ABSTRACT
Title of dissertation: HEALTH, AGRICULTURE
AND LABOR MARKETS
IN DEVELOPING COUNTRIES
Yeon Soo Kim, Doctor of Philosophy, 2010
Dissertation directed by: Professor Maureen Cropper
Department of Economics
Assistant Professor Jeanne Lafortune
Department of Economics
Rural households comprise a large share of the population in developing coun-
tries. This dissertation examines how the welfare of these households, whose eco-
nomic activity mainly relies on agriculture, is a ected by weather shocks and health
shocks in the context of West Africa and Vietnam.
In the second chapter of the dissertation, I use the variation in rainfall within
and across years at a detailed geographic level in West Africa to examine how rainfall
shocks might a ect the well-being of very young children. Variations in rainfall may
a ect not only income, but also the opportunity cost of time of parents, which may
negatively impact child welfare. I  nd that high long-term rainfall averages for a
particular location and month increase the probability of giving birth in the dry
season, whereas positive deviations from this long-term mean (\rainfall shocks")
have a small but statistically signi cant negative e ect on the probability of giving
birth in the rainy season. Further, contrary to what one might expect, rainfall shocks
do not appear to improve the survival chances of young children and shocks in the
 rst year of life have an adverse e ect on the survival of children that are born in
the rainy season. This result may be partly attributable to the  nding that rainfall
shocks signi cantly reduce the time mothers breastfeed their children, which could
be due to a trade-o with work. Breastfeeding is important for the health of young
children since it provides not only essential nutrients but also e ective protection
against various diseases.
In the third chapter, I examine the e ect of health shocks on the production
decisions of agricultural households in Vietnam. I look at whether malaria illnesses
experienced by the household have an e ect on their agricultural production deci-
sions. While I am not able to entirely overcome issues with endogeneity that are
persistent in this literature, results show that pro ts are negatively associated with
the share of household members experiencing malaria. This result is not explained
by the decrease in the total number of labor days the household employed. Rather,
households appear to change their crop choice to less labor-intensive, less pro table
crops in anticipation of these seasonal health shocks.
HEALTH, AGRICULTURE AND
LABOR MARKETS IN DEVELOPING COUNTRIES
by
Yeon Soo Kim
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial ful llment
of the requirements for the degree of
Doctor of Philosophy
2010
Advisory Committee:
Professor Maureen Cropper, co-Chair
Assistant Professor Jeanne Lafortune, co-Chair
Professor Mark Duggan
Assistant Professor Raymond Guiteras
Associate Professor Kenneth Leonard
c Copyright by
Yeon Soo Kim
2010
Dedication
To my mother, who would have been proud
To F, who always believed in me
ii
Acknowledgments
First and foremost, I would like to express my deepest gratitude to my advi-
sors, Maureen Cropper and Jeanne Lafortune, for their continuous encouragement
and support throughout this entire process. Their detail-oriented comments were
indispensable and this dissertation would not have been possible without their guid-
ance.
I would also like to thank Mark Duggan and Raymond Guiteras, who have
provided valuable comments at various stages of this dissertation. I am also grateful
to Kenneth Leonard who has agreed to serve as the Dean?s representative on my
dissertation committee. Thanks also to Ramanan Laxminarayan who kindly shared
the data used in the third chapter.
Finally, I acknowledge the technical support through the Maryland Population
Research Center. The handling of the large data involved in the second chapter has
been greatly facilitated through access to their computing facilities.
