Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Essays on development and health economics
(USC Thesis Other)
Essays on development and health economics
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
ESSAYS ON DEVELOPMENT AND HEALTH ECONOMICS
by
Dawoon Jung
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
August 2018
Copyright 2018 Dawoon Jung
Dedication
To my beloved fianc´ e, Chan Mee Koak
and my supportive family, Youngsoo Jung, Gisook Choi, and Yujin Jung
ii
Acknowledgement
I would like first to thank my numerous advisors for their invaluable guidance and support. I especially
thank my advisor Prof. John Strauss and Prof. Jeffrey Nugent. They have provided much feedback, support,
time, and patience, all of which have guided me to become a better economist. Sincere thanks to Prof. John
Strauss for his constant constructive training and insightful feedback, which helped me to become a critical
thinker. A hearty thank you to Prof. Jeffrey Nugent for instilling confidence in my research and his endless
encouragement. I am also thankful to Dr. Jinkook Lee who has led me to the field of aging economics and
provided me an opportunity of field research experience in India. I am incredibly grateful to her for constant
support and encouragement. I have also benefited greatly from discussions with Dr. Erik Meijer and want to
thank him for his time and keen insights.
This dissertation would not have been possible without my friend and colleague, Tushar Bharati. I am
hugely indebted to your support, feedback and selflessness. It has always been a wonderful time to have an
intellectual conversation with you, Tushar. I would also like to express my gratitude to Dr. Su Jin Lee for
your assistance. You have always been my great friend and mentor. To my true friend, Jae Chul Hwang,
who has always stood by me at all times, I really cannot express my gratitude enough.
I am extremely grateful to my family, Youngsoo Jung, Gisook Choi and Yujin Jung, who have always
been supportive and have given me the encouragement in whatever I have pursued. Without your unwavering
and unselfish love and support, this achievement would not have been possible. To my lovely fianc´ e, Chan
Mee Koak, thank you so much for your warm love that has always been there when I needed it most. I am
forever grateful to you for your support, patience, and encouragement. I could not have done this without
you.
Most of all, I thank and praise God, who is making a way in the wilderness and streams in the wasteland.
I thank God for his help and comfort throughout this entire journey.
iii
Contents
Dedication ii
Acknowledgement iii
List of Tables vi
List of Figures vii
Abstract viii
1 Introduction 1
2 Does Eduction Affect Time Preference? Evidence from Indonesia 3
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Existing Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Data and Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Indonesia Family Life Survey (IFLS) . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.2 INPRES school construction program . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.3 Time preference measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.6 Robustness Check and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 Plausible Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7.1 Mediation Analysis 1 (Imai et al. (2010b)) . . . . . . . . . . . . . . . . . . . . . . . 26
2.7.2 Mediation Analysis 2 (Huber (2014)) . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.7.3 Joint Bootstrap Analysis (Bennett et al. (Forthcoming)) . . . . . . . . . . . . . . . . 29
2.7.4 Mechanisms Related to Medical Findings . . . . . . . . . . . . . . . . . . . . . . . 30
2.8 Limitations and Concerns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3 Is 1+1 more than 2? Joint evaluation of two public programs in Tanzania 50
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2.1 Iodine Supplementation Program (ISP) . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2.2 Primary Education Development Program (PEDP) . . . . . . . . . . . . . . . . . . 54
3.3 Existing Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4 Data and Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.2 Iodine Exposure Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.4.3 Empirical Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.5.1 Impact of ISP on Educational Attainment and Joint Impacts . . . . . . . . . . . . . 62
iv
3.5.2 Delay in Starting Primary School . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.5.3 Dynamic Complementarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4 Revisiting the Effect of Retirement on Cognition: Heterogeneity and Endowment 76
4.1 Introdcution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.2 Existing Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.3 Econometric Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.3.1 Empirical Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.3.2 Instrumental Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.4 Data and Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.4.1 Health and Retirement Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.4.2 Cognition Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.4.3 Work and Cognitive and Physical Demands of the Job . . . . . . . . . . . . . . . . 88
4.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.5.1 First Stage Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.5.2 Cognition Effects of Retirement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.5.3 Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.5.4 Endowment of Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5 Conclusion 100
Appendix 102
A.1 Additional Figures and Tables in Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . 102
A.2 Time Preference Imputation in Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
A.3 ISP Treatment Definition in Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
A.4 Alternative Definitions of ISP Exposure in Chapter 3 . . . . . . . . . . . . . . . . . . . . . 112
A.5 Additional Tables in Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Bibliography 116
v
List of Tables
2.1 Summary Statistics 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.2 Summary Statistics 2 (IV sample) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3 Time preference and behavior (Association) . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4 Main Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.5 Robustness 1: Time preference category and imputed discount rate (Female) . . . . . . . . . 46
2.6 Robustness 2: Different level of fixed effect, controls and clustered robust standard error
(Female only) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.7 The effect of education on behavior measures (IV-2SLS) . . . . . . . . . . . . . . . . . . . 48
2.8 Mediation analysis of mechanism (Residual) . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2 Impact of Iodine Supplementation Program on completed years of schooling . . . . . . . . . 72
3.3 Impact of ISP and PEDP on completed years of schooling . . . . . . . . . . . . . . . . . . . 72
3.4 Impact of ISP and PEDP on primary school starting age . . . . . . . . . . . . . . . . . . . . 73
3.5 Conversion of an additional year into additional years of schooling . . . . . . . . . . . . . . 73
3.6 Impact of ISP on height of the child (Height-for-age) . . . . . . . . . . . . . . . . . . . . . 74
3.7 Within household impacts of ISP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.8 Impact of ISP and PEDP on hours worked . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.1 Summary statistics (men, polygenic risk score samples for total word recall) . . . . . . . . . 95
4.2 Summary statistics (women, polygenic risk score samples for total word recall) . . . . . . . 95
4.3 Main regression results for total word recall (men, balanced polygenic risk score sample) . . 96
4.4 Main regression results for total word recall (women, balanced polygenic risk score sample) 96
4.5 Sensitivity check using social security early retirement age (IV) (men, balanced polygenic
risk score sample) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.6 Sensitivity check using social security normal retirement age (IV) (men, balanced polygenic
risk score sample) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.7 Sensitivity check using social security early retirement age (IV) (women, balanced polygenic
risk score sample) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.8 Sensitivity check using social security normal retirement age (IV) (men, balanced polygenic
risk score sample) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.9 Relation between physical and cognitive demands and other regressors (men, unbalanced
polygenic risk score sample for total word recall) . . . . . . . . . . . . . . . . . . . . . . . 99
A1 The effect of education on negative time discounting . . . . . . . . . . . . . . . . . . . . . 105
A2 The first stage of IV-2SLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
A3 Suggestive Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
A4 ISP Coverage Variation (from Field et al. (2009)) . . . . . . . . . . . . . . . . . . . . . . . 110
A5 Probability of Protection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
A6 Robustness of ISP Exposure Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
A7 Impact of ISP on Vaccinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
A8 Across Household Impacts of ISP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
A9 Primary School Starting Age as a Plausible Mechanism . . . . . . . . . . . . . . . . . . . . 114
A10 Main regression results for total word recall (men, unbalanced polygenic risk score sample) . 115
A11 Main regression results for total word recall (women, unbalanced polygenic risk score sample)115
vi
List of Figures
2.1 Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 Time preference categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 Lowess plot of time preference and education . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 Time preference and education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5 Behavior and Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.6 Lowess of Inpres intensity and Female Time preference and Education (IV sample) . . . . . 38
2.7 Mediation Analaysis (Huber (2014)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.8 Bootstrapped correlation plot of time preference’s and mechanisms’ correlation with educa-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.9 Bootstrapped correlation plot of time preference’s and mechanisms’ correlation with educa-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1 Iodine Supplementation Program in Tanzania (from Field et al. (2009)) . . . . . . . . . . . . 70
A1 Distribution of INPRES intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
A2 Distribution of educational attainments by INPRES treatment status . . . . . . . . . . . . . 103
A3 Time preference and Age by education . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
A4 Distribution of time preference by INPRES treatment (IV sample) . . . . . . . . . . . . . . 104
A5 Distribution of imputed discount factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
vii
Abstract
This dissertation consists of three essays on development economics and health economics. The first chapter
examines the causal effect of educational attainments on individual time preference (patience) measured
by hypothetical questions from Indonesia Family Life Survey (IFLS). There are two theoretical hypotheses
regarding the relationship between education and individual time preference. More patient individuals may
pursue more schoolings while education makes people more patience. In this chapter, we test empirically
the second hypothesis. We overcome the endogeneity by using a primary school construction program
(INPRES) as an instrument variable for educational attainment to estimate the causal impact of education
on individual time preference.
The second chapter examined the joint effect of the Iodine Supplementation Program (ISP) and the
Primary Education Development Program (PEDP) on the educational attainment at the ages of 10-13 using
Kagera Health and Development Survey (KHDS). In this chapter, we provide an evidence of ‘dynamic
complementarity’ of two different positive shocks in early life periods. This study illustrates the importance
of early life environments and the importance of subsequent investments in adjacent periods.
The third chapter tests the cognitive reserve hypothesis that engaging in cognitively stimulating activities
reduces cognitive decline in people that are of an older age. In this chapter, we estimate the effect of retire-
ment on cognitive decline by the type of job an individual retires from using Health and Retirement Study
(HRS). We use multiple instrument variables to control for the endogeneity of retirement. We also control
for innate cognitive ability, using the genetic risk score of cognition and education, to test the ‘preserved
differentiation’ hypothesis that higher cognitive function in the later life is due to higher levels of cognitive
function throughout their lives.
viii
Chapter 1
Introduction
Understanding the role of education, health, and work environments in human capital formation (e.g., edu-
cational attainment, health status, cognitive and non-cognitive skills) has been an important topic in the field
of economics. This dissertation reveals that the heterogeneous early life and later life investments in human
capital accumulation can help explain the differences in socioeconomic outcomes (educational attainments
and cognitive skills).
The first chapter estimates the causal effect of education on individual time preference (patience) using
data from waves 4 and 5 of the Indonesia Family Life Survey. We use the INPRES primary school construc-
tion program between 1973-74 and 1978 in Indonesia to instrument years of schooling. The local average
treatment effect of the program is approximately an increase of seven percentage points in the patience of
females for every additional year of schooling. We also find consistent results in alternative specifications.
The results are robust in terms of alternative definitions of the time preference measure. This study con-
tributes to the existing literature by adding evidence of the heterogeneity and endogeneity of individual time
preferences, considered fixed and constant in neoclassical economics, by establishing the causal link be-
tween education and time preferences. We provide suggestive mechanisms through which education affects
patience using two mediation analyses and the bootstrap coefficient correlation method. We document that
cognition and total income are the plausible mechanisms.
The second chapter evaluates the joint effects of two public programs from Tanzania, the Iodine Sup-
plementation Program (ISP) and the Primary Education Development Program (PEDP) on the schooling
outcomes for the exposed cohorts. We find that ISP treatment was associated with lower schooling achieve-
ments for the exposed children at the age of 10-13. The effect operated through delays in enrollment. We
provide suggestive evidence why the exposed children experienced delays in enrollment. We also find sug-
gestive evidence of ‘dynamic complementarity’ between ISP and PEDP: those exposed to ISP were better at
converting years in school into completed years of schooling due to PEDP. That is, for each additional year
1
spent at school, ISP treated children convert them into higher years of completed schooling than those not
treated by ISP.
The third chapter provides empirical evidence on heterogeneous retirement effects by occupation. Re-
tirement effect on cognition has drawn significant research interests from economists in recent years. Es-
pecially with ongoing policy discussion on public pension reforms and increasing burden of dementia, it is
indisputably an important research question with significant policy implications. Since the seminal paper of
Rohwedder and Willis (2010), several economists have studied the relationship between retirement and cog-
nition. Building on this growing literature, our paper makes two important contributions. First, we explicitly
consider cognitive and physical demands of jobs. Coe et al. (2012) investigated potential hetereogeneity of
retirement effect between white-collar and blue-collar workers, and more recently, Mazzonna and Peracchi
(2017) found the differential effects of retirement by physical demands of jobs. While both of these earlier
investigations make important contributions by recognizing the hetereogeniety of jobs, it does not directly
examine cognitive and physical demands of jobs. As the primary explanation for potential adverse effect
of retirement is that cognition is better maintained through mental exercise (Salthouse (1991)), by investi-
gating the cognitive demands of the job one retires from we can directly test the hypothesized relationship.
Another important contribution of this paper is to control for endowed cognitive ability. While endowed ge-
netic differences in cognitive ability is an important omitted variable that can explain individual differences
in cognitive performance as well as selection into particular type of job, such inherited characteristic has not
been controlled for in prior literature. Taking advantage of polygenic risk score of cognition and education
(Davies et al. (2015)), we control for individual differences in genetic endowment and find empirical evi-
dence, supporting the heterogeniety of retirement effect. We find supporting evidence for differential effects
of retirement by physical demands, but not cognitive demands of jobs. We also find evidence for selection
bias that cognitive demands of jobs are associated with innate differences in cognition through educational
attainment.
2
Chapter 2
Does Eduction Affect Time Preference? Evidence from
Indonesia
2.1 Introduction
This paper analyzes the Indonesia Family Life Survey (IFLS), a nationally representative dataset to estimate
the causal effect of education on individual time preference. Time preference refers to the relative valuation
placed on a good delivered immediately compared to its valuation on a later date. The rate of time preference
(or time discounting)
2
islog(
T
), where
T
is the discount factor ranging from 0 to 1. It is considered
one of the key factors determining individual economic decision-making related to, for example, savings,
consumption, and investment in health and education (Bickel et al. (1999); Golsteyn et al. (2014); Non and
Tempelaar (2016)). It is closely related to factors that govern an individuals’ inter-temporal choices or trade-
offs over time. In economics, time preference has been the center of the theory of consumer optimization
and economic growth (Callen (2015)). It has also been paid growing attention in the field of development
economics because of its crucial role in investment decisions regarding the use of fertilizers and agricultural
products (Duflo (2006); Duflo et al. (2011)) and the allocation of resources within the households (Rubalcava
et al. (2009)).
NOTE: This chapter is coauthored by Tushar Bharati (USC) and Seungwoo Chin (USC). This chapter has benefited from feed back
provided by numerous economists. This chapter also has benefited from feedback by participants at Western Economic Association
International Conference (Hawaii, 2015), the Population Association of America Annual Meeting (Washington D.C., 2016), UNU
WIDER’s Human Capital Development Conference (Helsinki, 2016), and the seminar at California State University, Long Beach
(2017).
2
We use the terms (inverse of) time preference, time discounting and patience interchangeably. This, of course, is a simplification.
Readers interested in the distinction between the three terms should look at Frederick et al. (2002). In this study, we do not need clear
distinction between two concepts.
3
Since Samuelson’s discounted utility model, time preference has long been considered to remain stable
or fixed, so relatively little research has focused on the determinants and characteristics of time preference.
Several studies have tested the stability of individual preferences empirically, either by using observational
data or by conducting relevant lab and field experiments. However, we still have surprisingly little evidence
of the determinants of time preference formation (Anderson et al. (2004)).
This paper contributes to the growing literature of empirical analysis in this subject by investigating one
of the potential factors (education) affecting time preference formation. There are at least two theoretical
hypotheses with regard to the relationship between education and time preference. First, individuals who
are more patient might be more willing to sacrifice current labor market opportunities and instead decide
to obtain more schooling (Grossman (2006)) to increase future earnings. Second, schooling may infuse
more patience in individuals by improving their forethought and planning, and it may make them more goal-
oriented (Becker and Mulligan (1997); Oreopoulos and Salvanes (2011)). In psychology, as humans are
considered born impatient, it is assumed difficult for individuals to choose actions with a delayed reward
without education (Metcalfe and Mischel (1999); Doepke and Zilibotti (2008)). An empirical analysis of the
causal effect of education on time preference, therefore, entails challenges such as reverse causality of the
effect of time preference on education and selection into education. In this paper, our main goal is to provide
reliable estimates of the causal effects of education on patience. We also provide suggestive evidence of
plausible mechanisms through which education affects patience.
We document a causal effect of education on patience by leveraging the richness of the individual- and
household-level data from waves 4 (2007) and 5 (2014) of the IFLS, which include questions that attempt to
elicit information regarding individual time preference. We compare results from several different empirical
specifications such as ordinary least squares (OLS), household FE, individual FE, and the instrument variable
(IV) approach. For the IV analysis, we exploit the exogenous variation in the cost of schooling induced by
the INPRES primary school construction project launched in 1973-4. This project was the fastest and one
of the largest primary school construction programs in the world (Peters (1990)). We instrument years
of schooling with the intensity of the school construction program at the district level and variation from
the cohort level to overcome the endogeneity of years of schooling. Consistent across specifications and
robustness checks, we find that individuals with more education are relatively more patient. We demonstrate
4
that cognition measured by Raven’s test, the word recall test, and the serial sevens subtraction test, and total
income would be the potential mechanisms for this causal link.
The remainder of the paper is organized as follows. Section 2.2 describes the previous literature. Section
2.3 describes the background and the data. Section 2.4 explains the empirical strategies and section 2.5 dis-
cusses the results. Section 2.6 explores the potential mechanisms behind our findings. Section 2.8 discusses
the limitations and concerns. We conclude in section 2.9.
2.2 Existing Literature
In this section, we describe the literature with regard to individual preferences (time or risk preference,
mostly focusing on time preference). In the neoclassical theory, individual time preference is assumed an
economic primitive, fixed over time for individuals and, often, constant across individuals. Economists
have often questioned this assumption, and they have investigated the endogenous formation of individual
preferences (Becker and Mulligan (1997)). Becker and Mulligan (1997) introduced a simple framework
model to explain an endogenous time preference formation. An individual is maximizing a function of
u(c
0
) +
P
1
t=1
(e)
t
U(c
t
), where c is consumption and e is a matrix of potential factors affecting time
preference formation. A handful of empirical studies document heterogeneity of individual time prefer-
ences (Leigh (1986); Barsky et al. (1997); Rubalcava et al. (2009); Dohmen et al. (2011); Ng (2013)) along
different socioeconomic dimensions (‘e’ in the Becker and Mulligan (1997) framework model), such as edu-
cation, marriage, drinking/smoking behavior, age, gender, and household wealth. In addition, a recent study
by Meier and Sprenger (2015) identified the temporal instability of time preference within an individual
across a time horizon.
Individual preferences have been considered crucial in determining economic behavior and socioeco-
nomic status (Stigler and Becker (1977)). For example, investment in the stock market, entrepreneurial
aptitude, insurance and health care utilization, etc. - factors that may, in turn, determine such outcomes as
employment, income, health and wealth accumulation - are all affected by individual preferences. Figure
2.1 shows how individual time preference is well linked to various factors. At a macro level, individual time
preference or an aggregate level of time preference can partially explain the economic growth and inequality.
Dohmen et al. (2015) has highlighted the positive reduced form relationship of patience with income and
5
growth rate across different countries. They found that an increase in patience by one standard deviation
is correlated with an increase in log GDP (Gross Domestic Product) by two-thirds of a standard deviation.
On the other hand, at a micro level, a growing number of empirical papers is investigating the relationship
between time preference and behavior. For example, Lawrance (1991), Pender (1996), Tanaka et al. (2010)
and Golsteyn et al. (2014) have found that income is positively correlated with patience in the United States,
Vietnam, India and Sweden, respectively. Khwaja et al. (2007) used the self-reported measure of impul-
sivity
3
to identify the relationship between patience and smoking. They concluded that smokers are more
impatient. Further, Golsteyn et al. (2014) addressed that patience at age 13 can predict lifetime outcomes,
such as school performance, educational attainment, health and labor supply. They concluded that patience
is associated with better results in terms of such lifetime outcomes.
While most empirical studies have focused on investigating the relationship between individual prefer-
ences and socioeconomic outcomes, as presented on the right side of figure 2.1, several studies have also
investigated the determinants of individual preferences, as shown on the left side of figure 2.1. As addressed
in Dohmen et al. (2015), the variation in individual time preference can be explained by many factors, such
as natural disasters (Cassar et al. (2011); Cameron and Shah (2015); Callen (2015); Brown et al. (2017)),
conflict (V oors et al. (2012); Callen et al. (2014)), economic shocks (Guiso et al. (2013)) and cultural back-
grounds, such as religion (Becker and Mulligan (1997); H’Madoun and Nonneman (2012); Benjamin et al.
(2016)) and ethnicity (Benjamin et al. (2010)). For example, Callen (2015) has found that exposure to a
natural disaster, such as the Indian Ocean earthquake tsunami, increased patience in a sample of Sri Lankan
wage workers. By providing evidence that the least-educated people generated the greatest change in time
preference due to the tsunami, he argued that a tsunami experience might provide opportunities for the
least-educated people to learn the importance of delaying gratification.
Among many factors, we are interested in the role of education in time preference formation. After
Becker and Mulligan (1997) acknowledged the suggestive link between education and time preference, a
few studies have examined the correlation between education and time preference. For example, Oreopoulos
and Salvanes (2011) showed a positive correlation between schooling and patience among 25-45 year-old
Americans using General Social Survey questions (“Nowadays, a person has to live pretty much for today
and let tomorrow take care of itself”). Harrison et al. (2002) used lab/field experiments in Denmark to show
3
Question: I make hasty decisions. I do not control my temper. I act on impulse.
6
that a higher level of education is associated with a lower time discount rate. The association is not only
identified in developed countries, but also identified in low-income countries. The correlation magnitude is
likely more profound in low-income countries, as there may be more channels (such as reducing the role
of certain traditional beliefs) through which education makes individuals more patient (Bauer et al. (2009)).
Kirby et al. (2002) showed that a part of variation of the time discount rate can be explained by education in
Bolivia. In addition, Bauer et al. (2009) also found a positive association between education and patience.
Relative to the correlational studies, causal links between education and individual time preference
have rarely been explored. To the best of our knowledge, only two empirical papers have investigated
the causal effect of education on time preference.(Bauer and Chytilov´ a (2010); Perez-Arce (2017)). Bauer
and Chytilov´ a (2010) attempt to examine the causal effect of education on subjective discount rates for in-
dividuals in Uganda using instrument variables, and they found that education increased patience among
males. They used two different sources of variation in education for instrument variables: a variation in
access to schools at the district level and a cohort-level variation due to the Idi Amin regime change. How-
ever, the estimates from both their instruments are likely to be biased. Because they used only a variation
from one dimension (village-level or cohort-level), unobservable differences across villages (when using a
village-level instrument variable) and across cohorts (when using a cohort-level instrument variable) may
confound the true estimate. In addition, their instrument variables do not satisfy the exclusion restriction
because, for example, the Idi Amin regime change might have affected time preference through channels
other than education.
A recent paper by Perez-Arce (2017) used the exogenous variation in educational attainment by exploit-
ing random assignment to a public college in Mexico, and it was found that college education makes people
more patient. However, there are two major concerns regarding the robustness of Perez-Arce (2017)’s find-
ings. First, to be eligible, the applicants needed to have a high school diploma and a home in Mexico City.
Moreover, participation was voluntary. The college enrollment rate in Mexico around the time was close to
25% and it is possible that patient individuals were more likely to have completed their high school diploma
planning to enroll in a college. The sample, therefore, was highly selected and the results may not be ex-
ternally valid. Second, the survey was conducted through telephonic interviews with a significantly high
non-response rate of 55%. A non-response is most likely associated with the time preference of individuals,
and it can bias the causal estimates.
7
In this context, our study makes three important contributions. First, we contribute to the limited existing
literature by revisiting the causal effect of education on time preference. This study improves substantially
on Bauer and Chytilov´ a (2010) and Perez-Arce (2017) by applying different specifications and an already
agreeable instrument variable. Second, we examine the impact of education on patience at all levels of
education. With only a little over 30% of the world population having enrolled in tertiary education in
2014, (with as low as 3% in certain developing countries), (World Development Indicators, The World
Bank), much of the effect of education on time preference might be due to the completion of primary and
secondary school years, an aspect not captured in Perez-Arce (2017). Third, the causal effect of education
on time preference in developing countries deserves attention in its own right because relatively little is
known about the determinants of time preference in the context of developing countries (Hamoudi (2006);
Ng (2013)). Given the sub-par levels of public and private infrastructure in many developing countries and
the association of patience with positive health and consumption traits, the causal effect of education on
patience holds immense relevance to policy discussions and actions aimed at alleviating the poverty and
inequality in consumption, education and health in these countries.
By providing the empirical evidence of the effect of education on time preference, we challenge the
readily made assumption that time preference is constant across time among individuals. This has far reach-
ing implications for models of human capital formation that do not account for the reverse causality from
education to time preference. The results also shed light on the possible existence of a virtuous cycle where
investment in education leads to individuals becoming more patient and, in turn, investing more in educa-
tion. We also attempt to provide a clearer understanding of a less-explored channel through which education
policy affects individual behavior and outcomes in the long run. According to Heckman (2007) and Cunha
and Heckman (2007), human capital production function requires both cognitive and non-cognitive abilities
together. Whether education causally affects non-cognition is still under debate, but our results suggest that
education might affect human capital formation by affecting non-cognitive traits, including time preference,
in addition to or independent of its effect through cognitive channels.
8
2.3 Data and Measurement
2.3.1 Indonesia Family Life Survey (IFLS)
This paper uses waves 4 (2007) and 5 (2014) of the IFLS. The IFLS is an on-going longitudinal survey
by the RAND corporation that started in 1993, with follow-ups at least every seven years. The survey
respondents are representative of about 83% of the Indonesian population living in 13 of the 27 provinces
in the country and contains information on over 30,000 individuals in wave 4 and over 50,000 individuals
in wave 5 (Strauss et al. (2016)). It contains information on a wide variety of topics at the individual, the
household, and the community level. Data collected at the individual level contains information on health,
education, employment, migration, income, and asset holdings. Since wave 4, information on individual
preferences (time and risk preference) has also been collected using hypothetical questions. We explain
hypothetical questions in detail in the time preference measure section (section 2.3.3).
Table 2.1 provides the descriptive statistics for the sample we use for the OLS (entire sample) analysis.
Our study sample consists of 23,137 observations from the two waves.
4
The numbers of male respondents
are 5,202 and 5,122 in wave 4 and wave 5, respectively, while the numbers of female respondents are 5,122
and 6,286 in wave 4 and wave 5, respectively. The average number of years of schooling is 8.9 years for
males and 8.4 years for females in wave 4, and 9.7 years for males and 9.2 years for females in wave 5.
Most of the sample population live in rural areas (more than 90%). For the household FE and individual FE
analysis, we retain only those households or individuals that have appeared twice across these two waves.
In addition, for the individual FE analysis, our sample consists of only those who answered both the time
preference module in wave 4 and wave 5 and whose education level did not change or increase between
wave 4 and wave 5. Since older cohorts are unlikely to change their education level between the two waves,
we include individuals who are 20 years old or less in 2007 for the individual FE specification.
For the IV analysis, our sample consists of cohorts born in 1958-1962 and 1968-1972.
5
Following Duflo
(2001), those born in between 1968 and 1972 are considered to have been fully treated by the INPRES school
construction program and those born in 1958-1962 are considered to have been unaffected by INPRES
4
The time preference questionnaire sub-module (explained below) was not administered to respondents below 15 years of age. In
addition, we excluded respondents who did not answer either one or both of the time preference sub-module questions.
5
Duflo (2001) uses cohorts born between 1957 and 1962 as the control group and cohorts born between 1968 and 1972 as the treated
group. We drop those born in 1957 to keep equal number of cohorts in each group. The results presented here do not change even if we
use those born in 1957.
9
(control group). For the IV analysis, we make use of the information from the wave 5 when the samples
appear both in wave 4 and wave 5 or appear only in wave 5. Furthermore, we add samples from wave 4
when the samples appear only in wave 4. Table 2.2 presents a summary of the statistics for this sample. The
sample for the IV analysis is, as expected, older and has a lower level of education on average than the one
used for FE analyses.
In addition, we include controls for early life rainfall shocks (rainfall shock in the month of birth and
in the first five years after birth) since they are found to be important determinants of adult health and
education in Indonesia (Maccini and Yang (2009)). We use the University of Delaware Center for Climatic
Research’s “Terrestrial Precipitation: 1900-2008 Gridded Monthly Time Series (1900-2008) (Version 2.01)”
rainfall, which uses an interpolation algorithm based on the spherical version of Shepards distance-weighting
method to combine data from 20 nearby weather stations for every 0.5 latitude by 0.5 longitude grid. We
use the district level latitude-longitude information from Maccini and Yang (2009) to match the districts to
monthly rainfall information from all grids in a 200-kilometer radius around the district center. We calculate
a weighted average of rainfall across these grids, weighing rainfall information from each grid by the inverse
of its distance from the district center. For every month of every year, we calculate a district level mean
rainfall by averaging the month-district-specific rainfall in the previous 50 years. For the birth month, the
rainfall shock variables comprise the difference between rainfall in a particular district in a particular month
and the prior 50-year district specific average for that month.
6
For years after birth, we aggregate this
difference over the corresponding 12 months.
2.3.2 INPRES school construction program
Beginning in 1973-74, the Indonesian government launched the Sekolah Dasar INPRES school construction
program in an attempt to improve education levels in the country by improving access to primary schools.
Between 1973 and 1978, over 61,000 new primary schools were constructed. This construction intensity
differed across districts (Kabupaten) and time. The program was primarily intended to assist first the districts
with the lowest primary enrollment rates as per the 1971 census. As a result, there was a considerable degree
6
Alternative definitions of the rainfall shock variable, such as below the 25th percentile, above the median, and above the 75th
percentile levels of historical rainfall are explored in Bharati et al. (2017) and they all yield consistent results.
10
of variations in the school construction intensity across districts. We use the INPRES school construction
program as a source of exogenous education variation.
We borrow the information on intensity of school construction used in Duflo (2001). Intensity, here,
is the number of newly constructed schools per 1,000 children in the district. The distribution of school
intensity is presented in appendix figure A1. Most districts received around two newly constructed schools
per 1,000 students under the INPRES program. We match the individual level information from IFLS with
the the school construction intensity in the district in which an individual was born. We assume that an
individual went to school in the district in which he or she was born.
7
In Indonesia, primary school starts when individuals are 6-7 years old and continues until they are 11-12
years old. Therefore, while those born after 1968 received the maximum advantage from INPRES, given the
late enrollment in developing countries, it is possible that individuals who were born not long before 1968
might also have benefited in part from the primary school construction. For this reason, our control group
consists of individuals who were between 12 and 16 years of age in 1974 (born between 1958 and 1962), and
our treatment group consists of those who were between 2 and 6 years of age in 1974 (born between 1968
and 1972). We omit the cohorts in between because the extent of exposure to INPRES for these cohorts is
not clear. The INPRES treatment variable is the product of the district-level intensity of school construction
and the treatment status of the individual based on year of birth. The treatment variable takes the value 0 for
all who were born in or before 1963 and is equal to the intensity of the school construction in their district for
all those born in or after 1967. In appendix figure A2, we present the difference in distribution of education
level by treatment and control group. We observe a significant difference in education level between the
treated group and the control group.
