ABSTRACT
Title of dissertation: SHOCKS AND HUMAN CAPITAL
IN DEVELOPING COUNTRIES
Aaron Szott, Doctor of Philosophy, 2014
Dissertation directed by: Professor Judith Hellerstein
Department of Economics
I begin my dissertation by introducing the following two chapters. I start out
by describing the basic theory underlying economists’ historical interest in the e↵ects
of shocks in developing countries. I then briefly review the empirical literature on
household responses to shocks and outline how it is reasonable to expect that even
recurring, non-exotic shocks have substantial permanent e↵ects on a↵ected young
peoples’ completed human capital stocks. Next, I describe the contributions that
the following two chapters of this dissertation make and how they are similar with
respect to their use of nationally representative household survey data, their policy
relevance, and the way they use basic economic theory and contextual knowledge to
uncover meaningful impacts of shocks on di↵erent groups of young people. Finally,
I conclude this introductory chapter by considering whether or not it should be
regarded as a surprise that shocks in developing countries can be allowed by house-
holds to impact a↵ected children’s human capital stocks the way they do, despite
the large returns associated with human capital investments.
The second chapter considers the permanent e↵ects of rainfall shocks on adults’
completed human capital stocks. Existing research suggests that health and educa-
tion investments in children are a↵ected by aggregate income shocks, but there is
considerably less evidence on what the e↵ects of many years’ worth of such shocks
are on individuals’ completed human capital stocks. I contribute to this literature by
studying the e↵ects of the number and timing of all the unusually wet and dry years
over the first 21 years of rural West African individuals’ lives on their likelihood of
being literate and their completed educational attainment. I use historical rainfall
data merged with nationally representative household surveys conducted in 12 coun-
tries, and I find that both wet and dry years have substantial negative impacts on
women’s human capital and smaller positive impacts on men’s: The average e↵ect
of wet and dry rainfall shocks over the first 21 years of life is a 22 percent decrease
in females’ likelihood of being literate and a 10 percent increase in males’. I argue
that this pattern of results is consistent with existing research on how West African
females and males work and acquire human capital.
The third chapter provides evidence on how children in di↵erent types of house-
holds are a↵ected by food price shocks. While people in poor countries report that
inflation is one of their primary concerns, there is not much evidence on how they are
actually a↵ected by it. In particular, it is not clear how food (along with other) price
inflation a↵ects individuals in net food producing and net food consuming house-
holds. I use four highly comparable household surveys collected in Egypt to examine
how children’s weights-for-height evolved over time and in the face of the food price
crisis of 2008, and I utilize data on region of residence and parents’ occupation to
examine how changes over time di↵ered by household net food consumption sta-
tus. I find that despite several years’ worth of solid economic growth in the run-up
to 2008, most region- and parental occupation-based groups of Egyptian children’s
weights-for-height had not increased at all since 2005 (for example). Quantile re-
gression results reveal that the lightest children in 2008 were considerably lighter
than the lightest children in 2005, and also that children in those households most
likely to have been net food producing seem to have been more negatively a↵ected
by high food prices than children in most other kinds of households. High food
prices seem to have o↵set any possible beneficial e↵ects of growth for children in the
poorest households.
SHOCKS AND HUMAN CAPITAL
IN DEVELOPING COUNTRIES
by
Aaron Szott
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2014
Advisory Committee:
Professor Judith Hellerstein, Chair/Advisor
Professor Raymond Guiteras
Professor Jessica Goldberg
Professor Peter Murrell
Professor Kenneth Leonard
c  Copyright by
Aaron Szott
2014
Dedication
I dedicate this dissertation to Grace Lopez.
ii
Acknowledgments
First and foremost I’d like to thank my advisor, Professor Judy Hellerstein.
Any advisor in economics will continually push their students to think more carefully
about their work. This is much easier for a student to deal with when the advisor
is cool.
I would also like to thank my other committee members for all of the guidance
and feedback they’ve provided me. I am proud to have been trained to understand
every detail of my work and to endlessly try to get things right, and of course inter-
actions with economists are a crucial part of the training I’ve received. In particular,
I would like to thank Raymond Guiteras, Jessica Goldberg, Roger Betancourt, Peter
Murrell, Seth Sanders and Chris McKelvey for helping me with various aspects of my
work. Thank you also to Matthew Neidell for all of the opportunities he has given
me. Finally, thank you to The Economist for giving me a sense of what economics
was and how powerful a set of tools it is, as well as to my dad for recommending
that I pick up a copy.
I would like to thank my fellow graduate students who also became my friends:
Abby Alpert, Paul Bailey, Juan Bonilla-Angel, Randy Chugh, Carolina Gonzalez-
Velosa, Amy Knaup and Gabriel Lara-Ibarra. You guys are all great and I miss
you.
Thank you to my family, who are of course by far the greatest contributors to
anything I ever have or will achieve in my life. Thanks also to Grace Lopez, you’re
the best.
iii
Finally, I would like to thank the big homie Matthew Avedon. Actually, shouts
out to everyone who was at St. Vincent’s Hospital in New York City the day Matt
was born there. Check’s in the mail buddy, I’ll holler.
iv
Table of Contents
List of Tables vii
List of Figures viii
List of Abbreviations ix
1 Introduction 1
1.1 Shocks and Basic Economic Theory . . . . . . . . . . . . . . . . . . . 1
1.2 Shocks and Investments in Children’s Human Capital . . . . . . . . . 5
1.3 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Rainfall Shocks and Human Capital 15
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Background and Conceptual Issues . . . . . . . . . . . . . . . . . . . 21
2.2.1 The Agricultural Labor Cycle in West Africa . . . . . . . . . . 21
2.2.2 Conceptual Framework and Existing Evidence . . . . . . . . . 24
2.2.2.1 Shocks and School-Aged Children . . . . . . . . . . . 24
2.2.2.2 Shocks and Preschool-Aged Children . . . . . . . . . 27
2.3 Empirical Strategy, Data and Sample . . . . . . . . . . . . . . . . . . 31
2.3.1 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.3 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4 Descriptive Statistics and Results . . . . . . . . . . . . . . . . . . . . 42
2.4.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . 42
2.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4.2.1 Baseline Results . . . . . . . . . . . . . . . . . . . . 44
2.4.2.2 Shocks at Preschool Ages . . . . . . . . . . . . . . . 51
2.4.2.3 Shocks and Schooling Along the Extensive Margin . 55
2.4.2.4 Shocks Numbers Indicators Results . . . . . . . . . . 57
2.4.3 E↵ects by Agroecological Zone . . . . . . . . . . . . . . . . . . 60
2.5 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.5.1 Region-Specific Cohort Fixed E↵ects . . . . . . . . . . . . . . 62
v
2.5.2 Sample Selection Bias Robustness Check . . . . . . . . . . . . 64
2.5.2.1 Shocks, Migration and Mortality: Theory . . . . . . 64
2.5.2.2 Shocks and Total Changes in Neighborhood-Cohort
Sizes: Empirical Evidence . . . . . . . . . . . . . . . 67
2.5.2.3 Shocks and Mortality: Empirical Evidence . . . . . . 70
2.5.2.4 Sample Selection and Selection-Free Estimates . . . . 74
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3 Food Price Inflation and Children’s Health 99
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.2 Background and Conceptual Issues . . . . . . . . . . . . . . . . . . . 104
3.2.1 The 2008 Food Price Crisis in Egypt . . . . . . . . . . . . . . 104
3.2.2 Agriculture in Egypt . . . . . . . . . . . . . . . . . . . . . . . 107
3.2.3 Conceptual Issues and Existing Evidence . . . . . . . . . . . . 109
3.3 Empirical Analysis, Data and Sample . . . . . . . . . . . . . . . . . . 117
3.3.1 Empirical Approach . . . . . . . . . . . . . . . . . . . . . . . 117
3.3.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.4 Descriptive Statistics and Results . . . . . . . . . . . . . . . . . . . . 123
3.4.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . 123
3.4.2 Ordinary Least Squares Results . . . . . . . . . . . . . . . . . 124
3.4.2.1 Region-Specific OLS Results . . . . . . . . . . . . . . 124
3.4.2.2 Region and Parental Occupation-Specific OLS Results128
3.4.3 Conditional Quantile Regression Results . . . . . . . . . . . . 131
3.4.3.1 Region-Specific Quantile Regression Results . . . . . 131
3.4.3.2 Region and Parental Occupation-Specific Quantile
Regression Results . . . . . . . . . . . . . . . . . . . 139
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
A E↵ects of Shocks at Particular Ages 179
Bibliography 183
vi
List of Tables
2.1 Household Survey Countries and Years of Interview . . . . . . . . . . 86
2.2 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
2.3 Baseline Human Capital Results for Women and Men . . . . . . . . 88
2.4 Women’s Height Results . . . . . . . . . . . . . . . . . . . . . . . . . 89
2.5 Schooling Attainment (Extensive Margin) Results . . . . . . . . . . . 90
2.6 Shocks Numbers Indicators Results . . . . . . . . . . . . . . . . . . . 91
2.7 Baseline Results by Agroecological Zone . . . . . . . . . . . . . . . . 92
2.8 Region-Specific Birth Year Fixed E↵ects Results . . . . . . . . . . . 93
2.9 Rainfall Shocks and Numbers of Living Non-Migrants . . . . . . . . 94
2.10 Rainfall Shocks and (Sibling) Mortality Results . . . . . . . . . . . . 95
2.11 E↵ects of Shocks on Mortality and Implied E↵ects on Migration, Per
Average Rainfall Neighborhood . . . . . . . . . . . . . . . . . . . . . 96
2.12 Baseline Literacy Estimates and Their Sample Selection-Free Analogs 97
2.13 Baseline Highest Grade Completed Estimates and Their Sample Selection-
Free Analogs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3.1 Demographic and Health Survey Years and Months . . . . . . . . . . 160
3.2 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
3.3 Region-Specific OLS Results . . . . . . . . . . . . . . . . . . . . . . 163
3.4 Region and Parental Occupation-Specific OLS Results . . . . . . . . 165
3.5 Region-Specific Quantile Regression Baseline Results . . . . . . . . . 167
3.6 Region-Specific Quantile Regression Enhanced Specification Results . 168
3.7 Region-Specific Inter-Quantile Regression Results . . . . . . . . . . . 170
3.8 Region and Parental Occupation-Specific Baseline Quantile Regres-
sion Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
3.9 Region and Parental Occupation-Specific Enhanced Specification Quan-
tile Regression Results . . . . . . . . . . . . . . . . . . . . . . . . . . 173
3.10 Region and Parental Occupation-Specific Conditional Weight-for-Height
2008 - 2005 Di↵erences . . . . . . . . . . . . . . . . . . . . . . . . . 176
3.11 Region and Parental Occupation-Specific Inter-Quantile Regression
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
A.1 Shocks Ages Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
vii
List of Figures
2.1 Distribution of Women’s and Men’s Highest Grades Completed . . . 81
2.2 Average Rainfall Levels Over Time . . . . . . . . . . . . . . . . . . . 82
2.3 Mean Numbers of Annual Rainfall Levels Experienced Between Birth
and Age 21 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.4 Literacy Rates by Highest Grade Completed . . . . . . . . . . . . . . 84
2.5 Age Heaping for Women’s and Men’s Samples . . . . . . . . . . . . . 85
3.1 Kernel Density, Weight-for-Height Z-Score, By Year . . . . . . . . . 153
3.2 GDP Growth Per Capita and Food Price Inflation in Egypt, 1997-
2009 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
3.3 Weights-for-Height Over Time, By Region . . . . . . . . . . . . . . . 154
3.4 Weight-for-Height Changes for Urban Children, By Conditional Quan-
tile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
3.5 Weight-for-Height Changes for Lower Rural Children, By Conditional
Quantile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
3.6 Weight-for-Height Changes for Upper Rural Children, By Conditional
Quantile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
3.7 Weight-for-Height Changes for Urban Children, By Conditional Quan-
tile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
3.8 Weight-for-Height Changes for Lower Rural Children With Self-Employed
Farmer Parents, By Conditional Quantile . . . . . . . . . . . . . . . 157
3.9 Weight-for-Height Changes for Lower Rural Children With Agricul-
tural Employee Parents, By Conditional Quantile . . . . . . . . . . . 157
3.10 Weight-for-Height Changes for Lower Rural Children With Non-Agricultural
Worker Parents, By Conditional Quantile . . . . . . . . . . . . . . . 158
3.11 Weight-for-Height Changes for Upper Rural Children With Self-Employed
Farmer Parents, By Conditional Quantile . . . . . . . . . . . . . . . 158
3.12 Weight-for-Height Changes for Upper Rural Children With Agricul-
tural Employee Parents, By Conditional Quantile . . . . . . . . . . . 159
3.13 Weight-for-Height Changes for Upper Rural Children With Non-Agricultural
Worker Parents, By Conditional Quantile . . . . . . . . . . . . . . . 159
viii
List of Abbreviations
GDP Gross Domestic Product
ICRISAT International Crops Research Institute for the Semi-Arid Tropics
OLS Ordinary Least Squares
ix
Chapter 1: Introduction
1.1 Shocks and Basic Economic Theory
Aggregate shocks in developing countries have been shown to result in a wide
variety of e↵ects and responses. This thesis will present evidence of hitherto un-
known or underappreciated e↵ects of shocks, but the motivation for the associated
empirical analyses of these additional e↵ects is based on consideration of what is
already known about shocks. A natural starting point when thinking about the
consequences of shocks is to consider their contemporaneous e↵ects on utility. One
of the most commonly made and intuitively appealing assumptions in economics
is that utility functions are concave. In other words, marginal utility is a decreas-
ing function of the quantity consumed. A direct implication of this assumption is
that individuals would prefer quantities consumed to be highly similar across time
periods. When marginal utility decreases with consumption, the sum of low con-
sumption period utilities and high consumption period utilities will be less than the
sum of period utilities with smooth consumption.
Under this assumption, we might expect that the contemporaneous utility
e↵ects of shocks are particularly severe in developing countries. A characteristic
feature of developing economies is that certain markets are either incomplete or
1
missing. In particular, formal markets for credit or insurance are typically assumed
to be lacking or absent; the sorts of information and record-keeping necessary for
these kinds of markets to function has historically simply not been present in devel-
oping country settings. More specifically, Besley (1995) cites the specification and
enforcement of contracts, legal uncertainty, and inadequate communication-related
technology as binding constraints preventing the functioning of proper credit and
insurance markets. In the absence of informal responses such as the sales of valu-
able assets or borrowing from other households, then, individuals and households in
these settings would therefore be vulnerable to the sorts of shock-driven intertempo-
ral variation in quantities consumed described above. Also, although missing credit
or insurance markets are potentially large parts of the reason why shocks might
have serious consequences in developing countries, there are other features of de-
veloping economies that likely imply substantial impacts of shocks. For example,
almost by definition, opportunities to earn money outside of farming in rural areas
of developing countries will typically be limited. This of course implies that chances
to compensate for a disappointing harvest via increases in o↵-farm labor supply
will also be limited. Furthermore, the low baseline consumption levels of people in
developing countries coupled with the assumption of decreasing marginal utility will
mean that the welfare consequences of a given shock will be felt more dearly as well.
That is, when marginal utility is high, small changes to consumption quantities will
have larger impacts.
How large are the contemporaneous utility e↵ects of shocks likely to be in prac-
tice? Despite the absence of formal credit or insurance, empirical evidence shows
2
that household consumption in developing countries is at least partially protected
from the e↵ects of shocks. Townsend (1995) finds that while most of the rural In-
dian households studied using the ICRISAT dataset are e↵ectively insured against
idiosyncratic shocks, about one-sixth of them are not. The quality of insurance
against idiosyncratic shocks has also been found to vary by shock type, with Gertler
and Gruber (2002) and Fafchamps and Lund (2003) finding that a key household
member’s sickness is less well-insured than is their job loss or death. Also, Lim and
Townsend (1998) find that the poor are less well-insured than are the rich. Aggre-
gate shocks have also been found to result in substantial fluctuations in quantities
consumed (Frankenburg et al., 2003), even if consumption does not very closely
track aggregate income in some cases (Paxson, 1992). In sum, the general conclu-
sion is that consumption in developing country settings is only partly smooth in the
face of shocks, which implies in turn that contemporaneous utility will tend to be
negatively a↵ected to some extent.
But unsurprisingly in developing country settings, where markets are some-
times poorly functioning and the institutions and responses that pick up the slack
are many and varied, contemporaneous utility is not the only outcome a↵ected by
shocks. Indeed, poor households are believed to engage in costly ”income smooth-
ing”, or the ex ante avoidance of profitable but risky income-generating activities
(Morduch, 1995). It has been shown, for example, that the riskiness of various agri-
cultural production activities prevents poor households from engaging in those ac-
tivities, even when expected returns are higher (Rosenzweig and Binswanger, 1993).
Bardhan (1983) argues that ”tied labor” arrangements, whereby employees receive
3
low but steady wages for long periods of time, can be understood as a rational
response to a risky environment. Fink et al. (2014) find evidence that apparently
sub-optimally large allocations of o↵-farm labor are driven by the absence of seasonal
savings and credit.
Again, responses to risk and shocks in developing countries have been found
to be numerous and creative. Rosenzweig and Wolpin (1993) find evidence that
agricultural households in India rely on ”bu↵er stocks” in the form of bullocks to
cope with production shocks (although subsequent research by Fafchamps, Udry and
Czukas (1998) and Lim and Townsend (1998) failed to find evidence of this kind of
behavior). There is also a wide variety of substitutes for formal credit and insur-
ance, such as group lending programs where a number of loan recipients are jointly
liable for repayment (Stiglitz, 1990), credit cooperatives that borrow as a collective
and then distribute loan money to members (Banerjee, Besley and Guinnane, 1994),
and rotating savings and credit associations that allocate group savings to members
based on lottery or a bidding process (Besley, Coate and Loury, 1993). Udry (1994)
studies informal credit arrangements between friends and family members in Nigeria,
where lending and borrowing (and sometimes both simultaneously) are extremely
common. Fafchamps and Lund (2003) study risk sharing in the Philippines and
find that informal insurance networks largely consist of friends and relatives, while
Grimard (1997) finds that ethnic groups constitute the mutual insurance network
in Cote d’Ivoire. In addition to sales of jewelry, Frankenburg et al. (2003) find that
Indonesian households adjusted their membership sizes and members increased la-
bor supply in response to the economic crisis of 1998. Kochar (1999) also finds that
4
Indian farmers increase o↵-farm labor supply in response to harvest shocks. Rosen-
zweig and Stark (1989) finds that Indian daughters are married o↵ from households
in risky areas to households in areas where weather conditions typically di↵er, so
that sending and receiving households each have better transfer sources. Jacoby
and Skoufias (1997), Jensen (2000) and Bjorkman-Nyqvist (2013) each find that
children’s school attendance fluctuates with income. We have seen evidence that
shocks in developing countries tend to result in both contemporaneous utility fluc-
tuations as well as costly risk-avoiding productivity choices. Are there other types
of costs associated with this wide variety of ways that households in developing
countries have been found to respond to shocks?
1.2 Shocks and Investments in Children’s Human Capital
Existing research suggests that the last coping strategy described above, whereby
households temporarily pull children out of school in response to shocks, does not
result in permanently lower human capital stocks for a↵ected children (Jacoby and
Skoufias, 1997; Funkhouser, 1999). On the other hand, de Janvry et al. (2006)
hypothesize that if not being enrolled in school causes children to forget either the
material they last learned or how to learn more generally, then re-enrollment will be
su ciently costly that some children will not do so. Turning back to the possibility
that shocks could result in consumption fluctuations, then given that consumption
of basic goods like food for young children has an investment component, it follows
5
that this is another way in which shocks can a↵ect investments in children. 1
More generally, investments in children’s human capital requires resources just as
consumption or other investments do, so it is natural to wonder whether or not
they will be a↵ected by income shocks. Also, human capital stocks are sometimes
su ciently low and marginal utility su ciently high in developing country settings
that it seems reasonable to wonder whether or not investments would be sacrificed
for current consumption in response to shocks. In sum, it certainly seems reasonable
to expect that the costs of shocks could include decreased investments in children’s
human capital.
The existing evidence makes clear that investments in children’s human cap-
ital have potentially substantial permanent welfare consequences. Maccini and
Yang (2009) find that women’s adult height, schooling attainment, and a house-
hold durable good index are each positively a↵ected by rainfall in the first year of
life, and argue that this is because rural incomes increase with rainfall. In a strik-
ing illustration of the value of parental time inputs, Gertler et al. (2013) find that
one-hour weekly visits by health workers to the homes of stunted Jamaican children
that encouraged mothers to develop their children’s cognitive and personality skills
resulted in 42 percent increases in those children’s earnings 20 years later. Child
nutrition has also been found to have consequences for adult outcomes. Maluccio
et al. (2009) find that a high-protein energy drink given to Guatemalan children
under the age of 8 results in increases in cognitive ability amongst adults of 0.24
1In principle, children’s nutritional intake could be protected even if shocks decrease household
consumption. It is therefore an empirical question as to whether or not investments are a↵ected.
6
standard deviations. Hoddinott et al. (2008) study the same intervention and find
an increase in male wages of 46 percent as a result of nutritional supplementation
in the first two years of life. The disease environment during childhood is also im-
portant: Bleakley (2010) estimates that 100 percent malaria eradication in several
Latin American countries increased adult incomes by more than 40 percent. Moving
onto the returns to education, Duflo (2001) uses a large school construction program
in Indonesia to identify a return to education on men’s wages of between 6.8 and
10.6 percent. 2 Foster and Rosenzweig (1996) find that better educated farmers
profited more from the introduction of Green Revolution technologies than other
farmers did. The benefits of education are also believed to persist across genera-
tions: Thomas, Strauss and Henriques (1991) and Glewwe (1999) each find that
paternal education positively impacts child health, and Osili and Long (2008) find
that an additional year of schooling decreased Nigerian women’s number of births
before age 25 by 0.26 births. Many of the results in these papers are generated
from either plausibly exogenous or experimental variation in the key explanatory
variables, and it is therefore clear that human capital investments can have large,
extremely persistent e↵ects.
1.3 Contribution
It seems clear that the study of the e↵ects of shocks on a↵ected individuals’
human capital stocks is a worthwhile pursuit. This literature is incomplete with
2On the other hand, Uwaifo Oyelere (2010) uses a similar identification strategy and finds a
return of only 2.8 percent in Nigeria.
7
respect to both of the types of shocks that I study in the following two chapters
of this dissertation. The main chapter (2) studies the e↵ects of the number and
timing of rainfall shocks experienced while young on adults’ human capital stocks.
As mentioned above, there does exist research on the e↵ects of various kinds of
aggregate shocks on the contemporaneous enrollment in school of a↵ected children,
and there is also research suggesting that these lapses in education investments are
made up for later. 3 But there is less evidence on what the permanent e↵ects of an
entire childhood’s worth of those kinds of shocks might be. The following chapter
in this dissertation provides this kind of evidence by studying the e↵ects of rainfall
shocks experienced over the first 21 years of life on West African adults’ literacy
and highest grade completed at school. This chapter takes econometric identification
issues seriously and provides policy-relevant evidence on how an objective measure of
human capital is permanently a↵ected by a type of shock that is common throughout
the developing world.
The third and final chapter studies the food price crisis of 2008. While there
exists a substantial amount of work on the impacts food price shocks have on out-
comes like hunger and poverty, most of it is based on simulations; there is less
evidence that uses household survey data and less still that relies on objective mea-
sures of well-being from survey data. Moreover, as far as I know, there is no research
that uses the kind of anthropometric data I do to identify how di↵erent types of
households, and in particular net food producing and consuming households, seem
3In addition to the aforementioned papers by Jacoby and Skoufias (1997), Jensen (2000) and
Bjorkman-Nyqvist (2013) on the e↵ects of rainfall shocks on enrollment, Frankenburg et al. (2004),
Rucci (2003) and Funkhouser (1999) found that enrollment declined following macroeconomic
crises.
8
to fare in response to food price crises. This chapter cannot be considered evidence
of the causal e↵ects of high food prices on di↵erent households’ well-being, due to
the fact that the main analysis does not involve an estimation of how any group of
children’s weights-for-height would have evolved over time in the absence of high
food prices (i.e., the counterfactual scenario). But the chapter seems to be unique
in that it uses a relatively objective measure of children’s well-being–their weight-
for-height–to provide evidence of how food prices seem to have a↵ected children as a
function of their region of residence and parent’s occupations. Moreover, it does this
for children throughout the conditional weight-for-height distribution using quantile
regression techniques.
Both of these chapters use nationally representative household survey data col-
lected over multiple years to show that the shocks I study likely result in substantial
welfare e↵ects for certain individuals (and in the case of rainfall shocks, I show that
these e↵ects are permanent). The two chapters are also similar in how they deal
with shocks that are common. Again, rainfall shocks matter in any location where
peoples’ livelihoods depend at least in part on rainfed agriculture. Food price in-
flation is probably similarly familiar to the individuals throughout the developing
world who are a↵ected by it. These shocks are also both aggregate shocks, and
this distinction is worth noting because we would expect these types of shocks to
be more di cult to cope with. To the extent that informal insurance networks are
locally geographically concentrated, then since everyone in the network will have
been hit with the shock, there is no one to provide any kind of insurance ”pay-
out” (Townsend, 1994). The insurance value of ”bu↵er stock” savings is also likely
9
particularly low following aggregate shocks, due to the increase in the number of
households who would sell assets, which in turn would result in a ”fire sale” price
(Lim and Townsend, 1994). By the same logic, we would expect o↵-farm wages to
be particularly low following an aggregate shock, given the increase in labor supply.
There are simply fewer available coping mechanisms to draw on following aggregate
shocks.
The two chapters are also similar in that they use basic economic logic and
contextual understanding to motivate analyses of how di↵erent groups of children
might be a↵ected by shocks. The second chapter considers the permanent e↵ects
of rainfall shocks experienced while young on the completed human capital stocks
of women and men separately. One look at the di↵erences in literacy rates and
the average highest grade completed at school between the women’s and men’s
samples makes it clear that investment processes di↵er substantially along gender
lines. Men are more than three times as likely to be literate as women are, for
example. Female and male children in West Africa allocate their time di↵erently
outside of school as well. Existing research shows that girls work inside and outside
the household, whereas boys basically only work outside of it. The results I obtain
show that the prioritization of education investments in boys is actually exacerbated
by both wet and dry shocks, which each decrease completed human capital stocks
for females but (if anything) increase them for males. The third chapter is more or
less based on the idea that children in net food producing households are expected
to have been a↵ected by food price inflation di↵erently than children in net food
consuming households were. Since I do not actually observe net food production
10
status, I use region of residence (which determines rural-urban status, as well as
whether or not high-priced crops were being harvested in large quantities), parental
occupation (e.g., self-employed farmer, agricultural employee or other), and finally
a child’s implied place in the conditional weight-for-height distribution, as proxies
for net food producing household status. Existing research suggests that net food
producing households should benefit from high food prices, but in the case of the
2008 food price crisis in Egypt I do not find evidence in support of this. In fact, along
with the lightest (and, I assume, the most economically disadvantaged) children in
most region and parental occupation-based groups, it was the group of children that
I argue were most likely to have been members of net food producing households
who seemed to have been the most negatively a↵ected by food price inflation. (I
argue that the most likely explanation for this is the high cost of fertilizer and
other inputs.) In the case of both chapters, then, the allowance that the analyses
make for these di↵erent groups of children possibly being a↵ected di↵erently ends
up providing results that are somewhat surprising yet quite plausible.
Also, both chapters are policy-relevant, given the existing stock of research
and policy concerns. In the case of the second chapter, by complementing exist-
ing research on how some households respond to shocks by pulling children out of
school and providing evidence on the permanent e↵ects of this coping strategy, this
chapter speaks to outcomes that are at the heart of multiple Millennium Develop-
ment Goals (MDGs). In particular, the second and third MDGs call for universal
primary school enrollment and the promotion of gender equality, respectively, and
the chapter clearly speaks to both of these issues. Also, as debates about how much
11
development aid should be devoted to the management of the e↵ects of climate
change begin, evidence of the sort that I provide that ordinary weather variation
can have substantial permanent consequences will be useful. The food price infla-
tion chapter speaks to a topic that is likely to remain important for the medium
term; FAO (2011) concludes that food prices will likely exhibit higher volatility and
on average remain relatively high for the next several years at least. Both of the
following chapters therefore have to do with policy issues that will continue to be
important.
Finally, given the magnitude of the potential returns to investments in chil-
dren’s human capital, we might ask why parents ever sacrifice such investments (for
smoother consumption, presumably). In particular, we might first ask whether par-
ents are selfish rather than altruistic. 4 Existing economic research suggests that
parents are indeed altruistic, at least with respect to children’s time allocation to
work and school. Manacorda (2006) uses state-specific child labor laws to identify
who in the household benefits from child labor in early 20th century America. He
finds no evidence of any relationship between the proportion of working children in
a household and their parents’ labor supply, and that the returns from child labor
are redistributed to other children in the household in the form of less time work-
ing. Using data from Pakistan, Bhalotra (2004) finds that adult consumption and
children’s school participation covary positively, and concludes that the hypothesis
4This is not meant to suggest that parental selfishness alone would imply that parents would
forego human capital investments in their children for higher household income thanks to their
labor. Given the widespread tendency in the developing world for elderly parents to depend on
their children for support, it would seem to be in parents’ self-interest to invest in their children’s
education in a way that could help the parents themselves later.
