomugyenyi, R. (2000). In rural Ugandan communities the traditional kinship/clan system is vital to the success and sustainment of the African

35

Programme for Onchocerciasis Control. Annals of Tropical Medicine and Parasitology, 94(5), 485–495.

Keng, S.-H., & Sheu, S.-J. (2013). The Effect of National Health Insurance on Mortality and the SES-Health Gradient: Evidence from the Elderly in Taiwan. Health Economics, 22(1), 52–72.

Kenward, M. G., & Carpenter, J. (2007). Multiple imputation : current perspectives. Statistical Methods in Medical Research, 16, 199–218.

Koç, Ç. (2011). Disease-specific moral hazard and optimal health insurance design for physician services. Journal of Risk and Insurance, 78(2), 413–446.

Konde-Lule, J., Gitta, S. N., Lindfors, A., Okuonzi, S., Onama, V. O., & Forsberg, B. C. (2010). Private and public health care in rural areas of Uganda. BMC International Health and Human Rights, 10(1), 29.

Kuchenbecker, J., Jordan, I., Reinbott, A., Herrmann, J., Jeremias, T., Kennedy, G., … Krawinkel, M. B. (2015). Exclusive breastfeeding and its effect on growth of Malawian infants: results from a cross-sectional study. Paediatrics and International Child Health, 35(1), 14–23.

Kyomugisha, E. L., Buregyeya, E., Ekirapa, E., Mugisha, J. F., & Bazeyo, W. (2009). Strategies for sustainability and equity of prepayment health schemes in Uganda. African Health Sciences, 9 Suppl 2(59), S59-65.

Landmann, A., & Frölich, M. (2015). Can health-insurance help prevent child labor? An impact evaluation from Pakistan. Journal of Health Economics, 39, 51–59.

Levy, H., & Meltzer, D. (2004). What do we really know about whether health insurance affects health? In C. G. McLaughlin (Ed.), Health Policy and the Uninsured (pp. 179–204). Washington DC: The Urban Institute Press.

Levy, H., & Meltzer, D. (2008). The impact of health insurance on health. Annu. Rev. Public Health, 29, 399–409.

Longford, N. T. (1999). An experiment in primary health care provision. Journal of the Royal Statistical Society. Series A: Statistics in Society, 162(3), 291–302.

Longford, N. T., Ely, M., Hardy, R., & Wadsworth, M. E. J. (2000). Handling missing data in diaries of alcohol consumption. Journal of Royal Stastistical Society Series A: Statistics in Society, 163(4), 381–402.

Lu, C., Chin, B., Lewandowski, J. L., Basinga, P., Hirschhorn, L. R., Hill, K., … Binagwaho, A. (2012). Towards universal health coverage: An evaluation of Rwanda Mutuelles in its first eight years. PLoS ONE, 7(6).

Lu, C., Mejia-Guevara, I., Hill, K., Farmer, P., Subramanian, S. V., & Binagwaho, A. (2016). Community-based health financing and child stunting in rural Rwanda. American Journal of Public Health, 106(1), 49–55.

Makaka, A., Breen, S., & Binagwaho, A. (2012). Universal health coverage in Rwanda: a report of innovations to increase enrolment in community-based health insurance. The Lancet, 380, S7.

Marchenko, Y. V, & Eddings, W. (2011). A note on how to perform multiple-imputation diagnostics in Stata. Stata Journal, 1–9.

Marques, R. de F. da S. V., Taddei, J. A. de A. C., Konstantyner, T., Lopez, F. A., Marques, A. C. V., de Oliveira, C. S., & Braga, J. A. P. (2015). Anthropometric indices and exclusive breastfeeding in the first six months of life: a comparison with reference standards NCHS, 1977 and WHO, 2006. International Breastfeeding Journal, 10(1), 20.

Martinelli, C., & Parker, S. (2009). Deception and misreporting in a social program. Journal of the European Economic Association, 7(4), 886–908.

Mebratie, A. D., Sparrow, R., Yilma, Z., Abebaw, D., Alemu, G., & Bedi, A. (2013). Impact of Ethiopian pilot community-based health insurance scheme on health-care utilisation: a

36

household panel data analysis. The Lancet, 381, S92. Mebratie, A. D., Sparrow, R., Yilma, Z., Alemu, G., & Bedi, A. S. (2015). Dropping out of Ethiopia’s

community-based health insurance scheme. Health Policy and Planning, 30(10), 1296–1306. Mebratie, A., Sparrow, R., Alemu, G., & Bedi, A. S. (2013). Community-Based Health Insurance Schemes:

A Systematic Review (ISS Working Paper Series / General Series No. 568). Meshnick, S. R. (2015). Editorial: Malaria control: Tortoises and hares. American Journal of Tropical

Medicine and Hygiene, 92(5), 885–886. Morsink, K. (2012). Weathering the storm: the demand for and impact of microinsurance. University of

Twente, Twente. Mullahy, J. (1986). Specification and testing of some modified count data models. Journal of

Econometrics, 33(3), 341–365. Mullahy, J. (1997). Instrumental-Variable Estimation of Count Data Models: Applications to Models

of Cigarette Smoking Behavior. The Review of Economics and Statistics, 79(4), 586–593. Mullahy, J. (1998). Much ado about two: Reconsidering retransformation and the two-part model in

health econometrics. Journal of Health Economics, 17(3), 247–281. Mulupi, S., Kirigia, D., & Chuma, J. (2013). Community perceptions of health insurance and their

preferred design features: implications for the design of universal health coverage reforms in Kenya. BMC Health Services Research, 13(June 2016), 1–12.

Nabyonga Orem, J., Mugisha, F., Kirunga, C., MacQ, J., & Criel, B. (2011). Abolition of user fees: The Uganda paradox. Health Policy and Planning, 26(SUPPL. 2).

Namugumya, B., Sethuraman, K., Sommerfelt, E. A., Oot, L., Kovach, T., & Musimenta, B. (2014).

Reducing Malnutrition in Uganda : Estimates to Support Nutrition Advocacy Uganda PROFILES 2013. Washington DC; Kampala.

Nguyen, C. V. (2012). The Impact of Voluntary Health Insurance on Health Care Utilization and Out-Of-Pocket Payments: New Evidence from Vietnam. Health Economics, 21, 946–966.

Nguyen, H. T. H., Rajkotia, Y., & Wang, H. (2011). The financial protection effect of Ghana National Health Insurance Scheme: evidence from a study in two rural districts. International Journal for Equity in Health, 10(1), 4.

Nguyen, K. T., Khuat, O. T. H., Ma, S., Pham, D. C., Khuat, G. T. H., & Ruger, J. P. (2012). Impact of health insurance on health care treatment and cost in Vietnam: A health capability approach to financial protection. American Journal of Public Health, 102(8), 1450–1461.

Nyandekwe, M., Nzayirambaho, M., & Kakoma, J. B. (2014). Universal health coverage in Rwanda: Dream or reality. Pan African Medical Journal, 17.

Okwero, P., Tandon, A., Sparkes, S., McLaughlin, J., & Hoogeveen, J. G. (2010). Fiscal Space for Health in Uganda (World Bank Working Paper No. 186). Washington DC.

Palmer, M., Mitra, S., Mont, D., & Groce, N. (2015). The impact of health insurance for children under age 6 in Vietnam: A regression discontinuity approach. Social Science & Medicine, 145, 217–26.

Pan, J., Lei, X., & Liu, G. G. (2016). Health Insurance and Health Status: Exploring a Causal Effect from A Policy Evaluation. Health Economics, 25(11), 1389–1402.