2.3.3 Time preference measures
Hypothetical questions
The main outcome variable of interest, a measure of time preference, is constructed on the basis of two sets
of hypothetical questions, module A and module B, included in wave 4 and wave 5 of IFLS. The flowcharts
7
In Bharati et al. (2017), while analyzing the heterogeneity in impact of INPRES along with the amount of rainfall in the first
year after birth of individuals, we compare the estimates from the sample that makes the assumption that the treatment and control
individuals go to school in the district of birth with the estimates from the sample that consists of only those who report having lived in
their birth district at the age of 12. We do not find any qualitative difference.
11
in figure 2.2 describes how these questions were administered. The monetary amounts are in Indonesian
Rupiah (IDR).
Module A starts by asking respondents whether they will choose to receive 1 million Rupiah today or
1 million Rupiah in one year’s time. When they choose 1 million Rupiah today, they are asked whether
they would accept 3 million Rupiah in one year’s time as presented in figure 2.2. The respondents are asked
the same question with different monetary amounts and the same time horizon (one year). The IFLS has
another question module, B, with a similar structure as module A but with different monetary amounts and a
different time horizon. Module B starts by asking respondents whether they would choose 1 million Rupiah
today or 0.5 million Rupiah in five years’ time. When they choose 1 million Rupiah today, they are asked
whether they would choose 4 million Rupiah in five years’ time. As the flowchart shows, depending on the
responses at different levels, a respondent can be categorized into category 1, 2, 3, or 4 for each of the two
time sub-modules. Note again that sub-module A and B differ both in terms of the monetary rewards and
the delays in future rewards.
Respondents who choose immediate payment throughout a time sub-module (Time A or B = 4 in figure
2.2) are categorized as the most impatient individuals. The top panel in table 2.1 presents the distribution of
responses to the time preference questions in the OLS sample. In wave 4, 69% and 72% of the respondents
(male and female, respectively) are impatient as per time preference sub-module A and 79% and 81% of the
respondents (male and female, respectively) are impatient as per sub-module B. In wave 5, the proportion
of impatient respondents is smaller but very similar to that of wave 4. Consistent with previous studies,
figure 2.3 and figure A3 show that the measure is also closely associated with education and age (Dohmen
et al. (2011); Ng (2013)). There seems to be a negative association between education and impatience. The
distribution of responses for the IV sample, presented in table 2.2, is consistent with that in table 2.1. Among
the respondents, 71% (66%) and 83% (77%) of the males (females) respondents are impatient as per time
preference sub-module A and B, respectively. In appendix figure A4, we also present the distribution of time
preference by treatment and control groups. We observe that the treatment group (born between 1968 and
1972) has a smaller proportion of impatient people, suggesting the possibility that education makes people
more patient.
Before discussing the details of the estimation, a debate on the time preference measurement is worth
emphasizing. There are four different ways of collecting individual time preferences. First, as we used in this
12
study, a field survey collects individual time preferences, utilizing questions with hypothetical rewards. Sec-
ond, instead of using hypothetical rewards, a survey can collect time preference information using questions
with real rewards. Respondents would earn rewards according to their choices. Third, a field survey can use
actual inter-temporal choices, such as daily caloric intakes or food expenditures (Lawrance (1991); Shapiro
(2005)). As addressed in Frederick et al. (2002), the use of actual behavioral choices as a time preference
measure may reflect preferences better because they are revealed by real choices. However, individual time
preference estimates using this measure may be confounded due to many factors affecting time preference
and choices at the same time. Thus, it is preferable to use experimental settings, either with hypothetical
rewards or real rewards. Within experimental settings, there are several studies comparing time preference
between using hypothetical or real rewards. For example, Coller and Williams (1999), in their experiment
comparing the same monetary value of hypothetical and real rewards, found that discount rates inferred from
the two award procedures are not different once censored data, demographic differences and heteroscedastic-
ity are controlled for. This finding is well supported by other papers (Johnson and Bickel (2002); K¨ uhberger
et al. (2002); Madden et al. (2003)). Thus, a potential concern as to whether the responses elicited by these
hypothetical questions do not capture the true time preference of individuals compared to real reward ques-
tions is minimized, as both hypothetical and real money reward questions have their respective strengths
and weaknesses (Locey et al. (2011)). In addition, it has been shown that answers to hypothetical individual
preference survey questions are consistent with real-life behavior (Hamoudi (2006); Dohmen et al. (2011)).
We, too, find that the time preference measure is significantly associated with real-life behaviors, such as
smoking, contraception usage, savings, arisan participation, that, a priori, we expect to be correlated with
time preference. We summarize time preference and behavior in more detail in the next subsection 2.3.3
(table 2.3).
Another thing to discuss concerns the respondent’s choice in response to the first question of module
A and module B. For each of these modules, if an individual chooses to receive a smaller or equal amount
of reward in the future rather than today, the individual exhibits a zero or negative time discount rate (or
a discount factor 1). Virtually all analyses of inter-temporal choice in economics assumes a positive
time discount rate. This assumption has often been questioned (Koopmans (1960); Koopmans et al. (1964);
Olson and Bailey (1981)). In a world with positive real rates of return on savings and individuals with
infinite life spans and linear or weakly concave utility functions, a negative time preference is a possibility.
13
These assumptions, however, are rarely justified.
8
A negative time discount rate could signal a variety of
possibilities, from a strategic move of commitment to sheer irrationality to a simple inability to understand
the questions. Unfortunately, given the information in the data set, it is not possible to distinguish between
these possibilities. In any case, such behavior warrants separate analysis, and we dropped those respondents
from the sample. Such responses comprised 3.0% (module A) and 2.7% (module B).
9
Thus, our main
regressions estimate the average treatment effect or local average treatment effect conditional on respondents
having a positive discount rate.
The experimental procedure of asking the respondents to choose between a smaller, more immediate
reward and a larger but delayed award used in the IFLS time preference sub-modules, known as choice
tasks, is, by far, the most common method used to elicit time preference. There is, however, a downside to
using this method. We can, at best, establish bounds on the discount factor of respondents; it is not possible
to calculate point estimates for the discount factor. We work with this constraint and impute bounds on the
discount factor of respondents. The imputation process is explained in detail in appendix A.2.
Another shortcoming often found in experimental choice task studies is the anchoring effect; imputed
discount rates tend to be biased in the direction of the discount rates that would equate to the first round
options (Green et al. (1998)). However, in this study, we are not interested in point estimate of time discount
rates and estimate the effect separately for module A and module B, so the anchoring effect might not be a
big concern here. Furthermore, such a measurement error in dependent variable, as long as it is uncorrelated
with the independent variables, will not bias the estimated coefficients. Although there are such limits in
using imputed discount rates, in the robustness check to main empirical analysis, we use imputed discount
rate to examine the effect of education on time preference. We also use the categorical variables of time
preference as our outcome variables as for a robustness check.
Time preference and behavior measure
As we discussed, time preference measured by hypothetical questions has been found to be a good proxy
measure for true time preference (Hamoudi (2006); Dohmen et al. (2011)). In this section, we check whether
8
See Frederick et al. (2002) for the detailed discussion.
9
According to table A1, using our IV-2SLS sample, we show that education causally affects the likelihood of reporting a negative
discount rate A in the pooled sample with very low the first stage F statistics. While this finding is definitely interesting, it does not
help us in identifying which among the reasons cited above is behind these responses. Furthermore, we do not find any significant
relationship between education and negative time discount for males or females (module A) and all samples of module B.
14
hypothetical questions are closely related to behavior measures often considered as proxy for time prefer-
ence. For this, we look at the extent of association between our measures of time preference and behavior
that impatient individuals are supposed to exhibit.
10
Time discounting has been repeatedly found to be associated with smoking behavior (Fuchs (1982);
Khwaja et al. (2007); Ida and Goto (2009); Oreopoulos and Salvanes (2011); Kang and Ikeda (2014)). The
first three columns in table 2.3 present the association between impatience and whether an individual has
ever smoked or used tobacco in his life. An overwhelming proportion of males (73%) and a much smaller
proportion of females (2%) from our sample smoke.
11
Consistent with previous findings, impatient people
are much more likely to have smoked or used tobacco. Due to small proportion of females who are smoking,
we find no significant association between patience and smoking behavior.
DellaVigna and Paserman (2005) present a theoretical model of job search with alternative assumptions
of exponential and hyperbolic discounting. Using information from the National Longitudinal Survey of
Youth and the Panel Study of Income Dynamics for the United States, they show that impatience should lead
to less intensive job searching and more employment. Ben Halima and Ben Halima (2009) find the same
results for a French sub-sample from the European Household Panel Survey. On the basis of the previous
findings, we check the correlation between working status and time preference (table 2.3). We consider an
individual to be working if he or she reports that he or she has worked for pay in the previous week or if he
or she has received a wage/salary or made a non-zero business profit in the previous month. For both TP-A
and TP-B measures of impatience, we find that impatient females are less likely to work. By contrast, we
find no significant relationship for male respondents.
That time discounting affects savings was formalized as early as in 1928 (Ramsey (1928)). A recent
study by Sutter et al. (2013) also finds a close relationship between patience and saving behavior. This is
again observed in our study in column (7)-(9) in the first panel. Respondents who are more patient are
likely to save more. Not only is the saving behavior itself closely related to patience but saving amount
(intensive margin of saving) is also closely related to patience. Respondents who are more patient are likely
to save greater amounts. In developing areas of the world, where credit markets are not as developed, rotating
savings and credit associations (ROSCAS) are the most commonly found informal financial institutions used
10
In what follows, we will refer to those who are impatient (patient) as per our measure of time preference as impatient (patient)
individuals.
11
According to the World Bank, by 2012, 72% of Indonesian males and 4% of Indonesian females had smoked or used tobacco in
his or her life. Our sample looks representative of the Indonesian population in this regard.
15
for savings and insurance.
12
In Indonesia, the ROSCAS are known as arisans. We look at the association
between impatience and whether or not individuals participate in arisans and village savings. We find that
impatient individuals are less likely to participate in arisans and village savings, which is consistent with the
findings in the previous literature.
As already addressed in the literature review, income is positively associated with patience (Lawrance
(1991); Pender (1996); Tanaka et al. (2010); Golsteyn et al. (2014)). We identify a positive relationship
between patience and total income.
13
Time preference module B shows a stronger association with income
than does time preference module A.
Finally, we also check the correlation between amount of contraceptive use and time preference, for
female samples only. DellaVigna and Paserman (2005) addressed the correlation between time preference
and contraceptive use, with contraceptive use as a proxy for patience measure. We check this possibility in
this study as well. As shown in table 2.3, we find a significant association between time preference module
A and contraceptive use but no significant association is found for time preference module B.
It is interesting to examine the causal effect of education on behavior measures discussed here because
these measures may be a proxy for time preference. We present the figures describing the relationship be-
tween the behavior measures and time preference (figure 2.5). As expected, more education is negatively
correlated with smoking for males, and saving behaviors (saving, saving amount, arisan use, village savings)
are positively correlated with education. Income and contraceptive use also positively associated with edu-
cation. We do not find a clear graphical relationship between working status and education. We discuss the
regression results in the empirical result.
2.4 Empirical Analysis
We employ OLS, household FE,
14
individual FE specifications, and an IV-2SLS specification in our analysis.
We begin with the specification in equation (2.1) and gradually progress to stricter specifications.
Y
ijct
=+S
ijct
+
X
ijct
+
j
+!
c
+
t
+
ijct
(2.1)
12
ROSCAS are or were prevalent in both developing and developed countries around the world. For a detailed list of countries where
ROSCAS have appeared, see the anthropological studies by Ardener (1964), and Bouman (1977).
13
We define total income as the sum of wage salary, business profit, nonlabor income, and retirement pension income.
14
The results are available upon requests.
16
where Y
ijct
is a binary indicator of whether individual i, born in district j in year c surveyed in t, is
impatient. For all those individuals who choose the immediate payoff option throughout a module, Y
ijct
takes the value 1. It takes the value 0 for all other response paths in figure 2.2. Our main explanatory
variable is the individual’s completed years of formal education,S
ijct
.X
ijct
includes mother’s and father’s
education.
15
We also control for variables found in the previous studies to be important factors, such as
ethnicity (Benjamin et al. (2010)), religion (Becker and Mulligan (1997); H’Madoun and Nonneman (2012);
Benjamin et al. (2016)), and language spoken in daily life at home/language used in interview (Chen (2013)).
In order to control for any time-invariant unobservable differences across districts, we include a district of
birth-specific fixed effect and a district of current residence fixed effect,
j
, in all specifications. We also
control for differences across years of birth by including year of birth fixed effect and month of birth fixed
effect (!
c
). We include a survey wave fixed effect,
t
, to account for differences between the two waves
from which we draw our observations. There are also several important control variables that are worth
emphasizing. IFLS 5 collected time preferences slightly differently than did IFLS 4. IFLS 4 asked time
preference module A first followed by time preference module B, whereas IFLS 5 randomized the sequence
of the time preference module. In order to control for this, we simply add a dummy variable indicating
whether time preference module A or B asked first. We also control for interviewer fixed effect (Jakiela and
Ozier (2015)) and interviewer’s evaluation of accuracy of responses for robustness check. We conduct our
analysis separately for the two time preference sub-modules, A and B.
Even after controlling for many factors, however, time preference, could be shaped by unobservable
household characteristics or individual-level characteristics that also affects education level. For example,
failing to control for differences of household cultural values might bias our estimate. For this reason, we
augment equation (2.1) with household FE (results are available upon requests). It is still a concern even
after absorbing household-level differences, that prenatal and early life conditions across children from the
same parents might affect time preference formation. In addition, the variation in intra-household allocation
may capture some of variations in individual time preference, so it may confound the household FE results.
For this reason, we next employ an individual fixed effect (
i
) specification. We pool together information
from waves 4 and 5, both of which contain the time preference module. In order to exploit the variation in
the year of schooling, as mentioned in the previous section, we retain observations only for those individuals
15
We divide the mother’s and father’s education level into five categories; no education, some primary education, completed primary
school, completed junior high school, completed senior high school or above.
17
who see a positive change in their education level between the two waves. For this reason, we restrict the
samples to those aged below 21 in 2007, as mentioned earlier. As is clear from (2.2), we drop the vector of
covariates included in earlier specifications since their values do not change between the two waves, except
for current residence district FE, language used in interview, age at time of survey, wave FE, and a dummy
variable concerning whether time preference module A or B was asked first.
Y
ijct
=+S
ijct
+
i
+
j
+
t
+
ijct
(2.2)
While we have employed gradually stricter specifications from equation (2.1) to (2.2), increasingly elim-
inating confounding factors, there is still a possibility that there remains time varying unobservables that
drive changes in both schooling and time preference. We next employ an instrumental variable approach
to investigate the causal effect of education on time preference. We use the interaction of the variation in
the INPRES school construction intensity across districts and cohort-specific exposure to INPRES, detailed
in section 2.3.2, to instrument for completed years of schooling for an individual born in a particular year
in a particular district. For the first stage of equation (2.3), we employ a specification similar to that used
in Duflo (2001): we include controls for district of birth FE, year of birth FE, and their interaction with
the population aged 5 to 14 in 1971, enrollment rate in the region in 1971, and allocation of the water and
sanitation program, the second-largest INPRES program at the time, to account for omitted time-varying
and region-specific effects potentially correlated with the school construction program.
16
We also include
additional covariates - mother’s and father’s education as defined earlier; ethnicity FE; religion FE; current
residence district FE; language spoken in daily life at home; language used in interview; wave FE; a dummy
variable for whether time preference module A or B was asked; rainfall shock variables, as discussed in
section 2.3.1- controls that we believe are important for ruling out other confounding factors. As we briefly
discuss in the literature review, there are many potential factors affecting time preference, such as wealth,
income, labor market participation, and health. We understand the importance of these control variables to
estimate the effect, but we do not include them in the regression because these variables are endogeneous
and more likely to be directly affected by years of schooling. Accordingly, including these variables in the
regression may confound the results more severely.
Thus, we consider the following IV-2SLS regression model:
16
INPRES school construction was dependent on the children population in 1971 and enrollment rate in each region.
18
1
st
stage:S
ijc
=
1
+(T
c
P
j
)
1
+
1
X
ijc
+
j
+!
c
+
ijc
(2.3)
wherei is individual,j is the district,c is the cohort,P
j
denotes the intensity of the INPRES program in
the district of birthj measured by the number of school built per 1,000 children andT
c
is a dummy variable
concerning whether the individual belongs to the treated “young” cohort born between 1968 and 1972 or
the untreated “old” cohort born between 1958 and 1962. We estimate predicted schooling level using this
interaction IV .
The second-stage equation estimates the impact of education on the time preference:
2
nd
stage:Y
ijc
=
2
+
2
^
S
ijc
+
2
X
ijc
+
j
+!
c
+
ijc
(2.4)
Y
ijc
is a dummy variable that takes value 1 if the individual i’s response to the time preference sub-
modules are those reflecting the highest degree of impatience, but 0 otherwise, as discussed in the section
2.3.3. All the controls from the first stage are also included in the second stage. Thus, only the interaction
ofT
c
P
j
is excluded in the second-stage regression.
As addressed in Duflo (2001), after controlling for the interaction of year of birth dummies with child
population and school enrollment rate in 1971,
ijc
in the regression is less likely to be correlated with
IV , so the IV exclusion restriction is satisfied. However, we explore our IV characteristics to validate it.
As already noted, figure A1 shows sufficient variation in INPRES intensity and there are variations and
differences in education between treated and control groups (figure A2). The variation in INPRES intensity
can also explain the variation in time preference modules A and B. We also plot the residual of years of
schooling from the regression of years of schooling on control variables as a function of INPRES intensity
by treatment status in figure 2.6. In the left hand side of figure 2.6 (b), INPRES intensity positively affects
years of schooling for treated group while INPRES intensity has no impact on the control group in the right
hand side of figure 2.6 (b). From these figures, we find support for the validity of our instrument.
19
2.5 Empirical Results
Tables 2.4 presents the main results. We present the OLS results first and advance to the individual FE
and IV-2SLS results. The first three columns report the results of time preference module A (from here on,
TP-A), and the last three columns report the results of time preference module B (from here on, TP-B). We
report the results for pooled sample, male, and female samples separately.
Panel A reports the extent of correlation between years of schooling and time preference after including
all the control variables discussed in the previous section. We find that the educational attainment and
time preference measures are strongly correlated. One more year of schooling is associated with a 0.6 to
0.9 percentage points decrease in impatience as per TP-A and a 0.4 to 0.5 percentage point decrease in
impatience as per TP-B. This is consistent with figure 2.3 and figure 2.4, describing the negative correlation
between education and impatience.
The strong correlation does not imply a casual impact of education on time preference, however. To
probe the significance and magnitude of causal effect of education on time preference further, we tighten the
specification further (equation (2.2)), that is, the individual FE specification, to rule out potential confound-
ing factors at the individual level. The OLS estimates between education and time preference will likely be
biased if individual-level unobservables affect both schooling choice and time preference. In-utero and early
life environment, for example, could affect both the cognitive and non-cognitive traits of an individual. If
certain early-life factors make children smarter at school or at work and patient at the same time, our results
from the OLS estimates in Panel B will be biased even though we minimize this concern by controlling for
rainfall in early life and for many other observable control variables. For example, if individuals who may
hold a high preference for learning are likely to stay longer in school and a high preference for learning is
positively correlated with patience, then the coefficient of OLS is likely to be downward-biased. Absorbing
individual specific time invariant factors can remove this potential source of bias.
We document that education is significantly associated with patience in the individual FE specification
although we find statistical significance only for TP-B. One more year of schooling is associated with a 2.4
percentage point decrease in impatience for female cohorts who were 20 years of age or younger in 2007.
While the estimated coefficients from the TP-A regression of individual FE have a relatively higher p-value,
the signs are all in the same direction: more education makes people less impatient. The estimated coefficient
20
magnitudes are often larger than those in the OLS estimates in Panel B, even though the sample size is less
than a third, suggesting that the lowered significance is, most likely, due to larger standard errors. It is also
worth emphasizing that the magnitude of coefficients for TP-A specification is smaller than the coefficients
for TP-B. This can also be observed in the IV-2SLS results. We will discuss the potential reasons for the
difference in the next section.
From the estimation results using the entire sample and the individual FE sample, we can conjecture
that the effect of education on time preference may hold across different cohorts. That is, the OLS results
in Panel A use all individuals who answered both on TP-A and TP-B while the individual FE in Panel B
uses only those who were younger (20 years of age or less at the time of survey wave 4, 2007). If we agree
that these estimates are reliable, we can conclude that the effect of education on time preference is well
established even though the magnitude of the coefficients is different. Thus, these results may support the
external validity of our findings.
However, we still need a stronger identification strategy for causal analysis. Since individual FE is a more
rigorous specification, the estimates from individual FE are closer to the true causal effect of education on
the time preference than are other estimates from specifications, such as OLS and household FE. However,
it is still a concern that if the estimates are picked up by time-varying individual unobservables, individual
FE estimates will be biased. For example, individual health status may change over time during the survey
period, and this change may confound the true estimates. Since estimating the causal effect of education on
time preference is a primary goal of this paper, we next explore the results from the IV-2SLS specification.
Panel C in table 2.4 presents the results of the IV-2SLS specification. The sample of individuals used
for the IV analysis is different from the sample of individuals used for the OLS (Panel A) and the individual
FE (Panel B) analyses. In the first stage, we find a significant effect of INPRES school construction on
years of schooling, for females only. We find that for our sample of females, one school per 1,000 children
contributes to approximately 0.7 more years of schooling, whereas no significant first stage relationship is
observed for male samples (table A2). By contrast, Duflo (2001) finds a significant impact for males.
17
As explained in section 2.4 (Empirical Analysis), the specification we use is slightly different from the one
used in Duflo (2001), with additional controls. Moreover, while Duflo (2001) used the 1995 Intercensal
Population Survey of Indonesia for her analysis, we use a much smaller sample of individuals from the wave
17
Since she estimated the returns to schooling in the labor market, she focused on male samples only.
21
5 of IFLS. However, it is still a concern if this “seemingly” inconsistent result is due to misspecification.
A recent study by Ashraf et al. (2016) revisits the education effect of the INPRES school construction
program using the 1995 Indonesia Intercensal Data. The researchers find only weakly significant effect
on the completion of primary school for the male sample, while they find stronger and bigger effects for
the female sample when they control for ethnicity FE. Due to differences in cultural backgrounds across
ethnicity - for example, some ethnicities have the custom of bride price - ethnicity has the potential to be
an important factor in educational attainment. Following Ashraf et al. (2016) specification, our study finds
similar results in the magnitude of coefficients, although we find an insignificant effect for male samples.
This is possibly due to the much smaller sample size of the IFLS.
In addition, the higher impact for female education is consistent with the findings of Maluccio et al.
(2009), Maccini and Yang (2009), and Cutler and Lleras-Muney (2010). These studies put forth a number
of reasons why the higher impact arises, two of which seem especially relevant for the case of Indonesia.
First, girls in developing countries face a higher cost of schooling and lower levels of private investment in
education and health than boys do. Cost-reducing public interventions, thus, have much a larger impact for
girls (Maccini and Yang (2009)). Second, the economies of these developing countries are, often, brawn
intensive, and the comparative advantage of men in brawn implies that men have a higher returns on in-
vestment in health and that women have higher returns on investment in education (Pitt et al. (2012)). The
results, therefore, are not surprising.
Panel C reports the IV first stage F-statistics. Robust (Kleibergen-Paap) F-statistics accounting for clus-
tered residuals are 12.59, larger than 10, the conventional rule of thumb threshold, avoiding a weak instru-
ment variable problem. We use the predicted level of schooling from the first stage to estimate the causal
effect of education on time preference. An IV estimate yields point estimates of -0.07 (standard error 0.42)
on impatience (TP-B), which means that one more year of schooling causally decreases the level of impa-
tience by 7 percentage points. Although it is only weakly significant, this is consistent with the individual
FE results. Note that we do not find a significant effect for TP-A and that the magnitude of the coefficient is
much smaller than the coefficients for TP-B, which is also consistent with individual FE results. For the male
sample, the low IV first stage F-statistics for males makes it difficult to interpret the second stage results.
22
In sum, we find a significant effect of education on patience consistently across different specifications.
The OLS specification estimates the average treatment effect of education on patience, although the endo-
geneity of education may prevent us from interpreting a causal relationship. Stronger specification (indi-
vidual FE) also identifies the association between education and patience. Finally, IV-2SLS supports causal
evidence of the effect on time preference. In addition, we find a relatively larger magnitude of coefficients
for IV-2SLS.
18
The larger magnitudes of IV-2SLS can be explained by three possible reasons. First, this may
be because of the potential attenuation bias from a measurement error in the years of schooling (Angrist and
Keueger (1991)). However, although the sample population completed their education around two decades
before the survey, schooling attainments, in general, do not have substantial errors.
19
A second, more plau-
sible, explanation is that the local average treatment effect captured in the IV-2SLS coefficient estimates are
larger than the average treatment effect in the sample population captured in the OLS coefficient estimates
(Imbens and Angrist (1994); Card (2001)).
20
It seems from figure 2.3 that the association between education
and time preference starts at around the end of primary school. INPRES primary school construction, by
virtue of its nature, increased the successful completion of primary school and higher enrollment in sec-
ondary school. Therefore, it is plausible that the local average treatment effect was bigger than the average
treatment effect in our sample. Third, compared to previous OLS and individual FE estimation, it is possible
that extra years of education for the Indonesian population during the 1970s, when the average level of edu-
cation in the population was much lower, had very different impact than an extra year of education between
2007 and 2014. It is possible that extra education changed preference in a more fundamental manner during
the 1970s than it did during the 2000s.
On the basis of the consistent results across different specifications and sub-samples, we conclude that
education causally affects patience. We next perform robustness checks with differently defined dependent
variables.
18
We can provide OLS results for IV-2SLS samples upon requests.
19
For example, Angrist and Krueger (1999) find that the measurement error in self reported schooling during 1970-2000 for surveys
of the United States was around 10-15 %. While it could be higher for a developing country, measurement error alone cannot explain
the high differences between the FE and IV estimates.
20
Card (2001) surveys 11 studies estimating the labor market returns to schooling and find a systematically higher IV estimate than
the OLS estimate. He proposes that the most plausible explanation of this difference is difference between local average treatment
effect and the average marginal effect in the population being studied.
23
2.6 Robustness Check and Discussion
For the robustness check, we investigate the relationship between education and time preference measured
by categorical time preference measures and imputed time discounting rate.
21
Since we observe significant
effects for the female sample in the individual FE and IV-2SLS method, we perform the robustness analysis
only for the female sample. We explore the two other dependent variables by using same specifications:
OLS (Panel A), individual FE (Panel B) and IV-2SLS (Panel C). First, instead of using a dummy variable
of time preference, we use categorical responses as an our outcome variable (figure 2.2). Column (1) and
(3) in table 2.5 present the results of all specifications using categorical responses. We find only a weakly
significant effect for OLS but the sign and the magnitude are similar to the main results.
Next, we estimate the relationship using imputed time discounting rate as an outcome variable. As
we already discussed in the section 2.3.3, the IFLS hypothetical questions only allows us to calculate the
bounds on the time discounting rate. Following the approach by Bauer and Chytilov´ a (2010), we calculate
the midpoint of the range for each category of time preference. We explain the calculation in greater de-
tail in appendix section A.2. Using the imputed time discounting rate, we perform the robustness analysis.
Columns (2) and (4) in table 2.5 report the results using all the specifications used in this study. We find
that one more year of schooling is associated with a 4.2 percentage point decrease in imputed time discount-
ing rate using the individual FE approach while we find that one more year of schooling causally reduces
impatience more than 10 percentage points in the IV-2SLS approach although it is not significant.
We next perform additional robustness checks by applying a different level of fixed effect and cluster
standard error (table 2.6). For example, we perform the analysis by using year of birth * month of birth fixed
effect instead of using them separately to absorb any year-specific seasonal trend effect. We also perform the
analysis by applying different level of clusters, either using province level, district of birth level, or current
residence district level. The impact of education on time preference is similar to the main results except
for statistical significance. The final robustness check is to include interviewer FE. Since an interviewer
may affect how a respondent answers, especially for questions of individual preference (Jakiela and Ozier
(2015)), this may confound our results. We present the results in the column (5) and (10) in table 2.6. After
including the interviewer FE, we find results that are stronger and more significant.
21
We also check the robustness using behavior measures as outcome variables in table 2.7.
24
In summary, we report that more education is closely related to patience, on average. The effect is
robust to the different specifications and differently defined dependent variables. In the next section, we
discuss some of the potential channels through which education might have affected time preference for
these individuals.
Before turning to plausible mechanisms, it is noteworthy that we find the bigger impacts and significant
impacts only for TP-B in the individual FE and IV-2SLS approaches. Since our study sample consists of
respondents who answer both in TP-A and TP-B, we can compare the coefficients in the TP-A regression and
TP-B regression. As seen in table 2.1 and table 2.2, the proportion of impatient respondents is approximately
10 percentage points higher. This difference may drive the main results differently. The difference is due
to the different structure of the questions. The reward amount in TP-A is smaller than the reward amount
in TP-B, and the time horizon is shorter for TP-A (one year) than for TP-B (five years). It is possible that
the time horizon may affect the differences (Dohmen et al. (2012)) but this is less likely. Since the future
rewards in TP-A (3M) and TP-B (4M) are different from those in the first question, it is hard to apply a time
horizon theory to explain the differences. One potential possibility is that respondents may understand less
well how far and how large the future rewards would be in TP-B than in the TP-A questions. Thus, if task
choice questions are asked that are more difficult, many people are likely to choose the simple choice (1M)
now instead of comparing the choices patiently and consequently end up being measured as impatient. In
this sense, we can conjecture that patience may be closely related to cognitive ability, as already discussed
in the previous literature (Dohmen et al. (2010)). We check this possibility in the next section.
2.7 Plausible Mechanism
In this section, we examine some of the potential channels through which education might affect time pref-
erence. To estimate clear causal effects of the mechanism variables, we need to find exogenous variation or
valid instruments for each of these mechanisms, which, clearly, is a tough task. Instead, we rely on a number
of weaker methods and base our arguments on the consistent findings across these methods. The first and
second approaches are mediation analyses introduced by different papers. The third approach is a bootstrap
coefficients correlation check. The last approach involves comparing the coefficient changes after including
potential endogenous mediating variables one by one. We discuss these in more detail below.
25
2.7.1 Mediation Analysis 1 (Imai et al. (2010b))
First, we apply the method of causal mediation analysis proposed by Imai et al. (2010b) and Imai et al.
(2010a). This method makes the strong assumption of sequential ignorability, a nonrefutable assumption
but one that could easily be violated.
22
Therefore, the results from this method, we wish to emphasize, are
suggestive at best. Nevertheless, we believe that there is some value to this exercise.