12
of altruistic parents is not rejected. We might expect, in turn, that if parents would
prefer to send their children to school rather than work as long as they could a↵ord
to do so, then school attendance would be positively a↵ected by household wealth
and work would be negatively a↵ected by it. In fact, there is a substantial amount
of evidence in favor of these sorts of income e↵ects (Edmonds and Schady, 2012;
Attanasio et al., 2006; Edmonds, 2006; Filmer and Schady, 2008; Ravallion and
Wodon, 1998; Schultz, 2004; Cogneau and Jedwab, 2012). 5
It would seem, therefore, that parents would prefer to invest in their children’s
human capital when resources permit. Given the kinds of fluctuations in children’s
human capital stocks that shocks are known to cause, it seems reasonable to expect
that it is the inability to smooth consumption perfectly in the face of shocks that
drives these investment fluctuations. We might also note that under the assumption
of concave utility functions, the marginal utility of consumption will be particularly
high in poor places (especially following negative income shocks), which might help
rationalize the way valuable investments are sometimes sacrificed for the sake of
trying to maintain current consumption. Because as we will see in the following
chapter, the costs that are incurred in the attempt to smooth can indeed by high,
at least in the case of rural West African women who experienced higher numbers
5The positive relationship between household wealth and children’s school enrollment is not as
straightforward as one might suspect. Basu, Das and Dutta (2010) find that child labor supply
is low for children in landless rural Indian households, that it is high for children in households
that own a moderate amount of land, and that it is once again low for children in households with
very large amounts of land. This is consistent with their model, which shows that labor market
imperfections will mean children in landless households do not work, while children in households
with some but not a large amount of land will work (thanks to the fact that their land o↵ers them
tasks to perform). Even in this kind of situation, however, higher household wealth does eventually
decrease the amount of work children do.
13
of rainfall shocks before the age of 21.
14
Chapter 2: Rainfall Shocks and Human Capital
2.1 Introduction
Risk is a central aspect of life in developing countries, given the widespread
existence of incomplete formal insurance and credit markets. Perhaps the most
common types of shocks for households in agricultural areas of the developing world
are rainfall shocks, and given the extent to which informal insurance networks are
typically geographically concentrated, households will likely not be insured against
these shocks. The variability in income that rainfall shocks cause can in turn a↵ect
education and health investments in children, despite the potentially high returns to
such investments. In fact, there is evidence that the kinds of rainfall shocks I will be
analyzing have measurable e↵ects on (contemporaneous) human capital investments
for both school-aged as well as preschool-aged children. Investments in children are
thought to be a↵ected by rainfall shocks primarily through some set of competing
income and substitution e↵ects, the exact nature of which depend on the type of
investment (and the age of the potential recipient) under consideration. However, as
existing research suggests is the case, temporary investment shocks could be made
up for later, and it is still mostly an open question as to what the long-term human
15
capital costs of these rather commonplace shocks are. 1 This chapter attempts to
provide an answer using rainfall shocks in West Africa.
More specifically, I study the literacy and highest grade completed of rural
West African women and men who were at least 21 years old at the time of collection
of the household surveys on which I rely, and I consider the e↵ects on those outcomes
of the numbers and timing of all of the unusually wet and dry years experienced
over the first 21 years of life. (Years are considered wet or dry if the total amount
of rainfall is su ciently far above or below, respectively, the local annual average
over the previous 25 years.) In other words, I consider how the human capital
stocks of adults were a↵ected by rainfall shocks that occurred before the age of 21.
Given the di↵erences in the ways West African females and males allocate their
time, I study the e↵ects of shocks on the completed human capital stocks of women
and men separately. My main empirical analysis allows the e↵ects of additional
wet and dry years to vary depending on the age range in which they occurred
and controls for (among other things) time-invariant, location-specific factors and
(country-specific) temporal factors. Intuitively, my main analysis compares changes
in human capital outcomes across birth cohorts across di↵erent locations (where
rainfall patterns di↵ered over the first 21 years of individuals’ lives). Thus, here I
provide the first estimates of the permanent e↵ects of all of the annual rainfall shocks
that people experienced over the course of their educational career (and before) on
an objective measure of human capital (literacy). I do this by studying 39 cohorts
worth of women and men using as many as 26 nationally representative household
1See Section 2.2.2.1 below for a discussion of these existing findings.
16
surveys from 12 West African countries merged with (as far as I know) the most
geographically disaggregated precipitation data available.
This chapter is related to the aforementioned literature on the contempora-
neous investment e↵ects of covariate shocks (including rainfall shocks), which is
summarized by Ferreira and Schady (2008). One of the main conclusions of that
paper is that in poorer places, education and health investments in school-aged and
preschool-aged children, respectively, tend to be procyclical. I build on these results
by considering the lasting e↵ects of a whole childhood’s worth of covariate shocks.
This project is also related to research dealing with the long-term consequences
of shocks experienced at formative ages. For example, economists have shown that
lower rainfall in the first year of life (Maccini and Yang, 2009), pandemics expe-
rienced while in utero (Almond, 2006), being one of a set of twins (Behrman and
Rosenzweig, 2004; Royer, 2005), drought and civil war before the age of 3 (Alderman
et al., 2006) and exposure to political violence (Leon, 2009) each lead to worse adult
outcomes, including completed educational attainment. Shah and Millett Steinberg
(2013) also study the permanent e↵ects of rainfall shocks on Indian adolescents and
adults, but their main long-term outcomes are the highest grade completed at school
and adult wages rather than a direct measure of human capital. Most of these pa-
pers either focus on unusual shocks or on shocks at very particular ages only (or
both). In contrast, I consider the e↵ects of common weather shocks and how those
e↵ects di↵er over the entire period from infancy to young adulthood; this chapter
is about a basic, recurring climatic feature as opposed to the kind of shock that
only a↵ects individuals at critical ages. Understanding the e↵ects of rainfall shocks
17
therefore allows us to estimate what a↵ected adults’ human capital stocks would
be in the absence of a major source of aggregate productivity risk that (as we will
see) a↵ects individuals over the entire course of their educational career. The fact
that the main outcome under study here is literacy also helps set this chapter apart.
As Bleakley (2010) argues, it is possible for shocks to have particular e↵ects on the
highest grade completed and di↵erently-signed e↵ects on the kind of outcome we are
oftentimes ultimately interested in, human capital, and so it is important to have a
valid measure of the latter. Another contribution of my chapter is that it considers
the e↵ects of unusually rainy as well as dry seasons in West Africa. Wet shocks in
particular are likely to become a topic of greater interest if the predictions of the
Intergovernmental Panel on Climate Change turn out to be true and these shocks
become more common in the region over the next several decades (Christensen et
al., 2007).
My baseline results for females are that both additional wet and dry years
decrease the likelihood of being literate as well as the highest grade completed
(between birth and age 20, depending on the type of shock, for both outcomes).
An extra wet (dry) year between age 7 and age 13 (birth and age 6) decreases the
likelihood of being literate by 1.2 (0.6) percentage points, which are economically
significant e↵ects when considered alongside the women’s sample literacy rate of
about 17 percent. Also, an additional wet (dry) year between age 14 and age 20
lowers the highest grade completed by 0.120 (0.033), which are again substantial
compared to the women’s sample average highest grade completed of 1. Given the
average numbers of shocks experienced during the di↵erent age intervals and the
18
baseline estimates of the e↵ects of shocks, the set of all shocks experienced over the
first 21 years of life decrease women’s likelihood of being literate by 22.1 percent,
and their highest grade completed at school by 19.1 percent (on average). Note that
these estimated e↵ect sizes are seriously large, especially given the relatively small
number of women in the sample who are literate or who completed even one year
of school. For men, each additional wet (dry) year between the ages of 14 and 20
(birth and age 6) increases the likelihood of being literate by 1.5 (1.1) percentage
points (which might be compared to the men’s sample literacy rate of 44 percent).
Also, an additional dry year between birth and age 6 increases the highest grade
completed by 0.075 (with the sample mean for this outcome being 4.10). On average
and as a whole, shocks increase men’s likelihood of being literate by 9.9 percent and
their highest grades completed by 5.6 percent. For women especially, shocks have
large e↵ects on completed human capital stocks.
As will be made clear, however, many of these baseline estimates could be
a↵ected by sample selection bias. In particular, under strongly unfavorable assump-
tions the negative e↵ects of wet years on women’s human capital could be either
understated or overstated and the negative e↵ects of dry years could be overstated,
while the beneficial e↵ects of both types of shocks on men’s human capital stocks
could be overstated. For example, under strong assumptions about how many fe-
males live long enough to enter the sample because of wet (dry) shocks between the
ages of 7 and 13 (birth and age 6), each additional shock would decrease women’s
likelihood of being literate by only 0.007 (0.003) percentage points. The magnitudes
of these selection-free estimates are about 42 and 50 percent smaller, respectively,
19
than their baseline analogs. But even under highly unfavorable assumptions, there
remains solid evidence that women’s human capital stocks are substantially nega-
tively a↵ected by both wet and (to a lesser extent) dry years and that wet shocks
increase men’s human capital.
In an attempt to understand the channels through which preschool-aged shocks
are having their e↵ects, I consider the hypothesis that shocks a↵ect children’s health,
which in turn later a↵ects the kinds of education investments that parents make in
their children. I test for evidence in favor of this hypothesis by examining how
rainfall shocks a↵ect individuals’ height (using the same baseline methodology as
described above), which I take to be a measure of early-life health conditions (Bozzoli
et al., 2009) (and which I have data on for a subset of the women in my data only). I
find no evidence that shocks between birth and age 6 a↵ect women’s height. On the
other hand, I find that the negative (positive) e↵ects of preschool-aged dry shocks on
women’s (men’s) human capital are driven by shocks at ages 1-3 (2-4). While this is
not necessarily evidence in favor of the health shocks explanation of preschool-aged
shocks, it seems inconsistent with the alternative hypothesis that preschool-aged
shocks persistently a↵ect household income or access to productive resources, and
therefore parents’ ability to finance their children’s education when those children
are old enough to attend school. It is therefore not clear whether shocks at preschool
ages are a↵ecting individuals through a health channel or in another way.
I argue that this pattern of results is broadly consistent with existing evidence
on how young, rural West Africans allocate their time in general and in response to
rainfall shocks in particular. More specifically, the finding that additional wet years
20
decrease females’ human capital stocks can plausibly be explained by the stylized
fact that both on- and o↵-farm economic activities are higher following the kinds
of bumper harvests that wet years tend to result in. As for the negative e↵ects of
dry years on females’ human capital, I argue that the most likely explanation for
this finding is that parents are simply not willing to finance girls’ school enrollment
when household income is low (as it is following particularly dry rainy seasons). At
the same time, given the facts that boys seem to work mostly in agriculture and
that their parents seem more willing to finance their education in general, I argue
that the low labor demand that dry years cause o↵ers boys the chance to spend
more time learning rather than working. Finally, wet years likely mean dominant
income e↵ects for boys, with parents taking the opportunity to invest in the human
capital of their sons when income is higher.
Section 2.2 provides background on the agricultural labor cycle in West Africa
as well as considers conceptual issues. Section 2.3 describes the empirical strategy,
the data and the sample selection criteria. Section 2.4 presents descriptive statistics
and results. Section 2.5 explores robustness and Section 2.6 concludes.
2.2 Background and Conceptual Issues
2.2.1 The Agricultural Labor Cycle in West Africa
As mentioned above, I study the e↵ects of shocks in 12 West African countries.
The region is su ciently large that it naturally plays host to variation in rainfall
conditions. For instance, the wet season (which lasts from either March or May
21
until October or September, respectively) is generally shorter and annual rainfall
levels typically lower at higher latitudes. 2 Also, the onset of the rainy season
tends to be less predictable, and the distribution of precipitation over the course
of the rainy season less uniform, in areas further north. There is also considerable
variation in the types of crops that are grown across the region. For example, crops
that require less rainfall to thrive and that tend to be less sensitive to dry shocks
such as bulrush millet, cassava, groundnuts and sorghum are grown in areas to the
north where rainfall levels are lower and more erratic. To the south, where rainfall
levels are higher and dry shocks are less common, crops such as co↵ee, cocoa, yams
and maize are grown.
In important ways, however, agricultural production methods and labor mar-
ket conditions in general and over the course of the agricultural cycle do not tend
to vary across the region. For example, agriculture is the main sector in which
most individuals in West Africa work, with 60-80 percent of all people there being
engaged in agricultural work of some sort. Also, rainfall is a critical input on the
farm, as the vast majority of agriculture in the region is rain-fed (Hayward and
Oguntoyinbo, 1987). Child labor is also common throughout the region, with labor
force participation rates for African children between the ages of 5 and 14 of 41
percent (ILO, 1998). 3 Most child labor is believed to be family-controlled in the
2Roughly speaking, Benin, Cote dIvoire, Ghana, Liberia, Nigeria, Sierra Leone and Togo have
a longer rainy season (interrupted by a brief dry spell near the end of August) from March to
October, while Burkina Faso, Guinea, Mali, Niger and Senegal have a shorter wet season from
May to September.
3Although this result is based on data collected near the end of the time period that I am
concerned with, Bhalotra (2003) notes that it likely reflects child labor supply conditions in previous
decades, as Africa did not experience the kind of post-war secular decline in child labor that other
regions did.
22
sense that children work in family-run enterprises (Andvig, 2001).
Also, agricultural labor demand is higher during the wet season throughout the
region (Hayward and Oguntoyinbo, 1987; ICRISAT, 1980; Canagarajah and Skyt
Nielsen, 1999). During the wet season, crops are planted, weeded, tended to, and
eventually harvested (and it is important that each of these tasks be performed at
very particular times). After the harvest is complete and crops have been prepared
for consumption and sale, household food supplies and income levels-and through
multiplier e↵ects, o↵-farm economic activities such as petty trading-are at their
highest (Mortimore, 1989; Thomas, 1997). Moreover, labor demand (both on and
o↵ the farm) and household incomes are higher still during and after those wet
seasons where rainfall was plentiful. 4
In contrast, the dry season is characterized by relatively slack on- and o↵-
farm labor demand. Although tasks such as clearing farmland and preparing it for
planting need to be done before the beginning of the wet season, the time frame
within which these tasks have to be performed is not necessarily so tight that, for
example, child workers cannot a↵ord to wait and do them on the weekend (when
school is not in) (Robson, 2004). The dry season is also when seasonal migration
takes place, with men leaving rural areas in substantial numbers to earn money
in urban areas since earning opportunities are scarce in rural areas. Moreover,
seasonal migration is more common following those wet seasons where rainfall was
particularly scarce (Mortimore, 1989). These stylized facts are consistent with the
4See Jensen (2000), Webb and Reardon (1992), Sillah (2009), Bjorkman-Nygvist (2013), Bhalo-
tra and Umana-Aponte (2010), Bengtsson (2010), Dercon (2004), Levine and Yang (2006) and
Paxson (1992) for evidence of a positive relationship between rainfall and income. See Iliya (1999)
and Reardon (1997) for evidence of increased labor demand following wet shocks.
23
ideas that agriculture largely drives economic activity in rural areas and that rural
West Africans have a limited number of other sectors that they can work in.
2.2.2 Conceptual Framework and Existing Evidence
2.2.2.1 Shocks and School-Aged Children
I start out by considering how rainfall shocks a↵ect contemporaneous invest-
ments in school-aged children’s human capital and follow up by discussing how tem-
porary investment shocks could have permanent e↵ects. For school-aged children,
it is expected that dry shocks (for example) will lead to income and substitution
e↵ects. On the one hand, a dry year is expected to decrease household income. This
will increase the marginal utility of consumption of whatever goods a child’s labor
can buy, which pushes children to devote more time to working and less to attend-
ing school or learning. On the other hand, a dry year also means lower agricultural
labor productivity and so lower demand for labor in a setting where most people
work in agriculture. Thus, the opportunity cost of attending school or spending time
studying is lower, which of course pushes children to devote more time to learning.
We might also consider the direct costs of attending school, which can be
substantial in the countries I study. For instance, Jensen (2000) finds that in Cote
dIvoire in 1986, the median cash outlay associated with sending children to school
was about one-third the median household income per capita. We would expect
higher school costs to reinforce the income e↵ect of dry years: They should be more
dear in terms of lost utility from foregone consumption during dry years compared
24
to normal years. Of course, it is an empirical question as to which of these e↵ects-
substitution or income (augmented by school fees)-dominate in practice, i.e. this
simple framework does not o↵er predictions as to whether a negative shock like a
dry year will decrease or increase investment in young peoples’ human capital.
In addition to dry years I will also be considering the e↵ects of wet years, which
will also (tend to) have competing income and substitution e↵ects: For most house-
holds in what I have defined to be wet years, the marginal utility of consumption
will be lower (thanks to a bountiful harvest and the higher household incomes that
accompany it) which will draw children toward learning, but the demand for labor
will also be higher which will draw them toward work. For those households whose
farm plots are especially prone to flooding, however, wet years will have negative
income e↵ects (which, again, will encourage children to work). Also, to the extents
that agricultural labor markets are poorly functioning or that children work more
or less exclusively in household-run enterprises, wet years will have the same kinds
of substitution e↵ects as dry years for children from households negatively a↵ected
by floods in that demand for these children’s labor will be low. 5 Once again, the
basic theory does not o↵er a prediction as to how children’s human capital should
be a↵ected.
Ferreira and Schady (2008) summarize and interpret the results of the existing
literature on the e↵ects of covariate shocks on (school-aged) children’s contempora-
neous human capital investment. They note that school enrollment tends to decline
5There is evidence to suggest that labor was not commonly traded across farms in West Africa
over the time period being studied here (Binswanger and McIntire, 1987; Oyekale, 2009; ICRISAT,
1980; Yaro, 2006; Reardon et al., 1988; Bhalotra, 2003; Fafchamps et al., 1998).
25
during or following negative covariate shocks in poorer countries such as Cote dI-
voire (Jensen, 2000), Malawi (World Bank, 2007) and Indonesia (Thomas et al.,
2007), but for the most part tends to rise during negative shocks in less-poor Latin
American countries including Brazil (Duryea and Arends-Kuenning, 2003), Mexico
(Binder, 1999; McKenzie, 2003) and Peru (Schady, 2005). 6 (It should also be
noted that the depressed economic conditions in Cote d’Ivoire and Malawi were
driven by rainfall deficits.) The authors argue that this pattern of results is con-
sistent with the aforementioned basic theoretical framework, and that the reason
why enrollment is mostly procyclical in poorer nations and mostly countercyclical
in less-poor nations is because in the former-where marginal utilities of consumption
are higher-income e↵ects should be stronger and therefore more likely to dominate
substitution e↵ects. They also argue that people in poorer nations are more likely
to be subject to the kinds of credit market imperfections that prevent households
from being able to borrow for the sake of financing their children’s schooling when
incomes are temporarily low.
It is possible that these kinds of temporary shock-driven changes to invest-
ments do not lead to permanent e↵ects. For example, if investments are easily sub-
stitutable over time then we would expect the e↵ects of a temporary lapse to be un-
done by increased investment later. Indeed, Jacoby and Skoufias (1997), Funkhouser
(1999) and Bjorkman-Nyqvist (2013) each suggest that enrollment lapses are made
up for later, and that there are not substantial permanent human capital e↵ects
6Costa Rica, however, is a less-poor nation that does not fit this pattern. There, enrollment
was found to have decreased during the 1981-1983 recession (Funkhouser, 1999).
26
associated with shocks. But it is not clear that getting non-enrolled children caught
up would be so easy. De Janvry et al. (2006) hypothesize that re-enrolling in school
following a period of non-enrollment might be especially costly for a few reasons:
Children could learn to appreciate other aspects of life or lose some of their ability
to study when not enrolled, and children might remember su ciently little of what-
ever material they previously learned that enrolling becomes more trouble than it
is worth. More generally, it seems reasonable to suppose that the cost of enrolling
in school is decreasing in the amount of learning that has recently taken place, for
reasons similar to those just given. For instance, learning less in one year could make
an additional year in school particularly costly if children become less interested in
learning relative to pursuing other things, or if new material simply becomes too
hard to learn then. In these ways, temporary shocks could result in permanent
e↵ects on individuals’ human capital stocks.
2.2.2.2 Shocks and Preschool-Aged Children
Rainfall shocks that occur during preschool ages can also a↵ect individuals’
completed human capital stocks. In particular, rainfall shocks can a↵ect health
investment in young children in ways that have long-term health e↵ects, and if
parents decide how much to invest later in the education of children as a function of
how healthy those by-then school-aged children are (as in Maccini and Yang, 2009;
Alderman et al., 2006; Alderman et al., 2001; and Hoddinott and Kinsey, 2001),
27
then shocks at early ages can have permanent human capital e↵ects. 7 More
specifically, we might suppose that child health depends on: privately purchased
health-promoting goods (e.g., food, clothing and medicine); parental time inputs
(for things like preventive health care visits, breastfeeding and the collection of clean
water); and the disease environment, which can be directly a↵ected by rainfall (Wang
et al., 2009). Of course, rainfall shocks can a↵ect all of these factors. The household’s
ability to purchase health-promoting goods would be diminished following a dry
year, for instance, while the amount of time parents devote to maintaining children’s
health will depend on whether parents supply more or less labor during dry years.
In other words, parental time inputs will be subject to the same kinds of income
and substitution e↵ects that school-aged children’s labor supply is, and so it is in
turn an empirical question as to whether preschool-aged childrens health is helped
or harmed by dry years. So once again, the theory does not o↵er predictions as
to how a covariate shock might a↵ect contemporaneous investments in children’s
health.
The evidence on the e↵ects of negative covariate shocks on young children’s
health investments is once again summarized and interpreted in Ferreira and Schady
(2008). The authors argue that health outcomes are more likely to worsen during
bad economic times in poorer countries, where the e↵ect of the marginal dollar (not)
spent on health-promoting private goods is more likely be large enough to overwhelm
the e↵ect of any possible increase in the amount of time that parents devote to chil-
7Of course, shocks that occur while individuals are still in utero can also have substantial
permanent e↵ects (Almond, 2006). In results not shown, however, I find no evidence that shocks
in the 12 months preceding birth a↵ect either of the human capital outcomes I focus on here, for
either females or males.
28
dren’s health. And once again, the results of the empirical literature are consistent
with the authors’ argument: Negative aggregate income shocks lead to worse health
outcomes in poor countries such as Cote dIvoire (Jensen, 2000), Zimbabwe (Alder-
man et al., 2006), Ethiopia (Yamano et al., 2005), Tanzania (Alderman et al., 2009)
and Cameroon (Pongou et al., 2006), while children’s health improved in developed
countries such as the US (Chay and Greenstone, 2003; Dehejia and Lleras-Muney,
2004) during downturns.
Any possible persistent health shocks that rainfall shocks cause could later
result in either compensatory or reinforcing education investment decisions on the
part of parents. In particular, Rosenzweig and Zhang (2009) find evidence of rein-
forcing education investments in response to health shocks, while Liu et al. (2009)
and Bharadwag et al. (2010) find evidence of compensating education investment
behavior. It should also be kept in mind that rainfall shocks during preschool ages
could a↵ect later education investments in ways that have nothing to do with chil-
dren’s health. For example, rainfall shocks have been shown to have persistent
e↵ects on rural households’ consumption , which implies that shocks early on could
a↵ect the timing of initial enrollment in school if the direct costs of attending school
are substantial. 8 It is also possible that income shocks at preschool ages af-
fect eventual education investments through their persistent e↵ects on households’
access to agricultural inputs and capital. For example, it is believed that one of
the ways farmers attempt to smooth consumption following an agricultural output
8Dercon (2004) finds that 10 percent lower rainfall 4-5 years before decreases current consump-
tion growth by 1 percentage point for a sample of rural Ethiopian households.
29
shock is by selling livestock (Fafchamps et al., 1998; Mortimore, 1989). Thus, if
dry years are simultaneously responsible for persistently low incomes and decreased
livestock numbers, it is reasonable to expect that those numbers will remain low for
a sustained period following a shock. 9 Since it is typically the responsibility of
children (and in particular boys) to care for livestock (Bolwig and Paarup-Laursen,
1999; Robson, 2004), rainfall shocks at preschool ages could a↵ect children’s labor
supply even after those children are old enough to attend school. It could also be
that rainfall shocks a↵ect households’ abilities to purchase agricultural inputs such
as fertilizer years into the future, which, to the extent that such inputs and child
labor are complementary, would also a↵ect the supply of the latter. Also, it is pos-
sible that individuals’ educational careers could either start by a particular age or
not at all, perhaps because formal learning is more di cult when it begins at more
advanced ages. Or it could be that children who do not enroll as soon as possible be-
come su ciently productive outside of school that the opportunity cost of enrolling
is prohibitively high by the time households do have the resources to finance those
children’s enrollment. These possibilities imply that not enrolling early on because
of rainfall shocks could have permanent consequences.
Thus, we have reason to believe that temporary rainfall shocks at any age
over the course of individuals’ educational careers (and before) could end up af-
fecting those individuals’ completed human capital stocks. Also, while the papers
described above are important in that they show that temporary shocks-including
9In addition to sales-driven decreases in livestock counts, livestock also perish following dry
rainy seasons due to lack of pasture (Mortimore, 1989).
30
rainfall shocks-can have contemporaneous human capital investment e↵ects, we may
ultimately be more concerned with the long-term e↵ects of these types of shocks. I
now turn to the econometric strategy that I will use to estimate these permanent
e↵ects.
2.3 Empirical Strategy, Data and Sample
2.3.1 Empirical Strategy
I use the following specification to try and learn about how additional wet and
dry years at di↵erent ages a↵ect human capital outcomes:Hipt = ↵+ 3X
j=1
 j#wetyearsjpt + 3X
j=1
⇢j#dryyearsjpt +  p + ✓ct +  r(t) +X 0ipt + "ipt.
(2.1)
Here Hipt is the human capital outcome (literacy or highest grade completed) of
individual i from rainfall neighborhood (i.e., place) p born in year t. 10 The
#wetyearsjpt and #dryyearsjpt variables are equal to the numbers of wet and dry
years, respectively, that individuals from place p and born in year t experienced
during the jth age interval, j 2 {1, 2, 3}. These three di↵erent age intervals are as
follows: the birth year until the age of 6, age 7 to age 13 and age 14 to age 20. These
intervals correspond to the preschool period, a period when individuals’ attachment
to school might be relatively strong, and a time period when those attachments
10The term rainfall neighborhood refers to the fact that individuals are assigned rainfall shocks
data based on which geographic location I have historical rainfall data for that they live closest
to. Thus, all individuals who live closer to a given one of these locations than any other are said
to live in that location’s rainfall neighborhood.
31
are likely weaker. 11 The coe cients associated with these shocks variables are
meant to tell us what the e↵ects of experiencing an additional shock during each
age interval are.
We might also expect, for example, that places where rainfall shocks are more
common are either better or worse equipped to deal with those shocks. In this case,
if we ignore these sorts of unobservable time-invariant, location-specific factors and
simply regress a literacy indicator on the shocks variables, the resulting estimates
of the wet and dry years e↵ects could be biased. To deal with this kind of concern,
I include rainfall neighborhood fixed e↵ects (denoted by  p) in my specification.
Also, to identify wet and dry years e↵ects separately from time series e↵ects, I in-
clude country-specific birth year fixed e↵ects (✓ct) in my specification as well. In
an attempt to deal with the possibility that distinct changes in literacy tended to
happen at a sub-national level in places that became more or less dry over time, I
also include controls for region-specific linear time trends ( r(t)) (where each indi-
vidual in my samples lives in one of 55 regions). The X vector includes full sets
of dummy variables in my specification corresponding to individuals religion (Mus-
lim and Christian, with ’other’ being the omitted category) and ethnicity (of which
there are 71 in all). Standard errors are clustered at the rainfall neighborhood level.
Regressions are unweighted, which, under the assumption that the baseline
model is correctly specified, will lead to consistent estimates of the e↵ects of rainfall
shocks on human capital (Deaton, 1997). To test whether or not this assumption
11The qualitative features of my baseline results are robust to changes in the choice of age
intervals (results available upon request).
32
holds, I would have to augment the sample weights that the DHS provides in multiple
ways, since these only work to make respondents for each survey representative at
the rural level (for the year the survey was conducted only). I pool together surveys
from multiple countries and years, however, and so in order for my sample to be
representative of the rural population of the 12 West African countries that I study
as a whole, I would (at least) need to multiply the weights that the DHS provides
by the inverse of the rural sampling rate (again, for each survey).
Only after doing this would individuals in my sample be assigned weights such
that, for example, individuals in smaller countries would not count for too much
or too little relative to individuals in larger countries. Since I am not aware of
any available data on rural populations between the ages of 21 and 49 across years,
however, I use data from the World Population Prospects online database on the
total populations in that age range for each country I study. 12 These data
are only available for every year that is a multiple of 5, and I therefore assume
constant rates of population growth in the intervening years to obtain estimates of
populations sizes then. I then calculate sampling rates for every country- and year-
specific household survey I use data from, and multiply the within-country weights
that the DHS data contains by the inverse of these sampling rates. The hope is that
these steps give weights that can be used to make my sample roughly representative
of the rural population of the countries I study as a whole. (Indeed, it is these
augmented weights that I use to calculate the descriptive statistics in Table 2.2.)
It should be noted, however, that given the possibility of unobserved poten-
12Available at http://esa.un.org/wpp/Excel-Data/population.htm.
33
tially non-random migration, then even having exact data on rural sampling rates
would not be su cient to generate the kinds of weights that would make my sam-
ple of non-migrants representative of the entire cohorts to which they belong (since
these cohorts include migrants as well as non-migrants). In other words, given the
likelihood that rural migrants are di↵erent in unobservable ways from rural non-
migrants, it is not at all clear what other kinds of adjustments would need to be
made to the weights that the DHS provides–in addition to incorporating sampling
rate di↵erences across surveys–for the weighted sample of rural non-migrants to be
truly representative of rural individuals as a whole. Again, the hope is that the
sample weights I apply do a reasonably good job of making my sample more or less
representative of the rural population of the 12 countries I study.