Panda, P., Chakraborty, A., Dror, D. M., & Bedi, A. S. (2014). Enrolment in community-based health insurance schemes in rural Bihar and Uttar Pradesh, India. Health Policy and Planning, 29(8), 960–974.

Pariyo, G. W., Ekirapa-Kiracho, E., Okui, O., Rahman, M. H., Peterson, S., Bishai, D. M., … Peters, D. H. (2009). Changes in utilization of health services among poor and rural residents in Uganda: are reforms benefitting the poor? International Journal for Equity in Health, 8, 39.

Pizer, S. D. (2009). An intuitive review of methods for observational studies of comparative

37

effectiveness. Health Services and Outcomes Research Methodology, 9(1), 54–68. Prendergast, A. J., & Humphrey, J. H. (2014). The stunting syndrome in developing countries.

Paediatrics and International Child Health, 34(4), 250–65. Quimbo, S. A., Peabody, J. W., Shimkhada, R., Florentino, J., & Solon, O. (2011). Evidence of a

causal link between health outcomes, insurance coverage, and a policy to expand access: Experimental data from children in the Philippines. Health Economics, 20(5), 620–630.

Rajkotia, Y., & Frick, K. (2012). Does household enrolment reduce adverse selection in a voluntary health insurance system? Evidence from the Ghanaian National Health Insurance System. Health Policy and Planning, 27(5), 429–437.

Rashad, I., & Kaestner, R. (2004). Teenage sex, drugs and alcohol use: Problems identifying the cause of risky behaviors. Journal of Health Economics, 23(3), 493–503.

Rashad, I., & Markowitz, S. (2007). Incentives in Obesity and Health Insurance (NBER Working Paper Series No. 13113).

Ravallion, B. Y. M., & Chaudhuri, S. (1997). Risk and Insurance in Village India : Comment. Econometrica, 65(1), 171–184.

Raza, W. A., Van de Poel, E., Bedi, A., & Rutten, F. (2016). Impact of Community-Based Health Insurance on Access and Financial Protection: Evidence from Three Randomized Trials in Rural India. Health Economics, 25, 675–687.

Reshmi, B., Sreekumaran, N., & Unnikrishnan, B. (2016). A systematic review of community-based health insurance programs in South Asia. The International Journal of Health Planning and Management.

Roodman, D. (2009). Practitioners’ corner: A note on the theme of too many instruments. Oxford Bulletin of Economics and Statistics, 71(1), 135–158.

Royston, P. (2009). Multiple imputation of missing values: Further update of ice, with an emphasis on categorical variables. Stata Journal, 9(3), 466–477.

Rubin, D. B. (1987). Multiple Imputation for Nonresponse in Surveys. John Wiley & Sons. Sanson-Fisher, R. W., Bonevski, B., Green, L. W., & D’Este, C. (2007). Limitations of the

Randomized Controlled Trial in Evaluating Population-Based Health Interventions. American Journal of Preventive Medicine, 33(2), 155–161.

Schenker, N., & Taylor, J. M. G. (1996). Partially parametric techniques for multiple imputation. Computational Statistics and Data Analysis, 22(4), 425–446.

Schoeps, A., Lietz, H., Sié, A., Savadogo, G., Allegri, M. De, Müller, O., … Souares, A. (2015). Health insurance and child mortality in rural Burkina Faso. Global Health Action, 8, 1–8.

Semba, R. D., Pee, S. De, Sun, K., Sari, M., Akhter, N., & Bloem, M. W. (2008). Effect of parental formal education on risk of child stunting in Indonesia and Bangladesh: a cross-sectional study. The Lancet, 371(9609), 322–328.

Sepehri, A. (2014). How much do I save if I use my health insurance card when seeking outpatient care? Evidence from a low-income country. Health Policy and Planning, 29(2), 246–256.

Sepehri, A., Sarma, S., & Oguzoglu, U. (2011). Does the financial protection of health insurance vary across providers? Vietnam’s experience. Social Science and Medicine, 73(4), 559–567.

Singh, K., Osei-Akoto, I., Otchere, F., Sodzi-Tettey, S., Barrington, C., Huang, C., … Speizer, I. (2015). Ghana’s National Health insurance scheme and maternal and child health: a mixed methods study. BMC Health Services Research, 15(1), 108.

Smith, K. V., & Sulzbach, S. (2008). Community-based health insurance and access to maternal health services: Evidence from three West African countries. Social Science and Medicine, 66(12), 2460–2473.

Sood, N., Bendavid, E., Mukherji, A., Wagner, Z., Nagpal, S., & Mullen, P. (2014). Government

38

health insurance for people below poverty line in India: quasi-experimental evaluation of insurance and health outcomes. BMJ, 349(sep25 4), g5114–g5114.

Sood, N., & Wagner, Z. (2016). Impact of health insurance for tertiary care on postoperative outcomes and seeking care for symptoms: quasi-experimental evidence from Karnataka, India. BMJ Open, 6(1), e010512.

Sosa-Rubí, S. G., Galárraga, O., & Harris, J. E. (2009). Heterogeneous impact of the “Seguro Popular” program on the utilization of obstetrical services in Mexico, 2001-2006: A multinomial probit model with a discrete endogenous variable. Journal of Health Economics, 28(1), 20–34.

Spaan, E., Mathijssen, J., Tromp, N., McBain, F., Have, A. ten, & Baltussen, R. (2012). The impact of health insurance in Africa and Asia: a systematic review. Bulletin of the World Health Organisation, (January), 1–16.

Spears, D. (2012). Height and cognitive achievement among Indian children. Economics and Human Biology, 10(2), 210–219.

Staiger, D., & Stock, J. (1997). Instrumental Variables Regression with Weak Instruments. Econometrica, 65(3), 557–586.

Staub, K. E. (2009). Simple Tests for Exogeneity of a Binary Explanatory Variable in Count Data Regression Models. Communications in Statistics - Simulation and Computation, 38(9), 1834–1855.

Stevens, G. A., Finucane, M. M., Paciorek, C. J., Flaxman, S. R., White, R. A., Donner, A. J., & Ezzati, M. (2012). Trends in mild, moderate, and severe stunting and underweight, and progress towards MDG 1 in 141 developing countries: A systematic analysis of population representative data. The Lancet, 380(9844), 824–834.

Stock, J. H., Wright, J. H., & Yogo, M. (2002). A Survey of Weak Instruments and Weak Identification in Generalized Method of Moments. Journal of Business & Economic Statistics, 20(4), 518–529.

Stock, J. H., & Yogo, M. (2005). Testing for Weak Instruments in Linear IV Regression. In D. W. K. Andrews & J. H. Stock (Eds.), Identification and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg (pp. 80–108). New York: Cambridge University Press.

Strobl, R. (2016). Does Health Insurance Reduce Child Labour and Education Gaps? Evidence from Rwanda. The Journal of Development Studies, 0(0), 1–20.

Terza, J. V. (2017). Two-Stage Residual Inclusion Estimation in Health Services Research and Health Economics. Health Services Research, 1–10.

Terza, J. V., Basu, A., & Rathouz, P. J. (2008). Two-stage residual inclusion estimation: Addressing endogeneity in health econometric modeling. Journal of Health Economics, 27(3), 531–543.

Terza, J. V. (2016). Two-Stage Residual Inclusion Estimation: A Practitioners Guide to Stata Implementation. In STATA Users Conference, Chicago (Vol. 46202). Chicago: STATA.

Thornton, J. A., & Rice, J. L. (2008). Does extending health insurance coverage to the uninsured improve population health outcomes? Applied Health Economics and Health Policy, 6(4), 217–230.