For this analysis, we first regress our time preference measure and the schooling and mechanism vari-
ables, separately, on all the other controls included in equation (2.1). We, then, predict the residuals for these
three relationships, the part of the variation in these variables, not explained by any of the other controls. We
then run the mediation analysis proposed in Imai et al. (2010b) with the residual from the time preference
regression as the dependent variable, the residual from the schooling regression as the independent variable
and the residual from the mechanism variable equation as the mediating variable. We check for several
possible mechanism variables: cognition, measured using Raven’s test (fluid intelligence), simple arithmetic
questions, word recall test (immediate and delayed recall), and serial sevens subtraction test; total income;
risk aversion; self-reported health; and depression index (Center for Epidemiological Studies Depression
(CES-D) 10-items) scores. A version of Raven’s test and arithmetic questions were administered to adults
(older than 24) for the first time in wave 5.
23
These tests, administered as a part of the IFLS cognition
module, are a combination of Raven’s progressive matrices and some basic mathematical tests that are used
to evaluate the cognitive ability of the test taker. IFLS 5 administered two types of Raven’s test. We use the
second module of Raven’s test which was the harder test. We construct the score variable as a numerical
measure of the total number of correct responses separately for fluid intelligence test (score range of 0-8)
and simple arithmetic test (score range of 0-5). We also examine other cognition measures, such as word
recall test scores (delayed and immediate recall, score range of 0-20) and serial sevens subtraction scores
(score range of 0-5). The total income variable aggregates reported monthly salary/wage amounts, net prof-
its from a participant’s business, rental income, nonlabor income and retirement pension income if there is
any. Along with the time preference measure, the IFLS also collects information about risk preference using
22
A nonrefutable assumption is one that cannot be directly tested from observed data (Manski (2009)) The assumption of sequential
ignorability assumes that given the observed predetermined variables, the treatment assignment is independent of potential outcomes
and potential mediators and that given the observed treatment and pretreatment confounders, the mediator is independent of the out-
comes. Faced with the options of not exploring the potential mechanism and choosing to use this method with its strong assumption,
we chose to do the latter.
23
It was conducted for respondents aged 7-24 in wave 4 and wave 5.
26
hypothetical questions about lottery rewards. For our analysis, we categorize the response to the risk pref-
erence question into two groups, most risk averse and others, as with the time preference measure. Finally,
we check for health-related pathways using a self-health report (we include an indicator for the response of
“very healthy”) and depressive symptoms (CES-D 10-items, score range of 0-30).
The results are presented in Panel A (TP-A) and Panel B (TP-B) of table 2.8. For each of these mech-
anism variables, table 2.8 presents the direct effect that schooling has on time preference independent of
mediation, the effect of the mediating variable on the time preference and the percentage of the total effect
of schooling on time preference obtained through the mechanism variable. The higher the mediation per-
centage for a mechanism variable, the greater are the chances that the variable is the pathway through which
schooling affects time preference.
First, as is clear from the magnitude and significance of the coefficient on the direct effect of schooling,
education is strongly associated with time preference even after controlling for the effect through these
mediation variables. This suggests that there are multiple mechanisms through which education affects
time preference. Next, most of these mechanism variables are themselves strongly associated with time
preference. However, our interest is in the amount of effect of schooling on time preference that is mediated
through the mechanism variables, the magnitude of the “mediation % of total effect”. Of the mechanism
examined, cognition, as measured by the Raven’s test scores and arithmetic scores and, to a lesser extent,
by word recall score and serial sevens subtraction, and total income seem to be important channels through
which education affects time preference.
According to column (1), around 30-40% of the effect of schooling on time preference is produced
through improvements in the cognitive ability of individuals. Word recall score and serial sevens subtraction,
which are considered measures of certain dimensions of cognition, are also reasonably strong mediators.
This is consistent with findings in social psychology. According to Kahneman and Tversky (1981) and Read
et al. (1999), individuals with better cognition are more likely to be future-oriented and patient. According
to Dohmen et al. (2010), emotions and cognition are both important factors in decision making. When
cognition dominates emotions, individuals are likely to take a longer-run view and to be more risk neutral.
The association between education and Raven’s test scores is also well documented (Ceci and Williams
(1997)). Therefore, it looks plausible that education affects time preference by improving the cognitive
27
ability of individuals. This lends support to the findings that the significant relationship is clearly found in
the TP-B regression while no significant relationship is found in the TP-A regression.
A second important mechanism is the total income of an individual, which mediates a little over 10% of
the effect of schooling on time preference. Our findings are consistent with previous studies that find wealth-
ier individuals to be more patient (Becker and Mulligan (1997); Benjamin et al. (2013)). In comparison, risk
aversion, self health evaluations and CES-D scores do not have a strong mediation effect according to this
analysis. The lack of any mediation through health channels is consistent with the arguments of Cutler and
Lleras-Muney (2006).
2.7.2 Mediation Analysis 2 (Huber (2014))
Second, we apply a different mediation analysis proposed by Heckman et al. (2013) and Huber (2014).
The basic idea is similar to the mediation analysis suggested by Imai et al. (2010b). That is, the mediation
analysis here explains how much variation in mechanism variables can explain the total effect of education
on time preference. The difference is that Huber (2014) develops the model by applying inverse proba-
bility weighting for each potential mediating variables. While the key identifying assumption “sequential
ignorability” is same as that of Imai et al. (2010b), it relaxes the conditional independence of the mediator
assumption by calculating the inverse probability weights of each potential mediator and applying it to the
regressions. To calculate the inverse probability weights of each potential mechanism variables, first, we run
a dummy variable indicating whether he or she graduated from a primary school on the instrument variables
(intensity of the INPRES * cohort treatment dummy), mediators and control variables. We calculate the
probability of completing primary school and construct inverse probability weights by taking an inverse of
the probability. Then, we run two separate regressions, which applies the inverse probability weights. First,
we run the regression of time preference measure on years of schooling, mediators, control variables and
interaction of mediators and years of schooling. Second, we run the regression of time preference measure
and mediators on years of schooling, control variables and other mediators not used in the dependent vari-
able. The contribution of mediators to the total effect can be calculated by comparing coefficients on years
of schooling. This method estimates the degree to which impacts of education on time preference can be
explained by potential mechanism variables, along with the impacts of education, on these potential mech-
anism variables directly. Following this approach, we calculate the percentage contribution of mechanism
28
variables to the total effect of education on TP-A and TP-B in figure 2.7. In order to make the mediators
comparable, non-binary mediators, such as Raven’s test (fluid intelligence), arithmetic test, total word recall,
serial 7 subtraction, total income, and CES-D are standardized.
The eight potential mechanism variables account for approximately 10-15 percent of total effect of edu-
cation on time preference. This means the remaining 85-90 percent can be explained either by other mecha-
nism variables or by the direct effect of education on time preference. We do not emphasize the direct effect
much because education can affect various socioeconomic outcome variables. Understanding the remaining
mechanism variables will be for future work on understanding time preference formation.
Among the eight mechanism variables, we find that cognition measured by Raven’s test, fluid intelli-
gence, is the most significant factor linking education and patience. This is a finding consistent with the
previous mediation analysis. Although the percentage contribution is relatively small, another cognition test
measuring arithmetic skill is also an important mediator. However, in this mediation analysis, we do not find
a meaningful role for total income as a mechanism.
2.7.3 Joint Bootstrap Analysis (Bennett et al. (Forthcoming))
As a third test for these mechanisms, we use an alternative method used in Bennett et al. (Forthcoming). We
jointly bootstrap the effect of education on the time preference measure and the effect of education on the
mechanism variable. Figures 2.8 and 2.9 show the scatter plot of bootstrapped replication coefficients. The
estimated correlation coefficient from the regression of education on the time preference measure is reported
along the Y-axis and the estimated correlation coefficient between from the regression of education on the
relevant mechanism variable is reported along the X-axis. Figure 2.8 depicts the coefficients plot for cog-
nition variables. The first figure from figure 2.8 shows a negative slope using the fluid intelligence Raven’s
test as a mechanism factor. The negative relationship in figure 2.8 indicates that strong effects on patience
also tend to have strong effects on cognition. That is, when we draw repeated random samples from the
analysis sample, we find that individuals who have a high degree of association between education and fluid
intelligence also have a high degree of association between education and time preference. This suggests
that improvements in cognition, as measured by Raven’s test (fluid intelligence) score, may be an impor-
tant channel through which education makes people more patient. We find similar patterns for cognition as
measured by arithmetic test and the word recall test. By contrast, we find no significant relationship using
29
serial sevens subtraction scores. To conclude, using a bootstrapped coefficients plot, we can conjecture that
cognition skill can be an important mechanism for the link between education and time preference.
In figure 2.9, we also find significant relationships for total income, risk averseion and self-health reports.
Total income can also be an important mechanism, which is a consistent finding as in table 2.8. In this
analysis, using a bootstrapped coefficients plot, we find that risk aversion and health may also be a potential
mechanism. For the risk-preference figure, the graph is consistent with the coefficient signs in table 2.8.
Individuals with lower schooling are more likely to be impatient and are also more risk averse. However,
since a different mechanism check provides a puzzling result, we are skeptical of the role of risk preference
as a mediator. The plots for self-reported health is also puzzling so we are cautious in interpreting the health
measure as a mechanism variable. CES-D scores plots are very flat, which suggests the weak explanatory
power of the mechanism.
In appendix table A3, we use another test where we include the endogenous mechanism variables, one
by one, as regressors in specification (2.1) (MacKinnon et al. (2007); Maccini and Yang (2009)). While we
are aware of problems with this approach, we believe that this exercise might, nonetheless, be informative.
However, the results, presented in table A3, have to be interpreted with caution. If the inclusion of a par-
ticular mechanism variable dampens the main correlation coefficient of education and the time preference
measure substantially and is itself correlated with the time preference measure significantly, it suggests that
this mechanism variable is soaking away the variation in the time preference measure that covaries with ed-
ucation. If this is also accompanied by a rise in adjustedR
2
for the model, the likelihood of this possibility
as a mechanism becomes higher. The results are consistent with the two tests used above that cognition mea-
sured by Raven’s test (fluid intelligence) and arithmetic test, the word recall test and serial sevens subtraction
are most likely to be the mechanism.
However, as mentioned above, the results are suggestive at best. Further research should go into a lot
more detail of these mechanisms and should also explore other possible pathways.
2.7.4 Mechanisms Related to Medical Findings
From the mechanism analyses above, we find that cognition is the most plausible factor linking education
and patience. Many previous studies have found a positive association between education and cognition
(Black et al. (2011); Falch and Sandgren Massih (2011); Green and Riddell (2012); Carlsson et al. (2015)).
30
In addition, the positive relationship between patience and cognition has also been well established (de Wit
et al. (2007); Dohmen et al. (2011)). Thus, it can naturally be conjectured that cognition might play a role
in linking education and cognition. This conjecture is also supported by medical findings indicating that
dorsolateral prefrontal cortex, one of the brain domains that governs the executive functions, such as work-
ing memory, cognitive flexibility, planning (Kaplan et al. (2016); Miller and Cummings (2017)), abstract
reasoning, inhibition (Miller and Cummings (2017)), and decision making (Duncan and Owen (2000)), is
closely related to one’s cognition and patience. According to Duncan and Owen (2000), people with damage
in the dorsolateral prefrontal cortex area have problems calculating costs and benefits of alternative choices.
Damage in this area could also lead to cases where an individual is tempted to select sub-optimal options by
choosing the rewards delivered immediately (Knoch and Fehr (2007)).
Since dorsolateral prefrontal cortex is malleable and continues the development until adulthood, improv-
ing the activeness of this area throughout the lifetime may help keep cognition and patience at the optimal
level. Rzezak et al. (2015) found that educational attainment is one of the key factors for improving the
functioning of the dorsolateral prefrontal cortex. Thus, education causally affects patience through increased
cognitive function by improving the brain areas of dorsolateral prefrontal cortex. Our study contributes to
the existing medical findings using observational data to suggest how cognition plays a role in explaining
the relationship between education and patience.
2.8 Limitations and Concerns
Although this study provides suggestive evidence of the causal effects of educational attainments on individ-
ual time preference, the causal estimate needs to be interpreted with caution as a statistical significance of
the effect is weak in the main IV-2SLS analysis. However, we do not interpret it as no causal link between
education and time preference. Since our estimate captures the long-term effect of education, the explana-
tion power of the variation in educational attainments approximately 30 years ago for the variation in time
preference may become weaker relative to short-term stronger effect studied by Perez-Arce (2017). As we
find the stronger effect of educational attainment on behavior measures proxied for time preference, more
research using alternative measures of time preference is needed to investigate the long-term causal effect.
31
Another concern of this study is whether our estimate captures the true effect of educational attainments.
As Jensen and Lleras-Muney (2012) and Perez-Arce (2017) discussed in their study, educational attainment
affects many different socioeconomic outcome variables, such as, income, working status, marriage, and
health behavior, which are more likely to be correlated with individual patience as discussed in this study.
Thus, much of the effect of education on time preference may be explained by correlations between time
preference and these socioeconomic outcome variables. Due to this concern, we estimate the effect after
controlling for these factors to check if the estimate is dramatically changed, which we don’t find any evi-
dence. However, we don’t emphasize these results much as including these factors potentially affected by
educational attainment may confound the main estimate because it is, by itself, an outcome variable. To in-
terpret our result conservatively, our estimate does not capture the pure effect of education, rather it captures
the total effect of education on time preference.
Related to income and liquidity, it may be also of concern if our estimate depends on the liquidity
condition at the time of survey. If those who were patient may experience different liquidity conditions,
disregarding these endogeneity may also confound our main estimate. That is, although two respondents
have same true preference as ‘patience’, if one respondent is in the urgent need of money now, he may reveal
his preference as ‘impatience’ at the time of survey. This can definitely confound our estimate. However,
it is less likely in our setting. Since our main specification follows IV-2SLS, if our instrument satisfies the
exclusion restriction well, there is no a priori reason to believe that our IV systematically makes difference
in liquidity conditions between individuals at the time of survey. Thus, it minimizes the potential concerns
for this possibility.
2.9 Conclusion
According to Ramsey’s growth model, time discounting is closely related to saving and consumption be-
havior, and it can explain economic growth and inequality (Ramsey (1928)), as depicted in the conceptual
framework (figure 2.1). Despite its far-reaching implications, the factors affecting time preference are not
well understood. In this study, we try to bridge this gap; we find a significant causal effect of education
on time preference, as measured by hypothetical survey questions. This paper overcomes the endogeneity
problem of educational attainments by utilizing several specifications (OLS, individual FE, IV-2SLS). The
32
empirical results are robust to different specifications and different sub-samples from waves 4 and 5 of the
IFLS. According to the IV specification results, one more year of schooling makes women more patient by
approximately seven percentage points. We provide suggestive evidence of the mechanisms that link edu-
cation and patience. Although the rigorous mechanism analysis is econometrically complicated, we suggest
that cognition measured by Raven’s test scores, word recall, serial sevens subtraction and total income are
potential mechanisms by using four different mechanism analyses. Particularly, cognition explains a signifi-
cant portion of the relationship between education and patience. This implies that cognition can explain the
different results between regressions of TP-A and TP-B.
The results underscore the importance of policies aimed at improving educational outcomes. Since
improvement in education can make people more patient and, therefore, more likely to invest in human
capital further, such policies can trigger a virtuous cycle. They can also trigger positive behavioral responses,
such as higher savings and insurance purchases, and positive health practices, such as lower chances of taking
up smoking and a higher use of condoms. The results also contribute to the literature on determinants of
non-cognitive abilities, abilities that are important inputs into the human capital production function and that
are crucial for the overall development of individuals (Heckman (2007)).
More research must be done to understand better the determinants of time preference. As many empiri-
cal papers have found the correlations between potential factors and time preference, further research should
investigate how much these correlations can be explained by causal relationships. This may help in rethink-
ing the utility model to explain many economic behavior outcomes. Further, we cannot disentangle the
direct and indirect effects of education on time preference in this study. At best, we propose some potential
mechanism variables. For future research, experimental settings will help to control for potential mechanism
variables, while creating a variation in education level only to estimate the direct effect of education on time
preference.
33
Figure 2.1: Conceptual Framework
* This conceptual framework is slightly modified based on the figure in Dohmen et al. (2015).
34
(a) Time preference module A
(b) Time preference module B
Figure 2.2: Time preference categories 35
(a) Time Preference A (b) Time Preference B
Figure 2.3: Lowess plot of time preference and education
(a) Time Preference A (b) Time Preference B
Figure 2.4: Time preference and education
36
(a) Smoking (b) Working for Pay
(c) Saving (d) Saving Amount
(e) Arisan Participation (f) Village Saving
(g) Total Income (h) Contraception
Figure 2.5: Behavior and Education
37
(a) Time preference and INPRES intensity
(b) Years of schooling and INPRES intensity by treatment status
Figure 2.6: Lowess of Inpres intensity and Female Time preference and Education (IV sample)
38
(a) Time Preference A
(b) Time Preference B
Figure 2.7: Mediation Analaysis (Huber (2014))
39
(a) Raven’s test (Fluid intelligence) (b) Arithmetic test
(c) Total word recall (d) Serial 7 subtraction
Figure 2.8: Bootstrapped correlation plot of time preference’s and mechanisms’ correlation with education
40
(a) Total Income (b) Risk Averse
(c) Self health - very healthy (d) CES-D
Figure 2.9: Bootstrapped correlation plot of time preference’s and mechanisms’ correlation with education
41
Table 2.1: Summary Statistics 1
Wave 4 Wave 5
Male (N=5,202) Female (=6,527) Male (N=5,122) Female (N=6,286)
Mean S.D. Mean S.D. Mean S.D. Mean S.D.
Time Preference A
1(=Most patient) 0.074 0.083 0.086 0.096
2 0.057 0.061 0.085 0.083
3 0.172 0.135 0.156 0.168
4(=Most Impatient) 0.697 0.720 0.672 0.653
Time Preference B
1(=Most patient) 0.018 0.019 0.017 0.015
2 0.038 0.035 0.049 0.045
3 0.145 0.129 0.184 0.169
4(=Most Impatient) 0.799 0.818 0.750 0.771
Demographics
Age 36.36 9.760 34.35 10.20 41.15 10.14 39.79 10.33
Years of schooling 8.896 3.989 8.395 3.932 9.686 4.046 9.151 4.060
Mother has no education 0.392 0.488 0.353 0.478 0.304 0.460 0.268 0.443
Father has no education 0.288 0.453 0.256 0.437 0.184 0.387 0.156 0.363
Urban residence 0.079 0.269 0.080 0.271 0.087 0.282 0.087 0.282
Note: We define the two types of time preference using hypothetical questions (figure 2.2). We divide the answers into four group.
The people in the group 1 is most patient compared to other group and the people in the group 4 is the most impatient.
42
Table 2.2: Summary Statistics 2 (IV sample)
Wave 4 Wave 5
Male (N=249) Female (N=163) Male (N=1,292) Female (N=1,301)
Mean S.D. Mean S.D. Mean S.D. Mean S.D.
Time Preference A
1(=Most patient) 0.060 0.074 0.085 0.085
2 0.056 0.043 0.084 0.074
3 0.173 0.141 0.163 0.164
4(=Most Impatient) 0.711 0.742 0.668 0.676
Time Preference B
1(=Most patient) 0.008 0.018 0.015 0.016
2 0.024 0.049 0.042 0.049
3 0.137 0.104 0.168 0.143
4(=Most Impatient) 0.831 0.828 0.776 0.792
Demographics
Age 40.85 5.168 41.58 5.171 47.32 4.962 47.64 5.024
Years of schooling 8.799 4.224 8.491 4.382 9.526 4.154 8.605 4.258
Mother has no education 0.357 0.480 0.393 0.490 0.370 0.483 0.330 0.470
Father has no education 0.281 0.450 0.362 0.482 0.213 0.409 0.188 0.391
Note: IV sample consists of cohorts born between 1968 and 1972 (treated group by INPRES) and born between 1958 and 1962
(untreated group by INPRES). We mainly use the sample from IFLS 5 and include respondents who answered the questions in IFLS
4 but did not answer in IFLS 5.
43
Table 2.3: Time preference and behavior (Association)
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Smoking Working for pay in the last week Saving
Time Preference A 0.010 0.029*** -0.002 -0.019* 0.041 -0.031*** -0.044*** -0.059** -0.041***
(0.007) (0.010) (0.003) (0.011) (0.029) (0.011) (0.010) (0.023) (0.012)
Observations 18,748 8,356 10,392 6,546 710 5,836 8,620 1,672 6,948
R-squared 0.081 0.171 0.155 0.218 0.738 0.238 0.177 0.446 0.201
Mean of DV 0.336 0.730 0.0198 0.197 0.135 0.205 0.277 0.268 0.279
Time Preference B -0.013 0.020* -0.005 -0.014 -0.010 -0.022* -0.036*** -0.057** -0.034***
(0.008) (0.012) (0.003) (0.012) (0.033) (0.013) (0.012) (0.027) (0.013)
Observations 18,748 8,356 10,392 6,546 710 5,836 8,620 1,672 6,948
R-squared 0.081 0.170 0.155 0.218 0.737 0.237 0.176 0.446 0.200
Mean of DV 0.336 0.730 0.0198 0.197 0.135 0.205 0.277 0.268 0.279
Saving Amount Arisan participation Village savings (participation)
Time Preference A -0.682*** -0.965*** -0.620*** -0.028*** -0.034*** -0.024** -0.003 -0.003 -0.003
(0.145) (0.333) (0.162) (0.007) (0.008) (0.010) (0.003) (0.004) (0.004)
Observations 8,284 1,606 6,678 18,748 8,356 10,392 16,286 7,280 9,006
R-squared 0.189 0.468 0.213 0.201 0.284 0.259 0.098 0.169 0.132
Mean of DV 3.524 3.425 3.548 0.307 0.169 0.417 0.277 0.268 0.279
Time Preference B -0.615*** -1.043*** -0.591*** -0.013 -0.021** -0.018 -0.005 -0.011** 0.002
(0.163) (0.381) (0.182) (0.008) (0.009) (0.011) (0.003) (0.005) (0.004)
Observations 8,284 1,606 6,678 18,748 8,356 10,392 16,286 7,280 9,006
R-squared 0.189 0.468 0.212 0.200 0.283 0.259 0.176 0.446 0.200
Mean of DV 3.524 3.425 3.548 0.307 0.169 0.417 0.277 0.268 0.279
Log of Total Income Contraception
Time Preference A -0.240** -0.145 -0.234* -0.059***
(0.100) (0.104) (0.140) (0.021)
Observations 18,525 8,340 10,185 8,832
R-squared 0.123 0.168 0.172 0.197
Mean of DV 9.437 12.50 6.927 1.385
Time Preference B -0.425*** -0.249** -0.284* -0.029
(0.113) (0.116) (0.160) (0.023)
Observations 18,525 8,340 10,185 8,832
R-squared 0.123 0.169 0.172 0.196
Mean of DV 9.437 12.50 6.927 1.385
Sample Pooled Male Female Female
Note: The table reports the association between education and behavior measures. We control for district FE, year of birth FE, ethnicity FE, whether respondents
answer TP-A first or TP-B first.
*p< 0:1, **p< 0:05, ***p< 0:01
44
Table 2.4: Main Result
Time Preference A Time Preference B
(1) (2) (3) (4) (5) (6)
Panel A. Entire Sample
OLS
Years of schooling -0.007*** -0.009*** -0.006*** -0.005*** -0.005*** -0.004**
(0.001) (0.002) (0.002) (0.001) (0.002) (0.002)
R-squared 0.037 0.054 0.046 0.039 0.060 0.044
Observations 18,426 8,189 10,237 18,426 8,189 10,237
Mean DV 0.687 0.688 0.686 0.785 0.778 0.790
Mean YoS 9.105 9.395 8.873 9.105 9.395 8.873
Panel B. Individual FE sample
OLS
Years of schooling -0.020*** -0.022*** -0.020*** -0.016*** -0.013*** -0.019***
(0.004) (0.005) (0.004) (0.003) (0.004) (0.004)
R-squared 0.125 0.215 0.164 0.118 0.202 0.151
Individual FE
Years of schooling -0.008 -0.005 -0.011 -0.019* -0.016 -0.024**
(0.009) (0.023) (0.015) (0.010) (0.023) (0.011)
R-squared 0.634 0.679 0.667 0.635 0.676 0.664
Observations 4,550 1,993 2,557 4,550 1,993 2,557
Mean DV 0.613 0.626 0.603 0.726 0.727 0.725
Mean YoS 10.267 10.335 10.214 10.267 10.335 10.214
Panel C. IV-2SLS Sample
OLS
Years of schooling -0.008*** -0.016*** -0.000 -0.004*** -0.009*** -0.000
(0.002) (0.004) (0.004) (0.001) (0.002) (0.003)
R-squared 0.074 0.149 0.120 0.077 0.142 0.115
IV-2SLS
Years of schooling -0.002 0.480 -0.002 -0.045 0.367 -0.070*
(0.054) (1.127) (0.032) (0.067) (0.932) (0.042)
Observations 3,005 1,541 1,464 3,005 1,541 1,464
Mean DV 0.679 0.675 0.684 0.790 0.785 0.796
Mean YoS 9.011 9.409 8.592 9.011 9.409 8.592
IV First stage F 7.896 0.111 12.59 7.896 0.111 12.59
Sample Pooled Male Female Pooled Male Female
Note: IV sample consists of cohorts born between 1968 and 1972 (treated group by INPRES) and born between 1958 and 1962
(untreated group by INPRES). We control for parental education, ethnicity, religion, year of birth FE, month of birth FE, district of
birth FE, district of birth FE, current district FE, language used for interview, language commonly used in daily life, survey wave FE,
the order of TP module asked. For IV analysis, we control for children’s population in 1971, primary school enrollment rate in 1971
and sanitation program at the time of INPRES. We also control for rainfall at birth and early life period. For individual FE analysis,
we do not need to control time invariant observables. Robust standard errors are clustered at the province level.
*p< 0:1, **p< 0:05, ***p< 0:01
45
Table 2.5: Robustness 1: Time preference category and imputed discount rate (Female)
Category A Imputed Rate A Category B Imputed Rate B
(1) (2) (3) (4)
Panel A. OLS (N=10,237)
Years of schooling -0.006* -0.046*** -0.003 -0.008**
(0.003) (0.016) (0.002) (0.003)
Mean of DV 0.686 0.686 0.790 0.790
Mean of YoS 8.873 8.873 8.873 8.873
Panel B. Individual FE (N=2,557)
Years of schooling -0.016 -0.090 -0.021 -0.042**
(0.034) (0.140) (0.018) (0.019)
Mean of DV 0.603 0.603 0.725 0.725
Mean of YoS 10.214 10.214 10.214 10.214
Panel C. IV-2SLS (N=1,464)
Years of schooling -0.016 -0.023 -0.096 -0.126
(0.083) (0.302) (0.069) (0.077)
Mean of DV 0.684 0.684 0.796 0.796
Mean of YoS 8.592 8.592 8.592 8.592
IV F-stat 12.59 12.59 12.59 12.59
Note: Time preference category consists of four categories (Figure 2.2). The point at which a respondent switches from the current
option to future option offers the range of discount rate. We define the midpoint of the range as respondent’s time discounting rate.
Robust standard errors are clustered at the province level.
*p< 0:1, **p< 0:05, ***p< 0:01
46
Table 2.6: Robustness 2: Different level of fixed effect, controls and clustered robust standard error (Female only)
Time preference A Time preference B
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
OLS (N=10,237)
Years of schooling -0.005*** -0.005*** -0.005*** -0.005*** -0.006*** -0.005** -0.004** -0.005*** -0.005*** -0.005***
(0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.001) (0.002) (0.002)
Year of Birth FE Y N Y Y Y Y N Y Y Y
Month of Birth FE Y N Y Y Y Y N Y Y Y
Year * Month FE N Y N N N N Y N N N
Interviewer FE N N N N Y N N N N Y
Individual FE (N=2,557)
Years of schooling -0.011 -0.011 -0.011 -0.011 0.001 -0.024* -0.024* -0.024 -0.024 -0.021
(0.015) (0.015) (0.018) (0.018) (0.025) (0.011) (0.011) (0.016) (0.015) (0.021)
Year of Birth FE Y N Y Y Y Y N Y Y Y
Month of Birth FE Y N Y Y Y Y N Y Y Y
Year * Month FE N Y N N N N Y N N N
Interviewer FE N N N N Y N N N N Y
IV-2SLS (N=1,464)
Years of schooling -0.002 -0.012 -0.002 -0.002 -0.048 -0.070* -0.094*** -0.070 -0.070 -0.134**
(0.032) (0.045) (0.042) (0.045) (0.055) (0.042) (0.034) (0.047) (0.047) (0.066)
Year of Birth FE Y N Y Y Y Y N Y Y Y
Month of Birth FE Y N Y Y Y Y N Y Y Y
Year * Month FE N Y N N N N Y N N N
Interviewer FE N N N N Y N N N N Y
Clustered SE Province Province District of Birth District of current residence Province Province Province District of Birth District of current residence Province
IV F-stat 12.59 11.47 8.45 6.6 12.59 12.59 11.47 8.45 6.6 12.59
47
Table 2.7: The effect of education on behavior measures (IV-2SLS)
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Panel A. Smoking Working for pay in the last week Saving
Years of schooling -0.025 -0.181 -0.000 0.076*** 0.073 0.076*** 0.029 -0.213 0.044**
(0.038) (0.799) (0.010) (0.026) (0.344) (0.026) (0.044) (0.667) (0.020)
Observations 3,005 1,541 1,464 3,005 1,541 1,464 3,005 1,541 1,464
Mean of DV 0.399 0.743 0.037 0.074 0.026 0.125 0.139 0.042 0.240
Mean of YoS 9.011 9.409 8.592 9.011 9.409 8.592 9.011 9.409 8.592
IV F-stat 9.189 0.0511 13.47 9.189 0.0511 13.47 9.189 0.0511 13.47
Panel B. Saving Amount Arisan Participation Village savings (participation)
Years of schooling 0.036 -0.931 0.083 0.113** -0.148 0.117** -0.018 0.040 -0.030
(0.103) (3.090) (0.054) (0.052) (0.573) (0.057) (0.033) (0.150) (0.018)
Observations 3,005 1,541 1,464 3,005 1,541 1,464 2,645 1,376 1,269
Mean of DV 0.349 0.102 0.610 0.389 0.236 0.551 0.053 0.046 0.061
Mean of YoS 9.011 9.409 8.592 9.011 9.409 8.592 8.859 9.264 8.421
IV F-stat 9.189 0.0511 13.47 9.189 0.0511 13.47 7.135 0.380 7.516
Sample Pooled Male Female Pooled Male Female Pooled Male Female
Panel C. Log of Total Income Contraception
Years of schooling 1.460* 4.725 1.615*** 0.026
(0.800) (13.072) (0.475) (0.442)
Observations 2,970 1,538 1,432 952
Mean of DV 10.900 12.832 8.825 1.716
Mean of YoS 9.023 9.414 8.603 9.374
IV F-stat 9.195 0.0554 12.49 1.206
Sample Pooled Male Female Female
Note: We use INPRES school construction program as IV for education. We control for parental education, ethnicity, religion, year of birth FE, month of birth FE,
district of birth FE, district of birth FE, current district FE, language used for interview, language commonly used in daily life, survey wave FE, the order of TP
module asked. For IV analysis, we control for children’s population in 1971, primary school enrollment rate in 1971 and sanitation program at the time of INPRES.