In any case, the augmented weights in question are used to test whether or not
the baseline results are sensitive to the use of sample weights. The results are quali-
tatively similar if quantitatively di↵erent, and in particular sometimes substantially
larger in magnitude (as well as available upon request). For example, both wet and
dry shocks still negatively a↵ect women’s human capital (although during fewer age
intervals), but the magnitudes of the e↵ects are between 100 and 157 percent larger
relative to the baseline results. For men, results are unchanged except that wet
shocks increase the likelihood of being literate between the ages of 7 and 13 (but
not between 14 and 20).
The consistency of my estimates of the wet and dry years e↵ects rests on the
assumption that the rainfall neighborhood-specific deviations in numbers of shocks
around the long-term neighborhood average, after controlling for birth year e↵ects
34
and region-specific time trends, are independent from idiosyncratic errors. In other
words, I assume that the variation left over in the wet and dry years variables after
controlling for time-invariant rainfall neighborhood characteristics, cohort e↵ects
and regional trends is orthogonal to unobserved determinants of individuals human
capital outcomes. For example, if rainfall neighborhoods which became more dry
over time also simultaneously played host to residents with lower human capital
stocks (for reasons that had nothing to do with rainfall shocks), then the estimated
e↵ects of wet and dry years would be biased downwards. On the other hand, if
those places that became more dry over time would have played host to the same
human capital trends (had they not actually become more dry) as places that did
not become more dry, then these comparison neighborhoods would serve as good
counterfactuals for the treated neighborhoods, and the resulting wet and dry years
estimates would be unbiased.
2.3.2 Data
The household surveys I rely on are Demographic and Health Surveys (DHS)
collected in Benin, Burkina Faso, Cote dIvoire, Ghana, Guinea, Liberia, Mali, Niger,
Nigeria, Senegal, Sierra Leone and Togo between 1988 and 2008 (see Table 2.1
for the complete list of surveys used). These surveys are nationally representative
(at rural and urban levels) and they contain information on a wide assortment of
socioeconomic and demographic variables, including respondents’ literacy, highest
grade completed, year of birth, religion, ethnicity, and whether or not respondents
35
have ever migrated. I append the surveys with the aim of obtaining enough power to
identify the e↵ects of shocks separately from local, time-invariant factors and time
series e↵ects (and because, as described earlier, in many ways the circumstances
under which individuals’ human capital stocks were formed were broadly similar
across countries). In sampled households each woman between the age of 15 and 49
was selected to be interviewed, while some fraction of the men between the ages of
15 and 59 were selected (with that fraction varying by survey), so that my women’s
sample is substantially larger than my men’s sample. 13
The human capital outcomes I consider include an indicator equal to 1 if an
individual is literate (and zero otherwise) and a variable equal to the individual’s
highest grade completed. For surveys collected before 2000, individuals literacy
status is self-reported. In particular, individuals were asked if they could read
and understand a letter or a newspaper with ease, with di culty or not at all.
I consider individuals in these pre-2000 surveys to be literate if they reported at
least being able to read with di culty. For surveys collected in and after 2000,
individuals were instead asked to try and read a sentence printed on a card while
survey workers decided whether individuals could read the whole sentence, only parts
of the sentence or none of the sentence. 14 I consider individuals in these surveys
literate if they could read at least parts of the sentence. As mentioned previously,
I merge the pre-2000 and post-1999 surveys together for my main analyses as if
the data used to construct the literacy variable was collected the same way in both
13Those women and men who either usually lived in the household or were visitors who slept
there the previous night were eligible for interview (ICF Macro, 2011).
14Literacy data is missing for 0.32 percent of women and 1.49 percent of men as a result of
interviewers not having sentences printed in respondents spoken language.
36
periods, but I should note that the estimated e↵ects of the key coe cients in my
main specifications-those corresponding to the rainfall shocks covariates-are robust
to the inclusion of a post-1999 survey year fixed e↵ect. 15 The highest grade
completed variable is defined to be just what it sounds like: the highest grade that
an individual finished (regardless of whether or not they attempted but failed to
complete a higher grade).
I study how these human capital outcomes vary as a function of historical
rainfall shocks. To measure these, I rely on the availability (in the DHS data) of
the longitude and latitude of the geographic center of the primary sampling unit
in which respondents live, which is normally the respondent’s village in rural areas.
Knowing relatively precisely where individuals reside allows me to assign to them
historical rainfall data that likely does a good job of characterizing the rainfall shocks
conditions that they actually experienced while growing up. The rainfall data I use
comes from Wilmott and Matsuura’s Gridded Monthly Rainfall Time Series, version
2.01. 16 These climate researchers constructed their data set by using historical
rainfall data for each month from January, 1900 until December, 2008 from actual
rainfall measurement stations to predict rainfall levels (again, for each month over
the same time period) at each of 900-plus geographic locations (grid points) within
the 12 countries I study. They did this for each grid point by using data from
(on average) the nearest 20 rainfall measurement stations. The grid points are 0.5
degrees apart from one another in both the east-west and north-south directions,
15Note that this fixed e↵ect is added to account for survey-specific factors, as opposed to cohort-
specific factors (which the country-specific fixed e↵ects already control for).
16Available at http://climate.geog.udel.edu/ climate/.
37
which works out to about 35 miles on average. Thus, for all the respondents in my
data the nearest rainfall data set grid point is at most about 25 miles away.
This gridded rainfall data was constructed using climatologically aided as well
as more traditional interpolation techniques to estimate monthly total precipita-
tion quantities. The actual station data (for the West African region) that the
interpolation scheme relies on comes from the Global HIstorical Climatology Net-
work (GHCN) and Professor Sharon Nicholson’s archive of African precipitation
data. The number of monthly rainfall totals that underlie the gridded data varies
by year, with there being between at least 190 and 385 stations for the period and
region I study (only the GHCN data is publicly available so that its stations can be
counted). This quantity of stations implies that on average, no point was any more
than between about 71 and 50 miles from a station, respectively. 17
Finally, I will define wet and dry years. For each rainfall neighborhood, for
each year between 1940 and 1998, I compute the annual average level of rainfall over
the previous 25 years, and similarly for the standard deviation. A year is considered
wet if the level of rainfall in that year is at least 1 standard deviation above the
local mean level over the past 25 years, and a year is considered dry if rainfall that
year is at least 1 standard deviation below that level. The 1 standard deviation
threshold, which roughly translates into a 44 percent di↵erence relative to the mean
rainfall level on average, allows for enough variation in the numbers of shocks that
the e↵ects of those shocks can be identified but not so much variation that rainfall
17With this in mind, and given that 99 percent of the geographic coordinates of sample clusters
in the DHS data are no more than about 3 miles from the sample’s actual location, it seems highly
unlikely that this o↵set is responsible for a quantitatively important proportion of however much
measurement error exists.
38
shocks no longer warrant the name.
The 1 standard deviation threshold follows the examples of Jensen (2000) and
Grimard and Hamilton (1999), who both study the e↵ects of rainfall shocks in West
Africa. 18 It should be kept in mind, however, that the qualitative features of the
main results are robust to changes in the definition of what constitutes a shock. In
particular, results are mostly qualitatively unchanged for shock thresholds between
0.8 and 1.3 standard deviations of local, recent rainfall (and are available upon
request).
2.3.3 Sample Selection
The quality of the rainfall data that I use depends positively on how geo-
graphically close the actual rainfall measurement stations underlying it were to the
grid points: A large number of rainfall measurement stations nearby makes it more
likely that predicted monthly rainfall levels at a given location will be close to the
true levels. The rainfall stations supporting the data set were relatively numerous
until the late 1990s. Thus, since my main specification will consider the e↵ects of
18In both of these studies, however, the areas being studied were no larger than single countries,
whereas I study an entire region spanning multiple agroecological zones. Given this, it should
be noted that the baseline results are robust to an alternative shock definition that is inspired
by the Standardized Precipitation Index (SPI). The SPI fits historical rainfall data to a chosen
distribution, and the parameters associated with the fitted distribution are used to rescale rainfall
realizations so that they are in terms of the standard normal distribution (Edwards and McKee,
1997)). Shocks might then be defined to be rainfall levels that are below the 16th percentile of the
historical rainfall distribution or above the 84th percentile–thresholds that roughly correspond to
the points in the distribution that are 1 standard deviation below or above the historical mean,
respectively. In the case of the alternative definition that I use, wet (dry) shocks are simply
defined to be rainfall amounts that are more than (roughly) the 84th percentile (less than the 16th
percentile) of the local rainfall distribution over the previous 25 years. This definition allows for
the possibility of assymetrically distributed rainfall in a way that the simple 1 standard deviation
threshold that the baseline results rely on does not, and it remains in keeping with the SPI in the
sense that it defines shocks relative to the actual historical distribution of rainfall. The results are
qualitatively unchanged (and available upon request).
39
rainfall shocks at ages as high as 20, I will not be including in my samples women
or men born after 1978 (so that my youngest cohort turned 20 in 1998). Also, 1940
is nearly the earliest year in which anyone in my data was born, so that year will
bound my sample on the early birth year side. Given that I consider the e↵ects
of shocks that occurred during the birth year and the subsequent 20 years, a large
number of cohorts is necessary to generate within-location variation in the numbers
of shocks experienced at di↵erent ages.
In order to maximize the likelihood that su ciently many of the people in
my sample had actually completed their educational career by the time they were
interviewed, I only consider women and men who were at least 21 years old then. In
addition, given that rainfall shocks a↵ect residents of agricultural areas most directly,
I restrict my sample to women and men who lived in rural areas as children.
While the DHS data does allow me to observe how long a respondent has
lived in their current location (which, combined with age data, allow me to observe
whether or not they had lived in their current location their entire life), it contains
practically no information on where respondents previously lived. Thus, if I want
to assign people who are at least 21 years old the proper shocks variables, then my
sample needs to consist of non-migrants only. According to a simple Heckman sam-
ple selection-type model, if people chose to migrate-and therefore exit the universe
of potential respondents from which my sample is drawn-because of the kinds of
rainfall shocks conditions they experienced between birth and the time they turned
21, and these rainfall shock-induced migrants were either more or less likely to be
literate (for example), then my estimates of the e↵ects of shocks on the literacy for
40
my sample of non-migrants will be biased.
Finally, it is also possible that my estimates of the e↵ects of rainfall shocks on
human capital outcomes will be biased because of selective mortality. More specif-
ically, if rainfall shocks during the first 21 years of life a↵ect individuals’ mortality
then once again, amongst my sample (of living individuals), the numbers of rainfall
shocks experienced and error terms could be correlated. Given the evidence de-
scribed in Section 2.2 on the procyclicality of young children’s health outcomes in
developing countries (and since all of that evidence was based on the e↵ects of nega-
tive aggregate income shocks) we might expect that dry years during preschool ages
would be the type of shock most likely to have caused children to die before they
had a chance to end up in my sample. If we assume also that these deceased individ-
uals would have been particularly unlikely to end up being literate or well-educated,
then there will be a positive correlation between the numbers of dry shocks that
individuals in my sample experienced early on and their idiosyncratic error terms,
which would bias the estimated e↵ects of dry years on individuals’ human capital
to the right on the real number line. 19
In Section 2.5, I will discuss these issues in more detail and attempt to place
bounds on the extent to which my baseline estimates might be biased as a result
of selective, rainfall shock-induced migration and mortality. In all that follows,
however, it should be kept in mind that under strong assumptions on how strong
the survival e↵ects of shocks are, many of the baseline estimates could be overstated
19It should also be kept in mind that the baseline results are robust to the exclusion of individuals
born before 1950, i.e. the exclusion of the subset of the sample that we might expect to be most
heavily subject to mortality-based sample selection bias (results available upon request).
41
to a non-trivial extent. But as mentioned in the introduction, even under strong
assumptions selection bias is not large enough to alter the qualitative features of the
baseline findings, which are that both wet and dry years have large negative e↵ects
on women’s human capital stocks and (more modest) positive e↵ects on men’s.
2.4 Descriptive Statistics and Results
2.4.1 Descriptive Statistics
Descriptive statistics (calculated using the weights described in Section 2.3.1)
for the samples of women and men can be seen in Table 2.2. Sample literacy rates
(of 17 percent and 44 percent for women and men) and highest grades completed (of
1.71 and 4.10 for women and men) are low, particularly for women, with nearly 84
percent of women not having completed a single year of schooling (see Figure 2.1).
There is substantial variation in human capital stocks across countries, with higher
than average literacy rates and mean highest grades completed in Cote d’Ivoire,
Ghana, Liberia, Nigeria and Togo, but stocks are fairly low in all 12 countries
(again, particularly for women). Literacy rates are also presented separately for
respondents interviewed before 2000 (when literacy was self-reported) and respon-
dents interviewed in or after 2000 (when interviewers made their own decisions on
how literate respondents were). The birth year distributions show that the men’s
sample is a bit older than the women’s; this is due to the fact that, again, men as
old as 59 were interviewed while only women under the age of 50 were. My earlier
claim regarding the frequency of the types of shocks that I consider is also borne out
42
by the descriptive statistics: People experienced around 2 and 5.5 of what I have
defined to be wet and dry years over the first 21 years of their lives, respectively.
The gender discrepancies in numbers of shocks experienced can be explained by the
facts that the men’s sample is older and that regional precipitation levels began a
marked decline starting in the late 1960s (see Hulme et al., 2001, as well as Figure
2.2). Similarly, the fact that the birth year distributions of the women’s and men’s
samples are as skewed towards later years as they are explains the greater frequency
of dry years experienced relative to wet years experienced. 20
To get a sense of the kinds of rainfall conditions that generate my results, refer
to Figure 2.3, which displays the averages (across individuals) of the numbers of
rainfall shocks of di↵erent intensities that people experienced over the first 21 years
of their lives. As the figure makes plain, the great majority of annual rainfall levels
were within 2 standard deviations of the local mean (over the previous 25 years).
The figure also reflects the aforementioned sustained decline in annual rainfall levels
in the way drier-than-average years were more common than years when rainfall was
above average.
Also, as Figure 2.4 makes clear, becoming literate can be a long process. Com-
pleting another year of schooling seems to add about 10 percentage points to the
likelihood of being literate on average, for both women and men. Thus, even if a
child is enrolled every year starting from the year she turns 6, she might still be
learning to read at the age of 16. This implies that individuals’ likelihoods of be-
20Baseline results are not sensitive to the truncation of the men’s sample so that the sample
only includes men younger than age 50, similarly to the women’s sample (results available upon
request).
43
ing literate (as well as their highest grades completed) could be a↵ected by shocks
experienced throughout their teenage years.
Finally, Figure 2.5 provides suggestive evidence of the presence of age heaping.
In settings like rural West Africa, it is reasonable to expect that (for older cohorts in
particular) ages and birth years are not recalled with perfect accuracy. This would
have serious consequences for the analysis here if su ciently many people reported
being born in the same year (e.g., a year ending in ’0’ or ’5’). However, as Figure 2.5
shows, this is not obviously a first order problem. For example, while a particularly
large number of women report being born in 1960, it seems unlikely that the results
would change drastically if the distribution of birth years was more even around
that time, especially since the key independent variables change relatively slowly as
birth years increase.
2.4.2 Results
2.4.2.1 Baseline Results
Table 2.3 contains the baseline results for the literacy and highest grade com-
pleted outcomes, for women and men separately. 21 For women, both wet and dry
years (significantly) decrease the likelihood of being literate and the highest grade
completed (see columns 1 and 2, respectively). Literacy and the highest grade com-
plete are each a↵ected by wet shocks during the latter two age intervals and dry
21Once again, since rainfall shocks before age 21 might have a↵ected who ended up in my sample
of (living) non-migrants, it is possible that some of the results that will be discussed here and in
the following sections are subject to sample selection bias (see Section 2.5.2).
44
shocks during all three. The F-statistics corresponding to the test of the hypothesis
that the shocks coe cients are jointly zero are greater than 4 for both outcomes.
The magnitudes of the point estimates are (again) seriously large, with (for
example) additional wet (dry) years between the ages of 7 and 13 (birth and age 6)
decreasing women’s likelihood of being literate by 1.2 (0.6) percentage points. These
estimated e↵ect sizes are substantial compared to the women’s sample literacy rate
of 17 percent. Moreover, the implied average e↵ects of shocks are very large as well:
The set of all shocks experienced over the first 21 years of life decrease women’s
likelihood of being literate by 22.1 percent. The magnitudes for the highest grade
completed outcome imply that, for each additional wet (dry) year between the ages
of 7 and 13 (birth and age 6), roughly 1 in 13 (18) women completed 1 less year of
schooling. In percentage terms, these e↵ect sizes are considerably larger than those
implied by Shah and Millett Steinberg (2013), where the implied e↵ects of, say, an
additional wet shock between the ages of 6 and 13 imply a 2.4 percent decrease in
the highest grade completed outcome, as compared to the 4.4 percent decrease in
that outcome here following a wet shock between the ages of 7 and 13. 22 Finally,
given that, as Figure 2.4 shows, an additional year of schooling is associated with
a 10 percentage point increase in the likelihood of being literate, the magnitudes
of the wet and dry years estimates for the highest grade completed outcome seem
consistent with their analogs for the literacy outcome if a substantial proportion of
the literacy e↵ects is being driven through decreases in time spent in school. In
22Although Shah and Millett Steinberg (2013) define shocks di↵erently, this is unlikely to explain
the di↵erence cited here, since my baseline results are robust to reasonable changes in shock
definition thresholds (as discussed in Section 2.3.2).
45
particular, the highest grade completed e↵ects are reasonably close to 10 times as
large as the literacy e↵ects.
How might rainfall shocks a↵ect women’s human capital stocks? Here I con-
sider the e↵ects of shocks after the age of 6 (and separately analyze shocks at
preschool ages in the next sub-section). The negative e↵ects of wet years on women’s
human capital can be explained by the likelihood that wet years increase female la-
bor demand (both on and o↵ the farm as well as inside the home), and that girls
respond by working more and going to school less; in terms of the conceptual frame-
work outlined in Section 2.2, substitution e↵ects dominate income e↵ects for females
during wet years. This interpretation is consistent with existing findings on the ben-
eficial e↵ects of rainfall on agricultural output and household incomes (Webb and
Reardon, 1992; Jensen, 2000), which in turn drive economic activity in rural areas
(Iliya, 1999; Mueller and Zevering, 1969). Moreover, Bhalotra and Umana-Aponte
(2010) find that African women’s labor supply is procyclical, and other research
shows that young females in West Africa tend to work on the farm, o↵ the farm
in small family-owned firms, and at home doing chores and taking care of younger
siblings (Herz et al., 1991; Ajani, 2008; Canagarajah and Skyt Nielsen, 1999; Rob-
son, 2004; Bhalotra, 2003; Glick and Sahn, 2000). 23 All of this would also be
consistent with the results of Edmonds (2005), who finds that children work in the
household in ways that their comparative advantages would suggest they would.
But given the aforementioned evidence on the procyclicality of female labor
23This interpretation of substitution e↵ects dominating is consistent with that of Shah and
Millett Steinberg (2013), who find that wet shocks worsen Indian students’ test score performance
and decrease Indian adults’ highest grade completed and wages. In contrast to the findings here,
however, the substitution e↵ect dominates the income e↵ect for dry as well as wet shocks.
46
supply in Africa, it is likely that adolescent girls were working less during and after
dry shocks. Indeed, evidence suggests that opportunities for work in rural West
Africa are few and far between at such times (Fafchamps et al., 1998; Mortimore,
1988; ICRISAT, 1980). The most likely explanation for the negative e↵ects of dry
years on women’s human capital stocks, then, is that parents simply were not willing
to finance their daughters’ educations when harvests were poor and incomes were
low. 24 Indeed, the low levels of female human capital-in absolute terms as well
as relative to males-certainly suggest that girls’ school fees might have commonly
been one of the first expenses cut after negative income shocks.
Switching to the men’s results in columns 3 and 4 of Table 2.3, we see quali-
tatively di↵erent e↵ects of shocks. In particular, additional wet years between the
ages of 14 and 20 positively a↵ect men’s likelihood of being literate, and dry years
between birth and age 13 have beneficial e↵ects on both human capital outcomes.
The e↵ects of wet shocks on the highest grade completed before age 14 are less
straightforward: They are positive at preschool ages but negative between the ages
of 7 and 13. The hypothesis that shocks do not jointly a↵ect men’s human capital
stocks is rejected for both outcomes.
The magnitudes of the estimated e↵ects of shocks are not particularly large
for men, with an additional wet (dry) year from age 14 to age 20 (birth to age 6)
increasing the likelihood of being literate by 1.5 (1.1) percentage points. Given the
men’s sample literacy rate of 44 percent, these magnitudes imply that additional
24Again, these negative e↵ects of dry shocks on completed human capital stocks are in contrast
to the findings of Shah and Millett Steinberg (2013).
47
wet and dry shocks during the age intervals in question increase men’s likelihood of
being literate by about 3.4 and 2.5 percent, respectively. 25 Both wet and dry
shocks at preschool ages also significantly increase men’s highest grade completed,
with an extra wet (dry) year then meaning that about 1 in 16 (13) men eventually
complete an additional year of schooling.
How might wet shocks after age 6 be a↵ecting men’s literacy? From the
perspective of the conceptual framework, it would seem that substitution e↵ects
dominate income e↵ects between the ages of 7 and 13 (and that men spend less
time in school as a result), but that income e↵ects dominate afterwards until age 20.
Given that wet shocks between the ages of 7 and 13 do not decrease men’s likelihood
of being literate, however, perhaps it is men who were either already literate or who
would not have gone on to become literate whose enrollment was a↵ected. On the
other hand, since wet shocks between the ages of 14 and 20 have positive e↵ects on
men’s literacy but not their schooling attainment, these shocks seem to have caused
men to spend more time learning while enrolled in school rather than increased
the number of years they were enrolled. Perhaps wet shocks cause some younger
men to increase their labor supply and decrease their school enrollment before age
14, but afterwards result in the kinds of higher household incomes that o↵er a last
chance for other kinds of men to learn how to read. These positive human capital
e↵ects of wet shocks during the third age interval stand in contrast to the negative
e↵ects we have observed for females. Given the di↵erences in human capital stocks
25Given men’s average numbers of shocks experienced, average e↵ects are also moderately sized
relative to the sample average literacy rate: Shocks as a whole increase men’s likelihood of being
literate by 9.9 percent.
48
across genders, it is perhaps not surprising that young men’s time spent learning
would have increased (between the ages of 14 and 20) thanks to rainfall-induced
lower marginal utilities of consumption, even as young women responded to the
increased demand for labor by working more. After all, cheap labor is scarce during
wet years, so parents choosing to have girls work and boys study more then would
simply be another example of parents favoring boys’ education when resources are
scarce (Chibiko O↵orma, 2009).
More generally, the fact that shocks negatively a↵ect females’ literacy in a
way that they do not negatively a↵ect males’ is consistent with Bjorkman-Nyqvist
(2013), who finds that dry shocks decrease female school enrollment but not male
school enrollment. Also, de Vreyer et al. (2012) find that the completed schooling
attainment of girls is more severely negatively a↵ected by locust swarms than that
of boys was in Mali (one of the countries my sample is drawn from). The pattern of
results in this chapter is also consistent with the findings of Manacorda (2006), which
show that the main beneficiaries of co-resident children’s labor are other children
themselves. In other words, the fact that some children work frees up other children
from the responsibility to do so, as seems to be the case here (along gender lines
strictly). A final reason why it might not be surprising that shocks have e↵ects
on female human capital that are more negative than the e↵ects on males is that
most West African tribes are patrilocal. Parents can therefore expect to receive
more of the benefits of the human capital investments they make in their sons,
while daughters’ husbands’ families will be the chief beneficiaries of investments in
daughters.
49
It is important to keep in mind, however, that the results are merely con-
sistent with these explanations, and there might have been other phenomena at
work. For example, following wet shocks, and from the perspective of the concep-
tual framework, the (magnitudes of) substitution e↵ects relative to income e↵ects
are larger for females than they are for males. In principle, however, there could have
been a number of other factors at work following wet shocks: The opportunity cost
of teachers’ time–and therefore their attendance–could have been a↵ected; school
lunches could have increased with public budgets; large amounts of rain on tin roofs
could have made learning di cult; violent conflict could have become less likely;
and certain individuals’ migration propensities could have been a↵ected. Moreover,
these other e↵ects could have worked alongside competing income and substitution
e↵ects. For example, to the extent that large amounts of rainfall, say, negatively
a↵ected teacher attendance, then the e↵ects of wet shocks on both females’ and
males’ completed human capital stocks were each pushed further to the left on the
real number line. These other kinds of e↵ects could be working across years as well,
so that the estimated e↵ects of shocks are the result of events that occurred in the
period immediately after shocks occurred as well as other events in later periods.
For example, if shocks before age 7 result in the migration of household members
that in turn mean higher remittance income for the sending household years later,
then dynamic e↵ects over the medium term could be important. Unfortunately, the
retrospective nature of the analysis makes it very di cult to determine what the
most important mechanisms driving the permanent e↵ects of shocks might be. In
any case (and with this kind of limitation in mind), I next consider the mechanisms
50
through which rainfall shocks at preschool ages might be a↵ecting women’s and
men’s human capital stocks.
2.4.2.2 Shocks at Preschool Ages
As the conceptual framework demonstrates, rainfall shocks during preschool
ages can have permanent e↵ects on human capital if those shocks have persistent
health e↵ects which later influence parents’ education investment decisions. I test
for evidence of this mechanism by considering how rainfall shocks a↵ect women’s
height, which is commonly taken to be a measure of early-life health conditions.
Unfortunately, height data was only collected for a subset of the women in my
sample-which precludes the possibility of analyzing how shocks might have a↵ected
men’s health at preschool ages. Also, the women for whom height data was collected
were not always randomly selected. For example, in older surveys height data tended
only to be collected for women who had given birth in the last five years. Thus, it
is possible that rainfall shocks before the age of 21 a↵ected who was eligible to be
in my sample of women for whom height data was collected (if, for example, rainfall
shocks a↵ect fertility), which could lead to biased estimates of the e↵ects of shocks
on women’s height. To see if rainfall shocks did in fact a↵ect who could have ended
up in my sample of women with height data, I regressed an indicator for whether
height data was available on all of the covariates from my baseline specification
along with survey year fixed e↵ects and controls for the kinds of rainfall shocks
conditions women experienced between the year they turned 21 and the year they
51
were surveyed. 26 The results (available upon request) indicate that the only
shocks to have significant e↵ects on the likelihood of having had height data collected
were wet years between the ages of 14 and 20, with the estimate of the corresponding
coe cient equal to 0.005. 27 Thus, if wet shocks caused (for example) particularly
tall (short) women to enter the sub-sample of women with height data, then the
estimated e↵ects of wet shocks between age 14 and age 20 will be biased to the right
(left) on the real number line. In short, the third age interval wet shocks estimate
for the height outcome should be interpreted with caution.
To conduct the actual analysis, I simply regress a variable equal to an individ-
ual woman’s height (in millimeters) on the same set of regressors as in my baseline
analysis. The results can be seen in Table 2.4, with columns 1 and 2 containing lit-
eracy and highest grade completed results for the subset of women for whom height
data is available and column 3 containing the results for the height outcome. As
the first two columns show, the wet and dry shocks estimates are broadly similar to
their baseline analogs in Table 2.3, with both wet and dry years having significantly
negative e↵ects (leading to a rejection of the null hypothesis that shocks do not
jointly a↵ect literacy and schooling attainment).
The results in column 3 give very little evidence that rainfall shocks during
preschool ages lead to persistent health e↵ects: The magnitudes of both the wet and
26Since rainfall shocks are slightly serially correlated and rainfall shocks after age 21 could have
a↵ected whose height was measured (through their e↵ect on fertility), neglecting to control for
shocks conditions for the post-age 21 period could lead to biased estimates of the main covariates
of interest, the wet and dry shocks for the three age intervals between birth and age 20.
27One explanation for this could be that wet shocks increase the likelihood of becoming pregnant
(and so the likelihood of having had height data collected).
52
dry estimates are small and far from significant. 28 Moreover, as described below
in Section 2.5, the e↵ects of dry shocks before age 7 could be biased to the left on
the real number line due to survival- and migration-based sample selection, in which
case the true e↵ects of shocks on women’s height are if anything more positive than
the results in Table 2.4 suggest. There is therefore no evidence of a negative e↵ect
of shocks on women’s height, and the positive e↵ects of dry shocks before age 7 on
women’s survival discussed in Section 2.5 suggests that if anything, dry shocks result
in positive health shocks for girls. Also, according to results described in Appendix
A, the e↵ects of dry shocks between birth and age 6 on women’s human capital are
driven primarily by shocks during the first 4 years of life. This seems more consistent
with some type of health channel rather than, say, a lagged income e↵ect of shocks
at preschool ages preventing households from having the resources to send female
children to school later. 29 It is also possible that dry shocks early on did in fact
result in health shocks that subsequent educational investment decisions took into
account, but those health e↵ects were not permanent and therefore are not reflected
in adult women’s heights. If the negative e↵ects of dry shocks before age 7 on human
capital did in fact operate through a health channel, then it seems more likely that
dry shocks negatively a↵ected young girls’ health and that parents invested less in
the education of those females in response. 30 Thus, while these results do not
28The magnitudes of the shocks coe cients point estimates are each considerably less than one-
tenth of 1 percent of the size of the sample average height of about 1,596 millimeters.
29Existing evidence on the persistent income e↵ects of dry shocks suggests that shocks at
preschool ages could a↵ect women’s human capital stocks in a way that is similar to the way
shocks at more advanced ages do, namely through their (lagged) e↵ects on parents’ willingness to
finance girls’ school enrollment (see Dercon, 2004).