Titelman, D., Cetrángolo, O., & Acosta, O. L. (2015). Universal health coverage in Latin American countries: How to improve solidarity-based schemes. The Lancet, 385(9975), 1359–1363.

Townsend, R. M. . (1994). Risk and Insurance in Village India. Econometrica, 62(3), 539–591. Trujillo, A. J., Portillo, J. E., & Vernon, J. A. (2005). The Impact of Subsidized Health Insurance for

the Poor : Evaluating the Colombian Experience Using Propensity Score Matching. International Journal of Health Care Finance and Economics, 5(3), 211–239.

Twikirize, J. M. (2009). Community health insurance as a viable means of increasing access to health care for rural households in Uganda. University of Cape Town.

Twikirize, J. M., & O’Brien, C. (2012). Why Ugandan rural households are opting to pay community

39

health insurance rather than use the free healthcare services. International Journal of Social Welfare, 21(1), 66–78.

UBOS, & ICF International. (2012). Uganda Demographic and Health Survey. Kampala, Uganda; Calverton Maryland, Uganda and Calverton, Maryland: Uganda Bureau of Statistics; ICF International Inc.

Van Buuren, S. (2012). Flexible Imputation of Missing Data. CRC Press. Van Buuren, S., Boshuizen, H., & Knook, D. (1999). Multiple imputation of missing blood pressure

covariates in survival analysis. Statistic in Medicine, 18(6), 681–694. Vogl, T. S. (2014). Height, skills, and labor market outcomes in Mexico. Journal of Development

Economics, 107, 84–96. Vollmer, S., Bommer, C., Krishna, A., Harttgen, K., & Subramanian, S. (2016). The association of

parental education with childhood undernutrition in low- and middle-income countries: comparing the role of paternal and maternal education. International Journal of Epidemiology, dyw133.

Vyas, S., & Kumaranayake, L. (2006). Constructing socio-economic status indices: How to use principal components analysis. Health Policy and Planning, 21(6), 459–468.

Wagstaff, A., & Lindelow, M. (2008). Can insurance increase financial risk?. The curious case of health insurance in China. Journal of Health Economics, 27(4), 990–1005.

Wang, H., & Pielemeier, N. (2012). Community-Based Health Insurance : An Evolutionary Approach to Achieving Universal Coverage in Low-Income Countries. Journal of Life Sciences, 6, 320–329.

Wang, W., Temsah, G., & Mallick, L. (2014). Health Insurance Coverage and Its Impact on Maternal Health Care Utilization in Low and Middle Income Countries (DHS Analytical Studies No. 45). Rockvile, Maryland.

Wang, Y., & Chen, H.-J. (2012). Use of Percentiles and Z -Scores in Anthropometry. In V. R. Preedy (Ed.), Handbook of Anthropometry: Physical Measures of Human Form in Health and Disease (pp. 91–114). Springer Science+Business Media.

Wehby, G. L. (2013). Child health insurance and early preventive care in three South American countries. Health Policy and Planning, 28(3), 328–338.

Wilcox, A. J. (2001). On the importance--and the unimportance--of birthweight. International Journal of Epidemiology, 30(6), 1233–1241.

Wirtz, V. J., Santa-Ana-Tellez, Y., Servan-Mori, E., & Avila-Burgos, L. (2012). Heterogeneous effects of health insurance on out-of-pocket expenditure on medicines in Mexico. Value in Health, 15(5), 593–603.

Woldemichael, A., Gurara, D. Z., & Shimeles, A. (2016). Community-Based Health Insurance and Out-of-

Pocket Healthcare Spending in Africa : Evidence from Rwanda (IZA Discussion Paper Series No. 9922). Bonn.

Xie, Z.-X., Chou, S., & Liu, J.-T. (2016). The Short-run and Long-run Effects of Birth Weight: Evidence from Large Samples of Siblings and Twins in Taiwan. Health Economics, 19(11), 1300–1317.

Xu, K., Evans, D. B., Kadama, P., Nabyonga, J., Ogwal, P. O., Nabukhonzo, P., & Aguilar, A. M. (2006). Understanding the impact of eliminating user fees: Utilization and catastrophic health expenditures in Uganda. Social Science and Medicine, 62(4), 866–876.

Yilma, Z., Mebratie, A., Sparrow, R., Abebaw, D., Dekker, M., Alemu, G., & Bedi, A. S. (2014). Coping with shocks in rural Ethiopia. The Journal of Development Studies, 50(7), 1009–1024.

Yilma, Z., Mebratie, A., Sparrow, R., Dekker, M., Alemu, G., & Bedi, A. S. (2015). Impact of Ethiopia’s Community Based Health Insurance on household economic welfare. World Bank

40

Economic Review, 29(Supplement), S164–S173. Zheng, X., & Zimmer, D. M. (2008). Farmers’ health insurance and access to health care. American

Journal of Agricultural Economics, 90(1), 267–279.

41

ANNEX 1: Summary and other results tables Table 1: Summary statistics: two Sample t-tests for mean differences in insured and uninsured households

Variable Mean Insured (standard errors)

Mean Uninsured (standard errors)

mean difference (t-stat)

Stunting 0.488 (0.035) 0.372 (0.03) -0.116*** (-2.52) Age of child in months 29.404 (1.046) 30.707 (0.957) 1.302 (-0.91) Child is male 0.473 (0.035) 0.487 (0.031) 0.0137 (-0.29) Vitamin A supplement 0.783 (0.029) 0.762 (0.026) -0.0208 (-0.53) Birthweight 3.09 (0.034) 3.27 (0.033) 0.178*** (3.702) Taken all child immunisations 0.276 (0.031) 0.303 (0.028) 0.0268 (-0.63) Exclusive breastfeeding 0.626 (0.034) 0.64 (0.03) 0.0142 (-0.31) All U5 slept under LLIN 0.133 (0.024) 0.195 (0.025) 0.0624* (-1.79) Religion 0.340 (0.033) 0.582 (0.031) 0.242*** (5.33) Mother’s age 30.30 (0.508) 30.13 (0.441) -0.164 (-0.244) Parental education (at least one) 0.236 (0.030) 0.356 (0.03) 0.120*** (-2.8) Household >=4 0.404 (0.035) 0.410 (0.031) 0.006 (0.131) Wealth Index. Quintile 1 -1.219 (0.032) -1.324 (0.027) -0.105** (-2.274

Quintile 2 -0.766 (0.018) -0.794 (0.018) -0.026 (-1.029) Quintile 3 -0.268 (0.023) -0.307 (0.021) -0.039 (-1.265) Quintile 4 0.324 (0.037) 0.316 (0.040) -0.007 (-0.129) Quintile 5 1.983 (0.204) 2.352 (0.212) 0.368 (1.146)