We also control for rainfall at birth and early life period. Robust standard errors are clustered at the province level.
*p< 0:1, **p< 0:05, ***p< 0:01
48
Table 2.8: Mediation analysis of mechanism (Residual)
Raven (Fluid*) Arithmetic test Word Recall Serial 7s Total Income Risk Averse Self health CESD
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. DV: Time Preference A
Years of schooling -0.006*** -0.006*** -0.006*** -0.008*** -0.007*** -0.008*** -0.007*** -0.007***
(0.002) (0.002) (0.001) (0.002) (0.001) (0.002) (0.001) (0.001)
Mechanism Variable -0.014*** -0.024*** -0.004*** -0.018*** -0.012*** 0.065*** -0.017 -0.005***
(0.003) (0.003) (0.001) (0.004) (0.001) (0.004) (0.011) (0.001)
Observations 10,965 10,951 18,413 9,240 12,615 12,143 18,480 18,361
Mediation % of total effect 0.268 0.310 0.141 0.137 0.119 0.000 0.001 0.012
Mean of DV 0.667 0.667 0.686 0.667 0.680 0.617 0.686 0.686
Panel B. DV: Time Preference B
Years of schooling -0.003** -0.003* -0.004*** -0.004** -0.006*** -0.005*** -0.005*** -0.005***
(0.002) (0.002) (0.001) (0.002) (0.001) (0.001) (0.001) (0.001)
Mechanism Variable -0.007*** -0.017*** -0.003*** -0.007* -0.003* 0.112*** -0.007 -0.004***
(0.003) (0.003) (0.001) (0.004) (0.004) (0.009) (0.009) (0.001)
Observations 10,965 10,951 18,413 9,240 12,615 16,809 18,480 18,361
Mediation % of total effect 0.265 0.402 0.137 0.124 0.039 0.073 0.000 0.016
Mean of DV 0.767 0.767 0.784 0.763 0.775 0.785 0.785 0.784
Sample Pooled Pooled Pooled Pooled Pooled Pooled Pooled Pooled
Note: Fluid means fluid intelligence test. Robust standard errors are clustered at the province level. All regressions control for year of birth FE, month of birth FE, ethnicity
FE, religion FE, district FE, rainfall in different years of life, language used for interview, language used for daily life, parent’s education level, wave fixed effect and the
order of questions asked. Raven test measuring fluid intelligence scores between 0-8. Arithmetic test measuring mathematical skill scores between 0-5. Word recall score
between 0-20. Serial 7 score lies between 0-5. Total income variable is log of total income earned as wage or business profit, rental income, nonlabor income and retirement
pension income. Risk averse is a dummy indicating whether the respondent chose the most risk averse option throughout the risk preference module. Self health is a dummy
that takes value 1 if individuals report being in very good health. CESD ranges between 0-30.
*p< 0:1, **p< 0:05, ***p< 0:01
49
Chapter 3
Is 1+1 more than 2? Joint evaluation of two public
programs in Tanzania
3.1 Introduction
Health and education investments in-utero and early life periods have been observed to have long run impacts
on educational attainments (Sommer et al. (1986); Miguel and Kremer (2004); Almond (2006); Bleakley
(2007); Field et al. (2009); Almond and Currie (2011)). We revisit this topic by examining the effect of
health program (Iodine Supplementation Program) and education policy (Primary Education Development
Program) on educational attainments at the age of 10-13 in Kagera region in Tanzania. We further the
discussion to study joint effects of health and education investments made in adjacent years to study the
interactive role of health and education investments in determining the educational attainments, which have
been less investigated before.
Cunha and Heckman (2007) propose a theoretical framework that considers the formation of human
capital as a dynamic process describing the complementarity between investments at different stages of life.
They term this ‘dynamic complementarity’. For example, investments in human capital later in life are likely
to be more productive for individuals who receive higher investments in early life periods vis-a-vis those
who received lower investments. There is a growing body of research that tests empirically this theoretical
framework by using ‘lightening to strike twice’ (Almond and Mazumder (2013)) exogenous variations made
at the different stages of life (Bhalotra and Venkataramani (2012); Adhvaryu et al. (2014); Rossin-Slater and
W¨ ust (2016); Gunnsteinsson et al. (2016); Malamud et al. (2016)). However, satisfying the identification
assumption is difficult as the treatment uptake or compliance to the second program could be correlated with
the impact of the first treatment. Tanzania, one of the least developed countries in the world, is the ideal
NOTE: This chapter is coauthored by Tushar Bharati (USC) and Seungwoo Chin (USC).
50
region for applying this identification strategy because only few efforts in child investments at an individual
level are expected so public interventions in child investments are the main driving force in human capital
production. Since the individual level investments made in response to the public policy interventions could
confound the identification strategy if there are immense individual level investments, this study setting can
minimize this concern.
In this paper, we attempt to add another empirical evidence to ‘dynamic complementarity’. We use
information from Kagera Health and Development Survey (KHDS) to examine the joint effects of two public
programs - the Iodine Supplementation Program (ISP) and the Primary Education Development Program
(PEDP) - on educational attainments. ISP, which was a health investment program and targeted pregnant
women (consequently their baby in utero) in our sample, is expected to have a positive impact on schooling
by improving baseline levels of cognition of the study sample, as argued in Field et al. (2009). PEDP, which
was an education investment program, is also expected to have a positive impact on schooling by reducing
the cost of schooling. This study, however, finds that ISP, which treatment intensity was varied by districts
and years, has a negative and significant impact on on-going educational attainments measured at the age
10-13 while PEDP, which treatment status was varied by birth cohorts, has a positive and significant impact
on on-going educational attainments at the age of 10-13 as expected. Several explanations are addressed
to explain the negative impact of the ISP. First, since we estimate the educational attainments at the age
of 10-13, the impact of ISP on final completed years of schooling could turn out to be positive. Second,
heterogeneous children’s labor participation by ISP treatment status can explain the negative impact of the
ISP. Third, related to children’s labor participation, heterogeneous starting age of primary school can explain
the negative impact of ISP. Parents may have a different strategy or preference by children’s ISP treatment
status, which could lead to compensating or reinforcing behavior in the education market.
In addition, we find significant and negative joint effects of ISP and PEDP on educational attainments,
meaning that individuals who were treated by both ISP and PEDP had less educational attainments than indi-
viduals not treated either of the two interventions. The existing literature address the ‘dynamic complemen-
tarity’ by estimating the positive interaction coefficients of two exogenous interventions when studying the
effects on human/physical capital measures such as educational attainments and health stock. In this study,
however, we do not interpret the interaction estimates as evidence for or against ‘dynamic complementarity’
for school grade attainments. Instead, we propose a new measure to capture ‘dynamic complementarity’: in
51
addition to impacts of these programs on school grade attainments, we also estimate their effects on primary
school starting age and use the two estimates to compute how productive individuals are in school. That is,
we show how productive ISP treated individuals are at converting years at school into on-going educational
attainments. We show that ISP treated children made much better use of time in school due to PEDP than
ISP untreated children, suggesting ‘dynamic complementarity’ between iodine supplementation in-utero and
time in school later in life in the production of education.
Our contribution is two-fold. First, we re-evaluate the impact of ISP on schooling attainments for indi-
viduals in Kagera region of Tanzania. We present results that could help reconcile the divergent findings of
two previous studies (Field et al. (2009); Bengtsson et al. (2013)) that have evaluated the impact of ISP and
underline the importance of understanding the mechanisms or behavioral responses to ISP treatment to better
understand the reduced form impact of ISP. Second, we start the methodological discussion on identifying
‘dynamic complementarity’ by illustrating the need for more cautious interpretation of the reduced magni-
tudes of the interaction of two exogenous shocks as evidence for or against complementarity by proposing
a new measure. Consequently, we provide another empirical evidence of ‘dynamic complementarity’ in the
context of developing countries, which have rarely been studied.
The remainder of the paper is organized as follows. Section 3.2 provides a brief background of the ISP
and PEDP interventions. Section 3.3 discuss the existing literature. Section 3.4 describes the data and the
empirical strategy. Section 3.5 presents the main results. Section 3.6 concludes.
3.2 Background
3.2.1 Iodine Supplementation Program (ISP)
Lack of proper nutrient in utero or during early life is detrimental to the physical and cognitive development
of individuals (Barker (1990); Cao et al. (1994); Barker (1995); Barker et al. (2002); Zimmermann et al.
(2005)). Iodine, one of the important nutrient for a pregnant mother, is required for the synthesis of thyroid
hormones, the adequate supply of which are important for physical and mental development of the fetus.
2
Especially, sufficient iodine in the first trimester of a pregnant mother is considered extremely crucial for the
2
Severe iodine deficiency during pregnancy can lead to stillbirth, spontaneous abortion, increased perinatal and infant mortality, and
severe mental retardation (cretinism) (Lamberg (1991)).
52
cognitive development of the baby in-utero. Brain development is sensitive to minor adjustments in thyroid
hormone and mild maternal iodine deficiency can impair the full cognitive development of an individual
(Dugbartey (1998); Pop et al. (1999)).
People from Tanzania, like those from many African countries, traditionally suffered from high rates of
Iodine Deficiency Disorder (IDD). According to a Tanzania nationwide survey of iodine levels in the early
1970s, about 40% of the Tanzanian population lived in iodine deficient areas and 25% of the population
was estimated to have had IDD. The prevalence among pregnant and lactating women was as high as 52%
(Van der Haar et al. (1988)).
3
In response, the Tanzania Food and Nutrition Center (TFNC) started distributing iodized oil capsules
(IOC), named as ISP, to individuals in districts where more than 10% of school children had some symptoms
of goiter. The program was expanded to 27 districts covering 7.3 million population by 1994 (Peterson et al.
(1999)) with heterogenous ISP treatment intensity across districts and years (figure 3.1).
This program was one of the largest and most intensive iodine supplementation programs in the world
(Peterson (2000)). The program was scheduled to begin in 1988 and planned to distribute iodized oil capsules
containing 400 mg of iodine amongst males and females aged 2 to 45 years and a dose of 200 mg for children
aged 12 to 23 months (Peterson et al. (1999)). However, the time line was not strictly followed. Four districts
received the supplementation in 1986 and only 10 had received it by 1988 while two districts did not receive
it until 1992. The coverage rate was never perfect in any district and the average coverage rate was 64% (see
table A4). The delays in the program start date, in all cases, were due to administrative issues arising from
the logistical challenges of distributing IOC throughout the district (Peterson (2000)). However, despite the
delays, according to a conservative estimate from Peterson et al. (1999), the program protected 12 million
individuals from iodine deficiency (ID).
The program was considered a success, with several previous evaluations finding reduction in visible
and total goiter rate (VGR, TGR) attributable to ISP (Peterson (2000)). In the early 1990s, a success of ISP
led to the Universal Salt Iodization (USI) program initiated by Tanzanian government. After the USI was
introduced, ISP was used to complement USI, focusing on districts not yet reached by the USI. Thus, during
this period, the absence of the ISP program in a district does not necessarily indicate that the individuals in
the district are unprotected from IDD.
3
It was estimated that about 5 million individuals lived with goiter, 450,000 with cretinoidism and 160,000 with cretinism. (Van der
Haar et al. (1988); recited from Peterson et al. (1999)).
53
3.2.2 Primary Education Development Program (PEDP)
Tanzania school system consists of seven years of primary school, four years of secondary school and two
years of upper secondary school. There are two national examinations in primary school - one at the end
of the 4th and 7th grade (primary school leaving exam: PSLE (Ministy of Education and Culture (1995)).
Students need to pass the examination at the end of grade 4 to progress to grade 5 and the PSLE to advance
to secondary school. Children are expected to enroll in primary school at the age of 7 and complete primary
school by the age of 13 (Ministy of Education and Culture (1995)). However, delays in enrollment, dropouts
and grade repetitions are common.
In January 2002, Tanzanian government launched the Primary Education Development Program (PEDP)
wherein tuition fees and other mandatory cash contributions to schools were abolished (Chikoyo (2010)).
The primary purpose of PEDP was to ensure the enrollment of all 7 to 12 years old by 2005. The net
enrollment rate in primary school in the year preceding the launch of the PEDP was less than 50%. The
program began by targeting those who were 7 to 8 years old in 2002, individuals born in 1993 or 1994.
While the coverage of the program was extended to 11 and 12 years old in 2004 (9 and 10 years old in
2002), the effort and impact for these children was substantially lower and delayed. As a result, PEDP
treatment status was dependent on the year of birth that those born in 1993 and 1994 were fully exposed to
PEDP while those born before 1993 were partially or never exposed to PEDP. Due to PEDP, net enrollment
rates went up significantly from 66% in 2001 to 97.3% in 2007.
The program worked towards bringing down the cost of primary education by abolishing all tuition
fees. Moreover, a $10 capitation grant was also introduced and controlled by school committees. This was
intended to cover some of the additional school-based costs. However, substantial indirect costs, such as an
expense for instructional materials, remained, the provision of which has not been sufficient to date.
3.3 Existing Literature
As we mention in the introduction, our contribution is two-fold: we re-estimate the impact of ISP and
examine the ‘dynamic complementarity’. In this section, we discuss existing literature with regard to ISP
impact and ‘dynamic complementarity’.
54
Field et al. (2009) analyze the impact of ISP on educational attainments of the children (baby in-utero
when they were exposed to ISP) of ISP treated pregnant mothers using the Tanzania Household Budget
Survey (THBS) 2000. They find that ISP had significant positive impacts on completed years of schooling
of treated children. According to Field et al. (2009), treated children completed 0.345 more years, and the
effect of ISP was bigger for girls (0.59 years) than boys (0.19 years). In contrast, using multiple information
from the Demographic and Health Survey (DHS) 1999, 2004, and 2010, the National Panel Survey (NPS)
2008, and the THBS 2000, Bengtsson et al. (2013) find that the estimated impact of ISP on educational
attainments is not consistent across datasets and often negative in sign. The impact is positive and significant
only for the THBS 2000 sample as shown in Field et al. (2009). They use a slightly different model of iodine
depletion over time to calculate the ISP treatment intensity than the one in Field et al. (2009) and find much
smaller magnitudes for the impact of ISP on educational attainments. They also explore the robustness
of their findings across different definitions of ISP treatment and across different criteria for selecting the
sample and find no evidence of significant ISP impact consistently. Based on this gap between two studies,
we intend to evaluate the ISP impact again with different data set and try to reconcile their findings.
The difference between the current study and Field et al. (2009) is that we take into consideration the
migration history of the sample. A child’s ISP treatment intensity was dependent on a pregnant mother’s
treatment status by construction and a pregnant mother’s treatment status was dependent on where she lived
when her baby remained in-utero. Field et al. (2009) used the current district of residence as the child’s
district of birth and calculate the ISP treatment intensity based on this information. This can be problematic
as both within and across district migration is high in Tanzania (Kudo (2015)). Particularly, women tend
to move more often than men (Kudo (2015)). The migration might be substantial even at young ages and
therefore, the district of residence might not be the child’s district of birth. Due to this concern, Bengtsson
et al. (2013) attempt to sign the bias by using information from NPS 2008 that has information on district of
birth and find that not accounting for selective migration seems to overestimate the effect of ISP treatment
for females. However, since the NPS 2008 results are much different from the results of other data sources
that do not take into account the migration, Bengtsson et al. (2013) could not draw any definite conclusion
about migration issue.
Like Bengtsson et al. (2013), we exploit the migration information from KHDS to examine the ISP
impacts after controlling for migration. The panel nature of KHDS allows us to track individuals across the
55
six waves in 1991-1994, 2004, and 2010, which will be described more in detail in the next section. Using
information on migration across waves, we minimize inaccurate assignment of treatment probabilities. We
use information from the wave 2010 to check if there was selective migration on the basis of ISP treatment.
Using the wave 2010, we also check if the impact of ISP treatment persisted or changed in the long run.
Our paper is also related to ‘dynamic complementarity’. As we mention in the introduction, Cunha and
Heckman (2007) construct a simple theoretical framework of ‘dynamic complementarity’. Based on their
initial work, Almond and Mazumder (2013) discussed empirical evidence of ‘dynamic complementarity’ and
emphasized the potential empirical challenge in identification strategy. More specifically, a major concern
in examining whether later investments are complementary to investments at early stage is the endogeneity
of such investments. Thus, due to the difficulties in identification the empirical evidence for this model
of complementarity is rather scarce. Most studies have approached this issue by jointly evaluating two
different exogenous interventions, public programs or interventions that affect individuals at different stages
in their lives. For example, Adhvaryu et al. (2014) examine the interaction between early life adverse rainfall
conditions and conditional cash transfer program (PROGRESA) in Mexico. They find that the conditional
cash transfer program enabled individuals, who were otherwise lagging behind due to adverse rainfall shocks
in early life, to catch up individuals who benefited from positive rainfall shocks. Malamud et al. (2016)
estimate the joint effects of access to abortion facilities at the time of conception and access to better school
later on in Romania. Although they find no significant interaction effect, they acknowledge that it does
provide evidence against dynamic complementarity as behavioral responses by individuals or their parents
between first intervention and second intervention could dampen the joint effects. Rossin-Slater and W¨ ust
(2016) use administrative data from Denmark to estimate the long term interactive effects of nursing home
program in infancy and high quality preschool program on educational attainments, income, and survival
rate. They find that added value of high quality preschool is reduced if an individual was exposed to the
health intervention in infancy.
56
3.4 Data and Empirical Strategy
3.4.1 Data
We study a sample representative of the population of Kagera region from Tanzania. Located in the north-
western corner of Tanzania, Kagera is one of Tanzania’s 30 administrative regions. Kagera is Tanzania’s
fifteenth largest region and accounts for more than three percent of the country’s area (CIA (2010)). During
1980s, Kagera suffered from high rates of IDD and four of its seven districts were targeted by the Iodine
Supplementation Program. We use the data from Kagera Health and Development Survey (KHDS). KHDS
was designed to provide data to understand the long-run wealth dynamics of households and individuals in
Kagera. The households were originally interviewed in four waves from 1991 to 1994. Follow-up surveys
were then carried out in 2004 and 2010. KHDS is one of the longest-running African panel data set with an
impressive tracking rate of around 90% (Beegle et al. (2006); De Weerdt et al. (2012)). We use the waves
1991-1994 and 2004 for information on individual’s educational attainments, primary school starting age,
parental investments in the child and a variety of covariates. Summary statistics are displayed in table 3.1.
To calculate iodine exposure intensity, we use the district-year coverage rates from Field et al. (2009).
We match this coverage rate for each year for each district with the corresponding observation from KHDS
using the year and name of the district information contained in KHDS. We follow Field et al. (2009)
in calculating the probability that the mother of a child received iodine supplementation while pregnant
with this child conditional on whether and when the iodine supplementation program was implemented in a
district. In that, we make the assumption that mothers, throughout their pregnancy, lived in the district where
they delivered their child. The details about the method followed are provided in the next section 3.4.2. We
restrict our sample to the cohorts born between 1991 and 1994. We do not include cohorts born after 1994
because nation-wide iodine supplementation (USI) began in late 1994. In addition, since PEDP (started in
2002) fully affected both 7 years old and 8 years old, cohorts born after 1992 were treated by PEDP and
those born in 1991 and 1992 were not treated. We do not include the cohorts born before 1991 to avoid more
serious recall bias in the waves 1991-1994 and to balance between PEDP treatment and control group across
cohorts. Consequently, while the variation in ISP treatment is at the level of cohorts and districts, the PEDP
treatment varies only across cohorts.
57
As mentioned in the previous section, the main reason we do not use information from other nationally
representative surveys like the Tanzanian Demography and Health Surveys (TDHS) or the Tanzania House-
hold and Budget Surveys (THBS) is that the relevant waves from these surveys do not have information on
the district of birth for individuals. Since internal migration across regions in Tanzania is common (Kudo
(2015)), it is important to know the district of birth of our sample to assign ISP treatment probability. Kagera
region is an ideal setting in this context because migration outside Kagera is relatively low and KHDS con-
tain migration information with high tracking rate. We restrict our attention to the sample who report not
having moved in the last ten years and report being a part of the household in preceding years. While
children born in 1991 and 1992 were born twelve and eleven years before the 2004 wave, respectively, the
probability that they moved in the first two years of their birth is relatively small.
3.4.2 Iodine Exposure Treatment
As described before, sufficient levels of iodine are most crucial in the first trimester. Therefore, the child of
an iodine deficient mother who received an iodized oil capsule in the first month of any year would not be
protected unless the child was born in the eighth month of that year or later. Following Peterson et al. (1999)
and Field et al. (2009), we assume that the timing of the distribution was uniform over the months of any
year that the district received the supplementation. We also maintain the assumption in Field et al. (2009)
that, conditional on the starting month, it took three months to complete the distribution of these capsules.
Therefore, for a district that received the supplementation program in the first month of the year t, children
born in the first seven months in that district were not protected by the supplementation program. Research
shows that the body stock of iodine depletes at a certain rate after every such iodine supplementation. To
account for this depletion, we use the method used in Field et al. (2009). For those born in the eighth month
or later, protection, therefore, depended on whether the program started early enough to have reached their
mothers in time (first trimester or earlier) and whether their mothers had retained adequate amounts of iodine
throughout their first trimester after accounting for the depletion of body iodine stocks with time.
The detailed table of probability of protection calculation is reproduced from Field et al. (2009) in the
appendix table A5. For instance, we present here the calculation for those born in the eighth, the ninth and
the tenth month of year t. For those born in month 8, probability of protection is equal to the probability that
the program started in January that year (equal to 1/12 using uniform timing assumption) and their mothers
58
were reached in that very month (equal to 1/3 using three-month diffusion time assumption.) Therefore, it is
equal to 1/36. For those born in the ninth month, the program reached their mother in the first trimester if it
started in January (1/12) and reached them by February (2/3) or if it started in February (1/12) and reached
them in February itself (1/3), therefore, the probability of protection conditional on treatment in that year is
1/12. For those born in October, the program reached their mothers in time if it started in January (1/12)
and reached the mother by March (1) or if it started in February (1/12) and reached them by March (2/3) or
if it started in March (1/12) and reached them in March itself (1/3). Therefore, the probability of protection
conditional on treatment in that year is 1/6. Given the assumption on the rate of depletion made in Field et al.
(2009), one that we maintain here, the stocks of iodine retained in the body would be above the required
levels for 24 months after the administration of the pill. Therefore, one does not need to adjust for depletion
for these months. Finally, this calculated probability is multiplied by the coverage rate in a particular district
in a particular year to get the final treatment probability.
3.4.3 Empirical Specification
We begin by examining the direct impact of the iodine exposure on educational attainments at the age of
10-13 in 2004. We follow Field et al. (2009) and Bengtsson et al. (2013), where treatment is considered
to vary exogenously at the district-birth year level after we control for several observables and apply some
fixed effects.
Y
idb
=+ID
idb
+
X
idb
+
b
+!
d
+
idb
(3.1)
whereY
idb
is the years of schooling completed by an individuali born in districtd in yearb by 2004.
It depends onID
idb
, the probability that individuali’s mother was treated by the ISP program in the first
trimester of her pregnancy. This treatment probability is calculated as explained in the previous section.
X
idb
is a vector of covariates that include a dummy each for whether the mother and father have some
education or not, a dummy for gender of the individual, a dummy each for whether the individual belongs
59
to the majority tribe or religion, and total land ownership of the household to which individuali belongs.
4 5
Standard errors are clustered at the district-year level in order to allow for arbitrary correlation in the error
terms within each cohort in a district.
6
Our specification is closer to that used in Bengtsson et al. (2013) and differs from Field et al. (2009)
in that it is more parsimonious with respect to the control variables used. Since the treatment occurred
before birth, many of the potential controls run the risk of being impacted by the treatment. For example,
as pointed out by Becker and Tomes (1986) and empirically verified by Rosenzweig and Wolpin (1980),
fertility decisions are endogenous to the quality of children, a dimension that ISP treatment might have
affected. Therefore, we exclude controls for birth order, number of children, distance to secondary school
and health clinic, food security measures, home ownership, housing quality, as used in Field et al. (2009),
and a dummy for household’s urban residence, as used in Bengtsson et al. (2013).
Unlike Field et al. (2009), we do not use household fixed effect specification in our main analysis.
We believe that households where mothers gave birth twice within a span of four years are different than
those with one birth during the period and excluding household with only one child born during this period
will lead to substantial selection biases.
7 8
While using the month of birth information for the assignment of
treatment probability would lead to more accurate assignment, we do not have the month of birth information
for a fairly large number of individuals in our sample. Therefore, we follow Field et al. (2009) in our main
specification and assign treatment probabilities on the basis of month of birth. In section A.4, we check
the robustness of our results to the assignment of treatment probability on the basis of month of birth for a
smaller sample of individuals for whom we have the month of birth.
9
4
Land in rural areas were regulated by the Tanzanian government under the Village Land Act beginning in 1999. To check for
robustness, we replace the land ownership variable with the value of livestock owned in an alternative specification. The results, not
presented here, remain very similar.
5
Our sample has a very small percentage of HIV positive individuals. Out of 1784 individuals for whom we clinical diagnosis,
only four were diagnosed as HIV positive. Out of the 4,240 individuals for whom we have self diagnosis information, 10 diagnosed
themselves to be HIV positive. Given the small percentage of HIV positive individuals, we do not include HIV status as a control.
6
Since the number of districts is just seven, we prefer the district-year level clustering. Since the number of clusters is relatively
small, we repeat our analysis with cgm wildboot cluster method to correct for standard errors (Cameron et al. (2008). The results
(available on request) remain unchanged.
7
It is important to mention here that Field et al. (2009), too, focus on district fixed effect results to draw their conclusions.
8
We do, however, report the results for household fixed effect analysis in the table 3.7. While the results are not significant, the
selection is evident from the fact that the sibling sample is less than one third of our total sample. The sign of the ISP effects also flips
in this specification.
9
Bengtsson et al. (2013), too, use treatment assignment based on the year of birth in most of their analysis. Similar to our findings,
their results remain qualitatively unchanged when they assign treatment probabilities based on the month of birth for the sub-sample
for which they have this information.
60
Since the treatment status of the PEDP program is based on the year of birth, we cannot use birth
fixed effects in the specification that evaluates the impact of PEDP or the joint impact of ISP and PEDP.
Instead, we replace fixed effects in year of birth by a quadratic term in age. To show that this quadratic term
approximates the year of birth fixed effects closely, we re-estimate the equation (3.1) with the year of birth
fixed effect replaced by a quadratic in age (table 3.2). We, then, look at the combined impact of ISP and
PEDP on completed years of schooling using the specification:
Y
idb
=+
1
ID
idb
+
2
P
idb
+
3
ID
idb
P
idb
+
X
idb
+age+age
2
+!
d
+
idb
(3.2)
For individuali, living in districtd and born in yearb,ID
idb
represents the probability that individuali’s
mother received iodine supplementation during the first trimester of her pregnancy.P
idb
indicates individual
i’s exposure to PEDP and takes value ‘1’ if the individual was born in 1993 or 1994, ‘0’ otherwise. Our
specification includes district fixed effects (!
d
), a quadratic in age, a dummy each for whether the mother
and father have some education or not, a dummy for gender of the individual, a dummy each for whether
the individual belongs to the majority tribe or religion, and total land ownership of the household to which
individuali belongs.
1
and
2
represent the independent impact of ISP and PEDP on the schooling attain-
ments, respectively. Coefficient
3
is a measure of heterogeneity in the impact of PEDP by ISP exposure
status.
To provide suggestive evidence for the mechanism we propose, we begin by showing that the ISP and
PEDP treatment status and their interaction predict the primary school starting age and individual’s involve-
ment in household work or work on family farm in a manner that is consistent with the impact of these
three variables on years of schooling. For this we use a specification that is similar to the equation (3.2),
except now the outcome variable is the age at which the individuals start primary school, probability that the
individual worked on-farm or at home in the last week, or the number of hours worked on-farm or at home
in the last one week.
We estimate the rate at which children treated by one or both of these programs convert years at school
into completed years of schooling by taking a ratio of estimated impact of these two treatments and their
61
interaction on completed years of schooling and primary school starting age. We interpret these rates mea-
sures of the ‘dynamic complementarity’. A higher value of this ratio for those exposed to ISP compared to
those not exposed to ISP would imply that the former group makes better use of each year spent in school.
We will discuss this in the results section.
3.5 Results
3.5.1 Impact of ISP on Educational Attainment and Joint Impacts
Table 3.2 presents the impact of ISP exposure on educational attainments in 2004. The columns differ in
the controls used in the specification. For example, columns (1), (3), (5), and (7) include year of birth
fixed effects to account for time varying unobservable factors that might have impacted schooling levels in
those years. Columns (2), (4), (6), and (8) replace the year of birth fixed effect with a quadratic term in
age. It is clear from the comparison of the coefficient across columns that a quadratic term in age closely
approximates year of birth fixed effects in our analysis. Both the ISP coefficient magnitude and the R-
squared fit of the model remains virtually unchanged. Controlling for tribe, religion and total land makes
almost no difference to the estimated impact of ISP on grade attainment. When discussing the results, we
will prefer the coefficient estimates from the specification with the full set of controls and quadratic in age,
similar to the one used in column (8). According to the estimates, ISP exposure is associated with 0.70 fewer
years of completed schooling. Conditional on non-zero probability of exposure, the average probability of
exposure to ISP is 0.31. Therefore, those exposed to ISP, on an average, had completed 0.21 (0.70 * 0.31)
fewer years of completed schooling.
At a first glance, the negative impact of ISP on grade attained is puzzling. There is no a priori reason
to expect a negative impact on grade attainment of a supplementation which is expected to improve the
cognition of those exposed. It also seems to contradict the results from Field et al. (2009) that those exposed
to ISP had completed more schooling. However, once we examine the behavioral responses to ISP exposure,
the negative association between ISP exposure and grade attainment is no longer a puzzle.
In table 3.3, we examine the joint impact of the two programs on years of schooling completed by the
time of the survey in 2004. The coefficient estimates in column (3) suggest that for those who were not
exposed to PEDP the impact of ISP remained comparable to the estimates from table 3.2. However, there
62
was a significant level of heterogeneity in the impact of PEDP by the ISP exposure status. Those not exposed
to ISP but exposed to PEDP had completed 0.18 extra years of education by the time of the survey compared
to those not exposed to either of the two programs. However, those exposed to ISP and PEDP had, on
average, completed 0.19 fewer years of schooling (-0.71 * 0.31 + 0.18 - 0.47 * 0.31 = -0.19). They were
comparable to those exposed to only ISP and not PEDP who were lagging behind those who did not get
exposed to either of the two programs by 0.22 years (-0.71 * 0.31).