30It need not be the case that dry shocks negatively a↵ect female human capital, however. As
the conceptual framework states, it is possible that dry shocks result in fewer health-promoting
53
strongly support the possibility that shocks a↵ect completed human capital stocks
through e↵ects on early-life health, they are consistent with that possibility and
they do not seem consistent with some alternative explanations.
While wet shocks at preschool ages do not seem to a↵ect women’s height, the
estimated e↵ects of wet years between the ages of 7 and 13 and between 14 and 20
are significantly positive. During each age range, an additional wet year is estimated
to increase height by at least 2 millimeters. 31 It is possible that wet shocks at
these advanced ages result in the kinds of positive nutrition shocks that increase
women’s final, adult heights. This would be consistent with the results of Aguero
and Deolalikar (2012), who find that exposure to the Rwandan genocide (and the
negative nutrition shocks that followed) substantially negatively a↵ected the adult
heights of females as old as 18 at the time of the conflict.
For men, while the baseline results from Table 2.3 suggest that wet shocks
at preschool ages positively a↵ect schooling attainment, the results in Table A.1
suggest that these e↵ects are driven more or less entirely by shocks that occur at
private goods but more parental time inputs, and that the latter e↵ect dominates the former.
Also, as Skoufias et al. (2011) note, breeding conditions for vector-borne diseases are hindered by
insu cient rainfall. It is therefore possible that dry shocks improved young females’ health, and
parents invested less in these more health females as a result.
31It is possible that the sign and magnitude of these estimates could be driven to some extent by
two types of sample selection bias. First, there is the just-mentioned possibility that the estimates
of the third age interval wet shocks coe cient could be subject to sample selection bias. Also,
as Section 2.5.2 shows, wet shocks between the ages of 14 and 20 increase the numbers of non-
migrant women in my sample. If we make the usual assumption that the would-be migrants who
wet years kept from moving tend to be high types, and in particular that they tend to be taller than
the average non-migrant woman in my sample, then there will be a positive correlation between
the numbers of wet shocks experienced over these two age ranges and idiosyncratic error terms
(amongst the women in my sample). This would of course lead to estimates of the e↵ects of wet
shocks on height that are biased to the right on the real number line, which could help explain
the sign and size of the estimates in question. But according to the results in Section 2.5.2, if
anything, wet shocks between the ages of 7 and 13 are causing female ’high types’ from migrating
out of the sample and causing ’low types’ to survive. Thus, the positive e↵ects of wet shocks from
age 7 to age 13 on women’s height are potentially understated.
54
age 5 only. Also, the beneficial e↵ects of dry shocks between birth and age 6 are
driven by shocks at ages 2, 3 and 4. It is possible that following both types of
shock, material resources are concentrated on boys and their health improves, and
that parents invest more in boys’ education later. Alternatively, it could be that
shocks result in negative health shocks for boys but that parents compensate for this
later in the form of greater education investments. With no direct measure of men’s
early-life health conditions, however, the evidence in favor of shocks permanently
a↵ecting literacy or the highest grade completed at school through a health channel
is of course quite limited. But the data do not seem to support the possibility of
early-life shocks a↵ecting later school attendance through a lagged asset or input
availability channel.
2.4.2.3 Shocks and Schooling Along the Extensive Margin
Given the high proportion of the individuals in my sample who have not com-
pleted a single year of schooling, I look for evidence that certain women’s and men’s
educational careers never even began because of shocks. To do this, I regress an
indicator equal to 1 if an individual completed at least 1 year of schooling (and zero
otherwise) on my baseline covariates.
The results for women and men can be seen in columns 1 and 2, respectively,
of Table 2.5, where we see that dry years between birth and age 6 do in fact a↵ect
educational attainment at the extensive margin for both women and men. The
estimated e↵ect for women of an additional dry year before age 7 implies that each
55
additional dry year then decreases the likelihood of having completed at least one
year of schooling by six-tenths of a percentage point, which is substantial given that
only about 25 percent of the women in my sample completed at least 1 year of
schooling. For men, dry shocks at preschool ages increase the likelihood of having
completed at least 1 year of schooling by 1.1 percentage points, which, once again,
is not particularly large compared to the men’s sample average for this outcome of
about 49 percent.
There is also evidence that wet shocks between the ages of 14 and 20 signif-
icantly a↵ect educational attainment along the extensive margin for both women
and men, and similarly for dry shocks at that age for women. Once again, the
estimated e↵ects are negative for women and positive for men. An additional wet
shock during this third age interval decreases the likelihood of completing at least 1
year of schooling by 1.2 percentage points for women and increases that likelihood
by 0.9 percentage points for men. 32 An additional dry shock after the age of
13 decreases women’s likelihood of having completed one year of schooling by 0.4
percentage points. It seems surprising that rainfall shocks could a↵ect this outcome
at ages this advanced-we would normally expect the decision to attend school at all
or not to have been made at younger ages. One possible explanation for these results
is that, following some shock to school availability (which was unrelated to rainfall
32Note that these estimated e↵ect sizes are quite large, especially for females. In particular, out
of the roughly 25 percent of all women in the sample who completed at least 1 year of schooling,
the magnitude of the e↵ect of experiencing an additional wet shock between the ages of 14 and 20
implies that a surprisingly large proportion of women must have been only beginning their formal
education above the age of 13 (and that a very large proportion of these must have been a↵ected
by wet shocks). Also, again, under strong assumptions both the women’s and men’s estimates
of shocks e↵ects could be non-trivially biased because of self-selection into the sample of living
non-migrants.
56
shocks), demand for schooling was unusually high amongst currently teen-aged in-
dividuals eager to take advantage of previously unavailable education opportunities.
33 Obviously, in a situation like this individuals’ educational careers were poten-
tially just beginning, which would explain how the outcome in question could have
been a↵ected during the advanced stages of adolescence.
2.4.2.4 Shocks Numbers Indicators Results
The baseline results assume that the e↵ects of additional wet and dry years
are linear. In order to learn whether this linearity assumption might be justified
and whether more shocks always have more severe e↵ects than fewer, I will now
substitute series’ of dummy variables for each of the shocks variables in the baseline
specification. More specifically, for each of the three wet (dry) years shocks variables,
I construct three dummy variables that are equal to 1 when the number of wet (dry)
years experienced during the age range in question is equal to 1, 2 and at least 3,
respectively (and equal to zero otherwise). In other words, zero shocks experienced
will be the omitted category for both types of shocks and for each of the three age
ranges.
The literacy and highest grade completed results for women can be seen in
columns 1 and 2, respectively, of Table 2.6. The results show that more wet years
between birth and age 6 do not always seem to have more severely negative e↵ects
on the literacy outcome, and similarly for both outcomes between the ages of 7
33Osili and Long (2008) and Oyelere (2010) find that educational attainment was positively
a↵ected when school fees were waived as part of a Universal Primary Education program in Nigeria.
57
and 13. For example, while women who experienced 1 wet year between birth and
age 6 had a 0.9 percentage point lesser chance of being literate than women who
experienced no wet shocks then, there is no di↵erence in the likelihood of being
literate between women who experienced 2 wet shocks and women who experienced
none (statistically speaking). Also, F-tests reject the hypothesis that additional
shocks e↵ects have equally-sized e↵ects for wet shocks during both of the first two
age intervals. More wet shocks are consistently worse than fewer for both outcomes
between the ages of 14 and 20 (with the e↵ect sizes of each additional shock once
again being statistically indistinguishable from one another). For this third age
interval, the negative e↵ects of wet shocks seems to be driven in large part by the
third-or-higher category, with a↵ected women having a likelihood of being literate
that is 5.2 percentage points lower than the analog for women who experienced no
wet shocks then. 34
Turning to the women’s dry years results, we see that additional dry years
between birth and age 6 do not always have more severely negative e↵ects than
fewer for the literacy outcome, but that they do for the highest grade completed
(with, for example, a third-or-higher dry shock implying that the average highest
grade completed of a↵ected women was lower than that of women who experienced
zero dry shocks then by more than 13 percent). Given the possibility that dry shocks
early on a↵ect women’s human capital through a kind of health e↵ect, the relatively
34One possible explanation for the relative strength of higher-order wet shocks during the 14-
20 age range is that the cost of re-enrolling following some period of time spent out of school is
increasing in that time. In this case, if enough women spent su ciently long periods of time outside
of school thanks to multiple wet shocks, we would expect higher-order shocks to have particularly
large e↵ects.
58
large magnitude of the third-or-higher e↵ect here could reflect the possibility that
it takes multiple shocks for young girls’ health–and later education investments–to
be a↵ected. More dry years between the ages of 7 and 13 and between 14 and
20 also tend to be consistently worse than fewer for both outcomes. Consistent
with the baseline results, the e↵ects of dry years during this third age interval
are smaller than dry years e↵ects at preschool ages, with a third-or-higher shock
estimate implying that a↵ected women were 1.3 percentage points less likely to be
literate compared to women who experienced zero shocks. Once again, the e↵ects
of a third-or-higher shock seem relatively large, which could imply that it takes
multiple dry years during this age range for household incomes to fall su ciently
far that girls’ education investments are curtailed.
For men (see columns 3 and 4), we see that wet shocks during the first age
interval seem to a↵ect the highest grade completed in a very particular way, with
the estimate of the first shock e↵ect being positive and large, the second estimate
being positive and slightly larger than the first, and the third not being statistically
significant. From age 7 to age 13, there is no evidence that wet shocks substantially
a↵ect men’s literacy, but the signs and magnitudes of the point estimates for the
highest grade completed outcome show that more wet years then decrease schooling
attainment. Also, more shocks between the ages of 14 and 20 tend to have more
strongly positive e↵ects on both human capital outcomes than fewer (though none
of the estimates for this age interval are significantly positive for the highest grade
completed outcome). 35 The magnitude of the third-or-higher point estimate
35E↵ect sizes for the literacy outcome are, however, statistically distinguishable according to the
59
implies that going from experiencing zero wet shocks to experiencing at least 3
increases the likelihood of being literate by 5.5 percentage points. Also, the fact
that the e↵ect of the second and third-or-higher shocks are so much larger than the
e↵ect of the first could mean that it takes a few wet shocks for income e↵ects to
dominate and for men to actually start spending substantially more time learning;
perhaps men are only able to devote more tim to learning after household incomes
have been high for a number of years.
Similarly, more dry shocks always have stronger positive e↵ects on men’s hu-
man capital than fewer, with (for example) the likelihood of being literate estimated
to be 2.7 percentage points higher for those men who experienced at least 3 dry
shocks as compared to those men who experienced zero. For the second and third
age interval dry years e↵ects, the results in Table 2.6 are consistent with the baseline
results in that there is no good evidence that more dry years have stronger e↵ects
than fewer. Thus, for a substantial proportion of the (women’s and men’s) signifi-
cant shocks numbers estimates from the baseline results, we see evidence that more
shocks do in fact have stronger e↵ects than fewer. To sum up the baseline results, it
would seem that in contrast to what Jacoby and Skoufias (1997), Funkhouser (1999)
and Bjorkman-Nyqvist (2013) each suggest, there are substantial permanent e↵ects
of shocks on a↵ected individuals’ human capital.
results of an F-test.
60
2.4.3 E↵ects by Agroecological Zone
West Africa is a large place that includes multiple agroecological zones. In
particular, as mentioned in Section 2.2, there is substantial variation in average
rainfall levels across latitudes, with areas farther to the north typically being more
dry. It is possible that shocks in drier regions have di↵erent e↵ects than do shocks
in wetter regions. For example, it could be that wet shocks in dry areas do not
increase agricultural productivity, perhaps because of weaker relationships between
the crops typically grown there and quantities of water. Similarly, dry shocks in
drier regions could have less severely negative consequences for household incomes
if crop choices are such that any associated risks are minimized ex ante.
I explore how shocks e↵ect di↵er by agroecological conditions by interacting
the shocks numbers variables from the baseline specification with indicators equal to
1 if the individual’s rainfall neighborhood is in a region that is at least 50 percent arid
or semi-arid according to the Food and Agriculture Organization’s Agroecological
Zones Framework. 36 This includes all of Burkina Faso, Mali, Niger, Senegal
and (the more northern) regions of Benin and Nigeria. The results can be seen
in Table 2.7. It seems clear that for women, both wet and dry shocks have more
severely negative e↵ects in wetter and more humid regions (if anything). This is only
true in a statistical sense when it comes to the e↵ects of wet shocks between the
ages of 7 and 13, but the way the magnitudes vary by arid/semi-arid region status
for the other shocks main terms relative to the baseline results also appear to be
36See http://www.fao.org/wairdocs/tac/x5756e/x5756e0j.htm (accessed on November 27, 2014)
for more information.
61
consistent with this conclusion. For men, it appears that the baseline wet birth-age
6 results mask heterogeneity by agroclimatic zone, with men in more humid regions
benefiting from these shocks then and men in drier regions being negatively a↵ected
by them. Otherwise, it appears that, if anything, the human capital stocks of men in
more arid regions are more positively a↵ected by both wet and dry shocks (during
the third and first age intervals, respectively) than are men in less arid regions.
It appears that households in more arid regions are more likely to focus resources
exclusively on sons following shocks, perhaps because of higher baseline household
poverty rates, which (again) would be consistent with households choosing to invest
in boys only when resources are scarce.
Overall, these results are consistent with the possibility that in arid regions,
the crops typically planted are less dependent on or sensitive to rainfall. It is also
possible, however, that the e↵ects of shocks on women’s human capital seem to
be more severely negative in more humid regions because these are the only regions
where there is meaningful variation in the outcome variables. In particular, as Table
2.2 makes clear, human capital levels are particularly low in countries that are either
entirely or mostly arid or semi-arid. Women’s human capital stocks in these places
might have therefore been particularly unlikely to have been negatively a↵ected by
rainfall shocks.
62
2.5 Robustness
2.5.1 Region-Specific Cohort Fixed E↵ects
We might worry that, even after controlling at the country level for cohort-
specific human capital levels, the resulting rainfall shocks estimates are to some
extent still capturing place-specific human capital trends along with the e↵ects of
rainfall shocks. To address this, I estimate the baseline specification with region-
specific birth year dummies in place of their country-specific analogs. Thus, this
version of the specification will include as many as 55 complete sets of birth year
fixed e↵ects-one for each region-as opposed to the 12 (countrys’ worth) from the
baseline version of the specification. 37
The results for women’s literacy and highest grade completed can be seen
in columns 1 and 2, respectively, of Table 2.8. They are highly similar to the
baseline results, with both wet and dry years once again having significantly negative
e↵ects on both outcomes and the null hypothesis of shocks not jointly a↵ecting
human capital rejected (for both outcomes). Indeed, for the literacy (highest grade
completed) outcome, the F statistic associated with the test of the null hypothesis
that the shocks numbers coe cients are jointly equal across specifications is about
1.033 (1.803), and the null hypotheses are therefore not rejected at the 5 percent
significance level. 38 Increasing the flexibility of the baseline specification with
37Since there are no men from Cote d’Ivoire in my sample, the men’s sample is of course spread
over fewer than 55 regions.
38The null hypothesis that the shocks coe cients are jointly equal across specifications for the
highest grade completed is rejected at the 10 percent level, however.
63
respect to how time series e↵ects are controlled for does not result in substantial
(or significant) changes to the results for women in the sense that both wet and
dry years are still estimated to have substantial, negative e↵ects on women’s human
capital.
Turning to the men’s results in columns 3 and 4, we see that they are highly
similar regardless of how cohort e↵ects are controlled for. In particular, additional
wet shocks between age 14 and age 20 still increase the likelihood of being literate
and wet shocks between age 7 and age 13 still decrease the time spent in school. On
the other hand, the estimated e↵ects of additional dry years during preschool ages
have decreased in magnitude and are no longer significantly positive. This could
reflect the possibility that trends in men’s literacy formation between birth and age
6 were more positive in those regions that became more dry over time relative to
trends in other regions. Once again, however, the null hypothesis that shocks e↵ects
di↵er across specifications which control for birth year fixed e↵ects at a finer or
coarser geographic level is not rejected for either outcome at conventional levels. 39
2.5.2 Sample Selection Bias Robustness Check
2.5.2.1 Shocks, Migration and Mortality: Theory
As mentioned previously, since my sample consists of living non-migrants only,
then if rainfall shocks before the age of 21 caused people who were especially likely
or unlikely to be well-educated to migrate (or not), or survive (or not), then my
39The associated F statistics for the literacy and highest grade completed outcomes are about
0.968 and 1.242, respectively.
64
baseline results will be biased. To illustrate the point, I will rely on a framework
that is loosely related to a Heckman selection model (Heckman, 1979).
I begin by considering how rainfall shock-induced migration could result in
the baseline estimates being a↵ected by sample selection bias. Suppose that the
utility that individual i receives from migrating is denoted by Mi1 and the utility
they receive from not migrating is denoted by Mi0. Then individual i will migrate if
and only if Mi1 Mi0 > 0. Let Mi be an indicator equal to 1 if individual i migrates
and zero otherwise (so that Mi is the observable outcome associated with individuali’s migration decision). Also, let Mi1  Mi0 be given by the right-hand side of the
following equation:Mipt = a+ 3X
j=1
Bj#wetyearsjpt + 3X
j=1
Cj#dryyearsjpt +dp +hct + lr(t)+X 0iptG+ eipt,
(2.2)
where once again p indexes individuals’ rainfall neighborhoods and t their years of
birth. Migration-and therefore the chance that an individual could possibly be in
my sample of non-migrants-is modelled here to depend on the same rainfall shocks
variables, rainfall neighborhood fixed e↵ects, country-specific birth year fixed e↵ects,
regional time trends and demographic variables as in the baseline specification.
Assume also that, conditional on observables, individuals with larger human
capital stocks are more likely to migrate, i.e. that Cov("ipt, eipt) > 0. (Recall that"ipt denotes the idiosyncratic error term from the baseline human capital specifica-
tion.) This is the standard assumption that migrants tend to be particularly highly
motivated or especially able. This assumption is reasonable here since, for example,
65
migrant women are about 50 percent and migrant men 37 percent more likely to be
literate than their respective non-migrant analogs.
If individuals can only be in my sample if they have never migrated and they
migrate if and only if Mi1   Mi0 > 0, then my sample of non-migrants will only
consist of individuals for whom the following condition holds:eipt   a  3X
j=1
Bj#wetyearsjpt  3X
j=1
Cj#dryyearsjpt dp hct lr(t) X 0iptG. (2.3)
If, for example, dry years cause individuals to move (i.e., if Bj > 0 for some j 2
{1, 2, 3}), then individuals with high eipt terms will tend to have experienced fewer
numbers of dry years in order for this condition to still hold and for those individuals
to still have a chance of being in my sample of non-migrants. Intuitively, since
individuals with high eipt terms were always more likely to migrate and dry years
increase the likelihood of migrating, then individuals in my sample of non-migrants
with high eipt will tend to have experienced fewer dry years during age intervalj. In other words, conditional on all other independent variables the covariance
of #dryyearsjpt and eipt will be negative amongst the non-migrants that comprise
my sample. Then, under the assumption that Cov("ipt, eipt) > 0, Bj > 0 will also
imply that the covariance of #dryyearsjpt and "ipt (conditional on controls) will be
negative in my sample of non-migrants. This will lead to estimates of the e↵ects of
dry shocks during age interval j on human capital outcomes that are biased to the left
on the real number line. Obviously, shocks estimates will also be biased if instead
dry years decrease the likelihood of migrating or if wet years a↵ect individuals’
migration decisions (again, as long as my sample only consists of non-migrants and
66
we assume that migrants tend to be high types relative to non-migrants).
In theory, a similar type of sample selection bias could arise if rainfall shocks
a↵ect the survival of certain types of individuals. It is possible that rainfall shocks
experienced during the first 20 years of life a↵ect the likelihood that individuals
die before they are old enough to possibly end up in the data I use. Moreover, it
is reasonable to expect that any individuals whose survival would be a↵ected by
rainfall shocks are less likely to have turned out to have been literate. In this case,
then amongst the (living) individuals in my sample, there would be a correlation
between the numbers of shocks experienced and the idiosyncratic error term from
the baseline human capital specification. For example, if dry shocks kill low types,
then in my sample there will be a positive correlation between the numbers of dry
shocks experienced and error terms, which will bias dry shocks estimates to the
right on the real number line. Thus, in the same way that rainfall shock-induced
migration could create sample selection bias, shock-induced mortality could also do
so.
Finally, it is very important to note that under reasonable assumptions, the co-
occurrence of migration-based sample selection and mortality-based sample selection
could have particularly important implications for the direction and magnitude of
the net sample selection bias. For example, if a particular type of rainfall shock
simultaneously causes literate people to migrate and illiterate people to survive,
then each of these individual sample selection forces would be working to make the
sample less literate. As we will see below, however, the number of migrants will have
much stronger e↵ects on the extent of sample selection bias than will the number
67
of survivors. It will therefore be important to obtain estimates of the numbers of
individuals entering or exiting the sample via both migration- and mortality-based
channels.
2.5.2.2 Shocks and Total Changes in Neighborhood-Cohort Sizes:
Empirical Evidence
To do this, I will first obtain estimates of the net e↵ects of rainfall shocks
on rainfall neighborhood-cohort sizes, i.e. the sums of changes in neighborhood-
cohort sizes from both shock-induced migration (or its absence) and mortality (or
survival). Then, in the next sub-section, I will obtain estimates of the mortality
e↵ects of shocks, which can then be combined with the net e↵ects to obtain implied
estimates of the migration e↵ects of shocks. 40
One way to examine sample selection due to the sum of migration and mor-
tality would be to consider whether or not the number of living non-migrants (in
my sample) who live in a particular place and who were born in a particular year
depends on the kinds of rainfall shocks that someone born in that same time and
place would have been exposed to. I do exactly this by estimating the following
equation:NLNMpts = ↵+ 3X
j=1
 j#wetyearsjpt+ 3X
j=1
⇢j#dryyearsjpt+Z 0pts⇧+ p+✓ct+⌧s+ r(t)+"pts.
(2.4)
Here NLNMpts is defined to be the total number of living rural non-migrants in my
40Although slightly inelegant, this two-pronged approach possesses the virtue of being feasible
given readily available data (the DHS data itself).
68
sample who are from (and currently living in) rainfall neighborhood p, born in year t
and surveyed in year s. To be clear, if a rainfall neighborhood was sampled in more
than one survey year (which is quite common), then there will be more than one set
of observations for that neighborhood (since, as described below, this specification
also controls for rainfall shocks conditions up until the year before survey year s).
For instance, if a place was surveyed once in 1990 and again in 1995, then there
will be one set of 30 observations corresponding to that first survey year–one for
each birth year between 1940 and 1969, the last year for which there will be people
from that place who were at least 21 years old in the survey year–and another set of
35 corresponding to the second survey year for that place (one for every birth year
from 1940 to 1974).
The #wetyearsjpt and #dryyearsjpt variables are all defined as before, and
the estimates of the  j and ⇢j coe cients will indicate whether or not wet and
dry years tended to either push people to leave a place (or die) or cause them to
remain there (or survive). More specifically, the estimated magnitudes of these
variables will measure how the size of the typical cohort (in the typical rainfall
neighborhood) was a↵ected by wet and dry shocks during the three age intervals.
The rainfall neighborhood fixed e↵ects ( p) control for each rainfall neighborhoods
long-term average population level (amongst other location-specific, time-invariant
unobservables), while the (country-specific) cohort fixed e↵ects (✓ct) and regional
time trend controls ( r(t)) are meant to control for temporal factors common to all
locations (within each country and region, respectively). Again, most neighborhoods
were sampled in more than one of the household surveys that I rely on, so to control
69
for survey year-specific unobservables I include survey year fixed e↵ects (⌧s) as well."pts is a mean-zero error term.
The Zpts vector contains controls for rainfall shocks before cohort t was born
and after they turned 21. Specifically, the vector contains two indicators for whether
or not the year before cohort t’s birth was wet or dry, since rainfall conditions in
these years might have a↵ected fertility and rainfall shocks are (slightly) serially
correlated. The Z vector also controls for rainfall shocks conditions after age 20
and before survey year s in two ways. The first pair of controls simply consists of
variables equal to the numbers of wet and dry years between the year in which cohortt turned 21 and the year before survey year s, while another pair consists of variables
equal to the proportions of years in that same post-21 time period (before survey
year s) that were wet and dry. I estimate the specification separately for each of
these types of post-21 shocks controls. The identification assumption here is similar
to that of the baseline analysis: Trends in the numbers of living non-migrants in
rainfall neighborhoods which became more or less dry over time have to be assumed
to be the same as trends in places that did not become more or less dry over time
for the estimated e↵ects of rainfall shocks to be consistent.
The results can be seen in Table 2.9, with the two sets of womens results in
columns 1 and 2 and mens results in columns 3 and 4. In general, both wet and dry
shocks tend to decrease neighborhood-cohort sizes for women and men. Dry shocks
that occur between the ages of 14 and 20 are an exception for women, however. In
order to know how these net e↵ects of shocks on neighborhood-cohort sizes break
down into migration and mortality e↵ects, we need estimates of the e↵ects of shocks
70
on at least one of these. I seek estimates of the mortality e↵ects of shocks next.
2.5.2.3 Shocks and Mortality: Empirical Evidence
I attempt to identify the extent of mortality-based sample selection bias by us-
ing data available for a subset of the females in the DHS data on the mortality status
of their siblings. Data on siblings’ birth years, survival status, gender, birth order
and (if applicable) death year are used to study whether or not rainfall shocks a↵ect
individuals’ likelihood of dying before reaching each of the ages of 20, 30, and the
gender-specific average age of the individuals in my sample. (These are the ages of
35 and 39 for women and men, respectively.) Examining the e↵ects of shocks on the
likelihood of surviving to these particular ages only will not necessarily provide a set
of estimates that can precisely estimate how large or small mortality-based sample
selection bias might be; shocks can be expected to have smaller impacts on the mor-
tality of individuals younger than average and larger impacts on older individuals’
mortality, for example. In an attempt to compensate for this imprecision, therefore,
I will choose estimates of the e↵ects of shocks on mortality (or survival) from that
outcome that happens to present the most severe test for my baseline results (given
the assumptions I make on the human capital stocks of migrants/non-migrants and
survivors/the deceased). I will describe this choice process in further detail below.
The e↵ects of rainfall shocks on mortality will be analyzed using the following
71
specification:Mortipt = ↵+ 3X
j=1
 j#wetyearsjpt+ 3X
j=1
⇢j#dryyearsjpt+ p+✓ct+ r(t)+X 0ipt +"ipt.
(2.5)
Here Mortipt is an indicator equal to 1 if individual i from rainfall neighborhood p
and born in year t died before a particular age. (Again, there are separate mortality
outcomes for ages 20 and 30 for both women and men, as well as age 35 for women
and 39 for men.) All other controls are the same as in the baseline specification.
After once again conditioning on rainfall neighborhood and country-specific birth
year fixed e↵ects, plus region-specific linear trend controls, the  j and ⇢j coe cients
will indicate whether or not wet and dry years tended to either aid or harm survival.
Sibling mortality data is collected in DHS surveys in an attempt to understand
the prevalence and determinants of pregnancy-related deaths. 41 In the surveys
that contain sibling mortality modules, all women interviewed are asked to list all
children born to their mothers, to provide data on the birth year and mortality
status of their siblings, and if applicable, siblings’ death years. The sample consists
only of siblings of non-migrant women from the countryside who were born between
1940 and 1978 (who were also su ciently old by the time the data was collected for
the mortality outcome variables to be properly defined). 42
The results in Table 2.10 suggest that wet (dry) shocks between the ages of
7 and 13 (birth and age 6, and age 14 and 20) aid the survival of women, while
41Due to the di culty of collecting sibling mortality information, only a subset of DHS’ contain
the associated modules. The countries (and years) for which sibling mortality is available are these:
Benin (1996); Burkina Faso (1998); Cote d’Ivoire (1994); Guinea (1999, 2005); Mali (1995, 2001,
2006); Nigeria (2008); Niger (1992); Senegal (2008); Sierra Leone (1992, 2005); and Togo (1998).
42I assume that siblings grew up in the same places where the (non-migrant) women respondents
were surveyed by DHS survey workers.
72
wet shocks between birth and age 6 have significantly negative e↵ects on men’s
survival. Unfortunately, mortality estimates based on these data are also subject to
sample selection bias. As mentioned above, this data only allows us to analyze the
survival of siblings of living women. In other words, the sample of siblings is selected
similarly to the way that the baseline human capital sample is selected, namely with
respect to survival status (of the women who provided the sibling data).
In the interest of moving forward despite imperfect data, I first assume that
the results of this analysis in Table 2.10 are not so a↵ected by this type of mortality-
based selection bias that the signs of the point estimates are incorrect. This assump-
tion will end up implying that the baseline human capital e↵ects of wet shocks for
women–which are some of the most striking results I obtain–are too negative. It
should therefore probably be considered an unfavorable assumption. I also assume
that siblings’ idiosyncratic survival factors are positively correlated, i.e. if one sib-
ling survived then their brothers and sisters were more likely to survive. When this
assumption is combined with the first assumption that the signs of the mortality
e↵ects of shocks in Table 2.10 are correct, then this second assumption is also in
many ways unfavorable. Finally, I assume that age di↵erences between siblings are
not so great that the shocks that siblings experienced are insu ciently highly corre-
lated with the shocks that the women who provided the data experienced. In other
words, I assume that the women who provided the data on their siblings’ mortality
statuses tended to experience su ciently many of the same shocks that their siblings
did. Assuming otherwise would once again imply that the baseline human capital
estimates were less biased than I will conclude they could be.