Household diet diversity score 4.187 (0.091) 3.996 (0.078) -0.191 (-1.60) Insured neighbour 0.911 (0.02) 0.517 (0.031) -0.394*** (-10.01) Casual labourers 0.088 (0.018) 0.039 (0.014) 0.048** (2.09) Perception index 0.702 (0.114) -0.546 (0.101) -1.247*** (-8.18) Listening to radio 0.631 (0.034) 0.521 (0.031) -0.109** (-2.37) Aggregate group membership 2.415 (0.058) 1.374 (0.055) -1.041*** (-12.92) Differential distance 0.305 (0.032) 0.766 (0.026) 0.461*** (-11.17) Household altitude 1783.343 (7.079) 1671.151 (6.796) -112.2*** (-11.31) No. of insurance groups in village 7.818 (0.193) 2.13 (0.167) -3.940*** (-15.47) Village has a health centre 0.325 (0.033) 0.46 (0.031) 0.135*** (-2.96) Village has school 0.581 (0.035) 0.674 (0.029) 0.0930** (-2.07) Banana villages 0.276 (0.031) 0.249 (0.027) -0.0268 (-0.65) Forest villages 0.35 (0.034) 0.05 (0.013) -0.300*** (-9.01) Cost fairness perception 0.458 (0.035) 0.341 (0.029) -0.117*** (-2.58) Insurance dropout 0.074 (0.018) 0.184 (0.024) 0.110*** (-3.47) Cluster demand 0.458 (0.016) 0.502 (0.008) 0.0444*** (-2.6) Deliver at insurance facility 0.551 (0.034) 0.326 (0.029) -0.226*** (-5.01)

N 203 261 464

Standard errors for means in parentheses; t statistics for the mean difference in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

42

Histogram for the frequency of the number of years in insurance

Instrumental variable validity testing Table 2: OLS regression results for association of the instruments with the insurance status

(1) (2) (3) (4) (5) VARIABLES insure insure insure insure insure

Ever dropped out -0.231*** -0.233*** (0.0665) (0.0644) Cluster demand rate -0.325*** -0.281** (0.125) (0.120) Cost fairness 0.121** 0.131*** (0.0469) (0.0451) Deliver in insurance facility 0.228*** 0.221*** (0.0455) (0.0444) Constant 0.469*** 0.595*** 0.390*** 0.341*** 0.459*** (0.0245) (0.0646) (0.0294) (0.0296) (0.0673) Observations 464 464 464 464 464 R-squared 0.025 0.014 0.014 0.051 0.103

Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

05

01

00

150

200

250

Fre

que

ncy

0 5 10 15 20Number of years in insurance

43

Table 3: OLS regression results for the association between the instruments and treatment intensity – number of years in insurance

(1) (2) (3) (4) (5) VARIABLES Years in insurance

Ever dropped out -1.078** -1.102** (0.438) (0.430) Cluster demand rate -1.976** -1.772** (0.819) (0.802) Cost fairness 0.835*** 0.886*** (0.307) (0.301) Deliver in insurance facility 1.024*** 0.985*** (0.302) (0.297) Constant 2.364*** 3.171*** 1.890*** 1.783*** 2.456*** (0.161) (0.423) (0.192) (0.197) (0.450) Observations 464 464 464 464 464 R-squared 0.013 0.012 0.016 0.024 0.065


Table 4: OLS regression results for the association of the instruments on the outcome - stunting

(1) (2) (3) (4) (5) VARIABLES stunted stunted stunted stunted stunted

Ever dropped out 0.0806 0.0813 (0.0670) (0.0674) Cluster demand rate 0.103 0.0904 (0.125) (0.126) Cost fairness -0.0351 -0.0393 (0.0470) (0.0472) Deliver in insurance facility -0.0372 -0.0351 (0.0465) (0.0465) Constant 0.411*** 0.373*** 0.436*** 0.438*** 0.398*** (0.0247) (0.0647) (0.0295) (0.0303) (0.0706) Observations 464 464 464 464 464 R-squared 0.003 0.001 0.001 0.001 0.007


44

Table 5: Instrumental variable results for the causal relationship between insurance and stunting (1) (2) (3) (4) (5) VARIABLES stunted stunted stunted stunted stunted

insure -0.349 -0.290 -0.317 -0.163 -0.247* (0.292) (0.384) (0.392) (0.202) (0.143) Residuals 1 0.477 (0.295) Residuals 2 0.412 (0.386) Residuals 3 0.439 (0.394) Residuals 4 0.295 (0.207) Residuals combined 0.404*** (0.149) Constant 0.575*** 0.549*** 0.561*** 0.494*** 0.530*** (0.130) (0.170) (0.174) (0.0916) (0.0672) Observations 464 464 464 464 464 R-squared 0.019 0.016 0.016 0.018 0.029

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Table 6: Instrumental variable results for the causal association between years of insurance and stunting

(1) (2) (3) (4) (5) VARIABLES stunted stunted stunted stunted stunted

Years of insurance -0.0747 -0.0420 -0.0521 -0.0363 -0.0484* (0.0623) (0.0557) (0.0651) (0.0450) (0.0277) Residuals 1 0.0936 (0.0625) Residuals 2 0.0606 (0.0560) Residuals 3 0.0707 (0.0657) Residuals 4 0.0553 (0.0454) Residuals combined 0.0706** (0.0286) Constant 0.588*** 0.516*** 0.538*** 0.503*** 0.530*** (0.140) (0.126) (0.147) (0.103) (0.0664) Observations 464 464 464 464 464 R-squared 0.018 0.016 0.016 0.017 0.026


45

Longer version Empirical Results Table 7: First stage IV results

(1) (2) VARIABLES insure Years

insured

Ever dropped out -1.263*** 0.621

(0.275) (0.917)

Cluster insurance demand rate

-0.484 -0.100

(0.628) (1.463)

Perception of cost fairness

0.359* 0.929*

(0.192) (0.499)

Delivery in insurance facility

0.466** 0.158

(0.204) (0.530) Age of the child (in months)

0.0104 0.104

(0.0283) (0.0688) Age squared -0.000358 -0.00104 (0.000422) (0.00110) Gender of the child (1=male)

0.208 0.0855

(0.186) (0.476) Vitamin A supplementation

-0.0566 0.828

(0.273) (0.870) All child immunisations 0.0385 -0.602 (0.455) (0.912) Vitamin A ## Child immunisation

-0.0464 -0.508

(0.512) (1.091) All U-5 sleep in LLIN -0.370 1.351* (0.248) (0.741) Birthweight 0.0612 -0.0671 (0.195) (0.409) Exclusive breastfeeding 0.358* -0.174 (0.199) (0.546) Religion (base= Catholic) Protestant -0.365 0.692 (0.224) (0.580) Others -0.541 -0.0496 (0.332) (0.799) Parental secondary education

-0.367 0.329

(0.227) (0.575) Household <=4 -0.108 -

1.698*** (0.189) (0.480)

Wealth quintile (base= quintile 1)

Quintile 2 0.449 -0.572 (0.291) (0.687) Quintile 3 0.490 1.182 (0.309) (0.977) Quintile 4 0.302 0.225 (0.306) (0.716) Quintile 5 0.695* 0.766 (0.366) (0.820) Household Diet Diversity Score

0.0443 -0.312

(0.0814) (0.215) Have an insured neighbour

0.510** 0.484

(0.248) (0.807) Casual employment -0.205 0.307 (0.400) (1.406) Composite perception index

0.228*** -0.0468

(0.0643) (0.179) Composite satisfaction index

-0.0422 0.0523

(0.0345) (0.0794) Perceptions# satisfaction indices

0.00127 0.0622

(0.0218) (0.0481) Waiting time 0.00103 -

0.000776 (0.000902) (0.00180) Number of groups 1.635*** 1.638 (0.378) (1.187) Number of groups squared

-0.196** -0.209

(0.0821) (0.224) Differential distance -0.327 0.0379 (0.324) (0.961) Household altitude 8.00e-05 -

0.000914 (0.00165) (0.00353) Village has a school 0.656 1.207 (0.444) (1.248) Village has a health centre