The negative interaction effect is equally perplexing. Why will a reduction in the cost of schooling hurt
those who, most likely, have better cognition? One advantage of using KHDS for our analysis is that it
contains information on primary school starting age. By examining the primary school starting age, which
is a choice variable for individuals or their parents, we can investigate if there were any behavioral responses
to the policies. Table 3.4 reports the impact of the two policies and their interaction on primary school
starting age. The estimated coefficients seem to mirror the impact of the two policies and their interaction on
completed schooling. Those exposed to PEDP alone start school at a younger age than those not exposed to
either of the two programs. Those exposed to ISP only or both the programs enter school later than those in
the omitted category. On comparing table 3.3 and 3.4, it is clear that the association of ISP exposure, PEDP
treatment, and their interaction with educational attainemtns is, at least, partly explained by changes in
primary school starting age in response to these treatments. In the next section 3.5.2, we provide suggestive
evidence to further explain why such a response might have arisen.
3.5.2 Delay in Starting Primary School
Why might those exposed to ISP delay start of primary school more than those not exposed? Late entry into
primary school is very common in Tanzania (Bommier and Lambert (2000); Burke and Beegle (2004)) and
elsewhere in Africa (Glewwe and Jacoby (1995); De Vreyer et al. (1998)). Several hypotheses have been
proposed to explain this delay in enrollment - existence of liquidity constraints (Jacoby (1994)), malnutrition
problems (Glewwe and Jacoby (1995)), considering children too young to be in school (Burke and Beegle
(2004)), and pre-school training or child labor (De Vreyer et al. (1998)). There is no reason to believe
that ISP exposure of child was correlated with credit constraints that her family faced, especially because
ISP exposure depended on the timing of first trimester of pregnancy vis-a-vis supplementation and not
63
on supplementation alone. Therefore, we focus on the next two most important hypothesis - malnutrition
problem and pre-school labor force.
Delay due to Malnutrition
If those exposed to ISP delay starting school because they are malnourished, we might expect it to be
reflected in their height-for-age. Low height-for-age is an indicator of stunted growth reflecting a process
of failure to reach linear growth potential and is often associated with increased risk of early exposure to
adverse conditions such as illness and/or inappropriate feeding practices. Column (1) of table 3.6 presents
the association between an individual’s ISP exposure status and height for age. Those exposed to ISP,
indeed, are shorter in 2004. However, it is not clear why those exposed to ISP had worse growth. In-utero
iodine supplementation, especially in such low doses, has no adverse impacts on physical growth (Isa et al.
(2000)). Most individuals from the sample were interviewed at least once during the first four waves of
KHDS between 1991 and 1994 and height measurements were also taken. Unfortunately, the number of
individuals from each wave that we have height information on is small. However, since the selection for
being interviewed during any of these years was unrelated to ISP exposure status, examining association
between ISP exposure status and height during these waves might still be informative. Columns (2)-(4)
present the association between ISP and height for age during these waves. Even though the standard errors
are too large to interpret these coefficients, the height for age during the early years for individuals exposed
to ISP seems to be higher than for those not exposed while the initial height advantage of those exposed to
ISP was reversed in 2004.
A possible explanation for the reverse is that parents of those not exposed to ISP responded, either to
their exposure status or to their lower cognition, by making compensatory investment in them. However,
previous studies from developing countries have mostly found that parental response in such scenarios is
often to reinforce the advantage that their children might have (Rosenzweig and Schultz (1982), Li et al.
(2010), Adhvaryu and Nyshadham (2014)). Most of these studies use a sibling fixed effect specification to
check for reinforcement or compensation within families. In order to check parental investments, we also
use sibling fixed effect specifications. However, it needs caution to interpret the sibling fixed effect results
in our study because our main sample consists of children born in between 1991 and 1994 while households
where mothers gave birth twice or more within a span of four years are likely to be different than those with
64
one birth during the period. Therefore, excluding households with only one child born during this period
will lead to substantial selection biases.
The results are presented in table 3.7. Column (1) and (3) look at the association of ISP exposure status
with years of completed schooling by 2004 and height for age in 2004, respectively. Since, here, we are
interested in the impact of ISP alone and including those born earlier than 1991 could, potentially, reduce
some selection bias, we include sample of all those born in between 1989 and 1994 in columns (2) and (4).
10
However, the results do not indicate any compensatory response within the household. Most of estimated
coefficients, even though not significant, are positive, consistent with reinforcement and not compensation.
Accordingly, we do not draw any conclusion about malnutrition related to parental investments.
11
In ad-
dition, this suggests that the negative impact of ISP exposure on education and height are identified from
individuals with differential exposure to ISP born to households with the single birth during this period (table
A8).
Delay due to Child Labor
Next, we turn to the child labor hypothesis. According to the 2013 US Department of Labor, over 25% of
the Tanzanian children aged 5-14 were engaged in the worst forms of child labor in 2011. A little over 20%
of the children aged 7-14 were combining work and school. Using information from the wave 2000 of the
Tanzania Household and Budget Survey, Kondylis and Manacorda (2012) find that over 60% of children
aged 7-14 engaged in some form of productive activity and around 40% of children combine work and
school. Children worked on the family farm or did household chores. The average number of hours working
every week was a little over 25 and around 20% of those who did not attend school reported the reason of
absence as work or perceived uselessness of schools. Beegle et al. (2006) use information from the wave
2004 of KHDS and find that children aged 7-15 were found to have worked a little over 18 hours in the week
prior to their interview. Burke and Beegle (2004) find that 10-15 years old were working close to 9 hours on
farming activities and between 11-15 hours on household chores.
These findings are consistent with 2000-2001 integrated labor force and child survey by the Tanzanian
Ministry of Labor, Youth Development, and Sports under International Labor Organization’s International
Program on Elimination of Child Labor. According to the report, of the total number of children aged
10
Kagera first received the program in one of its district in 1989.
11
We also check parental investments using vaccination information but do not find significant association (table A7).
65
5-17, 39.6 % were involved in economic activities and 47.8% were engaged in housekeeping activities.
Amongst those engaged in economic activities, more than three quarter of them (78.8%) worked as unpaid
family members in their family farm or shamba and another 17.99% work as unpaid family members in non-
agricultural establishment. An estimated 34% of the total working children worked for more than 4 hours
per day or 30 hours per week. Beegle et al. (2008) use crop and rainfall shocks as instrumental variables for
child labor and find that child labor has negative effects on completed years of schooling. One of the ways
in which child labor affects educational attainment of children in Tanzania is through delayed enrollment.
Even though children in Tanzania are expected to enroll in primary school at the age of seven, enrollment is
almost always delayed by two, three, or even four years (Bommier and Lambert (2000); Burke and Beegle
(2004)).
Another explanation is related to the returns to schooling. The returns to schooling in Tanzania are
lower than other countries in the region (Knight and Sabot (1990); Mason and Khandker (1997)). Since
the country’s agricultural practices mostly use traditional production methods, the returns to education in
agriculture are low (Mason and Khandker (1997)). Our findings are, therefore, not very surprising. Burke
and Beegle (2004) find that children in the Kagera region were not attending school due to household demand
of child labor and high opportunity cost of schooling. De Vreyer et al. (1998), Bommier and Lambert (2000)
and Burke and Beegle (2004) argue that the main reason for delay in starting school in African countries,
and in Tanzania in particular, is the high opportunity cost of going to school. De Vreyer et al. (1998)
present a model where a household’s decision is similar to a portfolio choice among three assets - physical
assets, ‘general’ human capital accumulation for the children through schooling, and ‘specific’ human capital
accumulation for the children through participation in family economic activities. Bommier and Lambert
(2000) use information from the Human Resource Development Survey conducted by the World Bank,
the Dar-es-Salaam University, and the Tanzanian government in 1992-1993 on 5000 households to test the
model. They show that the parents send their children to school later and for a smaller period of time since
Tanzania had high returns from accumulation of the ‘specific’ human capital.
From the previous studies, ISP exposure is likely to make the exposed children smarter. This, in turn,
might have increased their opportunity cost of schooling more than those not exposed. As a result, the ISP
treated children might have chosen to start school later. Moreover, if they were aware that they were better
at converting years in school into completed years of schooling, this might have incentivized them further
66
to delay schooling. This would imply that those exposed to ISP were working more often than those not
exposed before they started school.
In 2004, at the time of the survey, all the individuals from the sample were in school. It would have
been ideal if we had information on the working status of these individuals before they started school.
Unfortunately, KHDS collected information on involvement in market and non-market labor activities only
for the week preceding the interview date. We assume that the number of hours worked in the week preceding
the survey is correlated with the number of hours worked every week in years preceding their enrollment
in primary school. Using information from THBS 2000, Kondylis and Manacorda (2012) find that enrolled
children from all over Tanzania spent close to forty hours in school every week. They report that hours
of work among children in school was approximately half that of children out of school. The alternative
assumption that ISP treated children who enrolled later worked less than those who were not exposed and
enrolled at younger age is less plausible.
Most children generally work at home and on the family farm. These activities include working in the
fields or tending livestock (categorized as farm activities in KHDS) or collecting water, fetching firewood,
cleaning the house, preparing meals, and time spent caring for other children or sick household members
(categorized as household chores in KHDS). Less than 0.5% of the children in our sample were engaged
in wage work outside the family farm. Therefore, we focus on work on family farm and household chores
only. Table 3.8 presents the association between program exposures and number of hours worked in different
activities preceding the survey. ISP exposure seems to increase the number of hours worked on both family
farms and unpaid family chores. The coefficients, even though insignificant for the activities separately, are
in the right direction and large in comparison to the average number of hours worked in these activities by
the individuals in the sample. Moreover, when we combine the activities, those exposed to ISP are working
over five hours extra more than those not exposed. The impact of PEDP, even though it is positive, is small
and insignificant. However, the coefficient for the interaction is large in columns (2) and (3), and significant
for hours spent on household chores. The results suggest that pre-school work experience can be one of the
reasons behind delayed enrollment of those exposed to ISP.
While we find the suggestive evidence in favor of more pre-school work experience for those exposed to
ISP as possible reasons behind delayed enrollment, the evidence is rather weak. Therefore, we are cautious
about claiming child labor as the only strongest mechanism. What is clear, however, is that ISP invoked
67
different behavioral responses from different subgroups and one needs to take those responses into consid-
eration when evaluating ISP and its interaction with PEDP.
3.5.3 Dynamic Complementarity
In this sub-section, we propose a new instrument to measure ‘dynamic complementarity’. To be able to
interpret the interaction coefficient as evidence in favor of or against dynamic complementarity, one has to
make the assumption that in the absence of dynamic complementarity, the impact of the second policy on
those who were exposed to the first policy must be equal to the independent impact of the second policy on
those who were not exposed to the first policy. However, since ISP changes the primary start school age,
the cost and benefit from PEDP for those exposed to ISP might no longer be same as for those who were
not exposed to ISP. The difference in cost and benefits from PEDP might, therefore, also invoke different
behavioral response. However, since the coefficient estimates of the impacts of these two programs and
their interaction on educational attainment and primary school starting age from tables 3.3 and 3.4 are still
unbiased, a ratio of coefficient estimates for each of these policies and their interaction will be an indicator
of how productive a particular subgroup is in school. That is.
@(yearsof schooling)
@(schoolstartingage)
=
@(yearsof schooling)
@(exposuretoprogramX)
[
@(primaryschoolstartingage)
@(exposuretoprogramX)
]
1
(3.3)
whereX2fISP;PEDP;ISPPEDPg. We compute the conversion rate for the subgroups in table
3.5. If those exposed to PEDP but not to ISP started school one year earlier, they would have have attained
0.38 extra years of completed schooling. In comparison, those exposed to ISP alone would have attained
0.96 extra years of schooling. This suggests that ISP exposure makes individuals more productive at school.
If we consider years in school as an input in the production function of human capital, the productivity of
this input is higher for those who benefit from the in-utero iodine supplementation. This is an evidence of
‘dynamic complementarity’. In conclusion, instead of using an interaction term of joint effect, we propose a
new measure of ‘dynamic complementarity’ to explicitly show how productive the second policy is when it
is interacted with the first policy.
68
3.6 Conclusion
There is now a broad consensus amongst economists, demographers, and sociologists that the diffusion of
modern economic growth to the developing regions requires human capital accumulation by the population
in these regions (Counts (1931); Inkeles (1969); North (1973); Davis et al. (1971); Rosenberg et al. (1986);
Easterlin (1981); Easterlin (2009)). A higher level of human capital is desirable in its own right (Pigou
(1952); Adelman (1975); Grant (1978); Grant (1978); Streeten et al. (1981); Sen (1984)). However, poor
access to information and quality infrastructure, low levels of incomes, and imperfect credit markets in
these regions limit the possibilities of private investment in human capital accumulation. State run policies,
therefore, are of extreme importance in ensuring higher levels of human capital (Easterlin (1981)). Given
the limited state budget, the decision of whether or not to roll out a particular program depends a lot on the
cost benefit analysis of the program. Traditionally, the cost benefit analyses of such development programs
is based on the evaluation of the single program as if it was implemented in isolation. However, if two
independent programs interact in important ways, a partial equilibrium analysis might greatly understate the
net benefits of such programs. In such scenarios, it becomes essential to jointly evaluate the impact of the
two (or more) programs, allowing for possible complementarity between the programs.
In this chapter, we evaluate the Iodine Supplementation Program and Primary Education Development
Program in Tanzania. We find that ISP treatment was associated with lower schooling achievements for the
exposed children in 2004. The effect operated entirely through delays in enrollment. We provide suggestive
evidence that this behavioral response of delaying enrollment was because those exposed to ISP spent more
time in working on the family farm or in the house. This, we conjecture, might have been due to their
improved cognition that made them better at these jobs. More importantly, we find that those exposed to ISP
were better at converting years in school into completed years of schooling due to PEDP, which is a sign of
‘dynamic complementarity’ using our new proposed measure of ‘dynamic complementarity’.
The result that government policies interact in important ways might explain why the short run impacts
of many programs dissipate in the long run. The results also underscore the need to raise the dimension-
ality of the policy spaces to be considered. However, it is impossible to evaluate the combined effect of
different policy exposures and determine how they interact. Perhaps better combinations of theoretical, non-
experimental, quasi-experimental and experimental methods need to be developed to handle the situation.
69
Figure 3.1: Iodine Supplementation Program in Tanzania (from Field et al. (2009))
70
Table 3.1: Summary Statistics
Outcomes Control Treatment
Ages10-11
Mean SD Mean SD
Years of Schooling 1.87 1.00 1.44 0.96
Primary school start age 7.98 1.02 8.22 1.15
School Progression 0.82 0.30 0.79 0.40
HAZ in 2004 130.50 8.99 127.77 7.61
Proportionwith
Vaccination card 0.92 0.85
Tb vaccination 1.00 0.97
Measles vaccination 1.00 0.92
Tetanus vaccination 0.46 0.39
Polio vaccination 0.52 0.55
Ages12-13
Mean SD Mean SD
Years of Schooling 3.09 1.38 2.78 1.42
Primary school start age 8.57 1.45 9.01 1.46
School Progression 0.79 0.25 0.81 0.24
HAZ in 2004 141.21 8.47 139.64 8.44
Proportionwith
Vaccination card 0.95 0.99
Tb vaccination 0.99 0.99
Measles vaccination 0.94 0.97
Tetanus vaccination 0.80 0.88
Polio vaccination 0.84 0.86
Independent variables Control Treatment
Ages10-11
Mean SD Mean SD
Protection due to ISP 0 0 14.26 17.17
Age 10.34 0.47 10.13 0.34
Mother has any education 0.95 0.21 0.92 0.27
Father has any education 0.92 0.27 0.92 0.27
Household land per capita 0.48 0.53 0.55 0.47
Proportion
Sex = Male 0.54 0.48
Tribe = Mhaya 0.93 0.37
Religion = Catholic 0.65 0.53
N 133 185
Ages12-13
Mean SD Mean SD
Protection due to ISP 0 0 70.05 27.72
Age 12.57 0.50 12.55 0.50
Mother has any education 0.97 0.17 0.89 0.32
Father has any education 0.95 0.21 0.88 0.32
Household land per capita 0.56 0.56 0.80 0.66
Proportion
Sex = Male 0.47 0.48
Tribe = Mhaya 0.94 0.01
Religion = Catholic 0.65 0.50
N 161 87
71
Table 3.2: Impact of Iodine Supplementation Program on completed years of schooling
(1) (2) (3) (4) (5) (6) (7) (8)
Years of education
Iodine Supplementation Program -0.70*** -0.70*** -0.69*** -0.70*** -0.69*** -0.70*** -0.69*** -0.70***
(0.12) (0.13) (0.12) (0.13) (0.12) (0.12) (0.12) (0.12)
Age fixed effect YES NO YES NO YES NO YES NO
Quadratic in age NO YES NO YES NO YES NO YES
Religion dummy NO NO YES YES NO NO YES YES
Tribe dummy NO NO NO NO YES YES YES YES
Land ownership control YES YES YES YES YES YES YES YES
Mean of dependent variable 2.20 2.20 2.20 2.20 2.20 2.20 2.20 2.20
Mean ISP treatment probability 0.31 0.31 0.31 0.31 0.31 0.31 0.31 0.31
Observations 518 518 518 518 518 518 518 518
R-squared 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at district-year of birth
groups. Other controls include a dummy each for whether the mother and father have some education or not, a dummy for gender of the
individual, and district fixed effects.
Table 3.3: Impact of ISP and PEDP on completed years of schooling
(1) (2) (3)
Years of education
Iodine Supplementation Program (ISP) -0.70*** -0.71***
(0.12) (0.11)
Primary Education Development Program (PEDP) 0.13 0.18**
(0.15) (0.07)
ISP * PEDP -0.47*
(0.26)
Mean of dependent variable 2.20 2.20 2.20
Mean ISP exposure probability 0.31 0.31
Observations 518 518 518
R-squared 0.36 0.35 0.36
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at district-year
of birth groups. Controls include a quadratic in age, a dummy each for whether the mother and father have some education or not, a
dummy for gender of the individual, a dummy each for whether the individual belongs to the majority tribe or religion, controls for
total land holdings of the household, and district fixed effects.
72
Table 3.4: Impact of ISP and PEDP on primary school starting age
(1) (2) (3)
Primary school starting age
Iodine Supplementation Program (ISP) 0.77*** 0.74***
(0.16) (0.15)
Primary Education Development Program (PEDP) -0.47** -0.47**
(0.19) (0.18)
ISP * PEDP 0.26
(0.21)
Mean of dependent variable 2.20 2.20 2.20
Mean ISP exposure probability 0.31 0.31
Observations 518 518 518
R-squared 0.17 0.16 0.17
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at district-year
of birth groups. Controls include a quadratic in age, a dummy each for whether the mother and father have some education or not, a
dummy for gender of the individual, a dummy each for whether the individual belongs to the majority tribe or religion, controls for
total land holdings of the household, and district fixed effects.
Table 3.5: Conversion of an additional year into additional years of schooling
Treatment School starting age Years of schooling
@(yearsofschooling)
@(schoolstartingage)
PEDP only 0:47 0:18 0:38
ISP only 0:74 0:71 0:96
ISP * PEDP 0:26 0:47 1:8
Notes:
@(yearsof schooling)
@(schoolstartingage)
=
@(yearsof schooling)
@(exposuretoprogramX)
@(exposuretoprogramX)
@(primaryschoolstartingage)
,
whereX2fISP;PEDP;ISPPEDPg
73
Table 3.6: Impact of ISP on height of the child (Height-for-age)
(1) (2) (3) (4) (5)
Height for age Z-score in
2004 1994 1993 1992 1991
ISP -0.46** -2.34 3.33** 1.79 3.80
(0.21) (1.45) (1.40) (1.51) (2.48)
Observations 501 102 118 128 145
R-squared 0.04 0.09 0.08 0.17 0.13
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are
clustered at district-year of birth groups. Controls a dummy each for whether the mother and father have some
education or not, a dummy each for whether the individual belongs to the majority tribe or religion, controls
for total land holdings of the household, and district fixed effects. We used the WHO Child Growth Charts and
WHO Reference 2007 Charts for our height for age analysis.
Table 3.7: Within household impacts of ISP
(1) (2) (3) (4)
Years of education Height for age in 2004
ISP 0.24 0.22 0.78 -0.39
(0.59) (0.46) (0.87) (0.47)
Observations 132 335 119 298
R-squared 0.86 0.86 0.66 0.67
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at the level of the
household. Controls include a dummy each for whether the mother and father have some education or not, a dummy for gender of the
individual, a dummy each for whether the individual belongs to the majority tribe or religion, controls for total land holdings of the household,
age and district fixed effects.
74
Table 3.8: Impact of ISP and PEDP on hours worked
(1) (2) (3)
In the last week, how many hours did you work [...]?
on the family farm in unpaid work in total
ISP 1.97 2.07 5.43***
(1.46) (1.75) (1.93)
PEDP 1.97 0.63 1.72
(1.18) (1.20) (1.08)
ISP * PEDP -1.85 10.33* 7.62
(3.07) (5.61) (8.69)
Mean of dependent variable 4.51 5.49 10.18
Mean ISP treatment probability 0.32 0.32 0.32
Observations 540 540 540
R-squared 0.04 0.05 0.04
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at district-year
of birth groups. Controls include a quadratic in age, a dummy each for whether the mother and father have some education or not, a
dummy for gender of the individual, a dummy each for whether the individual belongs to the majority tribe or religion, controls for
total land holdings of the household, and district fixed effects.
75
Chapter 4
Revisiting the Effect of Retirement on Cognition:
Heterogeneity and Endowment
4.1 Introdcution
The effect of retirement on cognition has drawn significant research interests from economists in recent
years. Especially with ongoing policy discussion on public pension reforms and the increasing burden
of dementia, it is indisputably an important research question with significant policy implications. The
underlying mechanisms behind this association are, however, not fully understood, although the primary
explanation is that cognition is better maintained through engagement in cognitively stimulating activities,
and a job provides such opportunities (Salthouse (2006); Stern (2012)). The idea that lifestyle can affect both
the level of cognitive function and the rate of age-related change in cognitive function has appealed broadly,
inspiring cognitive training to improve cognitive abilities of older adults (Smith et al. (2009); Rebok et al.
(2014); Ngandu et al. (2015)) and contributing to an ever-expanding $1.3 billion market of brain-exercise
products (Alsever (2017)).
Since the seminal paper of Rohwedder and Willis (2010), several economists have studied the relation-
ship between retirement and cognition. Building on this growing literature, our paper makes two important
contributions. First, we explicitly consider cognitive demands of jobs. Coe et al. (2012) investigated po-
tential hetereogeneity of retirement effect between white-collar and blue-collar workers, and more recently,
Mazzonna and Peracchi (2017) found the differential effects of retirement by physical demands of jobs.
While both of these investigations make important contributions by recognizing the heterogeniety of jobs,
they do not directly examine cognitive demands of jobs. As the primary explanation for potential adverse
effect of retirement is that cognition is better maintained through mental activities (Salthouse (2006)), by
NOTE: This chapter is coauthored by Dr. Jinkook Lee (USC, CESR) and Dr. Erik Meijer (USC, CESR).
76
investigating the cognitive demands of the job one retires from we can directly test this hypothesized rela-
tionship.
Specifically, taking advantage of the information available from the Occupation Information Network
(O*NET) database, we examine the hypothesis that the adverse effect of retirement on cognition is larger
when retiring from a more cognitively demanding job than from a less cognitively demanding job. Fisher
et al. (2014) found some suggestive evidence along these lines, but they did not account for the endogeneity
of retirement. As poor cognition can contribute to selection into retirement, reverse causation is an important
issue to consider, and in this paper, we address this issue by taking an instrumental variables approach that
is more consistent with the economic literature. By employing different sets of instruments, we further
examine the sensitivity of our findings to the choice of instruments.
Because individuals are not randomly assigned to occupations, they may differ in characteristics other
than job demands for mental activity (Salthouse (2006)). Particularly, the differences in innate cognitive
ability or desire for cognitive stimulation may lead to occupations with different levels of cognitive demands.
Salthouse (2006) evaluated the validity of the mental exercise hypothesis and explained the lack of empirical
evidence supporting this hypothesis by selection bias: individuals with higher levels of cognitive functioning
in earlier life pursue more education and choose occupations with higher levels of mental demands. This
explanation is called ‘preserved differentiation’, referring to the notion that those who are mentally active are
those who have higher levels of cognitive function throughout their lives (Fisher et al. (2014)), in contrast to
‘differential preservation’, which mental exercise posits (i.e., how much cognitive functioning is preserved
depends on one’s mental activity). We shed a light on these alternative hypotheses by examining the causal
effect of retirement from a cognitively or physically demanding job, while controlling for innate cognitive
ability, using the genetic risk score of cognition (Davies et al. (2015)).
Our main findings contribute to improving the understanding of the effect of retirement on cognition.
Using an instrumental variable approach, we demonstrate the adverse effect of retirement on cognition, and
that this effect is heterogeneous across types of occupation. Furthermore, supporting recent findings by
Mazzonna and Peracchi (2017), we find evidence of a positive effect of retirement for those retiring from
physically demanding jobs. However, we do not find evidence for the mental exercise hypothesis that retiring
from cognitively demanding jobs has a negative effect on cognition. Instead, we find evidence that cognitive
77
demand of the job is significantly associated with the genetic endowment for education, which is closely tied
to the genetic endowment for cognition.
The remainder of the paper is organized as follows. In section 4.2, we present the most relevant back-
ground literature. We explain our analytic strategy in section 4.3 and the data we use in section 4.4. We
discuss the empirical results in section 4.5. We then conclude in section 4.6.
4.2 Existing Literature
The mental exercise hypothesis was posited by the earliest researchers in cognitive psychology (e.g., Foster
and Taylor (1920)), and the view that keeping mentally active will maintain one’s level of cognitive func-
tioning, and possibly even prevent cognitive decline and the onset of dementia is pervasive and provides the
main conceptual argument for the adverse effect of retirement on cognition (Rohwedder and Willis (2010)).
Despite frequent assertions of the mental exercise hypothesis and its intuitive plausibility, empirical evidence
has been largely descriptive. In a thorough review of the mental exercise literature, Salthouse (2006) con-
cluded that the empirical evidence fell far short of supporting this hypothesis. The positive relation between
mental activity and cognitive function would also be obtained if one’s prior and current cognitive ability
determine the amount of mental activity one engages in. One way to distinguish between the two alternative
hypotheses involves examining the relation between retirement and cognitive performance as a function of
the amount of mental activity the job demands. That is, to the amount of cognitive demands of job is ex-
pected to moderate the relations between retirement and cognitive function. This prediction is referred as
the ‘differential preservation hypothesis’ because the degree to which cognitive performance is preserved is
postulated to differ according to level of mental activity (Salthouse et al. (1990); Salthouse (2006)). The
effects of retirement can be postulated to have more negative effects when one retires from more mentally
stimulating jobs than less mentally stimulating jobs. We also consider alternative hypothesis that people who
have more cognitively demanding jobs are likely to have had high levels of cognitive function throughout
their lives, and this hypothesis is termed as the ‘preserved differentiation’, as the differences in cognitive
performance are preserved across all of adulthood (Salthouse et al. (1990); Salthouse (2006)). The key
differences between these two hypotheses is whether mental activity protects cognitive ability.
78
The possibility of a connection between life experience and cognitive aging and the prevalence of demen-
tia has also been widely discussed in the epidemiological literature (Stern (2012)). Neuroimaging studies
have found individual differences in the susceptibility to age-related brain changes or Alzheimer’s disease-
related pathology: some people can tolerate more of these changes than others and still maintain cognitive
function. From such observations, the concept of cognitive reserve has emerged. According to the cogni-
tive reserve hypothesis, innate endowed intelligence and cognitively stimulating activities supply cognitive
reserve, which enables individuals to use brain neurons more efficiently and therefore to optimize cognitive
performance in the face of brain damage (Stern (2002)). While epidemiologic studies can only describe
correlations and do not directly test causation, they have motivated further work in this area, particularly
functional imaging studies to understand the neural basis of cognitive reserve (Stern (2012)). The cognitive
reserve hypothesis, in fact, presents more similarities to human capital theory than the mental exercise hy-
pothesis. Cognitive reserve or human capital determines cognitive functioning, and engagement in cognitive
activities will contribute to accumulation or maintenance of cognitive reserve or human capital. Both the
cognitive reserve hypothesis and human capital theory also recognize innate individual differences, mainly
from genetic endowment. In this study, we control for genetic endowment, using the genetic risk score for
general cognition and education (Davies et al. (2015)).
Starting with Rohwedder and Willis (2010), several economists have studied the effect of retirement on
cognition, reporting some mixed findings. Conceptually based on mental exercise hypothesis, Rohwedder
and Willis (2010) investigated the adverse effect of retirement on cognition based on a cross-sectional,
cross-country analysis, using the U.S. Health and Retirement Study (HRS), the English Longitudinal Study
of Ageing (ELSA), and the Survey of Health, Ageing and Retirement in Europe (SHARE). They compared
memory (a sum of 10-words immediate and delayed recall) of the retired with that of the not-retired in
2004 and, using national pension policies (the age of eligibility for early and full retirement benefits in the
public pension system) as instrument variables, reported a significant and negative effect of retirement on
cognition. It is important to note that they did not control for any potential covariates: while their analysis
sample included only those in the age range of 60 to 64, the second stage analysis did not account for key
covariates, such as education and cardiovascular diseases.
Several researchers have raised the issue of bias associated with omitted variables, and a few found
insignificant effects, once they controlled for key determinants. For example, Coe and Zamarro (2011)
79
estimated the effect of retirement on cognition using respondents age 50-69 from the 2004 SHARE. Using
the same cognitive test measure and instruments that Rohwedder and Willis (2010) employed, Coe and
Zamarro (2011) found an insignificant effect of retirement once they controlled for covariates, including
age, education, marital status, income, health, and activities.
Using two waves of SHARE, Mazzonna and Peracchi (2012) found significant, adverse effect of re-
tirement on cognition. Their empirical strategy was different from Coe and Zamarro (2011) significantly
in three aspects. First, using two waves of SHARE, they exploited both cross-country and within-country
variations associated with pension reforms. Second, they examined retirement duration instrumented by the
difference between the actual and legislated ages of eligibility for early and normal retirement age, instead
of binary variables of early and normal retirement age. Third, they did not control for key covariates, such as
marital status, income, health, and activities and only controlled for age, education, and country dummies.
Bonsang et al. (2012) and Coe et al. (2012) estimated the effect of retirement on cognition using the
1998-2008 HRS but reached to very different conclusions. Bonsang et al. (2012) argued that the retirement
effect would not be contemporaneous, but rather will take time to have an impact on cognition and therefore,
used a lagged retirement variable as their main regressor. They also used reaching early and normal pension
eligibility age as instruments, also lagged. After controlling for only age, Bonsang et al. (2012) reported a
significant and negative retirement effect for those aged 51 and 75. Coe et al. (2012), on the other hand,
using the same data source (but slightly wider age spectrum of 51 to 79), reached to opposite conclusion.
Coe et al. (2012)’s empirical strategy was different from Bonsang et al. (2012) in three different aspects.