iii
Table of Contents
List of Tables v
List of Figures vi
1 Overview 1
2 The Impact of Rainfall on Early Child Health 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Background and Literature Review . . . . . . . . . . . . . . . . . . . 9
2.2.1 Women, Agriculture and Child Care in West Africa . . . . . . 9
2.2.2 Existing Literature . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.1 Fertility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.2 Infant Mortality and Breastfeeding . . . . . . . . . . . . . . . 16
2.4 Data Sources and Descriptive Statistics . . . . . . . . . . . . . . . . . 20
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5.1 Does rainfall shift the timing of births? . . . . . . . . . . . . . 23
2.5.2 Reduced-form Impact on Infant Mortality . . . . . . . . . . . 26
2.5.3 Rainfall shocks, Opportunity cost and Breastfeeding . . . . . . 28
2.5.4 Falsi cation Test with Neonatal Mortality . . . . . . . . . . . 30
2.5.5 Further results by urban{rural . . . . . . . . . . . . . . . . . . 31
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3 Health Shocks and Production Decisions of Agricultural Households 56
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.3 Background on Malaria in Vietnam . . . . . . . . . . . . . . . . . . . 64
3.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.5 Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.5.1 Vietnam Living Standard Measurement Survey (LSMS) . . . . 70
3.5.2 Economic Reform and Land Title Data . . . . . . . . . . . . . 71
3.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.6.1 Description of the Data . . . . . . . . . . . . . . . . . . . . . . 73
3.6.2 Models of Malaria Incidence . . . . . . . . . . . . . . . . . . . 74
3.6.3 Malaria Illness and Farm Pro ts . . . . . . . . . . . . . . . . . 76
3.6.4 Malaria Illness and Use of Agricultural Labor . . . . . . . . . 79
3.6.5 The Impact of Malaria on Crop Choice . . . . . . . . . . . . . 80
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
iv
List of Tables
2.1 Descriptive Statistics (All households) . . . . . . . . . . . . . . . . . 34
2.2 Breastfeeding in months . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3 Fertility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.4 Fertility results with nonlinearity . . . . . . . . . . . . . . . . . . . . 37
2.5 Infant mortality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6 Infant mortality results with nonlinearity . . . . . . . . . . . . . . . . 41
2.7 Breastfeeding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.8 Neonatal mortality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.9 Fertility by urban{rural with mother  xed e ects . . . . . . . . . . . 47
2.10 Infant mortality by urban{rural with mother  xed e ects . . . . . . . 48
2.11 Breastfeeding by urban{rural . . . . . . . . . . . . . . . . . . . . . . 49
2.12 Complete list of surveys . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1 Descriptive Statistics (All households) . . . . . . . . . . . . . . . . . 83
3.2 Descriptive Statistics (Panel households) . . . . . . . . . . . . . . . . 84
3.3 Geographic distribution of households . . . . . . . . . . . . . . . . . . 85
3.4 Means by province, grouped by region . . . . . . . . . . . . . . . . . 85
3.5 Determinants of malaria illness . . . . . . . . . . . . . . . . . . . . . 87
3.6 Determinants of malaria illness (Households  xed e ects) . . . . . . . 89
3.7 Relationship between malaria illness and farm pro ts . . . . . . . . . 91
3.8 Relationship between malaria illness and farm pro ts, controlling for
interview month . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.9 Relationship between pro ts and malaria illness using alternative
malaria measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.10 Malara incidence and total agricultural labor use of household . . . . 96
3.11 Malaria incidence and total labor use of household members only
(excluding non-household labor) . . . . . . . . . . . . . . . . . . . . . 97
3.12 Total labor use regressions with alternative malaria measure . . . . . 98
3.13 E ect of malaria incidence at the household- and province-level in
1992 on crop choice in 1998 . . . . . . . . . . . . . . . . . . . . . . . 99
v
List of Figures
2.1 Fraction of births by birth month, by country . . . . . . . . . . . . . 53
2.2 Distribution of births by birth year, by country . . . . . . . . . . . . 55
vi
Chapter 1
Overview
Rural households comprise a large share of the population in developing coun-
tries and they face various types of shocks that might a ect their well-being. The
assessment of the impact of these shocks is complicated because unobservable char-
acteristics of the household could be correlated with the likelihood of receiving such
shocks.
A large body of literature has studied the vulnerability of agricultural house-
holds to shocks due to volatile weather conditions (e.g., Paxson (1992)) or illnesses
(e.g., Gertler and Gruber (2002)). In general, households can mitigate the adverse
e ect of income risk either by managing or coping with it. Strategies of risk man-
agement include diversi cation of crops (Dercon (1996)) or occupation, while risk-
coping measures include saving and borrowing behavior to smooth consumption
across time or across households through risk-sharing (Deaton (1992), Townsend
(1994), Grimard (1997); see Alderman and Paxson (1992) for an overview of the
literature). While most studies reject perfect insurance they  nd some evidence of
consumption smoothing.
Several studies examine the consequences of these shocks for children: Jensen
(2000) shows that adverse weather shocks led to lower investments in education and
health, measured by school enrollment, short-term nutritional status, and use of
1
medical services in C^ote d?Ivoire. Jacoby and Skou as (1997) concludes that school
attendance  uctuated with seasonal income in India. Foster (1995)  nds that severe
 oods had an adverse impact on children?s weight in rural Bangladesh.