-0.124 0.622

(0.406) (1.408) Number of insurance groups in village

0.0884 -0.440

(0.136) (0.320) Insurance coverage in 1.848* 8.084***

46

2010 (1.040) (2.689) Forest village 0.701 0.703 (0.534) (1.533) Banana village 0.755** 1.879* (0.382) (1.055) Constant -4.903 -0.931

(2.991) (5.931) Observations 464 203


47

Table 8: Effect of number of years (Categorical) in insurance on stunting reduction

(1) VARIABLES stunted

Base category (0 years) 1 Year -0.0378 (0.409) 2 Years -0.165 (0.407) 3 Years -0.613 (0.408) 4 Years -0.326 (0.525) 5 Years -0.543 (0.513) 6 Years -1.681*** (0.641) 7 Years -0.661 (0.643) 8 Years -1.694* (0.895) 9 Years -1.816* (0.933) 10 Years -1.952** (0.900) 13 Years -3.060*** (1.169) 20 Years -3.685** (1.586) Age of the child (in months) 0.0804*** (0.0252) Age squared -0.00125*** (0.000383) Gender of the child (1=male) 0.0972 (0.158) Vitamin A supplementation -0.281 (0.236) All child immunisations -0.104 (0.382) Vitamin A ## Child immunisation 0.339 (0.423) All U-5 sleep in LLIN 0.159 (0.202) Birthweight -0.338** (0.161) Exclusive breastfeeding 0.440**

(0.172) Religion (base= Catholic) Protestant 0.192 (0.201) Others -0.0857 (0.292) Parental secondary education -0.327* (0.193) Household <=4 -0.404** (0.174) Wealth quintile (base= quintile 1) Quintile 2 -0.372 (0.250) Quintile 3 -0.201 (0.283) Quintile 4 -0.433 (0.265) Quintile 5 -0.381 (0.301) Household Diet Diversity Score -0.0276 (0.0686) Have an insured neighbour 0.139 (0.210) Casual employment 0.0423 (0.331) Composite perception index -0.0181 (0.0560) Composite satisfaction index -0.00193 (0.0298) Perceptions# satisfaction indices 0.0127 (0.0175) Waiting time -0.000686 (0.000833) Number of groups 0.526** (0.259) Number of groups squared -0.0893 (0.0576) Differential distance -0.395 (0.269) Household altitude 0.000558 (0.00146) Village has a school 0.339 (0.387) Village has a health centre -0.0667 (0.304) Number of insurance groups in -0.0371

48

village (0.121) Insurance coverage in 2010 1.706 (1.151) Forest village 0.345 (0.459) Banana village 0.282

(0.333) Residuals 0.244*** (0.0910) Constant -1.472 (2.681) Observations 457

Standard errors in parentheses bootstrapped to 10000 replications.

*** p<0.01, ** p<0.05, * p<0.1

Table 9: Predictive mean probability of stunting at different years of insurance

No of years insured Margin

Delta-method Std. Err. Z P>z [95% Confidence Interval]

N

0 0.518292 0.0375426 13.81 0.000 0.44471 0.591874 261

1 0.498884 0.0848484 5.88 0.000 0.332585 0.665184 21

2 0.471088 0.0834942 5.64 0.000 0.307443 0.634734 24

3 0.339517 0.0685047 4.96 0.000 0.20525 0.473784 32

4 0.424383 0.0987962 4.3 0.000 0.230746 0.61802 17

5 0.359425 0.0785127 4.58 0.000 0.205543 0.513307 26

6 0.133245 0.045131 2.95 0.003 0.04479 0.221701 23

7 0.328831 0.1045062 3.15 0.002 0.124003 0.533659 17

8 0.127642 0.0767238 1.66 0.096 -0.02273 0.278018 7

9 0.114448 0.071347 1.6 0.109 -0.02539 0.254285 6

10 0.101856 0.0514039 1.98 0.048 0.001106 0.202606 17

13 0.027391 0.0220369 1.24 0.214 -0.0158 0.070582 8

20 0.010437 0.0157955 0.66 0.509 -0.02052 0.041396 5

49

Table 10: Local average treatment effect and linearity of the treatment effects (1) (2) VARIABLES stunted stunted

Years in insurance -0.236 -0.133 (0.229) (0.170) Years squared -0.00196 (0.0144) Age of the child (in months)

0.0700 0.0844***

(0.0597) (0.0232) Age squared -0.000995 -

0.00132*** (0.000892) (0.000356) Gender of the child (1=male)

0.0659 0.0930

(0.377) (0.147) Vitamin A supplementation

-0.502 -0.290

(0.565) (0.221) All child immunisations

0.0659 -0.156

(0.840) (0.361) Vitamin A ## Child immunisation

-0.00890 0.371

(0.919) (0.394) All U-5 sleep in LLIN 0.729 0.156 (0.582) (0.203) Birthweight -0.695* -0.314** (0.376) (0.151) Exclusive breastfeeding 0.634* 0.411** (0.381) (0.160) Religion (base= Catholic)

Protestant 0.301 0.237 (0.468) (0.191) Others -0.211 -0.0616 (0.674) (0.280) Parental secondary education

-0.0884 -0.317*

(0.425) (0.186) Household <=4 -0.533 -0.346** (0.452) (0.172) Wealth quintile (base= quintile 1)

Quintile 2 -1.012 -0.322 (0.681) (0.232) Quintile 3 -0.555 -0.154 (0.792) (0.267) Quintile 4 -0.594 -0.375 (0.673) (0.249)

Quintile 5 -1.116 -0.341 (0.872) (0.284) Household Diet Diversity Score

-0.0785 -0.0246

(0.156) (0.0649) Have an insured neighbour

0.169 0.122

(0.748) (0.201) Casual employment -0.172 0.0606 (0.832) (0.296) Composite perception index

-0.0115 -0.0112

(0.132) (0.0523) Composite satisfaction index

0.00598 -0.00292

(0.0714) (0.0269) Perceptions# satisfaction indices

0.0230 0.0108

(0.0410) (0.0161) Waiting time -0.00234 -0.000711 (0.00153) (0.000749) Number of groups -0.109 0.480* (1.260) (0.253) Number of groups squared

0.00891 -0.0815

(0.233) (0.0556) Differential distance -0.790 -0.358 (0.752) (0.250) Household altitude 0.00307 0.000475 (0.00328) (0.00138) Village has a school 0.835 0.254 (0.914) (0.353) Village has a health centre

-0.590 -0.0860

(1.707) (0.286) Number of insurance groups in village

-0.197 -0.0113

(0.278) (0.112) Insurance coverage in 2010

1.186 1.451

(2.903) (1.064) Forest village -0.221 0.256 (1.668) (0.426) Banana village -0.135 0.245 (1.458) (0.298) Residuals 0.270 0.185 (0.237) (0.188) Residuals (Years squared)

0.00188

(0.0152)

50

Constant -1.743 -1.421 (5.885) (2.462) First stage F-statistic 26.56*** 26.56*** AIC 1.434 1.334

Observations 203 464

Bootstrapped standard errors with 3000 replications in parentheses

*** p<0.01, ** p<0.05, * p<0.1

51

Table 11: Heterogeneous treatment effects on child age groups and mother’s age

(1) (2) VARIABLES Effect with

mother’s age Effect with child age

groups

Years in insurance -0.184** -0.262** (0.0849) (0.110) Mother age > 30.2 years 0.0454 (0.209) Mother age > 30.2 years #Years in insurance 0.0999** (0.0493) Age of child (months 0.0819*** (0.0239) Age squared -0.00130*** (0.000366) Age categories (base: 6-11 months)

12-23 months 0.119 (0.309)

24-35 months 0.498 (0.303)

36-47 months 0.380 (0.325)