First, Coe et al. (2012) stratified the sample into white-collar and blue-collar workers. Second, they used
whether respondents’ employers had offered them an early retirement window as their instrument to measure
retirement duration. Third, they controlled for age, education, race, ethnicity, and wave dummies, while
Bonsang et al. (2012) only controlled for age. Coe et al. (2012) found an insignificant effect of retirement on
cognition for white-collar workers, but a significant and positive effect of retirement on cognition for blue-
collar workers. This was the first paper on heterogeneous effects of retirement in the field of economics.
There are handful of papers in cognitive psychology and epidemiology literature, examining heterogeneous
effect of retirement by mental work demands (e.g., Fisher et al. (2014)).
Since then, two recent papers further investigated the heterogeneity of retirement effect. First, Mazzonna
and Peracchi (2017) stratified the sample by physical demands of jobs, using the first two waves of SHARE
80
and found that retiring from physically demanding job has immediate, positive effect on cognition, although
the longer retirement duration has adverse effect on cognition. They used the same instrument for retirement
duration as in Mazzonna and Peracchi (2012) and did not control for any covariates that are known risk
factors of cognitive decline, such as education, cardiovascular health, and physical activities.
Celidoni et al. (2017) explored potential heterogeneity associated with the timing of retirement, whether
one retire early or retire at the time of statutory retirement age. They found that different characteristics of
workers tend to retire early such that men in low-skill jobs tend to retire early and that normal retirees tend
to report more satisfaction and control over their jobs. Using instrumental variables approach, they found a
positive effect of retirement for the early retirees and an adverse effect of retirement for normal retirees.
Building on this literature, we further examine heterogeneity of retirement effect. While all of these
earlier investigations make important contributions by recognizing the hetereogeniety of jobs, it does not
directly examine cognitive demands of jobs. As the primary explanation for potential adverse effect of re-
tirement is that cognition is better maintained through mental exercise (Salthouse (1991)), by investigating
the cognitive demands of the job one retires from we can directly test the hypothesized relationship. In addi-
tion to cognitive demands, we also examine the potential heterogeneity associated with physical demands of
the job given earlier evidence. As cognitive and physical demands of jobs can be correlated, we examine both
of their moderating effect separately first and then togethers to investigate their relative importance. Specif-
ically, we take advantage of the information available from the Occupation Information Network (O*NET)
database on cognitive and physical demands of occupations, instead of relying on self-reported job char-
acteristics. Previously, Fisher et al. (2014) linked O*NET database to examine the moderating effect of
cognitive demands of job and found some suggestive evidence, but they did not account for the endogeneity
of retirement.
As poor cognition can contribute to selection into retirement, reverse causation is an important issue to
consider, and in this paper, we address this issue by taking an instrumental variables approach that is more
consistent with the economic literature. By employing different sets of instruments, we further examine the
sensitivity of our findings to the choice of instruments. The most popular instrument in assessing the effect of
retirement has been pension eligibility age (Rohwedder and Willis (2010)) and its deviations (e.g., delayed
effect of retirement by taking a lag to capture the impact to set in (Bonsang et al. (2012)) and retirement
duration by taking the distance to the individual’s pension eligibility age (Mazzonna and Peracchi (2012);
81
Mazzonna and Peracchi (2017)). There are also a few other instruments that have been used to capture the
effect of retirement including the provision of early retirement windows (Coe et al. (2012)). In this paper, we
pay close attention to potential variations introduced by the different instrumental variables that have been
used in the literature and evaluate their impacts.
Another important issue we aim to study in this paper is the bias associated with omitted variables.
Empirically, the two studies that include a more extensive set of covariates (Coe and Zamarro (2011); Coe
et al. (2012)) reported different results from those that did not control or include only a few selected covari-
ates (Rohwedder and Willis (2010); Mazzonna and Peracchi (2012); Bonsang et al. (2012)). Particularly,
education, cardiovascular health, physical activities, and social support are widely recognized risk factors
of late-life cognitive decline and impairment (Qiu and Fratiglioni (2015)). General cognitive function is
substantially heritable even in old age (Davies et al. (2015)), and therefore, endowed, genetic differences in
cognitive ability is an important omitted variable that can explain individual differences in cognitive perfor-
mance as well as selection into particular type of job, but such inherited characteristic has not been controlled
for in prior literature. Taking advantage of polygenic risk score of cognition (Davies et al. (2015)), we con-
trol for individual differences in genetic endowment in estimating the heterogeneous effect of retirement on
cognition. The proportion of variation in cognition accounted for by polygenic risk score of cognition was
estimated as 28% with standard error of 7% (Davies et al. (2015)).
We also further investigate the relationship between genetically endowment of cognition and cognitive
demands of job. As discussed earlier, individuals are not randomly assigned to occupations (Salthouse
(2006)). The differences in innate cognitive ability or desire for cognitive stimulation may lead to occu-
pations with different levels of cognitive demands. Using polygenic risk score of cognition, we examine
whether individuals with higher level of genetic endowment of cognition pursue more education and choose
occupations with higher level of mental demands. Recently, Okbay et al. (2016) reported genetic covariance
between increased educational attainment and increased cognitive performance. Considering such genetic
covariance, we further examine the association between polygenic risk score of education and cognitive de-
mands of occupation. As educational attainment is an important determinant of occupation, by examining
this association, we aim to shed a light to preserved differentiation hypothesis.
82
4.3 Econometric Strategy
4.3.1 Empirical Specification
LetCog
it
denote a measure of cognitive ability individuali in periodt. Our main econometric model is
Cog
it
=+Work
it
+
Work
it
PhysDem
it
+Work
it
CogDem
it
+X
it
+
it
(4.1)
where Work
it
is a dummy indicating whether the individual works in period t, PhysDem
it
and
CogDem
it
are the measures of physical and cognitive demand of the individual’s job in period t, X
it
is
a vector of controls, and
it
is the error term. Most of the literature has estimated an equation like (4.1)
without the interaction terms, focusing on the average effect of working on cognition,, whereas our focus
is on the heterogeneity of this effect, as reflected by the coefficients
and . We transform the physical
and cognitive demand scores such that their means are zero among workers, so that the interaction terms are
uncorrelated with the work dummy. This has the advantage that the coefficient is comparable across spec-
ifications with and without the interaction term, as well as comparable with the estimates in the literature.
It measures the effect of working in a job with average physical and cognitive demands, compared to not
working, whereas
and measure the effects of working in a job with higher physical or cognitive demand
versus lower demand. The effect of retiring from a job with physical demandc points higher than average
and cognitive demandd points higher than average is +
c+d and thus depends on the physical and
cognitive demands of the job. In addition to the equation (4.1), we include additional variables, polygenic
risk score for cognition and education, to control for initial endowment effects, as mentioned before. The
section in 4.5 discusses mostly the results from this specification:
Cog
it
=+Work
it
+
Work
it
PhysDem
it
+Work
it
CogDem
it
+polygenic score (cognition)+ polygenic score (education)+X
it
+
it
(4.2)
As will be discussed in the section 4.4.2, we study total word recall as a measure of cognition. The work
dummy is an indicator of whether the individual is working for pay, rather than a (subjective) classification
83
of whether the individuals consider themselves retired. This is in line with the literature, and is arguably both
more objective and more relevant for our purposes. The physical and cognitive demand scores are derived
from the occupation of the individual and the O*NET database, which contains ratings of occupations on a
large number of dimensions, among which several characteristics that reflect physical and cognitive demand,
which we combine in two indexes, as explained in more detail in the next section. Our control variables are
age, age squared, education, and several known or hypothesized correlates of cognition: cardiovascular dis-
eases and risks, physical activity, and social interaction (marital status and time spent with friends). Standard
errors are clustered at the individual level. All analyses are separate for men and women.
4.3.2 Instrumental Variables
One concern with estimating the equation (4.1) by OLS may be that individuals may retire because they
experience cognitive decline, which increases the burden of working. To address this endogeneity issue, we
estimate the model with instrumental variables (IV). Several instruments for retirement have been proposed
in the literature. We focus on the instruments employed by Rohwedder and Willis (2010). They define
two dummies, one for having reached early retirement age and one for having reached full (or normal)
retirement age. In the U.S., the Social Security early retirement age is 62, so the first dummy is 1 if the
individual is 62 or older and 0 otherwise. The Social Security full retirement age was traditionally 65, but
is gradually increased to 67, depending on year of birth (Social Security Administration (2017)). Using the
Social Security rules, the year and month of birth of the individual, and the year and month of the interview,
we construct a dummy variable indicating whether the individual’s age is higher than or equal to their full
retirement age.
As discussed in section 4.2, other authors have used different instruments and some have found different
results. To replicate this, and as a robustness check on our results, we also estimate our model with the
instruments used by Bonsang et al. (2012) (lagged versions of the Rohwedder and Willis (2010) instruments)
and Coe et al. (2012) (early retirement window).
These instruments are instruments for retirement, that is, not working. Because our regressor is a dummy
for working, we reverse-code the instruments, that is, 1 if not yet reached early or full retirement age or
expected retirement age or if not offered an early retirement window, and 0 otherwise.
84
As instruments for the interaction terms, we interact the instrument(s) for the work dummy with measures
of physical and cognitive demand of the job. However, because we necessarily do not have a contempora-
neous measure of the latter for individuals who do not currently work, we use the cognitive demand score
of the job measured at baseline. To avoid sample selectivity, we use the baseline data only to construct the
instrument. We do not include the data for cognition and work at baseline in the estimation sample. See
section 4.4 for details about the data and what the baseline wave is for each respondent. Individuals do not
switch jobs very often and if they do, they tend to switch to jobs with similar physical and cognitive demand
scores (in our sample of those working, 80% did not change their jobs.). Hence, the baseline physical and
cognitive demand scores should be highly correlated with the current physical and cognitive demand scores
for individuals who currently work, and be strongly related to the job an individual retired from if they do
not work. Hence, the product of this with the instrument for the work dummy should be highly correlated
with the interaction term. We take the validity of the instruments used in the literature as instruments for
Work
it
as given. For such an instrumentz
it
, this impliesE(z
it
it
) = 0. The validity of our instruments
then additionally requiresE(z
it
b
i
it
) = 0, whereb
i
is the physical or cognitive demand of individuali‘s
job at baseline and z
it
b
i
is the additional instrument for the interaction term Work
it
PhysDem
it
or
Work
it
CogDem
it
. A sufficient condition for the additional requirement isE(z
it
it
jb
i
) = 0, that is, the
instrumentz
it
is valid for each level of physical or cognitive demand of the job at baseline. While this is
stronger than the unconditional requirement, it is difficult to imagine a realistic situation in which the un-
conditional requirement holds but the conditional one does not, as this would imply certain specific relations
between the distribution ofb
i
in the population and the nonzero conditional expectationsE(z
it
it
jb
i
), for
which there does not appear to be a reasonable theoretical rationale.
For example, suppose that there were only two types of jobs, with low cognitive demand (b
i
= 0)
and high cognitive demand (b
i
= 1), respectively. Let E
0
= E(z
it
it
jb
i
= 0), E
1
= E(z
it
it
jb
i
= 1),
P
0
=Pr(b
i
=0), andP
1
=Pr(b
i
=1). The validity of the instrumentz
it
then impliesP
0
E
0
+P
1
E
1
=0.
A sufficient condition isE
0
=E
1
= 0. This condition is the one that implies the validity of our additional
instrumentz
it
b
i
, because by the law of iterated expectations,E(z
it
b
i
it
) = E[b
i
E(z
it
it
jb
i
)] = 0. Hence,
if z
it
is a valid instrument but z
it
b
i
is not, we must have P
0
E
0
+P
1
E
1
= 0 with E
0
6= 0 and E
1
6= 0.
Specifically,E
1
=E
0
=(P
0
=P
1
). There is no reason why the ratio of conditional expectations should be
equal to this ratio of probabilities. With a more general distributionF(b) ofb
i
, the condition generalizes to
85
R
E(z
it
it
jb)dF(b)=0 withPr[E(z
it
it
jb
i
)6=0]>0. Thus, this still implies a nontrivial relation between
the marginal distribution F(b) of b
i
and the conditional expectation function E(z
it
it
jb) as a function of
baseline physical or cognitive demand, for which no theoretical reason can be given. Hence, we conclude
that validity ofz
it
in practice also implies validity ofz
it
b
i
.
As mentioned above, in the main part of this paper, we take as given the validity of the instrumentsz
it
that have been used in the literature, and thus the validity of our empirical strategy. However, we do perform
analyses in which we compare results based on the different instruments used in the literature, because if
there is heterogeneity that is unaccounted for, IV analyses based on different instruments estimate differently
weighted effects (a different local average treatment effect or LATE).
4.4 Data and Variables
Our main data source is the Health and Retirement Study (HRS; Juster and Suzman (1995); HRS Staff
(2017)), which is a large-scale multidisciplinary panel survey of individuals over the age of 50 and their
spouses of any age. Our source of information on cognitive and physical demands of the job is the O*NET
database of the U.S. Department of Labor. We merge the two data sources by the occupation codes available
in the HRS. In this section we describe the two data sources, which variables we use from them, and how
we construct the measures we use in our analyses.
4.4.1 Health and Retirement Study
The HRS is the primary source of information about older individuals in the U.S. and as such plays a key role
in scientific studies of this age group and in the evaluation of public policies affecting them. The HRS started
in 1992 with individuals born 1931-1941 and their spouses. These respondents (and their new spouses, if
any) have been re-interviewed in 1994 and 1996. The sister study Assets and Health Dynamics of the Oldest
Old (AHEAD) started in 1993 with individuals born in 1923 or earlier and these respondents (and their new
spouses, if any) were re-interviewed in 1995. In 1998, the AHEAD study was merged into the HRS, and
panel members (and new spouses, if any) have been re-interviewed every two years. Also, since 1998, new
cohorts have been added every six years to keep the panel representative of the population over the age of 50.
HRS interviews are conducted in person or over the telephone. Thus, the baseline wave year for the original
86
HRS cohort (born 1931-1941 and their spouses) is 1992, for the AHEAD cohort (1923 and earlier), it is
1993, for the Children of the Depression Age (CODA; 1924-1930) and the War Babies (WB; 1942-1947) it
is 1998, for the Early Baby Boomers (EBB; 1948-1953) it is 2004 and for the Mid Baby Boomers (MBB;
1954-1959) it is 2010. See HRS Staff (2017) for a more detailed account of the sampling and interviewing
history of the HRS.
The data provide information on a broad range of topics. Variables that are of particular relevance to our
study are various measures of cognition, labor force participation, and occupation, as well as demographics,
health, education, and social activities.
If the target individual is unable to answer the questionnaire (because of physical or mental health prob-
lems), the HRS interviews a proxy respondent instead, typically the spouse or other family member, who
answers the questions instead for the target individual to the extent possible and relevant. Thus, HRS is
still able to obtain a large amount of essential information about such panel members. However, by their
very nature, the cognitive tests we rely on in our study are necessarily skipped in proxy interviews (although
proxy respondents answer some questions about the cognitive abilities of the target individuals). Hence, this
may cause some selection bias in our results.
For data from the core interview, we use the RAND HRS, version P (Bugliari et al. (2016)). The RAND
HRS combines data from all waves into a more user-friendly format, with consistent variable naming, de-
rived summary variables, and other enhancements. Of particular note is that the RAND HRS includes the
cognition scores, which have been imputed when missing (Fisher et al. (2017)). In the public release data,
occupation codes have been aggregated to a relatively small number of categories (a little over 20 in most
waves). Because characteristics of the individuals’ jobs are central to our study, we use the restricted occu-
pation data, which primarily use three- or four-digit Census codes, which gives several hundred categories.
We restrict attention to individuals who are 60-69 years old, which reflect the ages where most individuals
retire and where retirement is more related to the instruments. We analyze men and women separately, but
mostly focus on men. For our main analyses, the sample size for the main analysis of these data is 3,529
observations and 857 individuals for men, and 4,088 observations and 963 individuals for women. Table 4.1
presents summary statistics for the two primary samples of men, while table 4.2 presents summary statistics
for women. We use balanced sample for our main analysis, where it is defined as those panel members
having at least one post-baseline wave in which they were working and one in which they were not working.
87
4.4.2 Cognition Measures
The cognition measures available in the HRS are described in Ofstedal et al. (2005) and Fisher et al. (2014).
We study total word recall, the most frequently studied measure of cognition in order to compare our results
with the findings from prior literature.
The interviewer read a list of 10 common nouns to the respondent. There were four such lists, and a
respondent was randomly assigned one of them. However, spouses were always assigned a different list
to avoid learning effects. Immediately after this, the respondents were asked to report as many words as
possible from this list in a time-span of two minutes or less. The number correctly mentioned is the score
on the immediate word recall test, which thus has a range of 0-10. After approximately five minutes during
which they were asked other survey questions, respondents were again asked to repeat as many of the words
as they could remember in a time-span of two minutes or less. The number correctly mentioned is the score
on the delayed word recall test, which also has a range of 0-10. The word recall tests are intended to measure
episodic memory. Since the impairment of episodic memory is a significant hallmark of early Alzheimer’s
disease (Gallagher and Koh (2011)), word recall test scores may be a good indicator of early Alzheimer’s
disease as well.
4.4.3 Work and Cognitive and Physical Demands of the Job
Following much of the literature (e.g., Rohwedder and Willis (2010)), we use “whether working for pay”
as the regressor of interest. This is not exactly the same as “not retired yet”, but it is more objectively
measured than retirement, and is also the more relevant measure because it corresponds better to the theory
that working provides a cognitively stimulating environment. Nevertheless, in the age group that we study,
there is a strong correspondence between not working and being retired, if one includes other forms of being
out of the labor force without intention to re-enter (e.g., disability or being a homemaker without intention
to work). Hence, we interpret the results we find as being indicative of the effect of retirement. Some of
the literature argues that the effect of retirement accumulates gradually (e.g., Coe et al. (2012)) or with a lag
(Bonsang et al. (2012)), but in some preliminary analyses, we found the strongest and most consistent effect
with just a “currently working for pay” indicator, so that is the variable we use in our analyses.
88
For measuring cognitive demand of the job, we employ the Occupation Information Network (O*NET)
database. This contains information about worker characteristics and job characteristics by detailed occu-
pation. It is developed under the auspices of the Department of Labor. After some pretests, full-scale data
collection started in 2001. The database is regularly expanded and updated. Data are collected from samples
of workers in the occupations and from outside experts.
2
We use the version 19.0 of the O*NET, which was released in July 2014. It has 1,110 detailed occupa-
tions and it measures 256 characteristics of those occupations (or the workers in those occupations), although
not all characteristics are measured for each occupation. We use 16 elements (characteristics) describing in-
formation input, mental processes, and documenting and recording information. For each of these elements,
the O*NET includes a level (LV) and importance (IM) measure. We multiplied these two and then took the
average of the 16 elements to arrive at our measure of cognitive demand of the detailed occupation.
Through a series of crosswalk files, we then linked each O*NET occupation to 1980, 2002, and 2010
three- or four-digit Census occupation codes, which in turn determine the occupation codes used in the HRS
public release data. The latter aggregate occupations to 9, 17, 23, or 25 broad categories, depending on when
the occupational information of the respondent was collected. We computed the (unweighted) averages of
the cognitive demand scores across occupations in each broad category, resulting in an average cognitive
demand score for each occupational category in each of the four classifications. We merged these scores
to the HRS data, using the relevant occupation code available for each observation in the RAND HRS.
The resulting cognitive demand score in our data ranges from 7.99 to 17.81. We also calculated a physical
demand score in a similar way, where the average physical demand score is around 4.9.
Table 4.1 shows summary statistics for the work dummy, the cognitive demand score, and the physical
demand score for our three main samples. The distributions for cognitive demand (conditional on working)
again look very similar, but the fraction who are currently (as of the core interview) working is a bit lower
in the cognitive leisure sample than in the other two samples.
As mentioned in section 4.3, we subtract the average cognitive/physical demand score (for those who
are working) from the individual’s cognitive/physical demand score, so that the average cognitive/physical
demand score for those who are working becomes 0 and the product term in the model is uncorrelated
with the working dummy, which facilitates comparisons across the different specifications. Furthermore, as
2
The data are available from http://www.onetcenter.org.
89
mentioned there, in constructing the instrumental variables, we use the cognitive/physical demand score as
measured in the baseline interview, as opposed to the contemporaneous cognitive/physical demand score.
4.5 Empirical Results
4.5.1 First Stage Results
In this section, we report the empirical result from IV-2SLS estimation of the heterogeneous effect of re-
tirement on total word recall using the identification strategy presented in section 4.3. Before turning to
the estimated effects of retirement on cognition, we discuss the validity of two instruments: a dummy for
early retirement age and a dummy for normal retirement age. As discussed in the previous section 4.3.2,
we use two instruments in the main specification in table 4.3 and 4.4. We present the first stage F-statistic
for the joint significance of the excluded instruments. In table 4.3, the robust first stage F-statistic is 12.05
when we do not include interaction term of work*cognitive/physical demand and endowment variables and
the robust first stage F-statistic is 8.431 and 8.422 when we include interaction term and both interaction
and endowment variables, respectively. We observe large enough the first stage F-statistic in the column (4)
in table 4.3, which is above the conventional threshold for weak instruments. The first stage F-statistic in
column (5) and (6) show possible weak instrument problems because they are below 10. However, since the
first stage F-statistics are close to 10 and the main coefficients are qualitatively stable across specifications
in table 4.3, we include this empirical finding in the main analysis but interpret it with caution.
On the contrary, the first stage F-statistic for women are very low in table 4.4, which indicates the weak
instrument variable problems. There are two possible interpretations for the weak instrument variables. First,
it could be possible that the low proportion of women working/retiring than men may lead to low association
between our instruments and working status. As shown in the descriptive statistics in table 4.2, however,
the proportion of women who are working is 49 percent, similar to that of men (56%), which provides no
decisive evidence that the weak instrument problem is due to the low proportion of women who participate
in the labor market. Second, it is also possible that women are less likely to be affected by pension eligibility
rules. Women often retire for personal reasons so women are less dependent on the retirement logic of the
labor market (Hofacker et al. (2016)). Since we don’t have data to examine retirement reasons, this evidence
is only suggestive.
90
We next perform over-identification test of our model since we have two instrument variables, while the
number of endogenous variable is only one. Table 4.3 shows the p-value of Sargan-Hansen J test statistic
that examines the joint validity of all instrument variables. The test of overidentifying restrictions fail to
reject the validity of our instruments since its p-values are all well above the conservative threshold (0.25)
suggested by Roodman (2009). Given the validity of one of our instrument variables, for example, early
retirement pension age, the high p-value of Sargan-Hansen J test statistic means that the validity of the
second instrument, or normal retirement pension age, is not called into question.
4.5.2 Cognition Effects of Retirement
Since the first stage F-statistic is large enough only for balanced sample of men as shown in table 4.3, we
mostly focus on the cognition effects of retirement for male samples. Table 4.3 presents the main results. All
specifications include control variables mentioned in the section 4.3. The columns (1)-(3) present the OLS
results, which ignores the endogeneity of working/retirement status. Working status is positively, though
not strongly significant, correlated with the total word recall. That is, retirement is negatively associated
with the total word recall (column (1)). The point estimates on working status are stable across specifica-
tions with different sets of regressors: controlling for heterogeneous effects and endowment effects provides
nearly identical results. However, we don’t find significant effects on interaction terms, which suggests no
heterogeneous effects by occupation type.
Instrument variables estimates yield bigger coefficients than OLS, which are significant at 95 percent
confidence. Columns (4)-(6) report results for IV-2SLS specifications. The magnitude of IV-2SLS estimates
on working status slightly differ when we add more regressors (columns (5) and (6)). The significant IV-
2SLS estimates on working status suggest that retirement leads to decrease in cognition measured by total
word recall test. In addition, the coefficients on the interaction term of working and physical demand score is
negative and significant at 90 percent confidence, suggesting that retirement from physically demanding jobs
is conducive to cognitive function, which is consistent with Mazzonna and Peracchi (2017). This is robust
to additional controls of polygenic risk scores that may confound our results by affecting both cognition
and working status. However, we do not find evidence for the mental exercise hypothesis that retiring from
cognitively demanding jobs has a negative effect on cognition: the coefficients on the interaction term of
work and cognitive demand score is statistically insignificant. In conclusion, using IV-2SLS method, we
91
discover that there is heterogeneous effect of retirement by physical demands of the occupations but we do
not support the ‘differential preservation’ hypothesis.
Table 4.4 reports results for women. As mentioned in the previous section 4.5.1, since the first stage
F-statistics are very low, IV-2SLS estimates are difficult to interpret. However, we find similar patterns of
the IV-2SLS estimates on working status and interaction term of work and physical demands as shown in
table 4.3, again suggesting that there is heterogeneous effect of retirement but only for those retiring from
physically demanding jobs while we do not find significant effects on interaction term of work and cognitive
demands.
4.5.3 Sensitivity
To check sensitivity of our results to different sets of instrument variables, we perform our analysis using
early retirement age and normal retirement age separately. We also check sensitivity of our results to different
sample using unbalanced sample.
Table 4.5 and 4.6 report the IV-2SLS estimates using early and normal retirement age for men, respec-
tively. The coefficient on working status is positive and significant while the interaction term of working and
physical demand score is negative and significant, which is nearly identical to the main result in table 4.3.
However, we don’t find significant effects of retirement using normal retirement age as an instrument vari-
able. The heterogeneity in IV-2SLS estimates suggests that there may be heterogeneous complier groups to
different sets of instrument variables, which could produce heterogeneous effects across different subgroups.
In addition, although we do not emphasize the results of women sample, we also find similar patterns
of heterogeneity of IV-2SLS estimates in table 4.7 and 4.8, which lends support to different local average
treatment effects (LATE) by different sets of instrument variables. In appendix table A10 and A11, we also
present the results using unbalanced samples and find similar patterns as the main results although it loses
statistical significance.
92
4.5.4 Endowment of Cognition
In this subsection, we test the ‘preserved differentiation’ hypothesis that higher cognitive function in the
later life is due to higher levels of cognitive function throughout their lives (Fisher et al. (2014)). To ex-
amine this hypothesis, we use a polygenic risk score that is an aggregate of trait-associated alleles across
many loci and is widely known to be effective to predict health status (Agerbo et al. (2015)). In table 4.9,
we examine the association between polygenic risk score of cognition, education and physical/cognitive de-
mands of occupations. Polygenic risk score of education has a strong and significant association with both
physical demands and cognitive demands while polygenic risk score of cognition is not correlated with phys-
ical/cognitive demands scores. We can conjecture that those who have high polygenic scores of education
are more likely to work in cognitively engaging sectors. Our results support the ‘preserved differentiation’
hypothesis, indicating potential bias in estimates of the heterogeneous retirement effects due to selection
into occupations.
3
4.6 Conclusion
This chapter revisits an interesting question of heterogenous effects of retirement on cognitive function
by occupation type using Health and Retirement Study (HRS) and the Occupation Information Network
(O*NET) database that contains detail information about the occupation. We first test ‘mental exercise’ or
‘preserved differentiation’ hypothesis. Using early and normal pension eligibility age as instrument vari-
ables for work/retirement status, we find the heterogeneous effects of retirement when an individual retires
from physically demanding jobs. Retiring from more physically demanding jobs has a positive impact on
cognitive function measured by total word recall test. However, retiring from more cognitively demanding
jobs has no significant effect on cognitive function, providing no evidence on ‘preserved differentiation’
hypothesis. This finding is similar across samples and different specifications.
It is noteworthy that there are heterogeneous local average treatment effects (LATE) depending on the
different sets of instrument variables and subgroups of our sample. This observation is suggestive of mixed
3
As we addressed in the previous section, we support the validity of our interaction instrument variables by positing a simple
theoretical tool.
93
empirical findings of retirement effects in the previous studies (Rohwedder and Willis (2010); Coe and
Zamarro (2011)).
We also study the effect of genetic endowment for cognition and education on cognitive and physical
demand of the job using polygenic risk scores. The genetic endowment for education is significantly asso-
ciated with cognitive and physical demand of the job, suggesting that those who are cognitively active are
those who maintain higher levels of cognitive function throughout their lives.
Overall, the empirical results of this chapter provide empirical evidence on ‘differential preservation’
hypothesis while the results provide little insight into ‘preserved differentiation’ hypothesis. However, we
find the evidence that retirement from physically demanding job is good for cognitive function. Further
micro-empirical studies are needed to illuminate causal heterogeneous retirement effects as it is dependent
on different specifications and different data sets.
94
Table 4.1: Summary statistics
(men, polygenic risk score samples for total word recall)
Unbalanced Balanced
Variable N mean sd N mean sd
Age (years) 7,979 64.38 2.83 3,529 64.51 2.78
Married/coupled 7,979 0.84 0.37 3,529 0.85 0.36
Education (years) 7,979 13.56 2.67 3,529 13.41 2.49
Work 7,979 0.56 0.5 3,529 0.5 0.5
Cognitive demand of job 4,497 14.39 4.96 1,759 13.76 4.86
Physical demand of job 4,497 4.73 2.45 1,759 4.94 2.47
Polygenic risk score cognition 7,979 0 0.99 3,529 0.04 1.01
Polygenic risk score education 7,979 0.01 1.01 3,529 0.04 1
Vigorous activity 7,979 0.6 0.49 3,529 0.6 0.49
Cardiovascular disease 7,979 0.61 0.49 3,529 0.63 0.48
Social interactions per year 7,979 98.04 188.52 3,529 103.72 159.00
Total word recall 7,979 10.5 3.08 3,529 10.43 3.00
Table 4.2: Summary statistics
(women, polygenic risk score samples for total word recall)
Unbalanced Balanced
Variable N mean sd N mean sd
Age (years) 8,869 64.22 2.83 4,088 64.44 2.81
Married/coupled 8,869 0.66 0.47 4,088 0.64 0.48
Education (years) 8,869 13.26 2.26 4,088 13.28 2.19
Work 8,869 0.49 0.50 4,088 0.49 0.50
Cognitive demand of job 4,389 13.22 4.89 1,984 12.82 4.81
Physical demand of job 4,389 3.48 1.45 1,984 3.58 1.47
Polygenic risk score cognition 8,869 0.00 1.02 4,088 0.01 1.00
Polygenic risk score education 8,869 0.00 0.97 4,088 0.02 0.97
Vigorous activity 8,869 0.46 0.50 4,088 0.46 0.50
Cardiovascular disease 8,869 0.54 0.50 4,088 0.55 0.50
Social interactions per year 8,869 91.62 140.53 4,088 94.90 152.60
Total word recall 8,869 11.81 3.00 4,088 11.90 2.99
95
Table 4.3: Main regression results for total word recall
(men, balanced polygenic risk score sample)
Regressor/statistic OLS IV
(1) (2) (3) (4) (5) (6)
Work 0.191* 0.188* 0.187* 2.988** 2.129** 2.120**
(0.11) (0.11) (0.109) (1.418) (0.867) (0.87)
Work * Phys.demand -0.059 -0.059 -0.154* -0.152*
(0.037) (0.037) (0.092) (0.092)
Work*Cogn.demand 0.006 0.005 0.003 0.001
(0.02) (0.019) (0.052) (0.052)
Polygenic cognition 0.164** 0.165**
(0.07) (0.071)
Polygenic education 0.047 0.04
(0.077) (0.077)
Observations 3,529 3,529 3,529 3,529 3,529 3,529
Individuals 857 857 857 857 857 857
First stage F 12.05 8.431 8.422
Hansen J (P-value) 0.561 0.581 0.586
Robust standard errors in parentheses; controls included as discussed in the text.