The next two chapters of this dissertation examine how the welfare of house-
holds whose economic activity mainly relies on agriculture is a ected by exogenous
weather shocks and health shocks in the context of West Africa and Vietnam, re-
spectively.
In the second chapter, I exploit the exogenous variation in rainfall within and
across years at a detailed geographic level in West Africa to examine how rainfall
shocks a ect the health of very young children. Rainfall generates variation in
income through its e ect on agricultural output and more income is usually bene cial
for the survival of children. However, rainfall also a ects the opportunity cost of
time of parents and may thus have a negative impact. I match detailed rainfall
data from nearby weather stations to household data drawn from Demographic and
Health Surveys from nine West African countries. I  rst examine how short- and
long-run variations in rainfall a ect the timing of birth of children, possibly due to
the e ect on the opportunity cost of their time. I  nd that high long-term rainfall
averages for a particular location and month increase the probability of giving birth
in the dry season, whereas positive deviations from this long-term mean (\rainfall
shocks") have a small but statistically signi cant negative e ect on the probability of
giving birth in the rainy season. Further, contrary to what one might expect, rainfall
shocks do not appear to improve the survival chances of young children and shocks
in the  rst year of life have an adverse e ect on the survival of children that are
2
born in the rainy season. This result may be partly attributable to the  nding that
rainfall shocks signi cantly reduce the time mothers breastfeed their children, which
could be due to a trade-o between work and child care. Breastfeeding is important
for the health of young children since it provides not only essential nutrients but also
immune bene ts against respiratory, diarrheal and other infectious diseases. These
account for a substantial share of child mortality in this region. While breastfeeding
might be directly linked to infant mortality, it might as well be representative of the
parental care the child receives in early age. Since households residing in rural areas
are much more likely to engage in agricultural activities, one would expect these
results to be stronger for rural households. This is indeed con rmed: results are
driven by children in rural areas where households are more vulnerable to variations
in rainfall.
In the third chapter, I examine the e ect of health shocks on the produc-
tion decisions of agricultural households in Vietnam. Previous literature has found
little e ect of health shocks on pro ts, which has been attributed to rather well-
functioning input markets in the countries studied. I examine a particular type of
health shock from a disease that, because it is endemic, might be largely anticipated
by households on a recurring basis. Using household-level data from two rounds of
the Vietnam Living Standard Measurement Surveys, I look at whether malaria ill-
nesses experienced by the household have an e ect on agricultural pro ts. While
concerns remain about the endogeneity of the malaria illness variable, pro ts are
found to be negatively associated with the share of household members experiencing
a malaria episode. This result is not explained by the decrease in the total number of
3
labor days the household employed. Rather, households appear to change their crop
choice to less labor-intensive, less pro table crops in anticipation of these seasonal
health shocks.
4
Chapter 2
The Impact of Rainfall on Early Child Health
2.1 Introduction
In West Africa, one child out of  ve dies before age  ve and one out of ten dies
during the  rst year of life. According to the World Health Organization (WHO
(2008)), 75% of childhood deaths in Africa are due to infectious, diarrheal or respi-
ratory diseases.1 Low income, high disease prevalence and lack of access to proper
health care are among the factors that are known to contribute to high childhood
mortality. Moreover, interventions regarding timely immunization, provision of mi-
cronutrients, access to clean water, distribution of bed nets for protection against
malaria, and simple oral rehydration solutions are not yet reaching enough people,
even though they could easily avert many deaths. The climate in Africa is certainly
favorable to the transmission of communicable diseases and rainfall variability fre-
quently raises concern over food security and water availability.
In this chapter, I use data from nine West African countries and exogenous
variation in rainfall within and across years at a detailed geographic level to examine
the impact of rainfall shocks on early child mortality. West Africa has had histor-
ically high variation in rainfall with occasional severe droughts, and those weather
1The causes of death for children under  ve in Africa are: diarrheal diseases (16%), acute
respiratory diseases (21%), malaria (26%), other infectious/parasitic diseases (11%) and neonatal
deaths (25%) (Boschi-Pinto and Shibuya (2008)).