48-60 months -0.370 (0.353) Age category# Insurance 12-23 months # Number of years in insurance 0.143 (0.0980) 24-35 months # Number of years in insurance 0.0336 (0.0912) 36-47 months # Number of years in insurance 0.121 (0.0941) 48-60 months # Number of years in insurance 0.0838 (0.0986) Gender of the child (1=male) 0.0470 0.116 (0.146) (0.156) Vitamin A supplementation -0.378* -0.313 (0.219) (0.223) All child immunisations -0.0708 -0.0487 (0.364) (0.378) Vitamin A ## Child immunisation 0.388 0.281 (0.398) (0.413) All U-5 sleep in LLIN 0.115 0.102 (0.202) (0.200) Birthweight -0.326** -0.338** (0.150) (0.158) Exclusive breastfeeding 0.418*** 0.490*** (0.156) (0.164) Religion (base= Catholic) Protestant 0.252 0.217 (0.188) (0.191) Others -0.0345 -0.0879 (0.271) (0.276) Parental secondary education -0.291 -0.320* (0.178) (0.186) Household <=4 -0.230 -0.409** (0.177) (0.168)

52

Wealth quintile (base= quintile 1) Quintile 2 -0.341 -0.280 (0.230) (0.237) Quintile 3 -0.237 -0.120 (0.263) (0.258) Quintile 4 -0.403 -0.375 (0.247) (0.259) Quintile 5 -0.358 -0.318 (0.289) (0.283) Household Diet Diversity Score -0.00453 -0.0204 (0.0645) (0.0665) Have an insured neighbour 0.0971 0.134 (0.195) (0.208) Casual employment 0.00799 -0.0291 (0.307) (0.318) Composite perception index -0.0203 0.00252 (0.0515) (0.0528) Composite satisfaction index -0.0117 -0.00410 (0.0274) (0.0276) Perceptions# satisfaction indices 0.00821 0.0102 (0.0161) (0.0164) Waiting time -0.000710 -0.000773 (0.000717) (0.000760) Number of groups 0.471* 0.532** (0.245) (0.262) Number of groups squared -0.0817 -0.0841 (0.0555) (0.0575) Differential distance -0.381 -0.424* (0.251) (0.257) Household altitude 0.000674 0.000221 (0.00137) (0.00142) Village has a school 0.201 0.245 (0.354) (0.373) Village has a health centre -0.101 -0.0338 (0.282) (0.295) Number of insurance groups in village -0.0184 0.0102 (0.111) (0.116) Insurance coverage in 2010 1.249 1.411 (1.021) (1.094) Forest village 0.179 0.295 (0.417) (0.431) Banana village 0.227 0.297 (0.302) (0.308) Residuals 0.153** 0.239*** (0.0769) (0.0833) Constant -1.602 -0.135 (2.454) (2.543) First stage F-Statistic 26.77*** 26.88*** AIC 1.398 1.398 Observations 464 464


53

Table 12: With and without exclusive breast feeding (1) (2) VARIABLES With Excl.

Breastfeeding Without Excl. Breastfeeding

Years in insurance -0.154** -0.168** (0.0709) (0.0724) Age of the child (in months) 0.0846*** 0.0897*** (0.0232) (0.0224) Age squared -0.00132*** -0.00138*** (0.000355) (0.000347) Gender of the child (1=male) 0.0962 0.106 (0.146) (0.146) Vitamin A supplementation -0.294 -0.243 (0.214) (0.214) All child immunisations -0.150 -0.199 (0.361) (0.363) Vitamin A ## Child immunisation 0.369 0.431 (0.396) (0.397) All U-5 sleep in LLIN 0.149 0.151 (0.197) (0.190) Birthweight -0.316** -0.292* (0.153) (0.151) Exclusive breastfeeding 0.417*** (0.154) Religion (base= Catholic) Protestant 0.233 0.221 (0.187) (0.183) Others -0.0624 -0.136 (0.275) (0.269) Parental secondary education -0.319* -0.330* (0.181) (0.176) Household <=4 -0.338** -0.334** (0.160) (0.156) Wealth quintile (base= quintile 1) Quintile 2 -0.318 -0.283 (0.230) (0.226) Quintile 3 -0.167 -0.0908 (0.254) (0.250) Quintile 4 -0.369 -0.302 (0.247) (0.241) Quintile 5 -0.335 -0.256 (0.280) (0.282) Household Diet Diversity Score -0.0235 -0.0283 (0.0642) (0.0638) Have an insured neighbour 0.127 0.143 (0.199) (0.197) Casual employment 0.0531 0.00732 (0.296) (0.294) Composite perception index -0.00900 -0.0144 (0.0512) (0.0499) Composite satisfaction index -0.00345 -0.00143 (0.0275) (0.0271) Perceptions# satisfaction indices 0.0106 0.0107 (0.0160) (0.0158) Waiting time -0.000698 -0.000501 (0.000723) (0.000723)

54

Number of groups 0.491** 0.486** (0.244) (0.247) Number of groups squared -0.0831 -0.0773 (0.0541) (0.0538) Differential distance -0.365 -0.337 (0.245) (0.241) Household altitude 0.000480 0.000583 (0.00137) (0.00136) Village has a school 0.253 0.306 (0.354) (0.353) Village has a health centre -0.0842 -0.0255 (0.290) (0.287) Number of insurance groups in village -0.00872 -0.0340 (0.110) (0.109) Insurance coverage in 2010 1.430 1.616 (1.049) (1.034) Forest village 0.260 0.251 (0.423) (0.418) Banana village 0.245 0.312 (0.296) (0.303) Residuals 0.205*** 0.221*** (0.0796) (0.0815) Constant -1.439 -1.648 (2.452) (2.452) First stage Wald test 26.56*** 26.25*** AIC 1.325 1.336 Observations 464 464


55

Table 13: Robustness checks: Comparison of main model alongside IV with ZIP, IV with Poisson and IV-Probit.

Main model Comparison models

VARIABLES IV-2SRI IV with ZIP IV-PROBIT

Years in insurance -0.154** -0.149** -0.165* (0.0709) (0.0628) (0.0925) Age of the child (in months) 0.0846*** 0.0822*** 0.0737*** (0.0232) (0.0197) (0.0201) Age squared -0.00132*** -0.00125*** -0.00114*** (0.000355) (0.000300) (0.000310) Gender of the child (1=male) 0.0962 0.0653 0.0827 (0.146) (0.128) (0.123) Vitamin A supplementation -0.294 -0.292 -0.275 (0.214) (0.186) (0.180) All child immunisations -0.150 -0.153 -0.141 (0.361) (0.304) (0.285) Vitamin A ## Child immunisation 0.369 0.394 0.355 (0.396) (0.334) (0.320) All U-5 sleep in LLIN 0.149 0.160 0.108 (0.197) (0.169) (0.169) Birthweight -0.316** -0.305** -0.278** (0.153) (0.130) (0.124) Exclusive breastfeeding 0.417*** 0.336** 0.360*** (0.154) (0.135) (0.139) Religion (base= Catholic) Protestant 0.233 0.252 0.213 (0.187) (0.165) (0.155) Others -0.0624 -0.00564 -0.0971 (0.275) (0.230) (0.208) Parental secondary education -0.319* -0.282* -0.253* (0.181) (0.158) (0.149) Household <=4 -0.338** -0.282** -0.276** (0.160) (0.137) (0.124) Wealth quintile (base= quintile 1) Quintile 2 -0.318 -0.385* -0.295 (0.230) (0.199) (0.184) Quintile 3 -0.167 -0.264 -0.162 (0.254) (0.214) (0.229) Quintile 4 -0.369 -0.398* -0.308 (0.247) (0.211) (0.224) Quintile 5 -0.335 -0.424* -0.311 (0.280) (0.240) (0.252) Household Diet Diversity Score -0.0235 -0.0194 -0.0362 (0.0642) (0.0551) (0.0533) Have an insured neighbour 0.127 0.165 0.152 (0.199) (0.171) (0.153) Casual employment 0.0531 0.0484 0.0539 (0.296) (0.257) (0.241) Composite perception index -0.00900 -0.0380 -0.00374 (0.0512) (0.0432) (0.0447) Composite satisfaction index -0.00345 -0.00171 0.000332 (0.0275) (0.0236) (0.0222) Perceptions# satisfaction indices 0.0106 0.0106 0.0122 (0.0160) (0.0135) (0.0124) Waiting time -0.000698 -0.000703 -0.000592