*p< 0:1, **p< 0:05, ***p< 0:01
Table 4.4: Main regression results for total word recall
(women, balanced polygenic risk score sample)
Regressor/statistic OLS IV
(1) (2) (3) (4) (5) (6)
Work 0.118 0.125 0.116 -8.853** -3.317* -3.285*
(0.104) (0.104) (0.104) (4.508) (1.918) (1.908)
Work * Phys.demand -0.102* -0.086 -0.297*** -0.288**
(0.055) (0.055) (0.114) (0.114)
Work*Cogn.demand 0.012 0.013 0.026 0.020
(0.017) (0.017) (0.041) (0.041)
Polygenic cognition 0.107 0.173** 0.151*
(0.072) (0.081) (0.080)
Polygenic education 0.161** 0.140*
(0.070) (0.076)
Observations 4,088 4,088 4,088 4,088 4,088 4,088
N individuals 963 963 963 963 963 963
First stage F 2.717 2.073 2.068
Hansen J (P-value) 0.725 0.0642 0.0650
Robust standard errors in parentheses; controls included as discussed in the text.
*p< 0:1, **p< 0:05, ***p< 0:01
96
Table 4.5: Sensitivity check using social security early retirement age (IV)
(men, balanced polygenic risk score sample)
Regressor/statistic OLS IV
(1) (2) (3) (4) (5) (6)
Work 0.191* 0.188* 0.187* 2.947** 3.218** 3.214**
(0.110) (0.110) (0.109) (1.420) (1.504) (1.506)
Work*Phys.demand -0.059 -0.059 -0.216* -0.215*
(0.037) (0.037) (0.112) (0.112)
Work* Cogn.demand 0.006 0.005 -0.003 -0.005
(0.020) (0.019) (0.065) (0.065)
Polygenic cognition 0.164** 0.166**
(0.070) (0.073)
Polygenic education 0.047 0.037
(0.077) (0.080)
Observations 3,529 3,529 3,529 3,529 3,529 3,529
N individuals 857 857 857 857 857 857
First stage F 24.09 7.413 7.407
Robust standard errors in parentheses; controls included as discussed in the text.
*p< 0:1, **p< 0:05, ***p< 0:01
Table 4.6: Sensitivity check using social security normal retirement age (IV)
(men, balanced polygenic risk score sample)
Regressor/statistic OLS IV
(1) (2) (3) (4) (5) (6)
Work 0.191* 0.188* 0.187* 4.577 4.645 4.550
(0.110) (0.110) (0.109) (3.404) (3.019) (2.978)
Work*Phys.demand -0.059 -0.059 -0.011 -0.014
(0.037) (0.037) (0.185) (0.184)
Work* Cogn.demand 0.006 0.005 0.036 0.032
(0.020) (0.019) (0.069) (0.069)
Polygenic cognition 0.164** 0.166**
(0.070) (0.077)
Polygenic education 0.047 0.037
(0.077) (0.084)
Observations 3,529 3,529 3,529 3,529 3,529 3,529
N individuals 857 857 857 857 857 857
First stage F 3.839 1.616 1.622
Robust standard errors in parentheses; controls included as discussed in the text.
*p< 0:1, **p< 0:05, ***p< 0:01
97
Table 4.7: Sensitivity check using social security early retirement age (IV)
(women, balanced polygenic risk score sample)
Regressor/statistic OLS IV
(1) (2) (3) (4) (5) (6)
Work 0.118 0.125 0.116 -8.099* -7.770* -7.772*
(0.104) (0.104) (0.104) (4.613) (4.338) (4.344)
Work*Phys.demand -0.102* -0.086 -0.097 -0.088
(0.055) (0.055) (0.178) (0.179)
Work* Cogn.demand 0.012 0.013 0.114 0.108
(0.017) (0.017) (0.082) (0.082)
Polygenic cognition 0.107 0.231**
(0.072) (0.114)
Polygenic education 0.161** 0.141
(0.070) (0.095)
Observations 4,088 4,088 4,088 4,088 4,088 4,088
N individuals 963 963 963 963 963 963
First stage F 4.477 1.583 1.569
Robust standard errors in parentheses; controls included as discussed in the text.
*p< 0:1, **p< 0:05, ***p< 0:01
Table 4.8: Sensitivity check using social security normal retirement age (IV)
(men, balanced polygenic risk score sample)
Regressor/statistic OLS IV
(1) (2) (3) (4) (5) (6)
Work 0.118 0.125 0.116 -104.982 354.480 254.550
(0.104) (0.104) (0.104) (2,639.929) (25,722.924) (12,908.191)
Work*Phys.demand -0.102* -0.086 6.349 4.413
(0.055) (0.055) (480.528) (236.857)
Work* Cogn.demand 0.012 0.013 -1.353 -0.985
(0.017) (0.017) (98.776) (50.017)
Polygenic cognition 0.107 -3.530
(0.072) (184.203)
Polygenic education 0.161** 0.947
(0.070) (40.516)
Observations 4,088 4,088 4,088 4,088 4,088 4,088
N individuals 963 963 963 963 963 963
First stage F 0.002 0.000 0.000
Robust standard errors in parentheses; controls included as discussed in the text.
*p< 0:1, **p< 0:05, ***p< 0:01
98
Table 4.9: Relation between physical and cognitive demands and other regressors
(men, unbalanced polygenic risk score sample for total word recall)
Regressor/statistic Physical Demand Cognitive Demand
(1) (2) (3) (4) (5) (6)
Polygenic cognition -0.095 0.002 0.001 -0.177
-0.062 -0.055 -0.121 -0.106
Polygenic education -0.381*** -0.171** 0.708*** 0.341**
-0.054 -0.054 -0.113 -0.11
Controls No No Yes No No Yes
Observations 4497 4497 4497 4497 4497 4497
Individuals 1770 1770 1770 1770 1770 1770
F statistic 2.328 50.3 46.442 0 39.44 40.622
Robust standard errors in parentheses; controls included as discussed in the text.
*p<0:1, **p<0:05, ***p<0:01
99
Chapter 5
Conclusion
The objective of the first chapter is to discuss the determinants of individual time preference, which has rarely
been addressed in the previous studies. Individual time preference is an important factor in economic models,
where time preference explains a wide range of socioeconomic outcomes, such as health outcomes, financial
decision making, savings, wealth, and criminal behavior. Despite its immense importance in economics,
less is known about the determinants of individual time preference. This dissertation contributes to the
understanding of the role of education, one of the potential determinants, in individual time preference
formation using Indonesia Family Life Survey, a panel data set containing information about time preference.
The second chapter investigates the joint effects of two public programs in Tanzania. In this chapter, the
Iodine Supplementation Program (ISP) was found to have a negative impact on educational attainment, while
the Primary Education Development Program (PEDP) positively affected educational attainment. Further-
more, a negative interaction effect of ISP and PEDP on educational attainment was observed. This chapter
contributes to the methodological discussion around identifying ‘dynamic complementarity’ by illustrat-
ing the need for more cautious interpretation of the interaction of two exogenous shocks as evidence for
or against complementarity. We provide an evidence of ‘dynamic complementarity’ using an alternative
strategy.
The third chapter revisits the effect of retirement on cognition. We address the endogeneity concern of
retirement by taking an instrument variable approach that is more consistent with the economic literature.
Further, in addition to previous studies, we take advantage of detail information of occupation characteris-
tics from the Occupation Information Network (O*NET) database. Even though we minimize the concern
for endogeneity of retirement, the selection into occupations may weaken the validity of our identification
strategy. Since the differences in innate cognitive ability or desire for cognitive stimulation may lead to
occupations with different levels of cognitive demands, controlling for innate cognitive ability captured by
the genetic risk score of cognition could minimize the concern for the validity of our identification strategy.
100
This chapter contributes to the understanding of retirement effect by using more detail information about
occupations and more comprehensive discussion regarding the identification strategy.
101
Appendix
A.1 Additional Figures and Tables in Chapter 2
Figure A1: Distribution of INPRES intensity
102
Figure A2: Distribution of educational attainments by INPRES treatment status
103
(a) Time Preference A
(b) Time Preference B
Figure A3: Time preference and Age by education
Figure A4: Distribution of time preference by INPRES treatment (IV sample)
104
Table A1: The effect of education on negative time discounting
Negative Time Discount A Negative Time Discount B
(1) (2) (3) (4) (5) (6)
Years of schooling -0.100** 0.202 -0.020 -0.014 -0.227 -0.014
(0.047) (0.526) (0.017) (0.034) (0.603) (0.014)
Observations 3,130 1,558 1,572 3,130 1,558 1,572
Sample Pooled Male Female Pooled Male Female
Mean of DV 0.030 0.033 0.027 0.027 0.033 0.022
Mean of YoS 8.411 8.904 7.922 8.411 8.904 7.922
IV F-stat 2.284 0.0961 8.061 2.284 0.0961 8.061
Note: Negative Time Discount is a dummy variable, 1 indicating the negative time discounting while 0 indicating the normal
response. The sample cohorts are born between 1958 and 1962, 1968 and 1972, exactly same as IV-2SLS main analysis. We
use same instrument variable (INPRES intensity*cohort treatment dummy) to overcome endogeneity of years of schooling.
Control variables are the same as the main analysis. Robust standard errors are clustered at the province level.
105
Table A2: The first stage of IV-2SLS
Years of Schooling
(1) (2) (3) (4) (5) (6)
Intensity * Treated cohort dummy 0.325** -0.096 0.796*** 0.315 0.053 0.578*
(0.116) (0.288) (0.224) (0.230) (0.290) (0.281)
Our controls Y Y Y N N N
Duflo controls N N N Y Y Y
Sample Pooled Male Female Pooled Male Female
Observations 3,005 1,541 1,464 3,187 1,626 1,561
R-squared 0.579 0.675 0.671 0.274 0.333 0.351
Note: Intensity is the number of schools constructed per 1,000 students varied by district level. Treated cohort
dummy is an indicator for treatment by INPRES. Our controls include parental education, ethnicity, religion, year of
birth FE, month of birth FE, district of birth FE, district of birth FE, current district FE, language used for interview,
language commonly used in daily life, survey wave FE, the order of TP module asked and children’s population in
1971, enrollment rate in 1971 and the allocation of the water and sanitation program. Duflo controls include year
of birth FE, district of birth FE, and children’s population in 1971, enrollment rate in 1971 and the allocation of the
water and sanitation program.
106
Table A3: Suggestive Mechanism
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Raven (Fluid*) Arithmetic test Word Recall Serial 7s
Time Preference A
Years of schooling -0.008*** -0.006*** -0.008*** -0.006*** -0.007*** -0.006*** -0.009*** -0.008***
(0.002) (0.001) (0.002) (0.002) (0.001) (0.001) (0.002) (0.002)
Mechanism variable -0.013*** -0.024*** -0.004*** -0.018***
(0.004) (0.003) (0.001) (0.006)
Observations 10,934 10,934 10,920 10,920 18,359 18,359 9,186 9,186
R-squared 0.099 0.102 0.099 0.103 0.078 0.079 0.104 0.106
Mean of DV 0.667 0.667 0.667 0.667 0.686 0.686 0.667 0.667
Time Preference B
Years of schooling -0.005*** -0.003** -0.005*** -0.003* -0.005*** -0.004*** -0.004*** -0.004***
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Mechanism variable -0.007 -0.017*** -0.003*** -0.008
(0.004) (0.004) (0.001) (0.005)
Observations 10,934 10,934 10,920 10,920 18,359 18,359 9,186 9,186
R-squared 0.097 0.097 0.096 0.099 0.078 0.078 0.103 0.103
Mean of DV 0.767 0.767 0.767 0.767 0.784 0.784 0.763 0.763
Panel B. Total Income Risk Averse Self health CESD
Time Preference A
Years of schooling -0.008*** -0.007*** -0.007*** -0.008*** -0.007*** -0.007*** -0.007*** -0.007***
(0.001) (0.001) (0.002) (0.002) (0.001) (0.001) (0.001) (0.001)
Mechanism variable -0.012** 0.065*** -0.017 -0.002**
(0.005) (0.008) (0.014) (0.001)
Observations 12,577 12,577 12,101 12,101 18,426 18,426 18,307 18,307
R-squared 0.097 0.097 0.095 0.098 0.078 0.078 0.078 0.078
Mean of DV 0.680 0.680 0.617 0.617 0.686 0.686 0.686 0.686
Time Preference B
Years of schooling -0.006*** -0.006*** -0.005*** -0.005*** -0.005*** -0.005*** -0.005*** -0.005***
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Mechanism variable -0.003 0.112*** -0.007 -0.003***
(0.004) (0.008) (0.013) (0.001)
Observations 12,577 12,577 16,763 16,763 18,426 18,426 18,307 18,307
R-squared 0.098 0.098 0.084 0.093 0.078 0.078 0.078 0.079
Mean of DV 0.775 0.775 0.785 0.785 0.785 0.785 0.784 0.784
Sample Pooled Pooled Pooled Pooled Pooled Pooled Pooled Pooled
Note: Fluid means fluid intelligence test. Robust standard errors are clustered at the province level. All regressions control for year
of birth FE, month of birth FE, ethnicity FE, religion FE, district FE, rainfall in different years of life, language used for interview,
language used for daily life, parent’s education level, wave fixed effect and the order of questions asked. Raven test measuring fluid
intelligence scores between 0-8. Arithmetic test measuring mathematical skill scores between 0-5. Word recall score between 0-20.
Serial 7 score lies between 0-5. Total income variable is log of total income earned as wage or business profit, rental income, nonlabor
income and retirement pension income. Risk averse is a dummy indicating whether the respondent chose the most risk averse option
throughout the risk preference module. Self health is a dummy that takes value 1 if individuals report being in very good health.
CESD lies between 0-30.
107
A.2 Time Preference Imputation in Chapter 2
We describe the imputation process for time preference. Here, we assume exponential discounting for sim-
plicity. However, there is always possibility of hyperbolic discounting. As mentioned above, to keep things
simple, we include only those individuals who exhibit a discount factor between zero and one. Conditional
on
c
T
1, the discounted value comparison for a person who chooses IDR 1m today over IDR 3m one year
from now must be1
c
T
3. Therefore,
c
T
must be less than or equal to1=3. For those who choose IDR
3m one year from now over IDR 1m today, therefore, have
c
T
1=3. Those who chose IDR 1m in the first
round have to choose between IDR 1m today or IDR 6m in one year from today in the next round. If, here,
someone chooses IDR 1m today, it must be that1
c
T
6. This implies that
c
T
1=6. Consequently, those
who chose IDR 1m in the first round but chose IDR 6m in the second round must have 1=6
c
T
1=3.
Those who choose IDR 3m in the first round have to choose between IDR 1m today and IDR 2m tomorrow
in the second round. If they prefer IDR 2m here, they have
c
T
1=2, 1=3
c
T
1=2 otherwise. In
the remaining of the paper, we will call this the indifference point method. The bound calculations for sub-
module B are exactly same as sub-module A. The only difference is instead of using, sub-module B uses
5
because time horizon is five years. Combining these imputed values with the distribution of responses
present in Table 2.1 and 2.2 produces a reasonable descriptive picture of the distribution of the time discount
factor in the first panel of Figure A5. For our robustness check, we calculate the midpoint of the range of
time discount factor and construct the time discounting rate, the reverse of time discount factor.
108
Figure A5: Distribution of imputed discount factor
109
A.3 ISP Treatment Definition in Chapter 3
This section draws heavily from Field et al. (2009). Information has been reproduced for clarity in under-
standing of how the iodine treatment variables were defined.
Table A4: ISP Coverage Variation (from Field et al. (2009))
Region District Year 1 Coverage 1 Year 2 Coverage 2 Year 3 Coverage 3 Year 4 Coverage 4 Year 5 Coverage 5
Dodoma Mpwapwa 1990 0.65 1992 0.58
Arusha Monduli 1992 0.71
Arusha Arumera 1991 0.89
Kilimanjaro Rombo 1990 0.68
Mororgoro Ulanga 1988 0.73 1991 0.61 1992 0.34
Ruvuma Songea Rural 1987 0.91 1991 0.74 1995 0.85
Ruvuma Mbinga 1995 0.92
Iringa Mufundi 1986 0.41 1991 0.63 1995 0.54
Iringa Makete 1986 0.2 1991 0.62 1993 0.62 1996 0.49
Iringa Njombe 1989 0.76 1992 0.68 1995 0.64
Iringa Ludewa 1989 0.59 1992 0.62 1995 0.47
Mbeya Chunya 1990 0.49
Mbeya Mebya Rural 1986 0.44 1989 0.84 1990 0.9 1993 0.53 1997 0.53
Mbeya Kyela 1989 0.91 1993 0.57
Mbeya Rungwe 1986 0.35 1990 0.73 1993 0.49
Mbeya Ileja 1989 0.94 1992 0.71
Mbeya Mbozi 1989 0.67 1991 0.63
Rukwa Mpanda 1987 0.79 1991 0.6 1993 0.72
Rukwa Sumbawanga 1987 0.76 1990 0.89 1993 0.72 1996 0.51
Rukwa Nkansi 1987 0.89 1991 0.49
Kigoma Kibondo 1989 0.73 1992 0.75 1996
Kigoma Kasulu 1987 0.5 1990 0.66 1996 0.49
Kigoma Kigoma Rural 1991 0.91
Kagera Karagwe 1990 0.96 1994 0.85
Kagera Bukoba Rural 1994 0.78
Kagera Biharamulo 1990 0.96 1994 0.38
Kagera Ngara 1989 0.29 1994 0.51
Calculation of probability of protection: The treated mothers received and iodine dosage of 380 mg
via the IOC (Peterson (2000); Peterson et al. (1999)). However, as described in Field et al. (2009), Wolff
(2001), Jun and Jianqun (1985) and Untoro et al. (1998) provide a review of literature that finds that majority
of iodine stored in the fatty tissue is depleted rapidly within the first week and an hyperbolic rate thereafter.
Following Field et al. (2009), we assume that 85 percent (323) of the 380 mg dose was extracted away
immediately within the first month and the depletion followed the simple hyperbolic discounting formula
V = A=(1 +kt) after that, where k
1
is the half life of iodine in months. Using the observation from
Cao et al. (1994) and Eltom et al. (1985), which use similar dosages of IOC provides full protection for 24
months and that 6.5 mg is the minimum iodine requirement for one full month of protection, Field et al.
(2009) calculate the half life to be 3 months. This implied half life is consistent with other studies of the
approximate half lives of urine iodine excretion after oral iodine administration to human populations with
iodine deficiency (See Wolff (2001)). The probability of protection in a month of the first trimester, therefore,
110
is the probability that the program had started and reached the mother of the child by that month and the
stocks of iodine had not depleted to levels insufficient for protection (< 4:2mg as per Field et al. (2009)) in
that month.
1
A child is protected in the first trimester if she is protected throughout weeks 1 to 12 (roughly
three months)
Based on the information and assumption above, probability of protection from in utero IDD if the child
district of birth received the ISP in year t (by month of birth):
Table A5: Probability of Protection
Year Jan Feb March April May June July Aug Sep Oct Nov Dec Birth year average
t 0 0 0 0 0 0 0 0.028 0.083 0.167 0.250 0.333 0.072
t+1 0.417 0.5 0.583 0.667 0.75 0.833 0.917 1 1 1 1 1 0.806
t+2 1 1 1 1 1 1 1 1 1 0.998 0.991 0.977 0.997
t+3 0.955 0.927 0.891 0.849 0.802 0.749 0.69 0.627 0.559 0.488 0.419 0.353 0.668
t+4 0.292 0.237 0.189 0.148 0.112 0.082 0.057 0.037 0.022 0.011 0.004 0.001 0.099
1
The 6.5 mg and 4.2 mg figures are calculated based in the recommended daily allowance (RDA) for pregnant women
111
A.4 Alternative Definitions of ISP Exposure in Chapter 3
Recall that our definition of ISP exposure probability used the variation in the coverage rate across districts
and years and the probability of protection based on the availability of adequate amount of iodine in the
mother’s body which depends on the relative timings of supplementation and conception. Since we did not
have the exact date of supplementation and birth, we assumed that the probability of supplementation and
birth are uniform across the year. This assumption can, however, bias our results. We check the robustness
of our results to different definitions of ISP exposure in table A6. The results from the most preferred
specification (specification (8) in table 3.2) are reproduced in column (1) for comparison. In column (2),
we use the district-year specific coverage rate as the probability of ISP exposure. This avoids biases due
to differences in exact date of supplementation. In column (3), we use only the depletion formula for the
exposure probability calculation assuming that if a district was treated in a particular year, all individuals
from the district received the supplementation. This avoids biases due to measurement and reporting errors
in coverage rate. This is the main specification used in Field et. al (2009). In column (4), we use a dummy
indicator of whether or not the probability of exposure of an individual was non-zero. The results suggest
our findings are robust to the alternative ways of defining ISP treatment.
112
Table A6: Robustness of ISP Exposure Definition
(1) (2) (3) (4)
Years of Schooling
ISP -0.71***
(0.15)
ISP2: coverage only (no depletion) -0.52***
(0.14)
ISP3: depletion only (no coverage) -0.55***
(0.13)
ISP4: depletion dummy (= 1, if exposed at all) -0.40***
(0.10)
Religion dummy YES YES YES YES
Tribe dummy YES YES YES YES
Livestock Value YES YES YES YES
Mean of dependent variable 2.20 2.20 2.20 2.20
Mean ISP treatment probability 0.31 0.80 0.39
Observations 518 518 518 518
R-squared 0.3641 0.3624 0.3633 0.3627
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at geoage level where geoage groups are district-year of birth groups. Other controls
include a dummy each indicating whether the mother and the father of the child had some education, age, gender and primary enumeration area fixed effects.
Table A7: Impact of ISP on Vaccinations
(1) (2) (3) (4) (5) (6) (7)
Has a vaccination card Tetanus vaccine Polio vaccine Measles vaccine TB vaccine Score 1 Score 2
ISP 0.0330 -0.0408 -0.1127 -0.0235 -0.0097 -0.1537 -0.1866
(0.0840) (0.0996) (0.0954) (0.0768) (0.0729) (0.2550) (0.2119)
Religion dummy YES YES YES YES YES YES YES
Tribe dummy YES YES YES YES YES YES YES
Livestock Value YES YES YES YES YES YES YES
Mean of dependent variable 0.5597 0.3180 0.3392 0.4416 0.5724 2.1557 1.5960
Mean ISP treatment probability 0.3220 0.3228 0.3228 0.3228 0.3228 0.3220 0.3220
Observations 552 552 552 552 552 552 552
R-squared 0.6637 0.3947 0.4161 0.5865 0.6954 0.7121 0.6869
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at district-year of birth groups. Controls include a quadratic in age, a dummy each
for whether the mother and father have some education or not, a dummy for gender of the individual, a dummy each for whether the individual belongs to the majority tribe or religion, controls for
total land holdings of the household, and district fixed effects.
113
Table A8: Across Household Impacts of ISP
(1) (2) (3) (4)
Years of education Height for age in 2004
ISP -0.95*** -0.69** -0.76 -0.06
(0.28) (0.34) (0.50) (0.44)
Observations 386 398 355 356
R-squared 0.40 0.49 0.08 0.08
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at district-year of birth
groups. Controls include a dummy each for whether the mother and father have some education or not, a dummy for gender of the individual,
a dummy each for whether the individual belongs to the majority tribe or religion, controls for total land holdings of the household, age and
district fixed effects.
Table A9: Primary School Starting Age as a Plausible Mechanism
(1) (2) (3) (4) (5) (6)
Years of Schooling
ISP -0.7330*** -0.3029*** -0.7267*** -0.2937*** -0.7663*** -0.3074**
(0.1171) (0.1022) (0.1138) (0.1000) (0.1341) (0.1243)
PEDP 0.1840*** -0.0931 0.1840*** -0.0941 0.1801*** -0.0951
(0.0499) (0.0891) (0.0509) (0.0916) (0.0481) (0.0923)
ISP * PEDP -0.5187** -0.3362 -0.5063* -0.3141 -0.4910* -0.3093
(0.2428) (0.2253) (0.2510) (0.2265) (0.2641) (0.2328)
Primary starting age -0.5941*** -0.5950*** -0.5944***
(0.0380) (0.0389) (0.0395)
Mean of dependent variable 2.2023 2.2023 2.2023 2.2023 2.2023 2.2023
Mean ISP treatment probability 0.3199 0.3199 0.3199 0.3199 0.3199 0.3199
Religion dummy NO NO YES YES YES YES
Tribe dummy NO NO YES YES YES YES
Livestock Value NO NO NO NO YES YES
Observations 519 519 519 519 519 519
R-squared 0.3635 0.6290 0.3637 0.6298 0.3651 0.6299
Notes: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. The standard errors are clustered at geoage level where geoage are district X year of birth groups. Other controls
include a quadratic in age, total land ownership of the household in which the child was born, a dummy each indicating whether the mother and the father of the child had some education, gender and
primary enumeration area fixed effects.
114
A.5 Additional Tables in Chapter 4
Table A10: Main regression results for total word recall
(men, unbalanced polygenic risk score sample)
Regressor/statistic OLS IV
(1) (2) (3) (4) (5) (6)
Work 0.396*** 0.414*** 0.400*** 4.295 1.092 1.032
-0.088 -0.088 -0.087 -3.824 -0.857 -0.861
Work * Phys.demand -0.054** -0.051* -0.076 -0.073
-0.027 -0.027 -0.147 -0.147
Work*Cogn.demand 0.041*** 0.041*** 0.115 0.111
-0.012 -0.012 -0.073 -0.074
Polygenic cognition 0.104** 0.105**
-0.049 -0.049
Polygenic education 0.102** 0.081
-0.05 -0.052
Observations 7,979 7,979 7,979 7,979 7,979 7,979
Individuals 2329 2329 2329 2329 2329 2329
First stage F 1.973 4.529 4.515
Hansen J (P-value) 0.268 0.275 0.271
Robust standard errors in parentheses; controls included as discussed in the text.
*p< 0:1, **p< 0:05, ***p< 0:01
Table A11: Main regression results for total word recall
(women, unbalanced polygenic risk score sample)
Regressor/statistic OLS IV
(1) (2) (3) (4) (5) (6)
Work 0.249*** 0.255*** 0.250*** -4.060 -0.653 -0.643
(0.080) (0.080) (0.080) (3.541) (1.335) (1.328)
Work * Phys.demand -0.121*** -0.114*** -0.533** -0.522**
(0.038) (0.037) (0.226) (0.224)
Work*Cogn.demand 0.014 0.013 -0.016 -0.018
(0.012) (0.012) (0.056) (0.056)
Polygenic cognition 0.007 0.006
(0.046) (0.048)
Polygenic education 0.125*** 0.108**
(0.048) (0.050)
Observations 8,869 8,869 8,869 8,869 8,869 8,869
N individuals 2454 2454 2454 2454 2454 2454
First stage F 2.239 2.717 2.731
Hansen J (P-value) 0.723 0.260 0.281
Robust standard errors in parentheses; controls included as discussed in the text.
*p< 0:1, **p< 0:05, ***p< 0:01
115
Bibliography
Adelman, I. (1975). Development economics–a reassessment of goals. The American Economic Review,
302–309.
Adhvaryu, A., T. Molina, A. Nyshadham, and J. Tamayo (2014). Recovering from early life trauma: Dy-
namic substitution between child endowments and investments. Unpublished Manuscript.
Adhvaryu, A. and A. Nyshadham (2014). Endowments at birth and parents’ investments in children. The
Economic Journal.
Agerbo, E., P. Sullivan, B. Vilhjlmsson, and et al (2015). Polygenic risk score, parental socioeconomic
status, family history of psychiatric disorders, and the risk for schizophrenia: A danish population-based
study and meta-analysis. JAMA Psychiatry 72(7), 635–641.
Almond, D. (2006). Is the 1918 influenza pandemic over? Long term effects of in-utero Influenza exposure
in the post 1940 U.S. population. Journal of Political Economy 114(4), 672–712.
Almond, D. and J. Currie (2011). Killing me softly: The fetal origins hypothesis. The Journal of Economic
Perspectives 25(3), 153.
Almond, D. and B. Mazumder (2013). Fetal origins and parental responses. Annual Review of Eco-
nomics 5(1), 37–56.
Alsever, J. (2017). Can electric ‘brain training’ devices make you smarter? Fortune.
Anderson, C. L., M. Dietz, A. Gordon, and M. Klawitter (2004). Discount rates in Vietnam. Economic
Development and Cultural Change 52(4), 873–887.
Angrist, J. D. and A. B. Keueger (1991). Does compulsory school attendance affect schooling and earnings?
The Quarterly Journal of Economics 106(4), 979–1014.
Angrist, J. D. and A. B. Krueger (1999). Empirical strategies in labor economics. Handbook of labor
economics 3, 1277–1366.
Ardener, S. (1964). The comparative study of rotating credit associations. The Journal of the Royal Anthro-
pological Institute of Great Britain and Ireland 94(2), 201–229.
Ashraf, N., N. Bau, N. Nunn, and A. V oena (2016). Bride price and female education. Technical report,
National Bureau of Economic Research.
Barker, D. J. (1990). The fetal and infant origins of adult disease. BMJ 301(6761), 1111–1111.
Barker, D. J. (1995). Fetal origins of coronary heart disease. BMJ 311(6998), 171–174.
Barker, D. J., J. G. Eriksson, T. Forsen, and C. Osmond (2002). Fetal origins of adult disease: strength of
effects and biological basis. International Journal of Epidemiology 31(6), 1235–1239.
116
Barsky, R. B., F. T. Juster, M. S. Kimball, and M. D. Shapiro (1997). Preference parameters and behavioral
heterogeneity: An experimental approach in the health and retirement study. The Quarterly Journal of
Economics 112(2), pp. 537–579.
Bauer, M. and J. Chytilov´ a (2010). The impact of education on subjective discount rate in Ugandan villages.
Economic Development and Cultural Change 58(4), 643–669.
Bauer, M., J. Chytilov´ a, et al. (2009). Time discounting, education, and growth: Evidence and a simple
model. Czech Journal of Economics and Finance (Finance a uver) 59(1), 71–86.
Becker, G. S. and C. B. Mulligan (1997). The endogenous determination of time preference. The Quarterly
Journal of Economics, 729–758.
Becker, G. S. and N. Tomes (1986). Human capital and the rise and fall of families. Journal of labor
economics, S1–S39.
Beegle, K., J. De Weerdt, and S. Dercon (2006). Kagera health and development survey 2004 basic infor-
mation document. The World Bank.
Beegle, K., R. Dehejia, R. Gatti, S. Krutikova, et al. (2008). The consequences of child labor: evidence from
longitudinal data in rural Tanzania. World Bank Washington, DC.
Ben Halima, B. and M. A. Ben Halima (2009). Time preferences and job search: Evidence from France.
Labour 23(3), 535–558.