5
shocks a ected agricultural output directly. The literature has often used rainfall
as a shock to household income, which, if positive, would bene t children?s health.2
However, rainfall creates not only variation in income, but also variation in the op-
portunity cost of time of agricultural labor. This might be particularly relevant for
women who provide farm labor but are also responsible for non-market household
production and thus may need to reallocate their time away from these tasks during
busy times of the year. If women have to work on the  eld for most of the day, they
might, for example, breastfeed infants less frequently or wean them earlier, and also
spend less time preparing nutritious meals at home than they otherwise would. All
these factors may have a negative impact on the health of young children. Therefore,
the overall impact of rainfall shocks on child health remains ultimately an empirical
question.
I  rst examine how the timing of birth of children may depend on rainfall
variations across seasons and years. Given that a woman?s opportunity cost of time
is higher during the rainy season when most agricultural activities occur, one might
expect that women would attempt to time their delivery at times when they are less
likely to be busy. I  nd that average rainfall levels, as well as transitory shocks oc-
curring in the rainy season a ect households? fertility decisions, although the impact
of transitory shocks is much smaller. One would expect this since fertility can be
timed to occur in a given season but not based on anticipated short-run  uctuations.
2The fact that rainfall leads to  uctuations in income has been established and used extensively
in the literature. Levine and Yang (2006) show that more rainfall resulted in increased rice output
in Indonesia. Du o and Udry (2004) look at how income earned by men and women from di erent
crops is spent di erently in the household, exploiting the fact that di erent crops have a di erent
sensitivity to rainfall. See also Paxson (1992) and Dercon (2004).
6
These results are similar to those found by Pitt and Sigle (1997) in that they
highlight the importance of opportunity cost of time. This paper di ers from their
study and another, similarly motivated paper by Artadi (2005) in the following way:
in both papers, variations in opportunity costs a ect child mortality only through
their e ect on the timing of births, which in uences mortality through the seasonal-
ity in income and disease. Neither paper allows the opportunity cost of a woman?s
time to have a direct impact on mortality post-birth, for example by a ecting the
length of time a mother breastfeeds her child.
I subsequently estimate how rainfall shocks a ect mortality in the  rst year
of life, where rainfall \shocks" are de ned as deviations from a 20-year long-run
average. I  nd that rainfall shocks in the  rst year after birth have an adverse e ect
on the survival of young children that were born in the rainy season, with the e ect
being largely driven by children in rural areas. While increased disease prevalence
related to seasonal rainfall levels may be one of the causes of higher mortality, part
of the result may be attributed to the  nding that positive rainfall shocks signi -
cantly reduce the time mothers spend breastfeeding during times when their value
of time is higher, such as the rainy season. To the extent that there is a trade-o 
in the allocation of time between work and child care, a positive rainfall shock (or
higher-than-usual rainfall levels) may induce women to wean their children prema-
turely, as more income could potentially be forgone.3 Early weaning may weaken
the child?s immune system and lead to higher incidence of various diseases.
3In West Africa, extended kinship networks and the prevalence of polygyny provide for an
institutional setting where the raising of young children is shared among extended family members
or older siblings (Adepoju and Oppong (1994)).
7
I explore breastfeeding as a channel that links increases in mortality to higher
opportunity costs, especially in a context where child mortality from diseases mat-
ters considerably and could potentially be averted through proper breastfeeding.
The bene ts of breastfeeding have been extensively documented in the med-
ical literature (Victora et al. (1987); Kramer et al. (2001)). Breast milk provides
necessary nutrients to the infant as well as immune bene ts against diseases until
it fully develops its own immune system. In particular, exclusive breastfeeding in
the  rst six months reduces the incidence of and mortality from respiratory, diar-
rheal and other infectious diseases.4 Children under age two who are not breastfed
have a higher risk of death, and exclusive breastfeeding has been shown to have
greater bene ts than complementary breastfeeding which combines breast milk with
other foods and liquids (Black et al. (2003), WHO (2000)).5 Therefore, as long as
the child is breastfed, he or she will drink less water and thereby reduce exposure
to water-borne diseases. This could be another avenue through which bene ts of
breastfeeding might accrue. Further evidence suggests that the positive e ects of
breastfeeding on survival extend beyond infancy to children aged 12 - 36 months
(Briend et al. (1988)). Therefore, the negative e ect of early weaning on child health
may be exacerbated in places where the incidence of water- and food-borne diseases
4Immunoglobulin (IgA), which helps prevent diarrhea, is passed on to the infant during breast-
feeding (Clemens et al. (1997)).