73
Under these assumptions, then amongst the siblings in my sample, there will
be a positive correlation between each of the wet shocks and dry shocks variables
and the idiosyncratic error term in the mortality equation above. In other words,
in order for (females and males) to end up in the sample of siblings, the living
women who provided the data and who experienced fewer wet or dry shocks will
tend to have lower idiosyncratic mortality factors themselves. Assuming siblings’
idiosyncratic survival factors are positively correlated therefore implies that each of
the wet and dry shocks coe cients estimates in Table 2.10 will be biased to the right
on the real number line. For women, then, the survival e↵ects of wet and dry shocks
during these age intervals will be understated, while men’s estimated dry shocks
mortality e↵ects will be overstated. Therefore, in the interest of being conservative,
I will assume that the actual mortality e↵ects of shocks are 3 times larger than the
estimates I obtain in Table 2.10. 43 This is likely a strong assumption, and it will
end up being quite unfavorable.
2.5.2.4 Sample Selection and Selection-Free Estimates
Given the mortality estimates of shocks in Table 2.10 and the net e↵ects of
shocks on the numbers of living non-migrants in my sample in Table 2.9, I can back
out the migration e↵ects of shocks. The implied numbers of individuals who enter or
exit my sample because of the e↵ects of rainfall shocks on mortality and migration
can be seen in Table 2.11. As mentioned above, the values in Table 2.11 represent
43As the results will show, 3 is the largest integer for which the sample selection-free analogs to
many of the most important baseline estimates would still be significantly di↵erent from zero at
conventional levels (assuming no change in standard errors).
74
the combinations of mortality and migration estimates that present the most severe
test for the validity of the baseline results. In particular, for each type of shock and
each age interval, the mortality estimate of shocks (out of the 3 possible ones in
Table 2.10) is assumed to be the one with the largest magnitude. This is because,
first and foremost, all individuals whose migration choices and mortality are a↵ected
by shocks are assumed to be literate and illiterate, respectively, with certainty. 44
Between the two sources of sample selection bias under consideration here, the
type that is certainly more powerful is that which is driven by migration. This is
because the sample average human capital stocks are so low. Intuitively, while both
migration and mortality a↵ect sample selection through their e↵ects on, for example,
the literacy rate–the number of literate individuals relative to the total number of
individuals–mortality/survival only a↵ects the relatively large denominator and so
only results in fairly small changes to the literacy rate. Migration, on the other
hand, a↵ects the comparatively small numerator (as well as the denominator), and
so causes much more sample selection bias. The mortality estimates from Table
2.10 that are assumed to be the true mortality e↵ects of shocks are therefore those
that either minimize the number of individuals who do not migrate because of
shocks, or maximize the number of individuals who do migrate because of them.
In both of these cases, migration-based sample selection decreases the literacy rate
to the greatest extent. Similarly, given these chosen mortality estimates, the total
44Again, migrants in the DHS data are considerably more likely to be literate than non-migrants,
and it also seems reasonable to expect that individuals whose very survival depends to some extent
on shocks would turn out to be illiterate. Assuming that everyone in these two groups of individuals
are literate and illiterate, respectively, maximizes the extent of the sample selection biases, and
these are also likely strong assumptions.
75
numbers of females and males who enter or exit the sample due to either migration
or mortality in Table 2.9 are chosen so that non-migration is minimized or migration
is maximized.
Table 2.11 contains the results of these choices. The values in the table refer
to changes in the numbers of individuals in the average rainfall neighborhood in
response to shocks, and values are only calculated when there is (statistically sig-
nificant) evidence in either Table 2.9 or Table 2.10 that shocks cause individuals to
either enter or exit the sample of living non-migrants. Columns 1 and 2 give implied
changes in the number of individuals in the sample due to mortality and migration,
respectively, assuming that the mortality estimates in Table 2.10 are correct, while
columns 3 and 4 of Table 2.11 give the analogs of these under the assumption that
the true mortality e↵ects of shocks are 3 times as large as the estimates in Table
2.10.
Table 2.12 contains the resulting sample selection-free human capital estimates
of shocks. More specifically, the table contains the baseline literacy results for
comparison purposes (in column 1), information on the signs of estimated sample
selection e↵ects of mortality and migration (in columns 2 and 3), and the implied
selection-free estimates of the e↵ects of shocks on literacy (in columns 4 and 5). The
column 4 estimates assume that mortality e↵ects are only as large as the estimates
in Table 2.10, while the column 5 estimates assume that those estimates understate
the truth by a factor of 3. For women, the negative e↵ects of dry shocks experienced
between the ages of 14 and 20 decrease substantially in magnitude when possible
selection e↵ects are taken into account. On the other hand, wet shocks during
76
the latter 2 age intervals and dry shocks in the first 7 years of life would still be
significantly negative (at the 10 percent level at least, and assuming no changes in
standard errors) and economically significant, even under the assumption of strong
survival e↵ects of shocks. For men, the positive e↵ects of wet shocks from age 14 to
age 20 and dry shocks before age 7 are only slightly diminished by sample selection
bias under strong assumptions.
Table 2.13 contains the highest grade analogs to the information in Table 2.12.
The negative e↵ects of wet and dry shocks on women’s highest grade completed are
less robust than they are for the literacy outcome. In particular, while there are
only fairly small changes in the e↵ects of shocks under the assumption that mortality
e↵ects are only as large as the estimates in Table 2.10 (in column 4), assuming they
are larger results in substantial changes to the magnitudes and significance of wet
shocks from age 7-13 and dry shocks from birth-age 6 and from age 14-20 (column
5). For men, the positive e↵ects of wet shocks over the first 7 years of life are quite
sensitive to possible sample selection e↵ects; the magnitude of this point estimate
drops substantially after taking into account the potential e↵ects of shock-induced
migration. This is less true for the dry shocks from birth-age 6 estimate.
It is important to note that weakening the assumption on the percentage of
individuals whose migration decisions were a↵ected by shocks were literate results in
selection-free estimates of the e↵ects of shocks that are much closer to the baseline
estimates than the estimates in column 5 of each of Table 2.12 and 2.13. Note that
this is consistent with the aforementioned power of migration-based sample selection
in particular to potentially seriously a↵ect the baseline estimates. For example,
77
assuming that women who migrated in response to wet shocks experienced between
age 7 and age 13 were only as likely to be literate as the average migrant in the DHS
data implies a selection-free estimate on literacy of -0.011, even under the stronger
assumption on how large the survival e↵ects of these shocks are. 45 (Compare
this point estimate to the baseline estimate of -0.012 and the strong survival e↵ect
selection-free estimate in Table 2.12 of -0.007.) Finally, it is important to realize
how large the mortality e↵ects of shocks must be for the mortality e↵ects of shocks
to actually be 3 times as large as the point estimates in Table 2.10. For example,
the number of women who survive until age 35 has to be more than 5 percent higher
following an additional wet shock between the ages of 7 and 13 for the strong survival
assumption to be true. This would represent remarkably strong survival e↵ects of
shocks.
2.6 Conclusion
I have examined the e↵ects of annual rainfall shocks over the first 21 years of
women’s and men’s lives on their likelihood of being literate and their highest grade
completed. Given the wage (Glewwe, 1996; Psacharopolous, 1994), health (Strauss
and Thomas, 1995; De Walque, 2007) and intergenerational (Thomas et al., 1991;
Oreopolous, 2006) benefits of human capital, knowing whether and to what extent
rainfall shocks a↵ect individuals’ stocks of it means knowing something more about
how poverty might be perpetuated in developing countries. In addition to what
45Selection-free estimates under these weaker literacy assumptions for other shocks available
upon request.
78
it might teach us about households’ abilities to cope with everyday-type shocks,
the answer to this question will be of interest to policy makers; if rainfall shocks
have permanent e↵ects on peoples’ human capital stocks, then social policy might
optimally take rainfall conditions into account. Given that one of the Millennium
Development Goals is to achieve universal primary enrollment and another is to
increase gender equality, this chapter is clearly quite relevant for policy.
In the West African context, I find that unusually wet and dry years have
economically significant e↵ects on women’s and men’s human capital. An additional
wet (dry) year between the ages of 7 and 13 (birth and age 6) decreases women’s
likelihood of being literate by about 7 (3.5) percent, while an additional wet (dry)
year between age 14 and age 20 (birth and age 6) increases men’s likelihood of being
literate by as much as 3.4 (2.5) percent. The baseline shocks e↵ects estimates imply
that thanks to shocks, women are 22.1 percent less likely to be able to read, while
men are 9.9 percent more likely to be literate. While some of these estimates might
be a↵ected by sample selection bias, even under strongly unfavorable assumptions
there remains evidence that both wet and dry shocks have substantial negative
e↵ects on women’s human capital and that wet shocks have positive e↵ects on men’s
human capital. These results contradict the suggestions of some earlier research
papers that temporary lapses in human capital investments following shocks are
made up for later.
I have argued that the most plausible explanation for this pattern of results
is that, following wet years females work more and spend less time in school while
males likely do just the opposite, while dry years decrease both household income
79
(to the point where parents are no longer willing to finance girls school enrollment)
as well as boys opportunity cost of attending school. These results relate to the
sizable development economics literature on household responses to shocks and pro-
vide evidence that whatever consumption smoothing that households manage to
achieve comes at the expense of some a↵ected individuals’ completed human capital
stocks. In particular, this chapter complements existing research on how invest-
ments in children are a↵ected by temporary income shocks by showing that those
contemporaneous investment e↵ects are not necessarily undone later, and that the
permanent e↵ects can be substantial. Moreover, income smoothing strategies that
avoid risk at the expense of profits with respect to productivity or labor supply
choices are not the only cost associated with attempts to smooth consumption in
risky environments–investments in children can also be substantially a↵ected by
shocks.
The results here suggest that the welfare e↵ects of temporary shocks could be
extremely persistent and that policies designed to help particular household mem-
bers stay in school following rainfall shocks could have substantial returns. In par-
ticular, if local institutional phenomena and economic conditions and behavior are
such that rainfall shocks a↵ect the way children use their time, then it will be im-
portant to consider the potential human capital benefits of social insurance schemes,
weather insurance or conditional cash transfers whose payouts are triggered by rain-
fall shocks. On the other hand, as Shah and Millett Steinberg (2013) note, programs
that increase wages could have the e↵ect of decreasing human capital investments in
children. This would be consistent with the findings of Basu et al. (2010), who find
80
that child labor initially increases with household wealth, given the lack of labor
market opportunities for the poorest children and the way those opportunities at
first increase with wealth.
81
Figure 2.1: Distribution of Women’s and Men’s Highest Grades Completed
82
Figure 2.2: Average Rainfall Levels Over Time
83
Figure 2.3: Mean Numbers of Annual Rainfall Levels Experienced Be-
tween Birth and Age 21
84
Figure 2.4: Literacy Rates by Highest Grade Completed
85
Figure 2.5: Age Heaping for Women’s and Men’s Samples
86
Table 2.1: Household Survey Countries and Years of Interview
87
Table 2.2: Descriptive Statistics
88
Table 2.3: Baseline Human Capital Results for Women and Men
89
Table 2.4: Women’s Height Results
90
Table 2.5: Schooling Attainment (Extensive Margin) Results
91
Table 2.6: Shocks Numbers Indicators Results
92
Table 2.7: Baseline Results by Agroecological Zone
93
Table 2.8: Region-Specific Birth Year Fixed E↵ects Results
94
Table 2.9: Rainfall Shocks and Numbers of Living Non-Migrants
95
Table 2.10: Rainfall Shocks and (Sibling) Mortality Results
96
Table 2.11: E↵ects of Shocks on Mortality and Implied E↵ects on Mi-
gration, Per Average Rainfall Neighborhood
97
Table 2.12: Baseline Literacy Estimates and Their Sample Selection-Free Analogs
98
Table 2.13: Baseline Highest Grade Completed Estimates and Their
Sample Selection-Free Analogs
99
Chapter 3: Food Price Inflation and Children’s Health
3.1 Introduction
People in developing countries often list inflation as one of their chief concerns
(Easterly and Fischer, 2001). This is perhaps not surprising since typical household
strategies for maintaining adequate consumption levels, such as borrowing from
relatives or friends or increasing labor supply, might be unavailable or inadequate
when price increases are su ciently large. This is especially likely to be true when
the goods whose prices have increased do not have close substitutes and they are an
extremely important part of the typical consumption basket, as with food. Indeed,
the expected changes in consumption habits that accompany food price inflation–
including the consumption of smaller portions, fewer meals, or cheaper (per unit)
but less nutritious substitute foods (Cohen and Garrett, 2009)–could negatively
a↵ect health, perhaps permanently in the case of children (Alderman et al., 2006).
Indeed, the food price crisis of mid-2008 brought with it reports of increased rates of
malnutrition and under-nutrition (IFAD, 2008). In the case of Egypt (the setting of
this chapter), this is reflected in the 90 percent increase in the proportion of children
whose weight-for-height was at least 2 standard deviations lower than that of an
international reference group of children between earlier years and 2008 (see Figure
100
3.1). But despite all the attention from the press, policymakers and international
organizations that this crisis received, it is still not clear whether di↵erent kinds of
households are helped or hurt by this inflation (Headey, 2011).
In particular, little is known about whether urban or rural households are more
severely a↵ected or whether the latter are negatively a↵ected at all. While urban
households are typically richer than rural households and therefore better placed to
cope with food price inflation, urban households of course tend to be net food buying
households. 1 Rural households, on the other hand, are typically more vulnerable
due to their lower household incomes, but these households could benefit from higher
food prices if higher output prices make agriculture su ciently profitable. The
common expectation, however, is that urban households would decrease quantities
consumed (on average, at least) while at least some rural households would benefit
from higher food prices (Fujii, 2011; OECD, 2008; Cohen and Garrett, 2009; Ivanic
and Martin, 2014). Consistent with this, Ferreira et al. (2011), who study household
welfare in Brazil in 2008, find that the poorest net food buyers were substantially
negatively a↵ected by high food prices, while the rural poor actually benefited from
the higher food prices. On the other hand, Gibson and Kim (2013) find that given
consumers’ tendencies to switch to cheaper but equally nutritious rice, the elasticity
of calories with respect to rice prices is quite low. If quality adjustments are in
fact empirically important, then perhaps households can cope with high food prices
rather easily. This chapter examines the evolution over time of Egyptian children’s
1This is particularly true in Egypt, where only about 4 percent of workers classified as residing
in urban areas are employed in agriculture according to the 2008 Demographic and Health Survey.
101
weight in urban and rural areas to shed light on how children in di↵erent types of
households seem to actually be a↵ected by rapid food price increases.
More specifically, the evolution of Egyptian children’s weight-for-height for
children from urban, rural agricultural and rural non-agricultural households is an-
alyzed here. This is done using four highly comparable household surveys conducted
between 2000 and 2008, a year in which world food (and other) price inflation was
especially high. Data on parents’ occupation along with information on when and
where unusually high-priced crops were being harvested allow for consideration of
how the weights of children from what were likely net food selling households evolved.
In addition to examining the evolution over time of the weights-for-height of chil-
dren in di↵erent types of households, I also use quantile regression techniques to
examine changes in the (conditional) weight-for-height distribution as a function of
these same factors. This is highly appropriate given what are likely widely diverging
household coping capabilities across the weight-for-height distribution; while heavier
children might have lived in households that were able to maintain consumption de-
spite inflation, lighter children were likely more vulnerable to food price rises. Also,
we might naturally be more interested in the evolution of something like weight-for-
height for the lightest children, given their stronger need for some sort of assistance.
To be clear, I do not claim to identify the causal e↵ect of high food prices on the
weight-for-height of di↵erent children. Rather, I compare changes in di↵erent groups
of children’s weight-for-height over time under the assumption that food prices had
e↵ects that were powerful enough that these e↵ects will be large relative to other
time-varying factors, and it will therefore be clear which children were a↵ected most
102
severely and in which directions.
This chapter is related to the literature on the e↵ects of food price inflation
in developing countries. Most of the work in this literature relies on simulations
or, when household survey data is used, subjective measures of well-being (Headey,
2011). Thus, one of this chapter’s main contributions is to use a relatively objective
measure of child well-being from nationally representative household survey data
to examine who might have been a↵ected and how by the food price crisis of 2008.
This chapter is also novel in that it simply tracks the evolution of di↵erent groups’ of
children’s weights-for-height over time, in contrast to the papers based on simulation
results that it complements. Many of those papers rely heavily on assumptions and
do not account for the full range of factors that can change over time along with
food prices, for example. This chapter therefore provides a sort of record that
simulation results might usefully be compared to. I do this by using Demographic
and Household Survey data collected in Egypt–a country seriously a↵ected by the
crisis due to its heavy dependence on imported staple foods–in 2000, 2003, 2005 and
2008 (in April-June, the height of the crisis).
I find that the e↵ects of food price inflation tend to have been negative for
children at the bottom of the conditional weight-for-height distribution, and this
is true for most of the region- and parental occupation-based groups I consider.
While most of the decreases in weight-for-height at the first decile of the conditional
distribution of weight-for-height between 2005 and 2008 are fairly small, given the
growth of the Egyptian economy over this time period it is somewhat striking that
there were any decreases at all. (Indeed, lower rates of growth over a shorter period
103
of time resulted in substantial gains in weight-for-height in most cases by the run-up
to measurement in 2005.) I also find that, if anything, the set of children expected
to have been most likely to have benefited from high food prices seem to have been
substantially negatively a↵ected by them. While the change in weights-for-height
amongst these children between 2005 and 2008 is not statistically significant at
conventional levels, it is close to being so, and the magnitudes of the changes for
these children are large. More specifically, children from the region of rural Egypt
where high-priced staple crops were being harvested and sold around the time of
measurement in 2008, and whose parents were self-employed farmers, registered
what appear to be some of the largest decreases in weight-for-height between 2005
and 2008. On the other hand, several groups of children from net food buying
households actually gained weight over this time period, including heavier urban
children and rural children whose parents did not work in agriculture. Overall, these
results suggest that for the most disadvantaged Egyptian children, the negative
e↵ects of high food prices outweighed any possible beneficial e↵ects of economic
growth. Also, for rural net food selling households, the results are consistent with
the possibility that high input prices had negative e↵ects that were stronger than
the beneficial e↵ects of high agricultural output prices.
Section 3.2 provides background on the 2008 food price crisis in Egypt and
considers conceptual issues. Section 3.3 describes the empirical analysis, the data
and the sample selection criteria. Section 3.4 presents descriptive statistics and main
results. Section 3.5 concludes.
104
3.2 Background and Conceptual Issues
3.2.1 The 2008 Food Price Crisis in Egypt
Between 2005 and mid-2008, world prices of wheat, coarse grains and rice
all roughly doubled (OECD, 2008). The causes of these price increases included
droughts in major grain-producing areas, low baseline stocks of cereals, increased
demand for maize for the production of biofuels, rising oil prices and the continuing
devaluation of the US dollar (the currency in which prices for these commodities
are usually quoted). Although the (real) prices in mid-2008 of these staples were
far from unprecedented, the price of wheat, for example, had not been as high as it
was then since the early 1980s. According to the FAO, between 2006 and 2008, the
world prices of cereals and oils increased by as much as 140 percent, while wheat
prices rose by 250 percent. It was not only the prices of foods that were rising so
rapidly, however; petroleum and agricultural input prices (which includes fertilizer
prices) increased by 160 percent in the year to August, 2008. This is noteworthy,
since Egyptian farmers use nitrogenous fertilizers highly intensively (Kherallah et
al., 1999).
Egypt’s most important staple foods are wheat, maize and rice. Moreover,
Egyptian consumers are highly vulnerable to increases in the world prices of wheat
and maize, as the country is the world’s largest importer of the former and is a ma-
jor importer of the latter. At the household level, more than half of all household
consumption expenditure is on food. On the other hand, Egypt is a net exporter
105
of rice (Arab Republic of Egypt Initiative on Soaring Food Prices, 2008). The
food price crisis of 2008 was widely reported to have taken a serious toll on Egyp-
tian consumers. According to Egypt’s Central Agency for Public Mobilization and
Statistics, overall (annual) consumer price inflation peaked in August, 2008, at 25.6
percent, and year-on-year food price inflation then stood at 35.5 percent, with re-
portedly higher food inflation rates in rural areas (Arab Republic of Egypt Initiative
on Soaring Food Prices, 2008). The accuracy of these o cial inflation rates has been
called into question, however, and informal surveys suggest that flour and rice prices
might have risen by over 100 percent while vegetable oil prices might have increased
by over 200 percent over the period in question (Wikileaks Discussion Forum, 2011).
Also, according to the World Food Program, the prices of cereals in Egypt rose by
129 percent between 2006 and 2008 (WFP, 2008). It should be emphasized again,
however, that food prices were not the only ones to rise substantially in Egypt in
2008; the prices of petroleum and construction materials also exhibited large in-
creases (Wikileaks Discussion Forum, 2011).
Annual real per capita growth (as well as food price inflation) rates over the
time period in question can be seen in Figure 3.2. As the economy slowed down
in the early 2000s after several years of fairly high growth rates (with particularly
low annual growth rates in 2001 and 2002), so did employment levels. Since the
end of 2004, however, as growth picked up it a↵ected more Egyptians’ employment
prospects positively (Hasan and Sassanpour, 2008). Wage growth also increased
after 2003, so that growth likely began increasing household earnings for employed
Egyptians around this time. Unemployment statistics, however, are mostly informa-
106
tive with respect to the labor supply of young, relatively well educated Egyptians
seeking entry into the labor market. In fact, according to the 2005 Labor Force
Sample Survey, 92 percent of unemployed Egyptians were younger than 30 years of
age, while recorded unemployment for older Egyptians is very low. For many Egyp-
tian households headed by individuals old enough to have young children, therefore,
growth after 2003 would have brought increases in household income. Also, un-
employment was evenly spread between rural and urban areas in the early 2000s.
Finally, Ersado and Aran (2014) document substantial increases between 2000 and
2008 in access to and consumption of a wide variety of health care treatments for
young children in Egypt, which suggests that improved health care was working
alongside growth to improve children’s physical conditions.
The Egyptian government responded to the food price crisis by expanding its
food aid programs, increasing public sector wages and the value of pensions and
making changes to trade policies. Egypt’s ration card system (which o↵ers eligible
households a monthly quota of rice, flour, tea, sugar and oil) went from covering
11.1 million to 26.1 million Egyptians in the second quarter of 2008. However, this
program is not well-targeted to the poor and is reportedly riddled with corruption
(Wikileaks Discussion Forum, 2011). Existing evidence indicates that all but 20
percent of the poorest and most vulnerable Egyptians fail to take advantage of the
ration cards, primarily because poor Egyptians rarely possess the national identi-
fication cards required to sign up and they live in su ciently remote areas (Arab
Republic of Egypt Initiative on Soaring Food Prices, 2008). Also, the goods meant
to be sold to ration card holders were sold by merchants who sold the same goods
107
at market prices, and this reportedly presented too tempting an arbitrage oppor-
tunity to pass up (Wikileaks Discussion Forum, 2011). There is also a subsidized
bread program that is open to everyone. It is not clear, however, that this program
does much to bolster the food security of those Egyptians who need help the most.
Reportedly, as corn prices rose and the gap between the prices of subsidized and
unsubsidized flour widened, an increasing number of bakers and millers began selling
their quotas’ worth of subsidized flour on the black market, leading to shortages for
all would-be program beneficiaries (Wikileaks Discussion Forum, 2011).
The Egyptian government took other steps to help particular groups. In May,
2008, public sector workers’ wages were increased by 30 percent and public sector
retirees’ pension payments were increased by 20 percent. Also, like many other
developing countries, Egypt made changes to its trade policies in mid-2008 in an
attempt to shield its citizens from the e↵ects of higher food prices. In particular,
import tari↵s for food products were lowered to 6.5 percent on average (the lowest
level in over a decade) in the first half of 2008 and the government banned the export
of rice in April, 2008. The latter policy change is believed to have reduced domestic
rice prices to some extent.
3.2.2 Agriculture in Egypt
Agriculture in Egypt is highly geographically concentrated on about 3 percent
of the country’s land, much of which is near the Nile river. About 35 percent of all
Egyptians live in farm households (Siam, 2002). The most common type of farm
108
is slightly larger than one (highly productive) acre in size and is located in canal-
irrigated areas near the Nile. This Nile-fed irrigation allows Egyptian farmers to
harvest as many as 3 crops per year. Most Egyptians who work in agriculture either
work on small farms or work as tenants or shareholders, and the majority of village
residents either have no or little land, and cultivate as tenants or smallholders. In
addition, for many agricultural workers, farm work is merely one of multiple types of
casual labor performed over the course of a year (Arab Republic of Egypt Initiative
on Soaring Food Prices, 2008).
Climatic variations within Egypt imply that di↵erent crops tend to be grown
in di↵erent locations. For example, Lower Egypt (i.e., northern Egypt) is relatively
humid and therefore suitable for the cultivation of cotton and rice. Upper Egypt, on
the other hand, is hotter and drier and is more suited to the cultivation of sugarcane,
onions and lentils (Metz, 1990). (The terms Lower and Upper refer to the altitude
of the Nile River, which flows north from the East African highlands–and relatively
high-altitude southern Egypt–to the Mediterranean Sea.) More importantly for the
purposes of this chapter, while wheat is grown all over Egypt in the winter season
(which lasts from December to the beginning of April), the warmer temperatures
in Upper Egypt restrict wheat yields substantially while yields are relatively high
in Lower Egypt. Thus, beginning in April, 2008, farmers (who sold any wheat) in
Lower Egypt in particular should have begun benefiting from high wheat prices. As
far as the production of other staples goes, rice and maize are each sowed in May
and are not harvested until October. Thus, to the extent that farmers in Lower
Egypt (the only region where rice and maize are cultivated) previously stored their
109
surplus production to sell when prices for these crops increased sharply in early
2008, then the benefits of high food prices should be reflected in children’s weights
as of the second quarter of 2008 (when that year’s household survey was collected).
Farmers in Lower Egypt are therefore more likely to have benefited from high food
prices compared to other Egyptians, as they were more likely to be net sellers of
the staple crops whose prices increased the most. Moreover, in addition to having a
climate that is in many ways less hospitable toward the cultivation of crops, Upper
Egypt has historically been a poorer and more food-insecure region.
3.2.3 Conceptual Issues and Existing Evidence
Food price inflation would only be harmful if wage inflation was relatively
low and, in the event that it was not, households could not easily switch to some
substitute set of foods. Alas, while in principal nominal wages could have grown
fast enough for the real price of food to have remained stable, evidence suggests
that this did not occur. In particular, nominal wage growth in 2008 was roughly 25
percent (Al-Nashar, 2008), which is lower than the food price inflation rate.
The staples whose prices increased so rapidly between 2006 and 2008 are ex-
tremely important to the Egyptian diet. Kherallah et al. (1999) note that the
average annual consumption of wheat and wheat products is about 200 kilograms
per capita. It is not clear precisely how price elastic demand for cereals is; Fayyad et
al. (1995) reports price elasticities of wheat, corn and rice of -0.12, -0.24 and -0.45,
respectively, while Dawoud (2005) finds an own-price elasticity of cereals as a whole
110
of -0.58. The consumption of cereals therefore likely decreased to some extent. Also,
staple food prices were not the only food to become more expensive; the prices of
other, non-staple foods such as meat, poultry, fish, fruits and vegetables increased
substantially as well, and there is evidence that consumption of the first three of
these foods declined between 2005 and 2008 (WFP, 2008), which would be consistent
with the (relatively high) own-price elasticity estimates from Fayyad et al. (1995)
and Dawoud (2005) for these goods. Thus, it is not obvious that households could
have simply switched from more expensive staples to these other foods, and thereby
maintained su ciently high levels of nutritional intake. Of course, there are a wide
variety of other coping mechanisms that net food buying households could have
utilized to deal with these price changes. For example, there is anecdotal evidence
that Egyptian households cut back outlays on other essentials (such as health- and
education-related goods) (IFAD, 2008). It is an empirical question as to whether or
not these types of strategies proved su cient to maintain normal food consumption
levels and children’s weights.
There are reasons to expect that high food prices were most likely to have
negatively a↵ected the poorest households. First, food makes up a larger proportion
of consumption baskets for poorer households. Also, the economic growth that took
place in the years preceding 2008 is believed to have protected richer households
from the e↵ects of food price inflation in a way that it did not for poorer households.
Indeed, while economic growth decreased the proportion of Egyptians who fell under
the country’s poverty line fell from 23.4 percent in 2005 to 18.6 percent in 2008, the
proportion of Egyptians living in extreme poverty increased from 5.4 to 6.4 percent
111
(Arab Republic of Egypt Initiative on Soaring Food Prices, 2008). 2 Also, the
rate of wastage–i.e., the proportion of children whose weight-for-height was at least 2
standard deviations lower than that of an international reference group of children–
increased by more than 90 percent from 3.4 percent before 2008 to 6.5 percent during
it (see Figure 3.1).
Given this possibility that di↵erent types of households and parts of the income
distribution were a↵ected di↵erently by inflation, I will examine the evolution of the
conditional distribution of children’s weights-for-height over time as a function of
region- and parental occupation-based factors. Since one of the crops whose price
was considerably higher in 2008 (wheat) was being harvested around the time the
2008 data was collected in large quantities in (rural) Lower Egypt, the region of
residence should help determine who might have benefited from high food prices
and who might have been harmed by them. Similarly, parental occupation data,
and in particular data on whether or not children had a parent who was either a
self-employed farmer or an agricultural employee, allow for an even more focused
examination of potential beneficiaries and others. The hypothesis that children in
Lower rural Egypt with parents who were more likely to have sold greater amounts
of high-priced wheat were relatively heavy in 2008 will be tested. More specifically,
I assume that children in rural Lower Egypt with self-employed farmer parents at
the top of the conditional weight-for-height distribution were most likely to have
been members of net (high-priced) food selling households. Thus, it is this group
2Here, a household is considered poor if daily per capita consumption is less than 2 USD/day
at purchasing power parity, and a household is considered to be extremely poor if it cannot meet
its basic food needs (Arab Republic of Egypt Initiative on Soaring Food Prices, 2008).