56

(0.000723) (0.000627) (0.000593) Number of groups 0.491** 0.472** 0.492*** (0.244) (0.207) (0.186) Number of groups squared -0.0831 -0.0906** -0.0815* (0.0541) (0.0458) (0.0429) Differential distance -0.365 -0.337 -0.310 (0.245) (0.212) (0.194) Household altitude 0.000480 0.000398 0.000375 (0.00137) (0.00121) (0.00111) Village has a school 0.253 0.109 0.273 (0.354) (0.299) (0.304) Village has a health centre -0.0842 0.00979 -0.0211 (0.290) (0.248) (0.220) Number of insurance groups in village -0.00872 -0.00171 -0.0146 (0.110) (0.0958) (0.0907) Insurance coverage in 2010 1.430 0.969 1.440 (1.049) (0.834) (0.970) Forest village 0.260 0.260 0.341 (0.423) (0.365) (0.356) Banana village 0.245 0.250 0.271 (0.296) (0.257) (0.240) Residuals 0.205*** 0.198*** (0.0796) (0.0689) Constant -1.439 -1.162 -1.281 (2.452) (2.142) (1.931) First stage Wald test 26.56*** 57.17*** 18.65*** AIC 1.325 1.323 Wald test for exogeneity 2.72* Observations 464 464 464

2SRI Model standard errors bootstrapped with 10000 replications and other models, robust standard errors in parentheses. Reported significance levels *** p<0.01, ** p<0.05, * p<0.1

57

ANNEX 2: A note of imputation for birthweight

Normally, for conventional multiple imputation procedures, an assumption is made that data is

missing completely at random or simply missing at random (MAR). However, as Rubin(1987) notes,

it is rarely the case that data is MAR and the presumption that responders are systematically different

from non-responders is plausible. We therefore suspect that our missing data is missing not at

random (MNAR). To be certain of this, we conduct simple t-tests and study the mean differences on

a set of selected observables. From the table below, we can observe that in our data, there are

significant differences between the households which birthweight was observed and the ones which

it was not observed.

Examination of the differences between households with observed birthweight and missing

birthweight by selected observables

Observed birthweight

Missing birthweight

mean se mean se mean diff t-stat

SES 0.220 0.092 -0.273 0.078 -0.493*** (-3.97) Distance to facility 11.179 0.217 11.770 0.399 0.591 (1.37) Perception about health insurance 0.116 0.107 -0.144 0.123 -0.261 (-1.61) Satisfaction with health services 0.065 0.179 -0.081 0.217 -0.146 (-0.52) HDDS 4.245 0.082 3.874 0.083 -0.371** (-3.13) Parental education 0.374 0.030 0.217 0.029 -0.156*** (-3.68) Household size 5.062 0.120 5.763 0.156 0.701*** (3.61) ANC services 0.463 0.031 0.338 0.033 -0.125** (-2.74) All child immunizations 0.261 0.027 0.329 0.033 0.0678 (1.60) Iron supplementation 0.125 0.021 0.073 0.018 -0.0517 (-1.83) Child deworming 0.739 0.027 0.744 0.030 0.00466 (0.11) Vitamin A Supplementation 0.813 0.024 0.720 0.031 -0.0934* (-2.39) Treated water 0.565 0.031 0.478 0.035 -0.0864 (-1.85) Protected water source 0.708 0.028 0.609 0.034 -0.0995* (-2.26) Handwashing facilities 0.156 0.023 0.053 0.016 -0.103*** (-3.55) Stunted 0.385 0.030 0.469 0.035 0.0834 (1.81) Diarrhea 0.265 0.028 0.237 0.030 -0.0279 (-0.69) Fever 0.128 0.021 0.111 0.022 -0.0173 (-0.57) Cough (ARI) 0.514 0.031 0.493 0.035 -0.0209 (-0.45) Waiting time 92.553 7.084 83.739 7.142 -8.813 (-0.87) Insured neighbor 0.681 0.029 0.705 0.032 0.0244 (0.56) Observations 257 207 464

t statistics in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001

These differences appear in about eight of the twenty selected observables. Normally, if only a

handful of variables present statistically significant differences in the means of the observables, it can

58

be assumed that the differences are due to random chance (Angrist & Pischke, 2015, p. 21) and

hence MAR. However, in our case, more than one third of the observed variables have significant

difference as shown in the table above. So we cannot confidently make the MAR non-ignorability

assumption (Rubin, 1987, p. 202).

The imputation

In our imputation, we use STATA 14 MI impute procedure. STATA imputation procedure generally

assumes that data is MAR (Royston, 2009) so our imputation procedure needs additional

assumptions to incorporate for non-random missingness (Horton & Kleinman, 2007). Rubin (1987,

pp. 202–204) suggested some MNAR imputation procedures which include incorporating fixed

transformations that increase or decrease the values of MAR depending on the assumptions used, or

incorporating with in the imputation procedure a distortion in the probability of drawing values to

use in the imputation function. These procedures have been improved over the years (Eddings &

Marchenko, 2012; Van Buuren, Boshuizen, & Knook, 1999). For our model of imputation, we

specify a linear regression model as recommended by the Stata multiple imputation procedure.7 In

the end, what we get as the imputed values of the missing variable will be MAR + MNAR, where

MAR is the mean difference caused by the predictors in the imputation model and the additional

MNAR is from the constant or random variation (Van Buuren, 2012, p. 92)

In the predictive model of birthweight, we include a large number of predictive variables (including

our outcome variable stunting, and all other variables used in the analysis) in order to maximise the

predictive power of missingness (Kenward & Carpenter, 2007). This procedure is also

recommended in Stata’s multiple imputation procedure. For imputation of missing data of MNAR

nature, there is need to add a term that incorporates a term that accounts for the missingness. For

instance, Rubin (1987, pp. 155, 203) suggests that a constant term of 20% can be added on or

subtracted from the imputed values to account for the assumption that the missing values are 20%

lower or higher than the observed. Or that a constant value can be added or subtracted from half of

the imputed values as a distortion on the fixed transformation or introduce a distortion in the

probability of drawing values to impute using the function of the value to be imputed (Rubin, 1987,

p. 204). However, Rubin’s suggestions are thought to be arbitrary and instead adding a model

specific standardized mean difference is advised (Enders, 2010). In our case, after executing the

7 This linear regression model is also similar to the pattern mixture model (Carpenter & Kenward, 2013, p. 18)

59

linear regression, we replace the missing values with the predicted values plus a random noise

equivalent to a random number with mean of 0 and standard deviation being the standard error of

the linear prediction. This implies that each imputed observation is subjected to a particular random

procedure that accounts for the non-random missingness.