5Most studies in the past have relied on observational data (Victora et al. (1987); Betran et al.
(2003); Yoon et al. (1996)) and maternal-level unobservables may have biased the results. Recently,
randomized trials in the form of educational interventions have been conducted: see, for example,
Kramer et al. (2001); Morrow et al. (1999); Bhandari et al. (2003). They  nd similar bene ts of
breastfeeding in terms of reduced incidence of diseases, as well as some longer-term bene ts on
cognitive ability. Based on these  ndings, the WHO and the American Academy of Pediatrics
recommend that infants be exclusively breastfed for the  rst 6 - 12 months, with complementary
feeding up to two years of age.
8
is high, as it is in West Africa.
Estimates of the impact of rainfall shocks on the length of breastfeeding, mea-
sured in months, show that mothers respond to positive rainfall shocks after birth
by reducing the time they breastfeed their children, especially the ones born in the
rainy season. The magnitude of the e ect is even greater and statistically signi -
cant for shocks occurring in the second year of life. This is not surprising as children
would have been breastfed for at least a year by then. Early weaning in the middle
of the rainy season would be especially critical for survival since that is the time
when diseases are rampant. Alternatively, as a woman?s spare time available for
household production becomes scarce, the meals that she prepares in the limited
available time after a long day of  eld work might be of lower nutritional quality.
The remainder of the chapter is organized as follows: section 2.2.1 provides
background information on women?s role in agriculture in West African countries
and the potential consequences for child welfare; section 2.2.2 reviews the existing
literature on weather shocks and health outcomes in sub-Saharan Africa; section 2.3
outlines the empirical strategy; section 2.4 describes the data sources; section 2.5
presents the empirical results; and section 2.6 concludes.
2.2 Background and Literature Review
2.2.1 Women, Agriculture and Child Care in West Africa
Women in West Africa play a pivotal role in agriculture, providing a large
share of farm labor and making key decisions for many agricultural activities. They
9
are mainly responsible for the production of subsistence crops.6 The seasonality of
labor demand is very pronounced in this region because agriculture is mainly rain-
fed and thus farming activities are largely de ned by rainfall. The nine countries
examined in this chapter exhibit either a unimodal rainfall pattern with a short rainy
season lasting from May to September (Burkina Faso, Guinea, Mali and Niger), or a
bimodal rainfall pattern with two rainy seasons lasting from March to October, with
a short dry spell in between (Benin, C^ote d?Ivoire, Ghana, Nigeria and Togo). Since
the latter dry spell lasts in general only about two weeks, I treat these countries
as having one long rainy season for practical purposes in the empirical section.
This classi cation is also consistent with seasonal ups and downs in the demand
for agricultural labor. The lean season occurs right before the main harvest, either
around the middle or the end of the rainy season, depending on the rainfall pattern.
This is another reason why we would expect to see less births occurring around this
time.
During the rainy season, when labor requirements are high, women have to
prioritize labor and time allocation between agricultural activities and non-market
household production such as fetching water, cooking and child care. Women sow
seeds, weed, apply fertilizer and pesticides, and harvest and thresh crops. While
rural women in sub-Saharan Africa are responsible for a substantial part of food
production, access to time-saving agricultural tools and technology is limited. Some
even argue that those time-saving innovations have actually hurt women as \Tractors
6Major food crops that are grown in this region include maize, yam, cassava, rice, millet and
sorghum. Doublecropping is practiced in regions with a longer rainy season, as is intercropping in
some areas using cowpea and millet, or rice with cassava, maize or millet, for example.
10
and animal-drawn plows have been used by men to increase the acreage under
cultivation, leaving women to struggle with an increase in weeding and harvesting,
using only handheld tools" (World Bank (2009)). Therefore, the high burden of
women during rainy seasons may put constraints on how much time women can
devote to taking care of their children. More speci cally, if the woman is trying
to free up as much time as possible for  eld work during the day, she may not
 nd enough time to breastfeed a newborn infant as often as needed, for example.
Moreover, since there is little time left for food preparation, the infant may be
introduced to porridge or solid family food earlier than recommended, which may
lead to various food-borne diseases.