112
of children who should have had the highest weights-for-height in 2008 (relative
to observationally similar children in previous years). Additionally, since wheat is
cultivated in rural Upper Egypt, it is also possible that children with self-employed
farmer parents at the top of the conditional weight-for-height distribution in that
region were also relatively heavy (given height) in 2008. Given the less productive
climate of Upper Egypt for wheat cultivation, however, it is expected that fewer
children would have belonged to net wheat selling households in this region than
did in Lower Egypt. 3 In keeping with the general expectation from the literature
that high food prices are beneficial for at least some net food selling households,
the assumption here is that the beneficial e↵ects of high food prices on children’s
weight-for-height dominated the harmful e↵ects of having to pay more for inputs
such as fertilizer.
There are also net food buying households in rural Egypt to consider. On the
one hand, the standard approach to understanding the impacts of high food prices
would typically conclude that children in these households should have been nega-
tively a↵ected by high food prices, but for farm employee households in particular,
there are reasons to doubt this. In Lower Egypt, children with self-employed farmer
parents at the lower part of the conditional weight-for-height distribution, children
with agricultural employee parents, and children with parents employed outside of
agriculture could have all benefited to some extent from high food prices (despite
the fact that they lived in net food buying households). In particular, members of
3Also, children in what could have been net food selling households in rural Upper Egypt
would not have benefited in late 2007 and afterwards from higher prices for maize and rice like
their counterparts in Lower Egypt might have, since those crops are not cultivated in Upper Egypt.
113
these types of households in Lower Egypt might have either: sold some quantity of
high-priced wheat in the second quarter of 2008; earned (substantially) higher wages
as agricultural workers; or earned more working in the rural non-agricultural sector
(thanks to increased demand for non-farm goods and services). Again, the same
things are true for observationally similar households in rural Upper Egypt, but
are less likely to have occurred. Under the assumptions of the standard approach,
however, children in these types of households in rural Lower and especially Upper
Egypt are therefore expected to have been negatively a↵ected by high food prices
in 2008.
It is important to emphasize, however, that for children in all 3 of these
parental occupation-based groups in both rural regions, the poorest children could
have been negatively a↵ected by high food prices. This vulnerability is driven by the
prominence of (what in 2008 were) high-priced staple foods in poorer households’
consumption baskets. Again, it is the conditionally lightest children in these groups
who are assumed to have been the poorest ones, and we therefore should expect
these lightest children to be more likely to have been negatively a↵ected by high
food prices. Finally, given their net food buying status, urban children–and, once
again, urban children near the bottom of the conditional weight-for-height distribu-
tion in particular–are expected to been relatively negatively a↵ected by high food
prices in 2008.
To sum up these standard approach-based predictions, since high-priced wheat
was being sold around the time that children were being measured in 2008, and as-
suming the negative e↵ects of higher input prices were not su ciently large, children
114
in net food selling households are expected to have registered the largest increases in
weight-for-height between 2005 and 2008. In terms of the observable characteristics
in the data, these are the conditionally heaviest children of self-employed agricultural
workers in rural Egypt, and they are expected to have benefited from higher food
prices in 2008 in a way that children from (within-region) net food buying house-
holds did not. 4 This latter group includes the lighter children of self-employed
agricultural workers, children of agricultural employees, children of non-agricultural
workers, and urban children. Note that once again under the standard assumptions,
the fact that agricultural employees are members of net food buying households
will have stronger e↵ects than the increase in wages for these workers will, and
members of these households are expected to be negatively a↵ected by higher food
prices. A second prediction is that children from net food selling and net food buy-
ing households in Lower rural Egypt should have benefited from higher food prices
to a greater extent than their Upper rural counterparts, since higher-priced crops
were produced and sold in greater amounts in the former region. A third and final
prediction is that, in all regions and for all parental occupations, the conditionally
lightest children are considered particularly likely to have been substantially harmed
by high food prices.
The analysis of the conditional distribution of children’s weights-for-height will
be done using quantile regression analyses (in addition to ordinary linear regression
analyses of the same factors over time). The assumption motivating the choice to use
4As I will discuss shortly, and since I do not observe household wealth in my data, I assume
that children’s weight-for-height is positively correlated with household wealth.
115
quantile regression techniques, of course, is that child weight is positively correlated
with household income. In this case, then consistent with the evidence described
above, it could have been children at the left tail of the conditional weight-for-height
distribution who were a↵ected most severely by inflation. The household surveys
on which I rely do not contain income data. However, as the results in Tables
3.3, 3.4, 3.6 and 3.9 show, parental highest grades completed at school, which are
of course typically positively related to income, are also positively associated with
weight-for-height. I also assume rank preservation: The weight-for-height of a child
at a particular conditional quantile of the distribution in one year would also be at
that same conditional quantile if they were observed in any other year. Intuitively,
this assumption will allow the empirical results to be interpreted as indicative of
the evolution of weights-for-height for di↵erent groups of children (e.g., low-weight
children of self-employed farmers in Lower Egypt), rather than the evolution of the
di↵erent conditional quantiles (corresponding to the weights-for-height of children
who were only temporarily at that location of the conditional distribution).
Again, in keeping with the predominant existing view, I expect that those
agricultural households who were most likely to have been net sellers of high-priced
crops would have benefited from the sale of those crops in the second quarter of
2008. It is important to note, however, that food prices were not the only prices to
exhibit large increases in 2008: High petroleum costs reportedly pushed the prices
of fertilizer and diesel fuel up substantially (Wikileaks Discussion Forum, 2011),
which in turn would have pushed up the costs of agricultural production, harvesting,
processing and transportation. Also, along with many other agricultural products,
116
livestock feed became more expensive. Thus, in principle it is not obvious how this
group of rural agricultural households will have been a↵ected by the food price crisis
of 2008.
Most of the evidence on how di↵erent kinds of households in developing coun-
tries were a↵ected by food price increases in 2008 comes from simulation studies,
which normally conclude that the numbers of poor or hungry people increased by
tens of millions of people worldwide (Headey, 2011). The conclusions from these
types of studies normally rest on the assumptions that only food prices changed
and that household incomes remain fixed. Both of these assumptions are not ap-
plicable for the case of Egypt, however, as we have seen. Also, in practice, food
and oil prices commonly move together (Arshad and Hameed, 2009), and so it is
not clear that understanding the e↵ects of solely food prices on well-being is prefer-
able from a policymaking perspective than understanding the e↵ects of inflation
more generally, or how well-being is a↵ected during a typical high food (and other)
price episode. Headey (2011) is an exception to these simulation studies in that
it is based on household survey data (from a cross-section of countries). This pa-
per relies on a subjective measure of well-being in the form of self-reported food
a↵ordability. In particular, respondents were asked whether their household had
any problems a↵ording food and whether the household had experienced episodes
of hunger in the previous12 months. As the author notes, however, this kind of out-
come has been found to be subject to test-retest unreliability, and question phrasing
and placement have been shown to have large e↵ects on responses. Interestingly,
the author finds that self-reported food insecurity decreased rather than increased
117
between 2005 and 2008 (with most of Africa and Latin America being exceptions
to this finding). Ferreira et al. (2011) also relies on household survey data, and
estimates the welfare consequences of food price increases in Brazil in 2008 using
simulation techniques. The authors find that net food buyers–and in particular
the poorest of these consumers–experienced large, negative e↵ects, while the rural
poor benefited from the higher food prices. This chapter is intended to add to this
literature by relying on a relatively objective measure of well-being in the form of
children’s weights-for-height, and going beyond a focus on average outcomes in favor
of distribution-wide analyses.
3.3 Empirical Analysis, Data and Sample
3.3.1 Empirical Approach
As Section 3.2 makes clear, Egypt o↵ers a good opportunity to consider the
hypothesis that rising food prices benefit households that sell food and negatively
a↵ects households that do not. This will be tested by comparing changes in children’s
weights-for-height in regions where high priced food was being sold (rural Lower
Egypt) to changes in children’s weights-for-height where food was not being sold
(urban Egypt and, to a lesser extent than in rural Lower Egypt anyway, rural Upper
Egypt). It is not obvious, however, that we might learn something by comparing
weight-for-height changes over time across these regions. Put another way, economic
growth or other time-varying factors could have had di↵erent e↵ects on the di↵erent
regions regardless of what happened to prices in 2008. In this case, to the extent
118
that these (region-specific) changes would have a↵ected children’s weight-for-height,
a comparison of changes in the di↵erent regions would confuse the e↵ects of inflation
with what were merely region-specific changes. This would obviously make for a
flawed test of the hypothesis in question.
As Figure 3.3 shows, however, the assumption that children’s weight-for-height
evolved similarly in the di↵erent regions is plausibly not too strong. While levels
di↵er substantially across regions, for 2000, 2003 and 2005, changes appear compa-
rable. In particular, child weight-for-height levels for children in rural Upper Egypt
are the lowest of the three regions, which is consistent with existing research (Arab
Republic of Egypt Initiative on Soaring Food Prices, 2008). On the other hand,
it is surprising that levels for children in rural Lower Egypt are higher than levels
for urban children. But more importantly for the purposes of this chapter, the di-
rections of changes in weight-for-height between surveys are constant across all 3
regions. First, weight-for-height drops substantially in all three regions between the
2000 and 2003 surveys. This large drop is likely to have been driven mainly by drops
in GDP growth per capita, which had been fairly high for several years before 2002
(see Figure 3.2). A second, less important factor behind the drop in weights-for-
height in 2003 was likely the devaluation in January, 2003 of the Egyptian pound,
which resulted in non-trivial increases in the (relative) price of food and has been
found to have a↵ected poor households more strongly than other households (Kraay,
2007). Indeed, the seeming severity of the events of 2003 compared to the food price
crisis of 2008 is striking and is an interesting finding in its own right. In any case,
after 2003, economic growth rates picked back up and this is once again reflected
119
in children’s health in all three regions. Thus, for the years prior to 2008 anyway,
weight-for-height changes seem comparable, and this should perhaps ease concerns
over the usefulness of comparing child weight-for-height changes across regions.
My empirical analysis is based on the following baseline specification:WFHit = ↵+ 3X
t=1
 tUrbanti + 4X
t=1
⇢tLoEgyptti + 4X
t=1
✓tUpEgyptti +X 0it + "it. (3.1)
HereWFHit is the weight-for-height of child i observed in year t. Urbanti, LoEgyptti
and UpEgyptti are dummy variables corresponding to residence in either urban ar-
eas of Egypt or rural areas of Lower and Upper Egypt in year t, respectively. Note
that t 2 {2000, 2003, 2005, 2008} for the rural Lower and Upper Egypt year indica-
tors, but t 2 {2003, 2005, 2008} for the urban Egypt indicators since urban children
measured in 2000 are the omitted category. An additional specification replaces the
sets of dummies for the pair of rural regions with dummies for 3 categories of rural
Egyptian children (again, for all four survey years for the rural Egypt year indicators
and for three survey years for urban children): children with no parents who work in
agriculture, children with at least one parent who works as a self-employed farmer
and children with at least one parent who works as an agricultural employee (e.g., a
casual farm laborer). The sets of dummies from the two specifications are included
to identify changes in children’s weight-for-height for the region- and region and
occupation-based groups of interest, respectively. 5
If the hypothesis that high food prices increase income in net food selling
households and decrease quantities of food consumed in net food buying households
5For linear regression analyses, sample weights are not used. The results in Tables 3.3 and 3.4
are robust to the use of sample weights, however.
120
is true, then the coe cient estimates of the dummies equal to one for rural children
measured in 2008 should take on larger positive values than the estimates of the dum-
mies for urban children measured in 2008. Moreover, the LoEgypt2008 coe cient
should have exhibited a larger jump relative to analogous values for earlier years
than did the UpEgypt2008 coe cient, since wheat yields would have been relatively
high in Lower Egypt (due to more favorable climatic/growing conditions). Finally,
if the predictions set forth in Section 3.2 are correct and children’s weight-for-height
in 2008 is especially high (relative to earlier years) in rural Lower Egypt, then the
second version of the specification should show that these large positive changes
in weights-for-height were driven by children whose parents work as self-employed
agricultural workers (i.e., children who are likely to reside in net food selling house-
holds). In particular, the estimate of the rural Lower Egypt self-employed farmer
parent indicator for the upper conditional quantiles of the conditional weight-for-
height distribution should exhibit relatively high growth between 2005 and 2008.
The Xit vector contains controls for individual and household-level factors.
Child-specific controls include a dummy equal to one if child i is a male and zero
otherwise. I also include quadratic terms for mother’s and father’s ages; variables
equal to mother’s and father’s highest grades completed at school; variables equal
to the numbers of females and males in the household in each of several age groups;
and the household dependency ratio. 6 Finally, I include a dummy variable equal
to one if child i’s mother works.
6The age group categories are defined for householders between birth and age 6, age 7-age 15,
age 16-25, age 26-45, age 46-60 and age 61-plus. The household dependency ratio is defined to
be the number of householders between the ages of zero and 15 plus the number of householders
above the age of 45 divided by the number of householders between the ages of 16 and 45.
121
The specification above will be used to conduct linear regression as well as
quantile regression analyses. As the conceptual framework section makes clear, there
is reason to believe that lighter children in net food buying households were more
severely negatively a↵ected by food price inflation than other children in these house-
holds. Also, it could be that only agricultural households containing the heaviest
children benefited from high food prices. The specification above will be estimated
for various conditional quantiles of the conditional weight-for-height distribution
to explore these possibilities. For quantile regressions, standard errors are boot-
strapped.
3.3.2 Data
The household surveys I rely on are Demographic and Health Surveys (DHS)
collected in 2000, 2003, 2005 and 2008. These surveys were all collected in the first
two quarters of the year (see Table 3.1 for the months of data collection for each
survey). These surveys are nationally representative at the rural and urban levels
and they contain a wide variety of socioeconomic and demographic data, such as
anthropometric data for all children living in sampled households under the age of
6, region of residence, respondent occupation, highest grade completed, and the age
and gender of all household members. In sampled households each woman between
the ages of 15 and 49 was selected to be interviewed, and all of these women who
were married were also asked about the occupation and highest grade completed of
their husband. Thus, for all of the children in my sample, I observe the parental
122
variables of interest.
The sampling scheme of the four household surveys was designed to provide
information on ”various population and health indicators of interest for... six major
subdivisions” of the country”, which are aggregated here to urban, rural Lower and
rural Upper Egypt regions. Each survey also relies on the same type of three-stage
probability sampling scheme.
The outcome I consider is children’s weight-for-height. I consider this outcome
because I am interested in using a relatively objective measure of the short-term
e↵ects of food price inflation on children’s health and well-being. 7 A valid
measure of this is one that is potentially a↵ected one way or another quickly, and
weight-for-height is therefore preferable to another objective measure of child health
such as, say, height-for-age.
My sample consists of all children below the age of 6 who resided in sampled
households and whose parents’ occupational status and highest grade completed
data is not missing. This leaves me with a total of 40,233 children (out of a total of
40,573 children with non-missing weight-for-height data), 20,600 (51.2 percent) of
whom are male. 8
7While at least 97 percent of all of the children under age 6 had their weights measured in
each survey, a small proportion of those children–as much as six percent–were judged to have
”implausibly high or low” weights that were consequently not included in the version of the data
released to the public.
8A total of 451 children are missing weight-for-height data.
123
3.4 Descriptive Statistics and Results
3.4.1 Descriptive Statistics
Descriptive statistics for the sample of children can be seen in Table 3.2. Be-
ginning with Panel A, the mean (standard deviation) of weight-for-height for the
entire sample is about 0.142 (0.028). Also, when we examine mean weight-for-height
for children in urban and rural Upper Egypt, we see that both groups of children
have slightly lower weights-for-height relative to the sample average. But while ru-
ral children as a whole still weigh less (given their age) than urban children, the
fact that children in rural Lower Egypt are the heaviest of all three of these groups
is once again surprising. Evidently the usual finding that rural Egyptian children
weigh less than urban Egyptian children does not tell the whole story. Also, in
both rural Lower and Upper Egypt, we can also see that average di↵erences in
weights-for-height across parental occupation categories are small.
The sample average mother’s highest grade completed is about 6.7, while that
of father’s is moderately higher at 8.3. About 36 (6) percent of children’s mothers
(fathers) were between the ages of 15 and 25, about 50 (50 percent) percent were be-
tween the ages of 26 and 35, while the rest of the sample children’s mothers (fathers)
were older. The sample average number of very young and elderly householders rel-
ative to the number of prime-aged adults is about 0.83, and about 15 percent of all
sample children’s mothers work. Panel B of Table 3.2 contains descriptive statistics
across years for the proportions of children whose parents worked in di↵erent occu-
124
pations and whose weight-for-height data was missing. Also, there is data on the
numbers of children that mothers gave birth to in the previous 5 years, as well as
the numbers of children under the age of 5 who died in the previous 5 years. These
statistics make clear that there did not seem to be uncommonly large changes be-
tween 2005 and 2008, which might ease concerns over the possibility that the high
food price inflation of the latter year substantially a↵ected sample composition.
Returning to Figure 3.3, it is striking that three years’ worth of substantial
growth in real GDP per capita beginning in 2005 did not result in larger increases
in children’s weights-for-height by 2008. Moreover, it seems that it is rural (Lower
Egyptian) and not urban children who lost the most weight (relative to height) in the
three years preceding measurement in 2008. This figure is not obviously consistent
with the main hypothesis motivating this chapter–namely, high food prices were
beneficial for rural children and harmful for urban children. Next we move on to
the main results for a more careful examination.
3.4.2 Ordinary Least Squares Results
3.4.2.1 Region-Specific OLS Results
The main results for the version of the specification that considers region-
specific changes in weights-for-height can be seen in Table 3.3. The linear regression
results can be seen in columns 1 (which only additionally controls for child gender
and age) and 2 (which adds controls for parents’ education and age, household
composition, and an indicator for whether the child’s mother works). The results
125
are quite similar for both versions of the specification, and I will discuss the less
parsimonious specification’s results here.
The omitted group is male children in urban Egypt in 2000 who are less than
one year old. This group of children has an average weight-for-height of 0.115.
The results make clear that relative to these children, urban children were lighter
(given height) in 2003 compared to 2000, their relative weights were on the way
to recovery in 2005, and urban children had more than recovered their weights-for-
height by 2008. While the changes in average urban children’s weights-for-height are
commonly highly statistically significant, the magnitudes in question are normally
small. For example, holding all else equal, urban children in 2003 were -0.00794
kilograms (given height) lighter than their counterparts in 2000 (on average), which
represents a bit more than one-quarter of a standard deviation of the weight-for-
height variable. Note that this is an example of a relatively large inter-survey change
in weights-for-height, i.e. most changes in weight-for-height between surveys were
fairly modest. It is also worth noting that 3 years’ worth of solid economic growth
in the run-up to measurement in 2008 was not as beneficial, for example, as the
events leading up to measurement in 2003 were harmful (according to the results
of a t-test). These relatively weak gains between 2005 and 2008–which might be
contrasted with the gains between 2003 and 2005, when the economy was growing at
a moderate rate but inflation was much lower–are consistent with food price inflation
having had a substantially negative e↵ect on urban children’s weights-for-height.
Lower rural Egyptian children in 2000 were heavier (given height) than their
urban counterparts, and these children were hit relatively hard by the events leading
126
up to measurement in 2003. By 2005, however, these children were well on their
way to back to their (conditionally average) 2000 weights-for-height. But not only
did the economic growth between 2005 and 2008 not result in substantial increases
in Lower rural Egyptian children’s weights-for-height over that period, but (condi-
tional) weights-for-height were actually lower in 2008 than they were in 2005 (though
not significantly so, according to a t-test). Given that high-priced wheat was being
harvested and sold in relatively large quantities in Lower rural Egypt around the
time of measurement in 2008, children there might have been expected to benefit
more than other children from food price inflation. The observed decrease in aver-
age weights-for-height between 2005 and 2008 is therefore surprising. Of course, it
remains to be seen which types of children in Lower rural Egypt in 2008 are driving
this result.
In contrast to Lower rural Egyptian children, Upper rural Egyptian children
who were measured in 2000 were lighter (given height) than their urban counterparts.
These children were also considerably lighter (given height) in 2003, but had made
a moderate recovery by 2005. Again, however, the 3 years’ worth of substantial
economic growth that Egypt experienced before 2008 appears unable to have fully
counteracted other, negative e↵ects, and Upper rural children grew only slightly
(but statistically significantly) heavier between 2005 and 2008.
Female children had slightly lower weights-for-height than male children, and
the age dummy coe cients make it clear that weight-for-height increases substan-
tially with age. The magnitudes of the age coe cients are quite large relative to the
typical region-year indicator estimate. Weight-for-height is increasing with father’s
127
(but not mother’s) highest grades completed at school. 9 There is some evidence
that the more young–and likely less economically productive–household members
there are, the lower is children’s weight-for-height. On the other hand, holding all
else equal, an increase in the number of males between the ages of 26 and 45 is as-
sociated with higher weights-for-height. Finally, weights-for-height are higher when
the child’s mother works and, surprisingly, the higher is the household dependency
ratio. These results are somewhat consistent with weights-for-height being higher
when more household members engage in economically productive activities.
Overall, the results are consistent with Figure 3.3 in that they show that it
was rural Lower Egyptian children who seemed to have been hit hardest by the
events in the run-up to 2008. Again, this is somewhat surprising. Recall that
it was agricultural households in rural Lower Egypt who were expected to have
benefited from high food prices the most, as wheat was being harvested there in
relatively large amounts as the 2008 survey was being conducted. Thus, it is striking
to see that rural Lower Egyptian children’s weights-for-height in 2005 were well
on their way back up to their pre-2003 levels, but that those weights-for-height
were then actually lower (if anything) by 2008. In contrast, children’s weights-for-
height in urban and rural Upper Egypt increased between 2005 and 2008. Also,
out of all the groups of children considered, urban children were the only ones to
have weighed more (given height) in 2008 than 2000. This seems to contradict the
predictions set forth in the conceptual framework, which predicts that as net food
9Given the positive relationship between parental education and a wide variety of measures
of household socioeconomic status and well-being, the fact that father’s education is positively
related to weight-for-height suggests that the latter is a useful measure of children’s health and
well-being.
128
buying households, urban households should have been relatively severely a↵ected
by high food prices. The hypotheses that high food prices are better for rural or net
food selling households than for urban or net food buying ones (respectively) will
be more properly assessed, however, when we examine what happened to children’s
weights-for-height as a function of parents’ occupations as well as region of residence
(and all this at di↵erent conditional quantiles of the weight-for-height distribution).
3.4.2.2 Region and Parental Occupation-Specific OLS Results
We turn now to the results for the version of the specification that allows for
the region-specific e↵ects over time to vary with parents’ occupations (see Table 3.4).
Once again, the results are highly similar regardless of whether or not controls for
parental education, household composition and maternal labor supply are included
(column 2 contains these additional controls). Since the urban indicators do not vary
by parental occupation, the results for urban children’s weights-for-height are quite
similar to the results described in the previous sub-section. In particular, urban
children’s weights-for-height were substantially negatively a↵ected by the events
leading up to measurement in 2003, they were heading for recovery to their 2000
weights-for-height in 2005 but were not quite there yet, and they had more than
recovered to baseline levels by 2008.
Turning to Lower rural Egyptian children, we see more results that seem to
contradict what was expected. In particular, the children of self-employed agricul-
tural workers–the group of children considered the most likely to have benefited
129
from high food prices in 2008–started out heavier (given height) than their urban
counterparts in 2000, and they were hit quite hard by the events leading up to 2003
but had recovered by 2005. In 2008, however, these children’s average weights-for-
height decreased substantially, and were actually not statistically distinguishable
from their weights-for-height in 2003. For the children of agricultural employees
and non-agricultural workers, however, weights-for-height were on average higher
in 2000 than they were for urban children, the events leading up to 2003 took a
substantial toll, and they were (at least) well on their way to recovering to their
2000 weights-for-height by 2005. But while weights-for-height in 2008 were statisti-
cally indistinguishable from their 2005 levels (according to t-tests) for the children
of parents who did not work in agriculture, the children of agricultural employees
also lost weight between 2005 and 2008.
In other words, weights-for-height for the children of non-agricultural work-
ers in 2008 had merely not grown relative to 2005, but weights-for-height for the
children of self-employed agricultural workers and agricultural employees decreased
substantially. Thus, the group of children in Lower rural Egypt who were expected
to have benefited the most from high food prices in 2008 seem to have been harmed
by them. One possible explanation for this is that self-employed farmers had espe-
cially high input costs to attempt to manage, and therefore household consumption
might have decreased as a result.
Many aspects of the results for Upper rural Egyptian children will look familiar
by this point. In particular, the children of self-employed agricultural workers,
agricultural employees and non-agricultural workers in Upper rural Egypt each had
130
average weights-for-height that were roughly comparable to their urban counterparts
in 2000, those weights-for-height were substantially negatively a↵ected by events in
the run-up to measurement in 2003, and substantial recoveries were under way in
2005. However, while that recovery for the children of non-agricultural workers had
continued in 2008 (if not statistically so, according to a t-test of equality of the
2005 and 2008 coe cients), weights-for-height in 2008 were significantly lower for
self-employed farmers’ and agricultural employees’ children than in 2005. Thus,
agricultural workers’ children in Upper rural Egypt were harmed by the events
leading up to measurement in 2008 in a way that other children in that region were
not. Again, it is possible that higher input costs make agriculture less lucrative for
these children’s parents.
As for the control variables (in column 2 of Table 3.4), females and younger
children are (unsurprisingly) still lighter (and once again substantially so in the
case of younger children). Also, father’s highest grade completed at school is still
significantly positively related to children’s weights-for-height. Having more young
females in the household still tends to be negatively associated with weights-for-
height, while higher numbers of prime-aged males in the household is associated
with higher ones.
In conclusion, the linear regression results seem to contradict several of the
predictions set forth in the conceptual framework. In particular, the children re-
garded as most likely to have benefited from high food prices in 2008–the children
of self-employed agricultural workers in Lower rural Egypt–seemed to have been
hit rather hard by the events leading up to measurement then. Also, while urban
131
children’s weights-for-height did not exhibit a very large positive change between
2005 and 2008, they were statistically significantly higher. Again, this was the only
group of children considered whose weights-for-height were significantly higher in
2008 than they were in 2000. We turn next to an analysis of conditional quantiles
of the weight-for-height variable to see if the average changes over time described in
this section hold for di↵erent parts of the conditional distribution.
3.4.3 Conditional Quantile Regression Results
3.4.3.1 Region-Specific Quantile Regression Results
The conditional quantile regression results for the baseline set of covariates
can be seen in Table 3.5. The results in Table 3.5 correspond to the specification
that only controls for region-year e↵ects, child gender and child age. There are
not substantial changes in the magnitudes of the key right-hand side variables after
controls for household members’ education and age and household composition are
added to the model (see Table 3.6), and there are certainly not changes in the
qualitative nature of the findings. Quantile regressions were run for the 5th, 10th,
25th, 50th, 75th, 90th and 95th percentiles. Also, Figures 3.4-3.6 provide a graphical
summary of the results for the key independent variables (the region-specific year
indicators) in Table 3.6, and so the discussion below will focus on the results from
that table. The omitted group is male children in urban Egypt in 2000 were were
less than a year old, whose mother did not work (with average values of the other
control variables). The results described below will mostly describe changes over
132
time in other children’s weights-for-height relative to this omitted group of children.
Figure 3.4 displays the results for the urban indicators in 2003, 2005 and 2008.
The results in this figure should be interpreted as changes in conditional quantiles of
the weight-for-height distribution over time for urban children (relative to weights-
for-height in 2000, holding children’s gender and age and other factors constant). For
example, the 0.05 conditional quantile for urban Egyptian children measured in 2003
is 0.00580 units below its 2000 analog (see column 1 of Table 3.6), which implies that
this point of the conditional distribution of weight-for-height decreased moderately
for urban children between 2000 and 2003. This change represents slightly more
than one-fifth of the sample standard deviation of the weight-for-height outcome.
As Figure 3.4 shows, the OLS results mostly do a good job of characteriz-
ing the e↵ects of the events leading up to 2003, in the sense that the conditional
quantiles are not substantially di↵erent from one another then. (It might be worth
noting, however, that according to the results in Table 3.7, the 0.05 and 0.95 condi-
tional quantiles are statistically significantly di↵erent from one another, if not highly
economically significantly di↵erent.) Turning to urban children measured in 2005,
we see that the economic growth over the preceding couple of years resulted in the
conditionally heaviest children’s weights-for-height surpassing those of urban chil-
dren in 2000 (as evidenced by the positivity of the slope for the upper conditional
quantiles). On the other hand, while conditionally lighter urban children in 2005
were very slightly heavier than they were in 2003, they had certainly not recovered
to their 2000 weights-for-height like the heaviest children had. It would appear that
economic growth in the run-up to measurement in 2005 did not benefit condition-
133
ally lighter children in urban Egypt the way it did conditionally heavier children
there. This is certainly true in a statistical sense for the 0.05 and 0.95 conditional
(for example) according to the results in Table 3.7, which show that the di↵erences
between heavy and light children went from being significantly negative in 2003 to
being significantly positive in 2005.