In the table below, we present how the imputation behaves. The overall mean of birthweight

increases by 1.3% from 3.15 to 3.19 kilograms. The proportion of low birthweight also increases

slightly from 6.2% to 6.5%.

Examination of the imputation exercises for birthweight

obs mean min max sd Imputed birthweight 207 3.23 1.59 4.86 0.540 Observed birthweight 257 3.15 2 5.6 0.532 Complete birthweight 464 3.19 1.59 5.6 0.517

Low birth weight

Imputed low birth weight 207 0.0676 0 1 0.225 Observed low birth weight 257 0.0622 0 1 0.242 Complete low birth weight 464 0.0647 0 1 0.234

The k-density graph below shows the distribution of the observed values, imputed valued and completed values of birth weight. The graph shows that the distribution of completed values fit smoothly along the distribution of observed values. This is the desired scenario.

If the data were MAR or missing completely at random, we would expect our distribution to be

close to fully fitting into each other (Marchenko & Eddings, 2011).

0.5

11

.5

kd

en

sity for

birth

weig

ht dis

trib

utio

n

2 3 4 5 6birthweight

Observed Imputed Completed

MNAR imputation (with standard errors of predicted values)

60

Sensitivity checks for imputation

Normally, when imputation procedures assume that data is MAR, one method of sensitivity analysis

is to assume that the data is MNAR and compare results in relation to how close the imputations get

close to what is observed (Kenward & Carpenter, 2007). If the data are MAR, it would be expected

that the assumption of MNAR would yield results not so different from each other. In our case, we

are certain that our data is MNAR and so compare with the results if we assumed that the data was

MAR. Our expectation is that the results would be different and the MNAR – complete values

would be close to the observed values than the MAR results. We therefore take the MAR

assumption and execute the imputation using the Stata MAR multiple imputation procedure

(Eddings & Marchenko, 2012). To improve the MAR imputation, Horton & Lipsitz (2001) suggest

changing the choice of predictors and or increasing the number of imputations (Rubin, 1987; S Van

Buuren et al., 1999). Earlier imputation procedures suggested that with up to 50% of the data are

missing, 5 -10 imputations can deliver an efficiency of up to 95% and this is sufficient to obtain valid

inference, however with a large number of missing values, it has been recommended to undertake

even up to 200 imputations to arrive at unbiased results (Horton & Lipsitz, 2001; Kenward &

Carpenter, 2007). Following this recommendation of increasing the number of imputations, we carry

out our sensitivity analysis with 100 imputations under a MAR assumption and compare the 1st, 50th

and 100th imputation and a combination of all 100 imputations.

0.5

11

.5

kde

nsity

1 2 3 4 5 6birth weight (kgs)


1st imputation

0.5

11

.5

kd

en

sity

1 2 3 4 5 6Birth weight (kgs)


50th imputation

61

Our comparison of the generated values of missing birthweight – based on the kernel density graphs

above show that even after 100 imputations, the distribution of completed observations still

performs worse than in our assumption of MNAR. We find that the assumption of MAR is still

inadequate to deliver imputations close enough to observed values when compared to MNAR. Our

assumption on MNAR is therefore in line with Van Buuren (2012, p. 93) recommendation that the

sensitivity procedure should have better understanding of the scenarios behind the missing data.

This method of sensitivity test has been applied elsewhere in health services research (Longford,

1999; Longford et al., 2000).

0.5

11

.5

kd

en

sity



100th imputation

0.5

11

.5

kd

en

sity



100 imputations

62

Simple OLS Model for predicting birthweight

(1)

VARIABLES birthweight

stunted -0.0204

(0.0826)

Mother’s age 0.00298

(0.00716)

No of years in insurance -0.0122

(0.0164)

Male -0.0274

(0.0828)

Age of child -0.00329

(0.00282)

Vitamin A supplement -0.193*

(0.113)

All immunisation 0.0167

(0.0962)

LLIN U-5 usage -0.0396

(0.112)

Exclusive breastfeeding 0.162*

(0.0881)

Religion (base: catholic

protestant 0.135

(0.107)

other 0.140

(0.145)

Parental education -0.110

(0.0932)

Household <=4 -0.0118

(0.128)

Used fertilizer 0.0354

(0.164)

Used pesticide -0.114

(0.105)

Had a good harvest -0.0106

(0.0886)

In a farmer’s group 0.0629

(0.130)

Extension services -0.0736

(0.104)

Farmer training 0.0105

(0.126)

Improved bathroom 0.146

(0.287)

Improved toilet 0.214

(0.192)

Handwashing facility 0.215*

(0.123)

Perceptions on insurance:

Premiums -0.0956**

(0.0421)

Health beliefs 0.0325

(0.0417)

Social influence -0.0607**

(0.0254)

Scheme management -0.0106

(0.0401)

Quality of care -0.0276

(0.0363)

Financial protection 0.0161

(0.0221)

Scheme convenience 0.0458

(0.0357)

HDDS -0.00179

(0.0339)

Has an insured neighbour -0.0820

(0.101)

Casual employment 0.0623

(0.162)

Land size -0.0125

(0.0373)

Livestock 0.0137

(0.0223)

Listen to radio daily 0.0940

(0.0927)

No of group membership -0.0221

(0.0516)

Differential distance -0.162

(0.128)

Household altitude -0.00115

(0.000897)

Village has a:

Health centre -0.0781

(0.214)

Road 0.314

(0.250)

school 0.112

(0.292)

No of insurance groups in village 0.0954

63

(0.0687)

Village economy (base: pastoral

villages

Banana villages 0.434

(0.296)

Forest villages 0.683*

(0.395)

Trade villages 0.0257

(0.215)

Attended at least 4 ANC 0.0591

(0.106)

Household size 0.0101

(0.0376)

At least 3 meals a day 0.0208

(0.0815)

Housing endowments -0.0176

(0.0328)

Household assets -0.0285

(0.0307)

Rooms-people ratio 0.231

(0.175)

Married 0.187*

(0.107)

Satisfaction with:

Time with health workers 0.0538

(0.0734)

Waiting time 0.0622

(0.0396)

Explaining to health workers -0.0389

(0.0863)

Not rushed by doctors -0.0422

(0.0982)

Information from doctors 0.00899

(0.121)

Advice on your illness -0.0294

(0.119)

Information on tests done 0.000300

(0.0833)

Conduct of health workers 0.0152

(0.112)

Doctor’s explanations -0.0251

(0.104)

Care showed by health workers 0.103

(0.113)

Interest by health workers 0.0330

(0.115)

Treated with dignity 0.0264

(0.0979)

Conduct of other hospital staff -0.0372

(0.101)

Other staff were helpful -0.114

(0.103)

Other staff kind and courteous 0.0775

(0.119)

Cost was fair -0.0255

(0.0415)

High toilet coverage 0.471*

(0.264)

Protected water source -0.191

(0.177)

Connected to leaders -0.0360

(0.0960)

Lent in six months 0.0476

(0.0880)

Borrowed in six months 0.0799

(0.0918)

Iron supplement -0.149

(0.130)

deworming 0.194**

(0.0982)

Waiting time 0.000403

(0.000393)

Constant 3.234*

(1.643)

Observations 257

R-squared 0.279

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

ZEF Bonn · West Nile 37.7 37.8 Western 37.6 43.9 South Western 49.6 41.7 Karamoja 53.6 45.0...

Documents

Transcript of ZEF Bonn · West Nile 37.7 37.8 Western 37.6 43.9 South Western 49.6 41.7 Karamoja 53.6 45.0...