2.2.2 Existing Literature
This chapter builds on a literature that looks at the relationship between ad-
verse weather shocks and health outcomes, such as anthropometric indicators (e.g.,
height-for-age). Hoddinott and Kinsey (2001) and Dercon and Hoddinott (2003)
use data from Zimbabwe and  nd that exposure to drought at an early age reduces
height-for-age.7 The latter also  nds that catch-up growth is limited, that adult
women are a ected more than adult men, and that there is no di erential e ect
between boys and girls. Dercon and Krishnan (2000) examine adult Body Mass
Index (BMI), de ned as a person?s weight in kilograms divided by squared height in
meters, and  nd that the burden of shocks is borne disproportionately by women in
7Much of this literature has focused on the e ect of droughts because rainfall data at a disag-
gregated geographic level was di cult to obtain.
11
poor households. In addition, there are studies showing that weather shocks in early
life can have a permanent e ect on longer-term outcomes. Alderman et al. (2006)
show that pre-school exposure to drought results not only in lower height-for-age,
but also in delays in school enrollment and fewer grades completed. Maccini and
Yang (2009) examine the e ect of rainfall shocks around the time of birth in Indone-
sia and  nd that more rainfall leads to better health in adulthood, more completed
schooling, and higher asset accumulation for women.
The papers most closely related to this study are Pitt and Sigle (1997) and
Artadi (2005). Both papers are motivated by the seasonality of child survival and
the woman?s opportunity cost of time. Pitt and Sigle (1997) argue that in Senegal,
children are less likely to die when born in the dry season, which is also the time
when women?s opportunity cost is the lower. They estimate the e ect of rainfall up
to two years before birth on fertility and on survival until 24 months of age, and  nd
that more rainfall increases the probability that a birth occurs, resulting in an e ect
on the average season of birth and therefore the permanent quantity of surviving
children. Artadi (2005) constructs a country-level measure of the trade-o between
child survival and opportunity cost by computing the di erence in expected sur-
vival probabilities from being born in months with high and low demand for labor.
Months with higher income coincide with months with higher child survival in some
countries, while they do not coincide in others. She  nds that households residing
in countries with a high level of trade-o are more likely to give birth in months
that have, on average, a high child mortality rate.
This paper stresses the importance of the opportunity cost of time for early
12
child mortality with regards to the potential trade-o between work and child care,
by exploring breastfeeding as one channel that may lead to higher child mortality.
A small but parallel literature from developed countries identi es return to work
as a common reason for stopping breastfeeding (Lindberg (1996); Noble and The
ALSPAC Study Team (2001)). A problem in this literature is that the studies do
not account for maternal or child-level unobservables. To circumvent this problem,
Baker and Milligan (2008) exploit an increase in maternity leave entitlements in
Canada and  nd that the policy change led to an increase in the incidence and
duration of breastfeeding but had little e ect on indicators of child health.
2.3 Empirical Strategy
2.3.1 Fertility
While there are various factors that a household may take into account in their
family planning, my main interest is to explore how households? decisions depend
on long- or short-run variations in rainfall across months and seasons. The goal is
to estimate how rainfall, by varying the opportunity cost of women?s time, a ects
whether or not a birth occurs in a given month. To investigate this, I construct a
panel dataset which consists of observations at the woman-month level, starting in
the month the woman turns 15.8 This results in an unbalanced panel of all women
who gave birth to at least one child. The main dependent variable is then an indi-
8More than 99% of births occur after age 15.
13
cator variable for whether the woman gave birth in each month.9
The distribution of births by month is shown in Figure 2.1, for countries ex-
hibiting a single short rainy season and for countries that have two rainy seasons.
The dark-colored bars in the graphs correspond to rainy season months and the
light-colored bars to dry season months. The fraction of births occurring in each
month is quite di erent from the fraction that would correspond to a uniform dis-
tribution of births across months. While the timing of birth may be in uenced by
several factors, I focus on variation due to the fact that the opportunity cost of
women?s time is lower during months in the dry season, since the time-consuming
tasks of planting and harvest are over. More time is available to give birth, recover
from pregnancy and care for children. Furthermore, diseases from which parents
may want to shield their newly born children are more prevalent during the rainy
season. Food availability also varies across seasons, as food is likely to be more
plentiful starting at the end of the rainy season after harvests have occurred. This
should provide better nutrition for mothers and their children. The lean season,
which is when households run out of food, usually occurs right before the harvest
and women may want to avoid childbirth around that time of the year.