This is even more true for the events leading up to measurement of urban
children in 2008. In particular, despite 3 years’ worth of solid economic growth
between 2005 and 2008, the bottom decile of urban children was actually slightly
lighter (given height) in the latter year than they were in the former. This is clearly
consistent with food price inflation having had negative e↵ects on the poorest urban
children’s well-being. Also, the higher the conditional quantile, the greater were
the benefits of economic growth in the run-up to 2008, and the less important
were the negative e↵ects of high food prices then. Indeed, quantiles near the top
of the conditional weight-for-height distribution for urban children in 2008 were
substantially higher than their 2005 analogs, and were considerably higher than
their 2000 analogs as well. It would appear therefore that the net e↵ects of food
price inflation and economic growth were negative for the conditionally lightest
urban children and positive for other urban children in the run-up to 2008. Again,
this spreading out of the conditional weight-for-height distribution is confirmed by
the results of tests of the equality of the inter-quantile regression urban children’s
coe cients across years in Table 3.7, which show that the di↵erences between heavy
and light urban children became larger over time.
Finally, while Figure 3.4 makes it clear that the mean (i.e., OLS) e↵ect does
134
a poor job of describing changes in weight-for-height throughout much of the con-
ditional weight-for-height distribution in general, the inadequacy of the OLS result
is particularly severe in the case of urban children’s weights-for-height in 2008. As
the figure shows (and as we will see), the 0.05 and 0.95 conditional quantiles, for
example, for urban children in 2008 were simply much farther from one another than
was the case in earlier years (statistically speaking as well as with respect to magni-
tudes). These results are consistent with food price inflation and economic growth
each having qualitatively di↵erent e↵ects on the di↵erent ends of the conditional
weight-for-height distribution.
Turning now to the conditional quantile results for Lower rural Egyptian chil-
dren in Figure 3.5, although conditional weights-for-height for the heaviest kids were
statistically larger than those of the lightest kids, in substantive terms the OLS re-
sults are mostly quite informative for most of the conditional weight-for-height dis-
tribution for Lower rural children in 2000. In particular, Lower rural children were
slightly heavier (given height) throughout the conditional weight-for-height distribu-
tion in 2000 compared to urban children then. Figure 3.5 also shows that children in
this region throughout the conditional distribution were fairly seriously negatively
a↵ected by the events leading up to measurement in 2003. Lower rural Egyptian
children at the 0.75 conditional quantile, for example, were roughly 0.00833 units
lighter in 2003 than they were in 2000, which represents a substantial drop.
By 2005, however, most Lower rural children were well on their way back up
to their 2000 weights-for-height. Also, once again similarly to urban children, this
was only barely true for the conditionally lightest Lower rural children in 2005,
135
while the 0.95 conditional quantile in 2005 was already higher than this region’s
2000 analog, for example. (And weight gains over time were statistically larger
for heavier children, according to tests of the equality of the coe cients in Table
3.7, which displays inter-quantile regression results.) For much of the conditional
distribution in 2005, however, the OLS result is a somewhat accurate indicator of
what was happening to Lower rural children’s weights-for-height.
Figure 3.5 also displays conditional quantile results for Lower rural children in
2008, and shows first and foremost that like urban children then, the conditionally
lightest children were substantially negatively a↵ected by the events leading up to
the second quarter of 2008, and were lighter then than their counterparts were in
2005. Again, this is striking given that the Egyptian economy grew substantially
during this time period. Beginning at the conditional median, however, Lower rural
children were heavier (given height) in 2008 than in 2005. Importantly, this is with
the exception of the 0.95 conditional quantile, which decreased over this time period.
There is actually no statistically significant di↵erence between the 2005-2008 weight
changes that took place at the top and bottom of the conditional distribution–the
conditional weight-for-height distribution for Lower rural children stopped spreading
out in 2008, at least when comparing the 0.05 and 0.95 conditional quantiles.
The conditionally lightest Lower rural children therefore had substantially
lower weights-for-height in 2008 than their counterparts in 2005 did, which is con-
sistent with high food prices having negative e↵ects on the health and well-being of
members of the poorest households in the region. Moreover, the decrease in weights-
for-height between 2005 and 2008 in this region for the conditionally lightest children
136
was considerably larger and a↵ected a greater proportion of the conditional distri-
bution than was the case for urban children. In particular, the 0.05-0.25 conditional
quantiles for Lower rural children in 2008 were at lower levels than their 2005 coun-
terparts, while this was only true for the 0.05 and 0.10 conditional quantiles for
urban children. Otherwise, the beneficial e↵ects of economic growth in the run-up
to measurement in 2008 seem to have dominated any negative e↵ects associated
with high food prices for (most of) the rest of the conditional weight-for-height dis-
tribution of Lower rural Egyptian children. We see, therefore, that lower average
weights-for-height in 2008 compared to 2005 in this region were mostly driven by
changes at the bottom and top of the conditional distribution, and that this average
decrease occurred despite what happened over the rest of the conditional distribu-
tion. It remains to be seen whether these results hold for each of the three parental
occupation-based groups that we will consider in the next sub-section.
Turning to Upper rural children in Figure 3.6, we see that these children had
lower weights-for-height than urban children did in 2000 (conditional on gender and
age and other factors) and that the OLS estimate is mostly informative through-
out the conditional distribution (although, again, there are statistically significant
di↵erences in the weights-for-height of Upper rural children at the 0.95 and 0.05
conditional quantiles). In 2003, once again, every conditional quantile analyzed was
lower than it was in this region in 2000. By 2005, on the other hand, the conditional
distribution was higher everywhere, but recovery was considerably more pronounced
at the top of the conditional distribution. This can be seen in Table 3.7, which shows
that the di↵erence between the 0.95 and 0.05 conditional quantiles went from being
137
significantly negative in 2003 to being significantly positive in 2005.
In 2008, the conditionally lightest Upper rural children had lost weight, and
so the net e↵ect of food price inflation and economic growth in the run-up to mea-
surement was once again negative for the conditionally lightest children. Above
the 0.25 conditional quantile up, however, weights-for-height were higher in 2008
than in 2005, but the di↵erence is quite modest below the 0.90 conditional quantile.
Thus, while the conditionally lightest children were lighter in 2008 (despite 3 years’
worth of economic growth), and the conditionally heaviest children were substan-
tially heavier, most of the conditional distribution was not much heavier than it
had been in 2005. According to tests of equality of the Upper rural 2005 and 2008
inter-quantile di↵erences, moreover, the heaviest children gained significantly more
weight than did the lightest children.
We can see in Table 3.6 that the conditional mean e↵ect for females is mostly
informative throughout the conditional distribution. We also see that the condi-
tionally lightest of each cohort of children is heaviest relative to the omitted group,
infants. Thus, it seems that some of the variation in weights-for-height amongst
infants disappears over time. For all other estimates, the OLS results are nor-
mally quite informative. In particular, what are more likely higher numbers of
economically productive household members are positively associated with higher
weights-for-height.
Many of the main results described here are also reflected in Table 3.7, which
displays results from inter-quantile regressions (between the 0.05 and 0.95 quan-
tiles). The results in this table denote the estimated di↵erences between the top
138
and bottom of the conditional distributions of each of the covariates in the model.
For Lower rural and Upper rural Egypt, the coe cients corresponding to the year
2000 indicator variables have some of the smallest magnitudes, which reflects the
fact that di↵erences between the top and bottom of the conditional distributions
in these regions were small (relative to this inter-quantile di↵erence for the omitted
group, urban male infant children in 2000). Inter-quantile di↵erences in 2003 for
urban and Upper rural children were (statistically) negative and larger, which re-
flects the facts that the downturn during that period a↵ected di↵erent parts of the
conditional weight-for-height distributions di↵erently, and that it was conditionally
heavier children who lost the most weight in or before 2003. The 2005 coe cients
for all 3 regions are positive and significant, which reflects the fact that during the
economic recovery period following the downturn of 2003, it was the conditionally
heaviest children whose weights-for-height increased the most. The same is true for
each region’s 2008 coe cient estimates–they are each significant and positive–and
this likely reflects the relative benefits of economic growth for conditionally heav-
ier children and the relative costs of high food prices for the conditionally lightest
children.
Another striking feature of the 2008 estimates for each region is how large
they are relative to the estimates for most earlier years; the conditionally heaviest
children were especially heavy relative to the conditionally lightest children in 2008,
statistically and substantively speaking (in urban and Upper rural Egypt anyway).
Again, this is consistent with both economic growth and high food prices each
having di↵erent e↵ects on weights-for-height at di↵erent parts of the conditional
139
distribution. These results make it clear that the OLS results mask substantial
variation across conditional quantiles in estimate sizes, and in so doing justify the
use of quantile regression for the analysis.
3.4.3.2 Region and Parental Occupation-Specific Quantile Regres-
sion Results
The conditional quantile regression results for the more flexible, region- and
parental occupation-specific specification can be seen in Tables 3.8 (for the version
of the specification that only contains additional controls for child gender and age)
and 3.9 (for the enhanced specification’s results). The results in the two tables are
quite similar, and for simplicity’s sake I will focus on the results in Table 3.9. Once
again, quantile regressions were run for the 5th, 10th, 25th, 50th, 75th, 90th and
95th percentiles of the conditional weight-for-height distribution. Graphical depic-
tions of the results for the key coe cients can be seen in Figures 3.7-3.13. Also,
for each conditional quantile considered, and for each of the 7 parental occupation-
based groups of children considered (i.e., urban children, the 3 groups of rural Lower
children and the 3 groups of rural Upper children), Table 3.10 contains di↵erences
between the children measured in 2008 and the children measured in 2005. (The
stars next to some of the di↵erences correspond to the results of t-tests of equality
of the 2005 and 2008 estimates.) So, for example, children at the 0.05 conditional
quantile of the weight-for-height distribution who lived in urban areas in 2008 were
0.00267 units lighter than their counterparts in 2005, and this di↵erence is signifi-
140
cantly di↵erent than zero at the 0.01 level. Since this chapter is primarily concerned
with what happened to di↵erent types of children’s weights-for-height in 2008, Table
3.10 contains many of the most important results.
Figure 3.7 displays the results corresponding to the estimates for the various
conditional quantiles of the urban resident indicators for the years 2003, 2005 and
2008. (The omitted group is once again male, urban, infant children in 2000.) The
results here are (unsurprisingly) highly comparable to the results for these same
coe cients in the region-specific quantile regression results (from Tables 3.5 and 3.6
and Figure 3.4). In particular, while the conditionally lightest urban children had
barely gained any weight between 2003 and 2005, the conditionally heaviest had
more than recovered. Also, the events leading up to measurement in 2008 were once
again found to have decreased the weights-for-height of the bottom of the conditional
distribution while the top three-quarters of the conditional distribution were heavier
in 2008 than they were in 2005 (see Table 3.10). Although the magnitude of the
decreases in weight-for-height between 2005 and 2008 that the lightest urban children
sustained are not particularly large, they are striking given the Egyptian economy’s
growth over that period. These results are consistent with high food prices having
negative e↵ects on those urban households least able to cope with shocks, and the
beneficial e↵ects of economic growth dominating for heavier urban children.
We turn now to the 3 parental occupation-specific groups of Lower rural Egyp-
tian children: children with at least one parent who was a self-employed farmer, chil-
dren with at least one parent who worked as an agricultural employee, and children
with no parent who worked in agriculture. We begin by considering these groups’ of
141
children’s weights-for-height in the year 2000. As can be seen in Figures 3.8, 3.9 and
3.10, the OLS results for these indicators are mostly informative for the conditional
quantile-specific estimates, in the sense that the magnitudes of the latter are nor-
mally relatively close to one another. While each of these 3 groups of children are
heavier on average than urban children were in 2000, we see that di↵erent parts of
the distribution are driving this di↵erential for each group of children. In particular,
it is the conditionally heaviest children with self-employed farmer parents and the
conditionally lightest children with agricultural employee parents who are heaviest
relative to their urban counterparts, while it is the middle of the conditional dis-
tribution of children with parents who work outside of agriculture in Lower rural
Egypt who are relatively heavy (compared to urban children, for example).
We have seen that each of these 3 Lower rural Egyptian groups of children
were substantially lighter in 2003 than they were in 2000 on average. We see from
the results in Table 3.9 and in Figures 3.8-3.10 that it was in fact the conditionally
lightest children with self-employed farmer parents and (to a lesser extent) non-
agricultural sector parents who were least negatively a↵ected by the events leading
up to measurement in 2003. This is not true for the children of agricultural employ-
ees, however, as it was the lightest of these children that had lost the most weight
relative to the omitted group.
We also saw before that in 2005, children in Lower rural Egypt had nearly
recovered to their 2000 weights-for-height on average, that it was children with par-
ents working in the agricultural sector who were driving this result, and that it
was the conditionally heaviest Lower rural children who had gained the most weight
142
(relative to height) between 2003 and 2005. The results in Table 3.9 and Figures 3.8-
3.10 show that it was these conditionally heavier children in each of the 3 parental
occupation-based groups that gained the most weight between 2003 and 2005. Re-
gardless of parental occupation, conditionally heavier children in Lower rural Egypt
gained more weight between 2003 and 2005 than their lighter counterparts. This
is shown clearly in Table 3.10, which displays 0.95-0.05 inter-quantile regression
results, and shows that di↵erences between the heaviest and lightest children went
from being (not significantly) negative in 2003 to being significantly positive in 2005.
Previously it was shown that it was the children of parents who worked in the
agricultural sector who were responsible for the fact that Lower rural Egyptian chil-
dren were lighter on average in 2008 than they were in 2005. It was also shown that
it was the weight-for-height losses of the conditionally lightest and heaviest children
that drove this e↵ect. Beginning with the children of self-employed agricultural
workers (the group of Lower rural children who lost the most weight on average be-
tween 2005 and 2008), we can see in Table 3.10 and Figure 3.8 that, while estimates
for all conditional quantiles were lower in 2008 than in 2005, it appears that it was
the conditionally lightest and conditionally heaviest of these children who lost the
most weight. For example, the 0.05 (0.95) conditional quantile shrank by 0.01357
(0.00945) units between 2005 and 2008, which, again, is striking given the extent to
which the Egyptian economy grew over that period (the di↵erence for the heavier
children is not significant, however). (As Table 3.10 shows, the conditional third
quartile for the children of these self-employed farmers’ children only decreased by
0.00401 units, in contrast, which represents a significantly smaller decrease than
143
took place at the 0.05 conditional quantile but a di↵erence that is not statistically
di↵erent at conventional levels from that which took place at the 0.95 conditional
quantile.) Comparing these larger changes at the tails of the conditional distribution
to the sample standard deviation in the weight-for-height outcome (0.028) makes
it clear that these were substantial decreases for these children of self-employed
farmers.
Recall the prediction from Section 3.2.3 that the heaviest children of self-
employed farmers in Lower rural Egypt were considered most likely to have gained
weight between 2005 and 2008. Table 3.10 shows, however, that these children seem
to have fared worse between 2005 and 2008 than the entire conditional weight-for-
height distribution of urban children, all but the lightest Lower rural children of
self-employed farmers, and the entire conditional distributions for the Lower ru-
ral children of agricultural employees and non-agricultural workers. One possible
explanation for this surprising decrease in weights-for-height for the conditionally
heaviest of these self-employed farmers’ children is that, while agricultural output
was worth more and so should have helped children in what were most likely to
be net wheat selling households to gain (or at least not lose) weight, inputs such
as fertilizer and labor–which are heavily used in Egyptian wheat farming–were also
more expensive. Indeed, Kherallah et al. (1999) note that fertilizer application rates
in Lower Egypt are some of the highest in the world, and find that on average, 79
person-days of labor were used on the typical hectare dedicated to the cultivation
of wheat. Moreover, IFAD (2012) reports that in 2008 agricultural wages were 4
times as high as they were a year earlier. World fertilizer prices were also about
144
twice as high in the second quarter of 2008 than they were a year earlier. Under
these circumstances, it seems plausible that costs could have increased by a greater
factor than did revenues, and profits could have decreased. On the other hand,
lower profits in 2008 for wheat farmers would contradict another finding from IFAD
(2012), namely that medium- and large-scale farmers reported benefiting from high
food prices. Finally, given higher output and input prices, it is also possible that
farmers’ children’s weights-for-height decreased to the extent they did in 2008 be-
cause of increases in the opportunity cost of the time that would otherwise have
been taken to ensure that young children were well-nourished throughout the day.
As for the especially large losses that the conditionally lightest of these Lower
rural children of self-employed farmers sustained, IFAD (2012) finds that some
households with farms too small for positive net food sales to occur sold output (at
high prices) that they otherwise would have consumed. Members of these house-
holds reportedly consumed subsidized bread in place of home-grown wheat-based
products but they also cut back consumption of non-staple foods such as meat,
poultry and fruits (which were also reportedly more expensive in 2008). Also, these
children’s weight-for-height losses could be partly explained by their households’
initial poverty levels. In other words, it could be that the conditionally lightest of
these children come from relatively poor and more vulnerable households, and so
might have been particularly hard hit by high food prices. These lighter children of
self-employed farmers could have been hit by multiple forces, then–high input prices
and high food prices–and it is therefore plausible that the inflation observed in 2008
would have had especially severe negative e↵ects on them.
145
Initially high poverty levels are also potentially why the bottom 15 percent
of the weight-for-height distributions of children with agricultural employee parents
or non-agricultural worker parents seem to have lost at least as much weight be-
tween 2005 and 2008 as children at the rest of most of the conditional distributions
did (see Table 3.10). In particular, above the 0.15 conditional quantile (where, it
should be noted, losses were not significantly negative according to t test results),
most of the children of agricultural employees lost only a small amount of weight
(given height). Also, some conditional quantiles were even higher in 2008 than in
2005 (though not statistically so). Similarly, some of the conditionally lightest 15
percent of children with non-agricultural employee parents were significantly lighter
than their counterparts in 2005; but most higher conditional quantiles were actually
higher for these children in 2008. It should be emphasized again, however, that
it was the conditionally lightest children of self-employed farmers in Lower rural
Egypt who really drove the region-wide decrease at the bottom of the conditional
distribution described in Section 3.4.2.2; the magnitudes of the changes are consid-
erably larger for the lightest children of self-employed farmers than they were for
their agricultural employee and non-agricultural worker analogs. Also, along with
the conditionally heaviest children of agricultural employees, it was once again the
children of self-employed farmers who really drove the weight loss amongst heavier
children. Thus, it was in fact self-employed farmers’ children (at the bottom and
top of the conditional weight-for-height distribution) who were mainly responsible
for the region-wide decrease between 2005 and 2008.
Turning now to Upper rural Egyptian children, recall that in 2000, these chil-
146
dren were slightly lighter (given height) than urban children then, and that it was
the children of agricultural employees and non-agricultural worker parents who were
lightest. From Table 3.9 and Figures 3.11, 3.12 and 3.13, we can see that for the
children of self-employed farmers, it is the top quarter of the conditional weight-for-
height distribution that was lighter than their urban counterparts in 2000 were. For
the children of agricultural employees and non-agricultural workers, it is most of
the rest of the conditional weight-for-height distribution that is lighter than urban
children in 2000.
We saw earlier that Upper rural children throughout the conditional weight-
for-height distribution lost a substantial amount of weight (given height) between
2000 and 2003 on average, and that children with self-employed farmer parents
were hit hardest. According to Table 3.9 and Figure 3.11, for the children of self-
employed farmer parents, conditional quantiles throughout the entire conditional
weight-for-height distribution were substantially lower in 2003 than in 2000. Also,
across conditional quantiles there is not a lot of variation in the magnitude of these
inter-year di↵erences overall, but if anything, conditionally heavier children lost
more weight (given height) than other children. This is also true for the children of
agricultural employees and non-agricultural employee parents (see Figures 3.12 and
3.13); if anything, therefore, conditionally heavier children of these types of workers
lost slightly more weight (given height) between 2000 and 2003 than lighter children
did (though any di↵erences between the 0.95 and 0.05 conditional quantiles were
not statistically distinct, for example).
By 2005, Upper rural children–and in particular Upper rural children with
147
an agricultural employee parent–had made a moderate amount of progress back
up towards their 2000 weights-for-height on average. Moreover, it was generally
conditionally heavier (given height) Upper rural children who gained the most weight
between 2003 and 2005. Table 3.9 and Figures 3.11-3.13 show that for all 3 groups
of children, it was the conditionally lightest children who registered the smallest
increases in weight-for-height. Once again, conditionally heavier children gained
considerably more weight than did their lighter counterparts between 2003 and 2005,
and the di↵erences in weights gained between the 0.95 and 0.05 conditional quantiles
are significant for all 3 groups of children.
We have also seen that, while Upper rural children as a whole were heavier
(given height) in 2008 than in 2005 on average, this was only true for the children
of non-agricultural worker parents. Moreover, this same overall gain in weights-
for-height on average masks the fact that the bottom quarter of the conditional
distribution was actually lighter in 2008. According to Table 3.10 and Figures 3.11-
3.13, children from what were most likely to have been net food selling households in
Upper rural Egypt seem to have actually faired worse in the run-up to measurement
in 2008 than children from the most disadvantaged net food buying households,
as well as their heavier net food buying counterparts. In particular, the heaviest
children of self-employed farmers weighed considerably less in 2008 than their coun-
terparts did in 2005 (and statistically so), while the lightest children of self-employed
farmers, agricultural employees and non-agricultural workers lost a small-moderate
amount of weight. (Also, according to t test results, di↵erences in the inter-quantile
regression coe cients are significant and the heaviest children of self-employed farm-
148
ers lost significantly more weight than did the lightest children of these farmers.)
The heaviest children of non-agricultural workers gained a substantial amount of
weight between 2005 and 2008, however, and once again the heaviest urban children
gained weight as well. Thus, once again, we see that the group of children mostly
likely to have been members of net food selling households lost more weight than
other children in their region and urban children. Again, this could be driven by
the possibility that net food selling households were hit with substantially higher
input costs as well as higher output prices.
Also, for the children of agricultural employees and non-agricultural workers,
the bottom 10 and 5 percent (respectively) of the conditional weight-for-height dis-
tribution was significantly lighter in 2008. Indeed, the entire conditional distribution
of agricultural employees’ children’s weights-for-height were lower in 2008 than in
2005 (if not always statistically so). For example, children with agricultural em-
ployee parents at the 0.05 conditional distribution were 0.00521 units lighter (given
height) in 2008 (compared to an insignificant decrease of 0.00387 for the children of
self-employed farmers). These results are once again consistent with the negative
e↵ects of high food prices outweighing any beneficial e↵ects from economic growth
for the lightest (and, presumably, the poorest) children in Upper rural Egypt.
The conceptual framework also predicts that Lower rural net food selling and
net food buying households would have benefited from high food prices to a greater
extent than their Upper rural counterparts did. This is not the case for children with
self-employed farmer parents. Between 2005 and 2008, the changes in Lower rural
weights-for-height at all conditional quantiles were negative, while for Upper rural
149
conditional quantiles, they were sometimes positive (and those that were negative
were smaller than their Lower rural analogs). For the children of agricultural employ-
ees, the conditionally lightest Upper rural children and the conditionally heaviest
Lower rural children lost more weight than their analogs in the other region. For the
children of non-agricultural workers, results are mixed, with neither region clearly
having fared better. Overall, then, it appears that households in Lower rural Egypt
did not benefit from the more widespread cultivation of wheat there like they were
expected to.
Finally, the estimates of the controls for child gender and age in Tables 3.8
and 3.9 are similar to those in Tables 3.5 and 3.6. In particular, female children
are lighter (given height) than males throughout the conditional weight-for-height
distribution, and there is not much variation in estimated e↵ects across conditional
quantiles.
Table 3.11 displays inter-quantile regression results for the parental occupation-
specific set of covariates. Once again, for most of the parental occupation indicators
for the year 2000, the estimated di↵erences between the 0.05 and 0.95 conditional
quantiles are quite small (relative to the same di↵erence for the omitted group, ur-
ban male infant children in the year 2000). The inter-quantile di↵erences for urban
and all 3 groups of Upper rural children for the year 2003 are once again signifi-
cantly negative, which reflects the fact that it was conditionally heavier children in
these places whose weights-for-height decreased the most between 2000 and 2003.
The fact that it was conditionally heavier children who gained more weight between
2003 and 2005 in all 3 regions is also once again shown in Table 3.11. Indeed, in
150
Lower rural Egypt in particular, it is clear that it was conditionally heavier children
in all 3 parental occupation-based groups who drove this recovery for heavier chil-
dren. This is less true for Upper rural Egypt, where it was the conditionally heavier
children of agricultural employees who drove the earlier, region-wide result. The rel-
ative recovery of conditionally heavier children continued in 2008 for urban children
and all 3 types of Lower rural children, as well as the children of non-agricultural
worker parents in Upper rural Egypt. Thus, once again, for each of these groups
of children, the benefits of economic growth seem to have been strongest for condi-
tionally heavier children, while high food prices seem to have taken the largest toll
on the weights-for-height of lighter children.
3.5 Conclusion
I have examined changes in Egyptian children’s weights-for-height for di↵erent
region- and parental occupation-based groups to get a sense of how changes di↵ered
for (what I assume are) net food selling and net food buying households. Given
the prominence of food in the consumption baskets of poor households all over the
developing world and the arrival of what appear to be at least occasionally seriously
higher food prices, policymakers are naturally concerned about the e↵ects of food
prices on individual well-being. Research that speaks to these concerns that utilizes
actual household survey data is fairly scarce, however. In particular, there is not
enough evidence on what kinds of factors might help households cope with food
shocks or what kinds of households need help the most urgently.
151
I find evidence that, for many children, high food prices in 2008 seem to have
largely counteracted whatever beneficial e↵ects several years’ worth of respectable
economic growth might have brought. This is particularly true for Egyptian children
closer to the bottom of the weight-for-height distribution; while economic growth
seems to have helped heavier children weather the food price shock, children in
what were likely the poorest households were not so fortunate. Moreover, this is
despite the Egyptian government’s e↵ort to shield more households from the e↵ects
of high food prices by expanding previously existing social safety net programs. I
also find that those children in what are most likely to have been net food selling
households–at the point in the agricultural cycle when they were expected to be cash
rich–experienced what appeared to be some of the largest decreases in weight-for-
height. (Again, however, these decreases were not statistically di↵erent from zero.)
While these results do not provide us with a well-quantified relationship between
food prices and children’s weights-for-height, they do provide a sense of how powerful
a force food price inflation is compared to economic growth, and they suggest that
it is likely the poorest households that policy should target first and foremost. Also,
if agricultural input prices are high along with food prices, then perhaps policies
designed to help children in net food selling households should also be considered.
This conclusion would be in contrast to the conclusions of a paper like Ivanic and
Martin (2014), who find that high food prices are good for farming households in the
long run. One of the main novelties of this chapter, however, is its use of a relatively
objective measure of household well-being from survey data. Given this feature, the
chapter serves as a complement to existing simulation papers. The conclusion here
152
is also di↵erent. In particular, one of the main findings is that a child’s place
in the weight-for-height–and, presumably, household income–distribution is just as
important to consider as is their region of residence or their parents’ occupations,
when thinking about whether they should be targeted for aid.
153
Figure 3.1: Kernel Density, Weight-for-Height Z-Score, By Year
154
Figure 3.2: GDP Growth Per Capita and Food Price Inflation in Egypt,
1997-2009
Figure 3.3: Weights-for-Height Over Time, By Region
155
Figure 3.4: Weight-for-Height Changes for Urban Children, By Condi-
tional Quantile
Figure 3.5: Weight-for-Height Changes for Lower Rural Children, By
Conditional Quantile
156
Figure 3.6: Weight-for-Height Changes for Upper Rural Children, By
Conditional Quantile
Figure 3.7: Weight-for-Height Changes for Urban Children, By Condi-
tional Quantile
157
Figure 3.8: Weight-for-Height Changes for Lower Rural Children With
Self-Employed Farmer Parents, By Conditional Quantile
Figure 3.9: Weight-for-Height Changes for Lower Rural Children With
Agricultural Employee Parents, By Conditional Quantile
158
Figure 3.10: Weight-for-Height Changes for Lower Rural Children With
Non-Agricultural Worker Parents, By Conditional Quantile
Figure 3.11: Weight-for-Height Changes for Upper Rural Children With
Self-Employed Farmer Parents, By Conditional Quantile
159
Figure 3.12: Weight-for-Height Changes for Upper Rural Children With
Agricultural Employee Parents, By Conditional Quantile
Figure 3.13: Weight-for-Height Changes for Upper Rural Children With
Non-Agricultural Worker Parents, By Conditional Quantile
160
Table 3.1: Demographic and Health Survey Years and Months
161
Table 3.2: Descriptive Statistics
162
163
Table 3.3: Region-Specific OLS Results
164
165
Table 3.4: Region and Parental Occupation-Specific OLS Results
166
167
Table 3.5: Region-Specific Quantile Regression Baseline Results
168
Table 3.6: Region-Specific Quantile Regression Enhanced Specification Results
169
170
Table 3.7: Region-Specific Inter-Quantile Regression Results
171
Table 3.8: Region and Parental Occupation-Specific Baseline Quantile Regression Results
172
173
Table 3.9: Region and Parental Occupation-Specific Enhanced Specification Quantile Regression Results
174
175
176
Table 3.10: Region and Parental Occupation-Specific Conditional Weight-for-Height 2008 - 2005 Di↵erences
177
Table 3.11: Region and Parental Occupation-Specific Inter-Quantile Re-
gression Results
178
179
Chapter A: E↵ects of Shocks at Particular Ages
The purpose of this appendix is to examine in more detail the exact ages
at which wet and dry shocks have their e↵ects on human capital. The following
specification will be used to conduct this analysis:Hipt = ↵+ 20X
j=0
 jwetyearagejpt + 20X
j=0
⇢jdryyearagejpt +  p + ✓ct + r(t)+X 0ipt + "ipt.