Bengtsson, N., S. Peterson, and F. S¨ avje (2013). Revisiting the educational effects of fetal iodine deficiency.
Working Paper.
Benjamin, D. J., S. A. Brown, and J. M. Shapiro (2013). Who is ‘behavioral’? cognitive ability and anoma-
lous preferences. Journal of the European Economic Association 11(6), 1231–1255.
Benjamin, D. J., J. J. Choi, and G. Fisher (2016). Religious identity and economic behavior. Review of
Economics and Statistics 98(4), 617–637.
Benjamin, D. J., J. J. Choi, A. J. Strickland, et al. (2010). Social identity and preferences. The American
Economic Review 100(4), 1913–1928.
Bennett, D., S. A. A. Naqvi, and W.-P. Schmidt (Forthcoming). Learning, hygiene, and traditional medicine.
The Economic Journal.
Bhalotra, S. and A. Venkataramani (2012). Shadows of the captain of the men of death: Early life health
interventions, human capital investments. Technical report, Human Capital Investments and institutions.
Mimeo, University of Bristol.
Bharati, T., S. Chin, and D. Jung (2017). The heterogeneous effect of primary school construction by rainfall
at birth in Indonesia. Working Paper.
Bickel, W. K., A. L. Odum, and G. J. Madden (1999). Impulsivity and cigarette smoking: delay discounting
in current, never, and ex-smokers. Psychopharmacology 146(4), 447–454.
Black, S. E., P. J. Devereux, and K. G. Salvanes (2011). Too young to leave the nest? the effects of school
starting age. Review of Economics and Statistics 93(2), 455–467.
117
Bleakley, H. (2007). Disease and development: Evidence from hookworm eradication in the American
south. The Quarterly Journal of Economics 122(1), 73–117.
Bommier, A. and S. Lambert (2000). Education demand and age at school enrollment in Tanzania. Journal
of Human Resources 35(1), 177–203.
Bonsang, E., S. Adam, and S. Perelman (2012). Does retirement affect cognitive functioning? Journal of
Health Economics 31(3), 490–501.
Bouman, F. J. (1977). Indigenous savings and credit societies in the third world. a message/un enseignement:
Les societes indigenes d’epargne et de credit. Savings and development, 181–219.
Brown, R., V . Montalva, D. Thomas, and A. Vel´ asquez (2017). Impact of violent crime on risk aversion:
Evidence from the Mexican drug war. Technical report, National Bureau of Economic Research.
Bugliari, D., N. Campbell, C. Chan, O. Hayden, M. Hurd, R. Main, J. Mallett, C. McCullough, E. Meijer,
M. Moldoff, et al. (2016). Rand HRS data documentation, Version P. RAND Center for the Study of
Aging: Santa Monica, CA, USA.
Burke, K. and K. Beegle (2004). Why children arent attending school: The case of northwestern Tanzania.
Journal of African Economies 13(2), 333–355.
Callen, M. (2015). Catastrophes and time preference: Evidence from the Indian ocean earthquake. Journal
of Economic Behavior & Organization 118, 199–214.
Callen, M., M. Isaqzadeh, J. D. Long, and C. Sprenger (2014). Violence and risk preference: Experimental
evidence from Afghanistan. The American Economic Review 104(1), 123–148.
Cameron, A. C., J. B. Gelbach, and D. L. Miller (2008). Bootstrap-based improvements for inference with
clustered errors. Review of Economics and Statistics 90(3), 414–427.
Cameron, L. and M. Shah (2015). Risk-taking behavior in the wake of natural disasters. Journal of Human
Resources 50(2), 484–515.
Cao, X.-Y ., X.-M. Jiang, Z.-H. Dou, M. A. Rakeman, M.-L. Zhang, K. O’Donnell, T. Ma, K. Amette,
N. DeLong, and G. R. DeLong (1994). Timing of vulnerability of the brain to iodine deficiency in
endemic cretinism. New England Journal of Medicine 331(26), 1739–1744.
Card, D. (2001). Estimating the return to schooling: Progress on some persistent econometric problems.
Econometrica 69(5), 1127–1160.
Carlsson, M., G. B. Dahl, B.
¨
Ockert, and D.-O. Rooth (2015). The effect of schooling on cognitive skills.
Review of Economics and Statistics 97(3), 533–547.
Cassar, A., A. Healy, and C. V on Kessler (2011). Trust, risk, and time preferences after a natural disaster:
Experimental evidence from Thailand. Unpublished manuscript.
Ceci, S. J. and W. M. Williams (1997). Schooling, intelligence, and income. American Psychologist 52(10),
1051.
Celidoni, M., C. Dal Bianco, and G. Weber (2017). Retirement and cognitive decline. a longitudinal analysis
using SHARE data. Journal of Health Economics 56, 113–125.
118
Chen, M. K. (2013). The effect of language on economic behavior: Evidence from savings rates, health
behaviors, and retirement assets. The American Economic Review 103(2), 690–731.
Chikoyo, R. A. (2010). Assessment of the implementation of primary education development plan (2002-06)
in Mbarali district council, Mbeya region, Tanzania. Ph. D. thesis, Sokoine Universitry of Agriculture.
CIA, U. (2010). The world factbook. Retrieved August 20, 2010.
Coe, N. B., H.-M. von Gaudecker, M. Lindeboom, and J. Maurer (2012). The effect of retirement on
cognitive functioning. Health Economics 21(8), 913–927.
Coe, N. B. and G. Zamarro (2011). Retirement effects on health in Europe. Journal of Health Eco-
nomics 30(1), 77–86.
Coller, M. and M. B. Williams (1999). Eliciting individual discount rates. Experimental Economics 2(2),
107–127.
Counts, G. S. (1931). Education: History. Encyclopaedia of the Social Sciences 5, 403–414.
Cunha, F. and J. Heckman (2007). The technology of skill formation. The American Economic Review 97(2),
31–47.
Cutler, D. M. and A. Lleras-Muney (2006). Education and health: Evaluating theories and evidence. Tech-
nical report, National Bureau of Economic Research.
Cutler, D. M. and A. Lleras-Muney (2010). Understanding differences in health behaviors by education.
Journal of Health Economics 29(1), 1–28.
Davies, G., N. Armstrong, J. C. Bis, J. Bressler, V . Chouraki, S. Giddaluru, E. Hofer, C. A. Ibrahim-Verbaas,
M. Kirin, J. Lahti, et al. (2015). Genetic contributions to variation in general cognitive function: a meta-
analysis of genome-wide association studies in the charge consortium (n= 53 949). Molecular Psychia-
try 20(2), 183.
Davis, L. E., D. C. North, and C. Smorodin (1971). Institutional change and American economic growth.
Number 123. CUP Archive.
De Vreyer, P., L. Sylvie, and M. Thierry (1998). Educating children: a look at family behaviour in Cote
d’Ivoire. Unpublished manuscript.
De Weerdt, J., K. Beegle, H. B. Lilleør, S. Dercon, K. Hirvonen, M. Kirchberger, and S. Krutikova (2012).
Kagera health and development survey 2010: Basic information document. Technical report, Rockwool
Foundation Working Paper Series.
de Wit, H., J. D. Flory, A. Acheson, M. McCloskey, and S. B. Manuck (2007). Iq and nonplanning impulsiv-
ity are independently associated with delay discounting in middle-aged adults. Personality and Individual
Differences 42(1), 111–121.
DellaVigna, S. and M. D. Paserman (2005). Job search and impatience. Journal of Labor Economics 23(3),
527–588.
Doepke, M. and F. Zilibotti (2008). Occupational choice and the spirit of capitalism. The Quarterly Journal
of Economics 123(2), 747–793.
119
Dohmen, T., B. Enke, A. Falk, D. Huffman, U. Sunde, et al. (2015). Patience and the wealth of nations.
Technical report, Working Paper.
Dohmen, T., A. Falk, D. Huffman, and U. Sunde (2010). Are risk aversion and impatience related to
cognitive ability? The American Economic Review 100(3), pp. 1238–1260.
Dohmen, T., D. Huffman, J. Schupp, A. Falk, U. Sunde, and G. G. Wagner (2011). Individual risk atti-
tudes: Measurement, determinants, and behavioral consequences. Journal of the European Economic
Association 9(3), pp. 522–550.
Dohmen, T. J., A. Falk, D. Huffman, and U. Sunde (2012). Interpreting time horizon effects in inter-temporal
choice. Working Paper.
Duflo, E. (2001). Schooling and labor market consequences of school construction in Indonesia: Evidence
from an unusual policy experiment. The American Economic Review 91(4), pp. 795–813.
Duflo, E. (2006). Poor but rational? Understanding poverty, 367–78.
Duflo, E., M. Kremer, and J. Robinson (2011). Nudging farmers to use fertilizer: Theory and experimental
evidence from Kenya. The American Economic Review 101(6), 2350–2390.
Dugbartey, A. T. (1998). Neurocognitive aspects of hypothyroidism. Archives of Internal Medicine 158(13),
1413–1418.
Duncan, J. and A. M. Owen (2000). Common regions of the human frontal lobe recruited by diverse cogni-
tive demands. Trends in Neurosciences 23(10), 475–483.
Easterlin, R. A. (1981). Why isn’t the whole world developed? The Journal of Economic History 41(01),
1–17.
Easterlin, R. A. (2009). Growth triumphant: the twenty-first century in historical perspective. University of
Michigan Press.
Eltom, M., F. Karlsson, A. Kamal, H. Bostr¨ om, and P. Dahlberg (1985). The effectiveness of oral iodized
oil in the treatment and prophylaxis of endemic goiter. The Journal of Clinical Endocrinology &
Metabolism 61(6), 1112–1117.
Falch, T. and S. Sandgren Massih (2011). The effect of education on cognitive ability. Economic In-
quiry 49(3), 838–856.
Field, E., O. Robles, and M. Torero (2009). Iodine deficiency and schooling attainment in Tanzania. Ameri-
can Economic Journal: Applied Economics, 140–169.
Fisher, G., H. Hassan, J. Faul, W. Rogers, and D. Weir (2017). Health and retirement study imputation
of cognitive functioning measures: 1992-2014. Technical report, Survey Research Center, University of
Michigan, Ann Arbor, Michigan.
Fisher, G. G., A. Stachowski, F. J. Infurna, J. D. Faul, J. Grosch, and L. E. Tetrick (2014). Mental work
demands, retirement, and longitudinal trajectories of cognitive functioning. Journal of Occupational
Health Psychology 19(2), 231.
120
Foster, J. C. and G. A. Taylor (1920). The applicability of mental tests to persons over fifty years of age.
Journal of Applied Psychology 4(1), 39–58.
Frederick, S., G. Loewenstein, and T. O’donoghue (2002). Time discounting and time preference: A critical
review. Journal of Economic Literature, 351–401.
Fuchs, V . R. (1982). Time preference and health: An exploratory study. In Economic Aspects of Health, pp.
93–120. University of Chicago Press.
Gallagher, M. and M. T. Koh (2011). Episodic memory on the path to alzheimer’s disease. Current Opinion
in Neurobiology 21(6), 929–934.
Glewwe, P. and H. G. Jacoby (1995). An economic analysis of delayed primary school enrollment in a low
income country: the role of early childhood nutrition. Review of Economics and Statistics, 156–169.
Golsteyn, B. H., H. Gr¨ onqvist, and L. Lindahl (2014). Adolescent time preferences predict lifetime out-
comes. The Economic Journal 124(580).
Grant, J. P. (1978). Disparity reduction rates in social indicators. Overseas Development Council.
Green, D., K. E. Jacowitz, D. Kahneman, and D. McFadden (1998). Referendum contingent valuation,
anchoring, and willingness to pay for public goods. Resource and Energy Economics 20(2), 85–116.
Green, D. A. and W. C. Riddell (2012). Understanding educational impacts: The role of literacy and numer-
acy skills. Working Paper.
Grossman, M. (2006). Education and nonmarket outcomes. Handbook of the Economics of Education 1,
577–633.
Guiso, L., P. Sapienza, and L. Zingales (2013). Time varying risk aversion. Technical report, National
Bureau of Economic Research.
Gunnsteinsson, S., A. Adhvaryu, P. Christian, A. Labrique, J. Sugimoto, A. A. Shamim, and K. P. West Jr
(2016). Resilience to early life shocks. Technical report, Technical report.
Hamoudi, A. (2006). Risk preferences in households and families. Unpublished Manuscript.
Harrison, G. W., M. I. Lau, and M. B. Williams (2002). Estimating individual discount rates in denmark: A
field experiment. The American Economic Review 92(5), 1606–1617.
Heckman, J., R. Pinto, and P. Savelyev (2013). Understanding the mechanisms through which an influential
early childhood program boosted adult outcomes. The American economic review 103(6), 2052–2086.
Heckman, J. J. (2007). The economics, technology, and neuroscience of human capability formation. Pro-
ceedings of the National Academy of Sciences 104(33), 13250–13255.
H’Madoun, M. and W. Nonneman (2012). Time preference, immigrant background and religion. Working
Paper.
Hofacker, D., H. Schroder, Y . Li, and M. Flynn (2016, 01). Trends and determinants of work-retirement
transitions under changing institutional conditions: Germany, england and japan compared. Journal of
Social Policy 45(1), 39–64.
121
HRS Staff (2017). Aging in the 21st century: challenges and opportunities for americans. Ann Arbor:
University of Michigan, Institute for Social Research.
Huber, M. (2014). Identifying causal mechanisms (primarily) based on inverse probability weighting. Jour-
nal of Applied Econometrics 29(6), 920–943.
Ida, T. and R. Goto (2009). Simultaneous measurement of time and risk preferences: Stated preference
discrete choice modeling analysis depending on smoking behavior. International Economic Review 50(4),
1169–1182.
Imai, K., L. Keele, and D. Tingley (2010a). A general approach to causal mediation analysis. Psychological
Methods 15(4), 309.
Imai, K., L. Keele, T. Yamamoto, et al. (2010b). Identification, inference and sensitivity analysis for causal
mediation effects. Statistical Science 25(1), 51–71.
Imbens, G. and J. Angrist (1994). Identification and estimation of local average treatment effects. Econo-
metrica 62(2), 467–475.
Inkeles, A. (1969). Making men modern: On the causes and consequences of individual change in six
developing countries. American Journal of Sociology, 208–225.
Isa, Z., I. Z. Alias, K. A. Kadir, and O. Ali (2000). Effect of iodized oil supplementation on thyroid hormone
levels and mental performance among Orang Asli schoolchildren and pregnant mothers in an endemic
goitre area in Peninsular Malaysia. Asia Pacific Journal of Clinical Nutrition 9(4), 274–281.
Jacoby, H. (1994). Borrowing constraints and progress through school: Evidence from Peru. Review of
Economics and Statistics 76(1), 151–60.
Jakiela, P. and O. W. Ozier (2015). The impact of violence on individual risk preferences: evidence from a
natural experiment. Working Paper.
Jensen, R. and A. Lleras-Muney (2012). Does staying in school (and not working) prevent teen smoking and
drinking? Journal of Health Economics 31(4), 644–657.
Johnson, M. W. and W. K. Bickel (2002). Within-subject comparison of real and hypothetical money rewards
in delay discounting. Journal of the Experimental Analysis of Behavior 77(2), 129–146.
Jun, W. and L. Jianqun (1985). Metabolism of iodized oil after oral administration in guinea pigs. Nutrition
Reports International 31(5), 1085–1092.
Juster, F. T. and R. Suzman (1995). An overview of the health and retirement study. Journal of Human
Resources, S7–S56.
Kahneman, D. and A. Tversky (1981). The simulation heuristic. Technical report, DTIC Document.
Kang, M.-I. and S. Ikeda (2014). Time discounting and smoking behavior: Evidence from a panel survey.
Health Economics 23(12), 1443–1464.
Kaplan, J. T., S. I. Gimbel, and S. Harris (2016). Neural correlates of maintaining one’s political beliefs in
the face of counterevidence. Scientific Reports 6, 39589.
122
Khwaja, A., D. Silverman, and F. Sloan (2007). Time preference, time discounting, and smoking decisions.
Journal of Health Economics 26(5), 927–949.
Kirby, K. N., R. Godoy, V . Reyes-Garcia, E. Byron, L. Apaza, W. Leonard, E. Perez, V . Vadez, and D. Wilkie
(2002). Correlates of delay-discount rates: Evidence from Tsimane’ Amerindians of the Bolivian rain
forest. Journal of Economic Psychology 23(3), 291–316.
Knight, J. B. and R. H. Sabot (1990). Education, productivity, and inequality: The East African natural
experiment. Oxford University Press.
Knoch, D. and E. Fehr (2007). Resisting the power of temptations. Annals of the New York Academy of
Sciences 1104(1), 123–134.
Kondylis, F. and M. Manacorda (2012). School proximity and child labor evidence from rural Tanzania.
Journal of Human Resources 47(1), 32–63.
Koopmans, T. C. (1960). Stationary ordinal utility and impatience. Econometrica, 287–309.
Koopmans, T. C., P. A. Diamond, and R. E. Williamson (1964). Stationary utility and time perspective.
Econometrica: Journal of the Econometric Society, 82–100.
Kudo, Y . (2015). Female migration for marriage: Implications from the land reform in rural Tanzania. World
Development 65, 41–61.
K¨ uhberger, A., M. Schulte-Mecklenbeck, and J. Perner (2002). Framing decisions: Hypothetical and real.
Organizational Behavior and Human Decision Processes 89(2), 1162–1175.
Lamberg, B.-A. (1991). Endemic goitre-iodine deficiency disorders. Annals of Medicine 23(4), 367–372.
Lawrance, E. C. (1991). Poverty and the rate of time preference: evidence from panel data. Journal of
Political economy 99(1), 54–77.
Leigh, J. P. (1986). Accounting for tastes: Correlates of risk and time preferences. Journal of Post Keynesian
Economics 9(1), pp. 17–31.
Li, H., M. Rosenzweig, and J. Zhang (2010). Altruism, favoritism, and guilt in the allocation of family
resources: Sophie’s choice in mao’s mass send-down movement. Journal of Political Economy 118(1),
1–38.
Locey, M. L., B. A. Jones, and H. Rachlin (2011). Real and hypothetical rewards. Judgment and Decision
Making 6(6), 552.
Maccini, S. and D. Yang (2009). Under the weather: Health, schooling, and economic consequences of
early-life rainfall. The American Economic Review 99(3), pp. 1006–1026.
MacKinnon, D. P., C. M. Lockwood, C. H. Brown, W. Wang, and J. M. Hoffman (2007). The intermediate
endpoint effect in logistic and probit regression. Clinical Trials 4(5), 499–513.
Madden, G. J., A. M. Begotka, B. R. Raiff, and L. L. Kastern (2003). Delay discounting of real and
hypothetical rewards. Experimental and Clinical Psychopharmacology 11(2), 139.
123
Malamud, O., C. Pop-Eleches, and M. Urquiola (2016). Interactions between family and school envi-
ronments: Evidence on dynamic complementarities? Technical report, National Bureau of Economic
Research.
Maluccio, J. A., J. Hoddinott, J. R. Behrman, R. Martorell, A. R. Quisumbing, and A. D. Stein (2009).
The impact of improving nutrition during early childhood on education among Guatemalan adults. The
Economic Journal 119(537), 734–763.
Manski, C. F. (2009). Identification for prediction and decision. Harvard University Press.
Mason, A. and S. Khandker (1997). Household schooling decisions in Tanzania. World Bank (mimeo).
Mazzonna, F. and F. Peracchi (2012). Ageing, cognitive abilities and retirement. European Economic
Review 56(4), 691–710.
Mazzonna, F. and F. Peracchi (2017). Unhealthy retirement? Journal of Human Resources 52(1), 128–151.
Meier, S. and C. D. Sprenger (2015). Temporal stability of time preferences. Review of Economics and
Statistics 97(2), 273–286.
Metcalfe, J. and W. Mischel (1999). A hot/cool-system analysis of delay of gratification: dynamics of
willpower. Psychological Review 106(1), 3.
Miguel, E. and M. Kremer (2004). Worms: identifying impacts on education and health in the presence of
treatment externalities. Econometrica 72(1), 159–217.
Miller, B. L. and J. L. Cummings (2017). The human frontal lobes: Functions and disorders. Guilford
Publications.
Ministy of Education and Culture (1995). Education and training policy. Technical report, Dar es Salaam:
United Republic of Tanzania.
Ng, J. (2013). Elicited risk and time preferences: the role of demographics, cognition, and interviewers.
University of Southern California Job Market Paper.
Ngandu, T., J. Lehtisalo, A. Solomon, E. Lev¨ alahti, S. Ahtiluoto, R. Antikainen, L. B¨ ackman, T. H¨ anninen,
A. Jula, T. Laatikainen, et al. (2015). A 2-year multidomain intervention of diet, exercise, cognitive
training, and vascular risk monitoring versus control to prevent cognitive decline in at-risk elderly people
(finger): a randomised controlled trial. The Lancet 385(9984), 2255–2263.
Non, A. and D. Tempelaar (2016). Time preferences, study effort, and academic performance. Economics
of Education Review 54, 36–61.
North, D. C. (1973). The rise of the western world: A new economic history. Cambridge University Press.
Ofstedal, M. B., G. G. Fisher, and A. R. Herzog (2005). Documentation of cognitive functioning measures
in the health and retirement study. Ann Arbor, MI: University of Michigan 10.
Okbay, A., J. P. Beauchamp, M. A. Fontana, J. J. Lee, T. H. Pers, C. A. Rietveld, P. Turley, G.-B. Chen,
V . Emilsson, S. F. W. Meddens, et al. (2016). Genome-wide association study identifies 74 loci associated
with educational attainment. Nature 533(7604), 539.
124
Olson, M. and M. J. Bailey (1981). Positive time preference. Journal of Political Economy 89(1), 1–25.
Oreopoulos, P. and K. G. Salvanes (2011). Priceless: The nonpecuniary benefits of schooling. The Journal
of Economic Perspectives, 159–184.
Pender, J. L. (1996). Discount rates and credit markets: Theory and evidence from rural India. Journal of
Development Economics 50(2), 257–296.
Perez-Arce, F. (2017). The effect of education on time preferences. Economics of Education Review 56, 52
– 64.
Peters, K. (1990). Indonesia: strategy for a sustained reduction in poverty. Washington: World Bank.
Peterson, S. (2000). Controlling iodine deficiency disorders: studies for program management in Sub-
Saharan Africa. Ph. D. thesis, Acta Universitatis Upsaliensis.
Peterson, S., V . Assey, B. C. Forsberg, T. Greiner, F. P. Kavishe, B. Mduma, H. Rosling, A. B. Sanga, and
M. Gebre-Medhin (1999). Coverage and cost of iodized oil capsule distribution in Tanzania. Health
Policy and Planning 14(4), 390–399.
Pigou, A. C. (1952). Essays in economics. London: Macmillan.
Pitt, M. M., M. R. Rosenzweig, and M. N. Hassan (2012). Human capital investment and the gender division
of labor in a brawn-based economy. The American Economic Review 102(7), 3531–3560.
Pop, V . J., J. L. Kuijpens, A. L. van Baar, G. Verkerk, M. M. van Son, J. J. de Vijlder, T. Vulsma, W. M.
Wiersinga, H. A. Drexhage, and H. L. Vader (1999). Low maternal free thyroxine concentrations during
early pregnancy are associated with impaired psychomotor development in infancy. Clinical Endocrinol-
ogy 50(2), 149–155.
Qiu, C. and L. Fratiglioni (2015). A major role for cardiovascular burden in age-related cognitive decline.
Nature Reviews Cardiology 12(5), 267.
Ramsey, F. P. (1928). A mathematical theory of saving. The Economic Journal 38(152), 543–559.
Read, D., G. Loewenstein, M. Rabin, G. Keren, and D. Laibson (1999). Choice bracketing. In Elicitation of
Preferences, pp. 171–202. Springer.
Rebok, G. W., K. Ball, L. T. Guey, R. N. Jones, H.-Y . Kim, J. W. King, M. Marsiske, J. N. Morris, S. L.
Tennstedt, F. W. Unverzagt, et al. (2014). Ten-year effects of the advanced cognitive training for inde-
pendent and vital elderly cognitive training trial on cognition and everyday functioning in older adults.
Journal of the American Geriatrics Society 62(1), 16–24.
Rohwedder, S. and R. J. Willis (2010). Mental retirement. Journal of Economic Perspectives 24(1), 119–38.
Roodman, D. (2009). A note on the theme of too many instruments. Oxford Bulletin of Economics and
statistics 71(1), 135–158.
Rosenberg, N., L. E. Birdzell, and G. W. Mitchell (1986). How the West grew rich. Popular Prakashan.
Rosenzweig, M. R. and T. P. Schultz (1982). Market opportunities, genetic endowments, and intrafamily
resource distribution: Child survival in rural India. The American Economic Review, 803–815.
125
Rosenzweig, M. R. and K. I. Wolpin (1980). Testing the quantity-quality fertility model: The use of twins
as a natural experiment. Econometrica 48(1), 227–240.
Rossin-Slater, M. and M. W¨ ust (2016). What is the added value of preschool? long-term impacts and
interactions with a health intervention. Technical report, National Bureau of Economic Research.
Rubalcava, L., G. Teruel, and D. Thomas (2009). Investments, time preferences, and public transfers paid to
women. Economic Development and Cultural Change 57(3), 507–538.
Rzezak, P., P. Squarzoni, F. L. Duran, T. d. T. F. Alves, J. Tamashiro-Duran, C. M. Bottino, S. Ribeiz, P. A.
Lotufo, P. R. Menezes, M. Scazufca, et al. (2015). Relationship between brain age-related reduction in
gray matter and educational attainment. PloS one 10(10), e0140945.
Salthouse, T. A. (1991). Theoretical Perspectives on Cognitive Aging. Psychology Press.
Salthouse, T. A. (2006). Mental exercise and mental aging: Evaluating the validity of the “use it or lose it”
hypothesis. Perspectives on Psychological Science 1(1), 68–87.
Salthouse, T. A., R. L. Babcock, E. Skovronek, D. R. Mitchell, and R. Palmon (1990). Age and experience
effects in spatial visualization. Developmental Psychology 26(1), 128.
Sen, A. (1984). The living standard. Oxford Economic Papers, 74–90.
Shapiro, J. M. (2005). Is there a daily discount rate? evidence from the food stamp nutrition cycle. Journal
of public Economics 89(2), 303–325.
Smith, G. E., P. Housen, K. Yaffe, R. Ruff, R. F. Kennison, H. W. Mahncke, and E. M. Zelinski (2009). A
cognitive training program based on principles of brain plasticity: results from the improvement in mem-
ory with plasticity-based adaptive cognitive training (impact) study. Journal of the American Geriatrics
Society 57(4), 594–603.
Social Security Administration (2017). Normal retirement age.
Sommer, A., E. Djunaedi, A. Loeden, I. Tarwotjo, K. West, R. Tilden, L. Mele, A. S. Group, et al. (1986).
Impact of vitamin a supplementation on childhood mortality: a randomised controlled community trial.
The Lancet 327(8491), 1169–1173.
Stern, Y . (2002). What is cognitive reserve? theory and research application of the reserve concept. Journal
of the International Neuropsychological Society 8(3), 448–460.
Stern, Y . (2012). Cognitive reserve in ageing and alzheimer’s disease. The Lancet Neurology 11(11), 1006–
1012.
Stigler, G. J. and G. S. Becker (1977). De gustibus non est disputandum. The American Economic Review,
76–90.
Strauss, J., F. Witoelar, and B. Sikoki (2016). The fifth wave of the Indonesia family life survey: Overview
and field report.
Streeten, P., S. J. Burki, U. Haq, N. Hicks, and F. Stewart (1981). First things first: meeting basic human
needs in the developing countries. New York Oxford Univ. Press 1981.
126
Sutter, M., M. G. Kocher, D. Gl¨ atzle-R¨ utzler, and S. T. Trautmann (2013). Impatience and uncertainty:
Experimental decisions predict adolescents’ field behavior. The American Economic Review 103(1), 510–
531.
Tanaka, T., C. F. Camerer, and Q. Nguyen (2010). Risk and time preferences: Linking experimental and
household survey data from Vietnam. The American Economic Review 100(1), 557.
Untoro, J., W. Schultink, R. Gross, C. E. West, and J. G. Hautvast (1998). Efficacy of different types of
iodised oil. The Lancet 351(9104), 752–753.
Van der Haar, F., P. Kavishe, and M. G. Medhin (1988). The public health importance of IDD in Tanzania.
The Central African Journal of Medicine 34(3), 60–65.
V oors, M. J., E. E. Nillesen, P. Verwimp, E. H. Bulte, R. Lensink, and D. P. Van Soest (2012). Violent
conflict and behavior: a field experiment in Burundi. The American Economic Review 102(2), 941–964.
Wolff, J. (2001). Physiology and pharmacology of iodized oil in goiter prophylaxis. Medicine 80(1), 20–36.
Zimmermann, M. B., I. Aeberli, T. Torresani, and H. B¨ urgi (2005). Increasing the iodine concentration in the
Swiss iodized salt program markedly improved iodine status in pregnant women and children: a 5-year
prospective national study. The American Journal of Clinical Nutrition 82(2), 388–392.
127
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Three essays on health & aging
PDF
Essays on health and aging with focus on the spillover of human capital
PDF
Essays on economics of education and private tutoring
PDF
Essays on development economics
PDF
Essays on the empirics of risk and time preferences in Indonesia
PDF
Essays on health economics
PDF
Essays on education and institutions in developing countries
PDF
Three essays in health economics
PDF
Essays in development and experimental economics
PDF
Essays on health insurance programs and policies
PDF
Three essays in education and urban economics
PDF
Essays on human capital accumulation -- health and education
PDF
Selection and impacts of early life events on later life outcomes
PDF
Essays in political economy and mechanism design
PDF
Essays in development economics
PDF
Beyond revealed preferences: how gender, relative socioeconomic status, and social norms drive happiness and behavior
PDF
Three essays on behavioral economics approaches to understanding the implications of mental health stigma
PDF
Three essays on economics of early life health in developing countries
PDF
Essays on labor and development economics
PDF
Essays on development and health economics: social media and education policy
Asset Metadata
Creator
Jung, Dawoon
(author)
Core Title
Essays on development and health economics
School
College of Letters, Arts and Sciences
Degree
Doctor of Philosophy
Degree Program
Economics
Publication Date
06/21/2018
Defense Date
04/12/2018
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
Development,education,health,Indonesia,Iodine,OAI-PMH Harvest,Retirement,tanzania,time preference
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Strauss, John (
committee chair
), Lee, Jinkook (
committee member
), Meijer, Erik (
committee member
), Nugent, Jeffrey (
committee member
)
Creator Email
dawoonju@gmail.com,dawoonju@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c40-506189
Unique identifier
UC11266862
Identifier
etd-JungDawoon-6340.pdf (filename),usctheses-c40-506189 (legacy record id)
Legacy Identifier
etd-JungDawoon-6340.pdf
Dmrecord
506189
Document Type
Dissertation
Format
application/pdf (imt)
Rights
Jung, Dawoon
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Tags
education
time preference