Whether a woman gives birth in a given month is likely a ected by the above
considerations and also by factors that a ect her nutrition before conception and
during pregnancy. The latter may a ect whether a child is carried to term. I there-
fore model the impact of rainfall on the probability that a woman gives birth during
9For example, if a woman is 20 years and 2 months old at the time of survey, there are 4 x 12
+ 2 = 50 observations, one for each month since the month she turned 15.
14
month t using the following equation:
Birthjt =  + Mj + 1RainConcepjt + 2RainPregjt + Rainyjt
+[ 01RainConcepjt + 02RainPregjt]  Rainyjt
+[ 01LongConcepjt + 02LongPregjt]  Rainyjt
+ 1LongConcepjt + 2LongPregjt + Countryj + Yearj + jt
(2.1)
where the subscript j indexes the mother and t = 1;2;:::Tj, where Tj is equal to
the number of months since the woman?s 15th birthday. The dependent variable
Birthjt is a dummy that takes on the value one if woman j gave birth in month t.
The rainfall variables used in this regression are constructed in the following
way. For each month of observation t, I  rst construct long-run rainfall variables for
the period corresponding to the months before conception (speci cally, t 10, t 11
and t 12) and during pregnancy (t 1 through t 9). So for example, for the month
of October 1970, two long-run rainfall variables are created, one corresponding to
the 20-year average of October through December, to measure the persistent trends
of rainfall of the months leading up to conception, and another long-run rainfall
variable that corresponds to the 20-year average of January through September, to
measure long-running trends in the months during pregnancy. In contrast to these,
rainfall \shock" variables are created that measure any deviations from these trends:
RainConcepjt, which measures the deviation of rainfall in the three months before
the child was conceived and RainPregjt, which measures the deviation of rainfall
15
in the nine months of pregnancy.10
Demographic controls include mother?s characteristics, such as mother?s age
at birth (a linear and a quadratic term), education (dummies equal to one if mother
received either primary education or secondary education or above), and dummies
for whether place of residence is urban. Rainyjt is a country-speci c dummy variable
that indicates whether each month of observation falls in the rainy season. In
addition, all rainfall variables are interacted with Rainyi to examine whether any
e ects of rainfall shocks di er by season of birth. Countryj and Yearj are vectors
of country and year dummies that are intended to control for any country- or time-
speci c factors that are not observed in the data. Models with country  xed e ects
are estimated in the baseline model, and them replaced with region- and mother
 xed e ects. Finally,  i is the usual idiosyncratic error term.11 Standard errors are
robust to heteroskedasticity and clustered at the sampling cluster level to account
for possible correlation of errors within the same sampling cluster.
2.3.2 Infant Mortality and Breastfeeding
There are at least three channels through which rainfall may a ect infant
mortality. First, rainfall has a direct e ect on the agricultural income of rural
households. Second, more rainfall is likely to increase the prevalence of infectious
10This variable would measure whether rainfall shocks during pregnancy a ect the probability
that a birth occurs. One way to interpret this would be to consider it as the e ect of rainfall shocks
on miscarriages. This is all assuming that the child was carried to full term because there is no
data available on prematurity. Only later surveys asked whether the child was born prematurely,
but the reported incidence is extremely low. Additionally, all rainfall variables are measured in
meters.
11In speci cations with mother dummies, any variable that is time-invariant is dropped from the
regression.
16
diseases (e.g., malaria). Finally, in an agricultural setting, household members?
opportunity cost of time  uctuates within and across years with varying rainfall
levels. This may have implications for child health if women need to prioritize
among di erent tasks during busy times and are not able to spend as much time
taking care of their children. For example, women may not have enough time to
breastfeed their newborn child frequent enough and may introduce the child to solid
food early. It is well known that breast milk provides immunity to infectious diseases
through the transfer of antibodies from mother to child and through improved child
nutrition. If rainfall increases the opportunity cost of time, women might be induced
to wean the children prematurely, which may have an adverse e ect on child health.
While the income channel has a positive impact on child health, the latter two exert
a negative e ect. Therefore, whether the overall e ect of rainfall on child health is
positive or negative is an empirical question and I explore breastfeeding as a possible
channel explaining the empirical relationship between rainfall and infant mortality.