(A.1)
The outcomes being analyzed here are once again the likelihood of being literate and
the highest grade completed at school. The wetyearagej (dryyearagej) variable
denotes an indicator equal to one if individual i from rainfall neighborhood p and
born in year t experienced a wet (dry) shock at age j, j 2 {0, 20}. All other control
variables are the same as the variables in the baseline specification. The estimates
for the wetyearagej and dryyearagej variables will therefore identify the e↵ects of
experiencing shocks at age j, as long as these shocks variables are not correlated with
the idiosyncratic error terms after conditioning on time-invariant, location-specific
factors, country-specific cohort factors, and regional linear trends.
The results can be seen in Table A.1. The results for women mostly reflect
the baseline results from Table 2.3: Wet shocks for females aged 7 and over and
180
dry shocks at preschool ages have negative e↵ects on human capital. 1 Shocks
e↵ects are not significantly negative at every one of these ages, but estimates are
normally negative, including between the ages of 7 and 13 in the case of dry shocks.
Importantly, dry shocks e↵ects are significantly negative and substantially sized at
ages 1-3. This seems more consistent with the type of health channel explanation
for the e↵ects of dry shocks early on described in Section 2.4.2.2 than with a lagged
income e↵ect explanation.
For men, the results are consistent with the baseline results in the sense that
wet shocks at relatively advanced ages (19 and 20) and dry shocks at preschool
ages (2-4) are sometimes significantly positive. 2 Again, the latter set of e↵ects
is consistent with the possibility that dry shocks resulted in positive health shocks
that are complemented with more education investments later (or negative health
shocks that are more than compensated for later).
1These point estimates are subject to the same kinds of sample selection bias as described
in Section 2.5.2. In particular, wet shocks from age 7-13 and dry shocks before age 7 might be
too large in magnitude, while (if anything) the e↵ects of wet shocks from age 14 to age 20 are
understated.
2Again, these point estimates are potentially too large in magnitude due to sample selection
bias, as discussed in more detail in Section 2.5.2.
181
Table A.1: Shocks Ages Results
182
183
Bibliography
[1] Aguero, J.M. and Deolalikar, A. (2012). ”Late Bloomers? Identifying Critical
Periods in Human Capital Accumulation. Evidence from the Rwanda Geno-
cide,” Presented at 9th Midwest International Economic Development Confer-
ence, University of Minnesota.
[2] Ajani, O. and Yetunde, I. (2008). ”Gender Dimensions of Agriculture, Poverty,
Nutrition and Food Security in Nigeria,” IFPRI Nigeria Strategy Support Pro-
gram Background Paper No. 005.
[3] Al-Nashar, S. (2011). ”Nominal Wage and Price Dynamics in Egypt: An Em-
pirical Analysis,” Working Paper No. 163.
[4] Alderman, H., Hoddinott, J. and Kinsey, B. (2006). ”Long-term Consequences
of Early Childhood Malnutrition.” Oxford Economic Papers. 58(3): 450-474.
[5] Alderman, H., Hoogeven, H. and Rossi, M. (2008). ”Preschool Nutrition and
Subsequent School Attainment: Longitudinal Evidence from Tanzania.” Eco-
nomic Development and Cultural Change. 57(2): 239-260.
[6] Alessie, R., Baker, P., Blundell, R., Heady, C. and Meghir, C. (1992). ”The
Working Behavior of Young People in Rural Cote d’Ivoire.” The World Bank
Economic Review. 6(1): 139-154.
[7] Almond, D. and Currie, J. (2011). ”Killing Me Softly: The Fetal Origins Hy-
pothesis.” Journal of Economic Perspectives. 25(3): 153172.
[8] Andvig, J.C. (2001). ”Family-Controlled Child Labor in Sub-Saharan Africa:
A Survey of Research,” World Bank Social Protection Discussion Paper No.
0122.
184
[9] Arab Republic of Egypt Initiative on Soaring Food Prices. (2008). ”Mission
Findings and Recommendations.”
[10] Arshad, F.M. and Abdel Hameed, A.A. (2009). ”The Long Run Relationship
Between Petroleum and Cereals Prices.” Global Economy and Finance Journal.
2(2): 91-100.
[11] Attanasio, O., Fitzsimons, E., Gmez, A., Lpez, D., Meghir, C. and Mesnard,
A. (2006). Child Education and Work Choices in the Presence of a Conditional
Cash Transfer Programme in Rural Colombia, IFS Working Paper No. W06/01.
[12] Banerjee, A., Besley, T., Guinnane, T. (1994). ”Thy Neighbors Keeper: the
Design of a Credit Cooperative with Theory and a Test.” Quarterly Journal of
Economics. 109(2): 491515.
[13] Bardhan, P. (1983). ”Labor-Tying in a Poor Agrarian Economy: A Theoretical
and Empirical Analysis.” Quarterly Journal of Economics. 98(3): 501-514.
[14] Basu, K., Das, S. and Dutta, B. (2007). ”Child Labor and Household Wealth:
Theory and Empirical Evidence of an Inverted U.” Journal of Development
Economics. 91(1): 8-14.
[15] Bengtsson, N. (2010). ”How Responsive is Body Weight to Transitory Income
Changes? Evidence from Rural Tanzania.” Journal of Development Economics.
92(1): 53-61.
[16] Besley, T. (1995). Nonmarket Institutions for Credit and Risk Sharing in Low-
Income Countries. Journal of Economic Perspectives. 9(3): 11527.
[17] Besley, T., Coate, S. and Loury, G. (1993). The Economics of Rotating Savings
and Credit Associations. American Economic Review. 83(4): 792810.
[18] Bhalotra, S. (2003). ”Child Labor in Africa,” OECD Social, Employment and
Migration Working Paper No. 4.
[19] Bhalotra, S. (2004). ”Parent Altruism, Cash Transfers and Child Poverty,”
University of Bristol Discussion Paper No. 04/562.
[20] Bhalotra, S. and Umana-Aponte, M. (2010). ”The Dynamics of Women’s Labor
Supply in Developing Countries,” IZA Discussion Paper No. 4879.
185
[21] Bharadwaj, P., Eberhard, J. and Neilson, C. (2010). ”Do Initial Endowments
Matter Only Initially? Birth Weight, Parental Investments and Academic
Achievement in School,” UCSD Working Paper No. 9-16-2010.
[22] Binder, M. (1999). ”Schooling Indicators During Mexico’s ’Lost Decade’.” Eco-
nomics of Education Review. 18(2): 183-99.
[23] Binswanger, H. and McIntire, J. (1987). ”Behavioral and Material Determinants
of Production Relations in Land-Abundant Tropical Agriculture.” Economic
Development and Cultural Change. 36(1): 73-99.
[24] Bjrkman-Nyqvist, M. (2013). Income shocks and gender gaps in education:
Evidence from Uganda. Journal of Development Economics. 105(c): 237-253.
[25] Bleakley, H. (2010). ”Malaria Eradication in the Americas: a Retrospective
Analysis of Childhood Exposure.” American Economic Journal: Applied Eco-
nomics. 2(2): 1-45.
[26] Bolwig, S. and Paarup-Laursen, B. (1999). ”Nature, Work, Culture: Labor Uti-
lization in Agriculture and O↵-farm Employment Among the Fulani in Northern
Burkina Faso.” Danish Journal of Geography. Special Issue 2: 27-41.
[27] Bozzoli, C., Deaton, A. and Quintana-Domeque, C. (2009). ”Adult Height and
Childhood Disease.” Demography. 46(4): 647-669.
[28] Canagarajah, S. and Skyt Nielsen, H. (1999). ”Child Labor and Schooling in
Africa: a Comparative Study.” World Bank Social Protection Discussion Paper
No. 9916.
[29] Chay, K. and Greenstone, M. (2003). ”The Impact of Air Pollution on Infant
Mortality: Evidence from Geographic Variation in Pollution Shocks Induced
by a Recession.” Quarterly Journal of Economics. 118(3): 1121-1167.
[30] Chibiko O↵orma, G. (2009). ”Girl-Child Education in Africa.” The Federation
of the University Women of Africa. Lagos, Nigeria.
[31] Christensen, J.H., Hewitson, B., Busuioc, A., Chen, A., Gao, X., Held, I.,
Jones, R., Kolli, R.K., Kwon, W. T., Laprise, R., Magaa Rueda, V., Mearns,
L., Menndez, C.G., Risnen, J., Rinke, A., Sarr, A. and Whetton, P. (2007).
”Regional Climate Projections.” In: Climate Change 2007: The Physical Sci-
ence Basis. Contribution of Working Group I to the Fourth Assessment Report
of the Intergovernmental Panel on Climate Change [Solomon, S., D. Qin, M.
Manning, Z. Chen, M. Marquis, K.B. Averyt, M. Tignor and H.L. Miller (eds.)].
186
Cambridge University Press, Cambridge, United Kingdom and New York, NY,
USA.
[32] Cohen, M. and Garrett, J. (2009). ”The Food Price Crisis and Urban Food
(In)security,” UNFPA Human Settlements Working Paper Series. Available at
http://www.iied.org/pubs/display.php?0-10574IIED.
[33] Cogneau, D. and Jedwab, R. (2012). ”Commodity Price Shocks and Child
Outcomes: the 1990 Cocoa Crisis in Cte d’Ivoire.” Economic Development and
Cultural Change. 60(3): 507-534.
[34] Cunha, F. and Heckman, J. (2008). ”Formulating, Identifying and Estimating
the Technology of Cognitive and Noncognitive Skill Formation.” Journal of
Human Resources. 43(4): 738-782.
[35] Dammert, A. (2008). ”Child Labor and Schooling Response to Changes in Coca
Production in Rural Ecuador.” Journal of Development Economics. 86(1): 164-
180.
[36] De Janvry, A., Finan, F., Sadoulet, E. and Vakis, R. (2006). ”Can Conditional
Cash Transfers Serve as Safety Nets in Keeping Children at School and From
Working When Exposed to Shocks?” Journal of Development Economics. 79(2):
349-373.
[37] Dawoud, S. (2005). ”An Analysis of Food Consumption Patterns in Egypt.”
Kiel University Dissertation.
[38] De Vreyer, P., Guilbert, N. and Mespl-Somps, S. (2012). ”The 19871988 Locust
Plague in Mali: Evidence of the Heterogeneous Impact of Income Shocks on Ed-
ucation Outcomes,” Universit Paris Dauphine Document de Travail DT/2012-
05.
[39] Deaton, A. (1997). ”The Analysis of Household Surveys: A Microeconometric
Approach to Development Policy.” Johns Hopkins University Press, 1997.
[40] Dehejia, R. and Lleras-Muney, A. (2004). ”Booms, Busts and Babies’ Health.”
Quarterly Journal of Economics. 119(3): 1091-1130.
[41] Dercon, S. (2004). ”Growth and Shocks: Evidence from Rural Ethiopia.” Jour-
nal of Development Economics. 74(2): 309-329.
[42] Duflo, E. (2001). ”Schooling and Labor Market Consequences of School Con-
struction in Indonesia: Evidence from an Unusual Policy Experiment.” Amer-
ican Economic Review. 91(4): 795813.
187
[43] Duryea, S. and Arends-Kuenning, M. (2003). ”School Attendance, Child Labor
and Local Labor Market Fluctuations in Urban Brazil.” World Development.
31(7): 1165-1178.
[44] Edmonds, E. (2005). ”Does Child Labor Decline with Improving Economic
Status?” Journal of Human Resources. 40(1): 77-99.
[45] Edmonds, E. (2006). Understanding Sibling Di↵erences in Child Labor. Journal
of Population Economics. 19(4): 795-821.
[46] Edmonds, E., Pavcnik, N. and Topalova, P. (2010). ”Trade Adjustment and
Human Capital Investments: Evidence from Indian Tari↵ Reform.” American
Economic Journal: Applied Economics. 2(4): 42-75.
[47] Edmonds, E. and Schady, N. (2012). Poverty Alleviation and Child Labor.
American Economic Journal: Economic Policy. 4(4): 100-124.
[48] Edwards, D. and McKee, T. (1997). ”Characteristics of the 20th Cen-
tury Drought in the United States at Multiple Time Scales,” Col-
orado State University Atmospheric Science Paper No. 634. Available at
ccc.atmos.colostate.edu/edwards.pdf.
[49] El-Zanaty, F. and Way, A. (2001). ”Egypt Demographic and Health Survey
2000.” Cairo, Egypt: Ministry of Health, El-Zanaty and Associates, and Macro
International.
[50] El-Zanaty, F. and Way, A. (2004). ”Egypt Demographic and Health Survey
2003.” Cairo, Egypt: Ministry of Health, El-Zanaty and Associates, and Macro
International.
[51] El-Zanaty, F. and Way, A. (2006). ”Egypt Demographic and Health Survey
2005.” Cairo, Egypt: Ministry of Health, El-Zanaty and Associates, and Macro
International.
[52] El-Zanaty, F. and Way, A. (2009). ”Egypt Demographic and Health Survey
2008.” Cairo, Egypt: Ministry of Health, El-Zanaty and Associates, and Macro
International.
[53] Ersado, L. and Aran, M. (2014). ”Inequality of Opportunity Among Egyptian
Children,” World Bank Policy Research Working Paper No. 7026.
[54] Fafchamps, M. and Lund, S. (2003). Risk-Sharing Networks in Rural Philip-
pines. Review of Economic Studies. 71(2): 261287.
188
[55] Fafchamps, M., Udry, C. and Czukas, K. (1998). ”Drought and Saving in West
Africa: Are Livestock a Bu↵er Stock?” Journal of Development Economics.
55(2): 273-305.
[56] FAO. (2011). ”OECD-FAO Agricultural Outlook 2011-2020.”
[57] Fayyad, B., Johnson, S. and El-Khishin, M. (1995). ”Consumer Demand for
Major Foods in Egypt,” Iowa State University Center for Agricultural and
Rural Development Working Paper No. 95-WP 138.
[58] Ferreira, F. and Schady, N. (2008). ”Aggregate Economic Shocks, Child School-
ing and Child Health,” World Bank Policy Research Working Paper No. 4701.
[59] Filmer, D., and Schady, N. (2008). Getting Girls Into School: Evidence from
a Scholarship Program in Cambodia. Economic Development and Cultural
Change. 56(3): 581617.
[60] Fink, G., Kelsey Jack, B. and Masiye, F. (2014). ”Seasonal Credit Constraints
and Agricultural Labor Supply: Evidence from Zambia,” National Bureau of
Economic Research Working Paper No. 20218.
[61] Foster, A. and Rosenzweig, M. (1996). ”Technical Change and Human Capital
Returns and Investments: Evidence from the Green Revolution.” American
Economic Review. 86(4): 931-53.
[62] Frankenberg, E., Smith, J. and Thomas, D. (2003). Economic Shocks, Wealth,
and Welfare. Journal of Human Resources. 38(2): 280321.
[63] Fujii, T. (2011). ”Impact of Food Inflation on Poverty in the Philippines,”
Singapore Management University Economics and Statistics Working Paper
No. 14-2011.
[64] Funkhouser, E. (1999). ”Cyclical Economic Conditions and School Attendance
in Costa Rica.” Economics of Education Review. 18(1): 3150.
[65] Gertler, P. and Gruber, J. (2002). Insuring Consumption Against Illness. Amer-
ican Economic Review. 92(1): 5170.
[66] Gertler, P., Heckman, J., Pinto, R., Zanolini, A., Vermeersch, C., Walker, S.,
Chang, S. and Grantham-McGregor, S. (2013). Labor Market Returns to Early
Childhood Stimulation: a 20-year Follow up to an Experimental Intervention
in Jamaica,National Bureau of Economic Research Working Paper No. 19185.
189
[67] Gibson, J. and Kim, B. (2013). ”Quality, Quantity, and Nutritional Impacts of
Rice Price Changes in Vietnam.” World Development. 43(C): 329-340.
[68] Glewwe, P. (1999). Why Does Mothers Schooling Raise Child Health in Devel-
oping Countries? Journal of Human Resources. 34(1):124-159.
[69] Glewwe, P. and Kremer, M. (2006). ”Schools, Teachers and Educational Out-
comes in Developing Countries”, in Eric A. Hanushek and Finis Welch, editors,
Handbook of the Economics of Education: Volume 2.
[70] Glick, P. and Sahn, D. (2000). ”Schooling of Girls and Boys in a West African
Country: the E↵ects of Parental Education, Income and Household Structure.”
Economics of Education Review. 19(1): 63-87.
[71] Grimard, F. (1997). Household Consumption Smoothing Through Ethnic Ties:
Evidence from Cote dIvoire. Journal of Development Economics. 53(2): 391422.
[72] Grimard, F. and Hamilton, B. (1999). ”Estimating the Elderly’s Returns on
the Farm: Evidence from Cote d’Ivoire.” Journal of Development Economics.
58(2): 513-531.
[73] Hasan, M. and Sassanpour, C. (2008). ”Labor Market Pressures in Egypt: Why
is the Unemployment Rate Stubbornly High?” International Conference on The
Unemployment Crisis in the Arab Countries.
[74] Hayward, D. and Oguntoyinbo, J. (1987). ”Climatology of West Africa.” Barnes
and Noble, 1987.
[75] Headey, D. (2011). ”Was the Global Food Crisis Really a Crisis? Simulations
versus Self-Reporting,” IFPRI Discussion Paper No. 1087.
[76] Heckman, J. (1979). ”Sample Selection Bias as a Specification Error.” Econo-
metrica. 47(1): 153-161.
[77] Herz, B., Subbarao, K., Masooma, H. and Raney, L. (1991). ”Letting Girls
Learn: Promising Approaches in Primary and Secondary Education,” World
Bank Discussion Paper No. 133.
[78] Hoddinott, J., Maluccio, J.A., Behrman, J.R., Flores, R. and Martorell, R.
(2008). ”E↵ect of a Nutrition Intervention During Early Childhood on Eco-
nomic Productivity in Guatemalan Adults.” The Lancet. 371(9610): 411-416.
190
[79] Hulme, M., Doherty, R., Ngara, T., New, M. and Lister, D. (2001). ”African
Climate Change: 1900-2100.” Climate Research. 17(2): 145-168.
[80] ICF Macro. (2011). ”Demographic and Health Survey Interviewer’s Manual.”
MEASURE DHS Basic Documentation No. 2. Calverton, Maryland, USA: ICF
Macro.
[81] ICRISAT. (1999). ”Proceedings of the International Workshop on Socioe-
conomic Constraints to Development of Semi-Arid Tropical Agriculture.”
ICRISAT.
[82] IFAD. (2008). ”Soaring Food Prices and the Rural Poor: Feedback from the
Field.” Available at http://www.ifad.org/operations/food/food.htm.
[83] IFAD. (2012). ”Impact of High Food Prices on Farmers in the Near East.”
[84] Iliya, M. (1999). ”Income Diversification in the Semi-Arid Zone of Nigeria: a
Study of Gigane, Sokoto, Northwest Nigeria,” Afrika-Studiecentrum Working
Paper No. 39.
[85] Ivanic, M. and Martin, W. (2014). ”Short- and Long-Run Impacts of Food Price
Changes on Poverty,” World Bank Policy Research Working Paper No. 7011.
[86] Jacoby, H. and Skoufias, E. (1997). Risk, Financial Markets and Human Capital
in a Developing Country. Review of Economic Studies. 64(3): 311-336.
[87] Jensen, R. (2000). ”Agricultural Volatility and Investments in Children.” Amer-
ican Economic Review. 90(2): 399-404.
[88] Kherallah, M., Minot, N. and Gruhn, P. (1999). ”Adjustment of Wheat Pro-
duction to Market Reform in Egypt,” Markets and Structural Studies Division
Discussion Paper No. 32.
[89] Kochar, A. (1999). Smoothing Consumption by Smoothing Income: Hours-of-
Work Responses to Idiosyncratic Agricultural Shocks in Rural India. Review of
Economics and Statistics. 81(1): 50-61.
[90] Kraay, A. (2007). ”The Welfare E↵ects of a Large Depreciation: The Case of
Egypt, 2000-05,” World Bank Policy Research Working Paper No. 4182.
[91] Kruger, D. (2007). ”Co↵ee Production E↵ects on Child Labor and Schooling
in Rural Brazil.” Journal of Development Economics. 82(2): 448-463.
191
[92] Leon, G. (2009). ”Civil Conflict and Human Capital Accumulation: The Long
Term E↵ects of Political Violence in Peru,” BREAD Working Paper No. 256.
[93] Levine, D. and Yang, D. (2006). ”A Note on the Impact of Local Rainfall on
Rice Output in Indonesian Districts.” Mimeo.
[94] Lim, Y. and Townsend, R. (1998). General Equilibrium Models of Financial
Systems: Theory and Measurement in Village Economies. Review of Economic
Dynamics. 1(1): 58118.
[95] Liu, H., Mroz, T. and Adair, L. (2009). ”Parental Compensatory Behaviors and
Early Child Health Outcomes in Cebu, Philippines.” Journal of Development
Economics. 90(2): 209-230.
[96] Maccini, S. and Yang, D. (2009). ”Under the Weather: Health, Schooling, and
Economic Consequences of Early-Life Rainfall.” American Economic Review.
99(3): 1006-1026.
[97] Maitra, S. (2007). ”Dowry and Bride Price.” International Encyclopedia of the
Social Sciences, 2nd Edition.
[98] Maluccio, J. (2005). ”Coping with the ’Co↵ee Crisis’ in Central America: The
Role of the Nicaraguan Red de Proteccion Social,” IFPRI Food Consumption
and Nutrition Division Discussion Paper No. 188.
[99] Maluccio, J., Hoddinott, J., Behrman, J., Martorell, R., Quisumbing, A. and
Stein, A. (2009). The Impact of Improving Nutrition During Early Childhood
on Education Among Guatemalan Adults. Economic Journal. 199(537): 734-
763.
[100] Manacorda, M. (2006). Child Labor and the Labor Supply of Other Household
Members: Evidence from 1920 America. American Economic Review. 96(5):
1788-1801.
[101] McKenzie, D. (2003). ”How Do Households Cope with Aggregate Shocks?
Evidence from the Mexican Peso Crisis.” World Development. 31(7): 1179-
1199.
[102] Metz, H. (1990). ”Egypt: A Country Study.” Washington: GPO for the Li-
brary of Congress.
[103] Miguel, D., Satyanath, S. and Sergenti, E. (2004). ”Economic Shocks and Civil
Conflict: An Instrumental Variables Approach.” Journal of Political Economy.
112(4): 725-753.
192
[104] Mortimore, M. (1989). ”Adapting to Drought: Farmers, Famines and Deser-
tification in West Africa.” Cambridge University Press.
[105] Morduch, J. (1995). Income Smoothing and Consumption Smoothing. Journal
of Economic Perspectives. 9(3): 103-114.
[106] Mueller, P. and Zevering, K. (1969). ”Employment Promotion through Rural
Development: A Pilot Project in Western Nigeria.” International Labor Review.
100: 111.
[107] OECD. (2008). ”Rising Food Prices: Causes and Con-
sequences,” OECD Working Paper Series. Available at
http://www.oecd.org/trade/agriculturaltrade/40847088.pdf.
[108] Osili, U.O. and Long, B. (2008). Does Female Schooling Reduce Fertility?
Evidence from Nigeria. Journal of Development Economics. 87(1): 57-75.
[109] Osorio, P., Carlos, J. and Wodon, Q. (2010). ”Gender, Time Use, and Labor
Income in Guinea: Micro and Macro Analyses,” MPRA Paper No. 28465.
[110] Oyekale, A. S. (2009). ”Climatic Variability and its Impacts on Agricultural
Income and Households’ Welfare in Southern and Northern Nigeria.” Electronic
Journal of Environmental, Agricultural and Food Chemistry. 8(7): 443-465.
[111] Oyelere, R.U. (2010). ”Africa’s Education Enigma? The Nigerian Story.”
Journal of Development Economics. 91(1): 128-139.
[112] Paxson, C. (1992). ”Using Weather Variability to Estimate the Response
of Savings to Transitory Income in Thailand.” American Economic Reviews.
82(1): 15-33.
[113] Pongou, R., Salomon, J. and Ezzati, M. (2006). ”Health Impacts of Macroe-
conomic Crises and Policies: Determinants of Variation in Childhood Malnu-
trition Trends in Cameroon.” International Journal of Epidemiology. 35(3):
648-656.
[114] Ravallion, M. and Wodon, Q. (2000). Does Child Labor Displace Schooling?
Evidence on Behavioral Responses to an Enrollment Subsidy. Economic Jour-
nal. 110(462): 158-176.
[115] Reardon, T. (1997). ”Using Evidence of Household Income Diversification to
Inform Study of the Rural Nonfarm Labor Market in Africa.” World Develop-
ment. 25(5): 735-747.
193
[116] Reardon, T., Matlon, P. and Delgado, C. (1988). ”Coping with Household-level
Food Insecurity in Drought-a↵ected Areas of Burkina Faso.” World Develop-
ment. 16(9): 1065-1074.
[117] Robson, E. (2004). ”Children at Work in Rural Northern Nigeria: Patterns of
Age, Space and Gender.” Journal of Rural Studies. 20(2): 193-210.
[118] Rosenzweig, M. and Binswanger, H. (1993). Wealth, Weather Risk and the
Composition and Profitability of Agricultural Investments. Economic Journal.
103 (416): 56-78.
[119] Rosenzweig, M. and Stark, O. (1989). ”Consumption Smoothing, Migration
and Marriage: Evidence from Rural India.” Journal of Political Economy.
97(4): 905-926.
[120] Rosenzweig, M. and Wolpin, K. (1993). Credit Market Constraints, Consump-
tion Smoothing, and the Accumulation of Durable Production Assets in Low-
Income Countries: Investments in Bullocks in India. Journal of Political Econ-
omy. 101(2): 223-244.
[121] Rosenzweig, M. and Zhang, J. (2009). ”Do Population Control Policies Induce
More Human Capital Investment? Twins, Birth Weight and China’s One-Child
Policy.” Review of Economic Studies. 76(3): 1149-1174.
[122] Rucci, G. (2003). ”Macro Shocks and Schooling Decisions: The Case of Ar-
gentina,” Mimeo, Economics Department, University of California at Los An-
geles.
[123] Schady, N. (2004). ”Do Macroeconomic Crises Always Slow Human Capital
Accumulation?” The World Bank Economic Review. 18(2): 131-54.
[124] Schultz, T. P. (2004). School Subsidies for the Poor: Evaluating the Mexi-
can PROGRESA Poverty Program. Journal of Development Economics. 74(1):
199250.
[125] Sen, A. (1999). ”Development as Freedom.” Knopf, New York.
[126] Shah, M. and Millett Steinberg, B. (2013). ”Drought of Opportunities: Con-
temporaneous and Long Term Impacts of Rainfall Shocks on Human Capital,”
National Bureau of Economic Research Working Paper No. 19140.
[127] Siam, G. (2002). ”Agricultural Sector Model of Egypt,” CIHEAM Institut
Mediterraneen De Montpellier Working Paper Series.
194
[128] Sillah, B. (2009). ”An Econometric Analysis for the Impacts of Climate
Change on Cash and Food Crop Production in The Gambia.”
[129] Skoufias, E., Vinha, K. and Conroy, H. (2011). ”The Impacts of Climate Vari-
ability on Welfare in Rural Mexico,” World Bank Policy Research Working
Paper No. 5555.
[130] Stiglitz, J. (1990). ”Peer Monitoring and Credit Markets.” World Bank Eco-
nomic Review. 4(3): 351366.
[131] Strauss, J. and Thomas, D. (1995). ”Human Resources”, in J.Behrman and
T.N. Srinivasan, editors, Handbook of Development Economics: Volume 3.
[132] Thomas, D., Beegle, K., Frankenburg, E., Sikoki, B., Strauss, J. and Teruel,
G. (2004). ”Education in a Crisis.” Journal of Development Economics. 74(1):
53-85.
[133] Thomas, D., Strauss, J. and Henriques, M. H. (1991). ”How Does Mother’s
Education A↵ect Child Height?” Journal of Human Resources. 26(2): 183-211.
[134] Townsend, R. (1995). ”Consumption Insurance: an Evaluation of Risk-Bearing
Systems in Low-Income Countries.” Journal of Economic Perspectives. 9(3):
83-102.
[135] Udry, C. (1994). Risk and Insurance in a Rural Credit Market: an Empirical
Investigation in Northern Nigeria. Review of Economic Studies. 61(3): 495-526.
[136] Uwaifo Oyelere, R. (2010). Africas Education Enigma? The Nigerian Story.
Journal of Development Economics. 91(1): 128-139.
[137] Wang, L., Kanji, S. and Bandyopadhyay, S. (2009). ”The Health Impact of
Extreme Weather Events in Sub-Saharan Africa,” World Bank Policy Research
Working Paper No. 4979.
[138] Webb, P. and Reardon, T. (1992). ”Drought Impact and Household Response
in East and West Africa.” Quarterly Journal of International Agriculture. 31(3):
230-246.
[139] WFP. (2008). ”Vulnerability Analysis and Review of the Food
Subsidy Program in Egypt,” Working Paper Series. Available at
http://home.wfp.org/stellent/groups/public/documents/ena/wfp198346.pdf.
195
[140] Wikileaks Discussion Forum. (2011). ”Egypt: Frustration Over the Bread Cri-
sis, Rising Food Prices and Corruption.” Available at http://www.wikileaks-
forum.com/index.php?topic=6415.0.
[141] World Bank. (2007). ”Malawi Poverty and Vulnerability Assessment: Investing
in Our Future.” Washington, D.C.
[142] Yamano, T., Alderman, H. and Christiaensen, L. (2005). ”Child Growth,
Shocks and Food Aid in Rural Ethiopia.” American Journal of Agricultural
Economics. 87(2): 273-288.
[143] Yaro, J.A. (2006). ”Is Deagrarianisation Real? A Study of Livelihood Ac-
tivities in Rural Northern Ghana.” Journal of Modern African Studies. 44(1):
125-156.
196