The impact of rainfall on the probability that a child dies before reaching age
one and on the number of months breastfed are estimated using the following model:
Mortalityi or Breastfeedingi =  + Ci + Mi + LongRaini + Rainyi
+[ 0RainShock0;i +:::+ kRainShockk;i] + 0LongRaini  Rainyi
+[ 00RainShock0;i +:::+ 0kRainShockk;i]  Rainyi + Countryi + Yeari + i
(2.2)
17
where i indexes the child, and Ci and Mi are sets of child and mother controls. Ci
includes sex and birth order of the child, and whether the child was born in a multiple
birth. When the outcome of question is length of breastfeeding, additional dummy
variables are included that indicate whether the child was still being breastfed at
the time the survey was carried out and whether the mother stopped breastfeeding
because the child died. Similar to equation 2.1, a 20-year long-run rainfall variable
is constructed that varies by cluster, month and year, however, now there is only one
long-run variable that measures rainfall before birth. For example, for the month of
April 1980, long-run rainfall is the average rainfall from April 1960 to March 1980.
The mean of this variable, which is measured as a monthly average across 20 years,
is around 82 mm (about 3.3 inches), which would correspond to an annual rainfall
level of 984 mm (about 39.4 inches), with a standard deviation of 86 mm.
Rainfall shock variables are created as deviations from this long-run mean, and
then aggregated to represent shocks for di erent periods around birth. RainShockt;i
measures the deviation of rainfall from the long-run average, where the subscript
t = 0;:::;k indexes the year relative to the birth year of the child that the rainfall
shock occurs. The index t runs from zero (which indicates the year before birth) to
one, if outcome is infant mortality, and from zero to two, if outcome is the length of
breastfeeding, to capture rainfall shocks in the  rst and second year after birth. For
example, rainfall shock before birth is measured as the average deviation of rain-
fall in the year before birth from long-run average rainfall. Although breastfeeding
obviously starts after birth, rainfall shock before birth is also controlled for when
estimating the impact of rainfall on the length of breastfeeding because rainfall is
18
correlated with agricultural labor demand across time. That is, a positive rainfall
shock by the beginning of the season likely increases labor demand throughout the
season. Rainfall shock in the  rst and second year of life is measured similarly.12
Similarly, rainfall shock before birth is also included in the mortality regression
because rainfall a ects agricultural output and therefore income, not only in the
seasons before birth when the rainfall actually occurs, but also in the following sea-
sons, since it takes time for crops to grow during the rainy season and since harvests
last until the next season.
Rainyi is a country-speci c dummy variable that now indicates whether the
child was born in the rainy season. For example, any child born in the months be-
tween May and September in Burkina Faso would be characterized as being born in
the rainy season and be assigned a value equal to one for the Rainyi dummy. This
variable therefore varies by country and birth month now. The rest of the variables
are de ned as previously.
After estimating the baseline model, the country dummies are gradually re-
placed with region and mother dummies in the mortality regression. These ad-
ditional dummies are intended to control for any time-invariant and unobservable
country-, region- or mother-speci c characteristics that might in uence the probabil-
ity of the child surviving beyond its  rst birthday. In the breastfeeding regressions,
only one additional speci cation with region dummies is estimated. This is because
the questions on breastfeeding were asked only for children that were born within
 ve years of the survey, and most women in the sample had one or two children
12These do not correspond to calendar years and allow for more variation in the variables.
19
at most during this time frame so there is not enough variation to estimate models
with mother dummies.
2.4 Data Sources and Descriptive Statistics
The data used in this analysis come from the Demographic and Health Sur-
veys (DHS). These are nationally representative household surveys that contain a
wide range of information on demographic characteristics, and the health and nu-
trition of women and children in particular. For countries with no reliable census
data, the DHS has been used to perform indirect estimation of important indicators
of population dynamics, such as fertility and mortality. More recently, GIS data
on the household sample cluster was collected along with household data for some
surveys. I use 13 surveys from nine countries in West Africa that include data on
the geographic coordinates of each sample cluster.13 Clusters are usually census
enumeration areas, villages in rural areas or city blocks in urban areas, and DHS
cluster points were taken at the center of each cluster. The number of clusters per
country ranges from 140 to 400, and each cluster originally sampled 30 to 40 women
in rural areas and 20 to 25 women in urban areas.14 Information was collected
on the entire retrospective birth history of women aged 15 - 49, including data on
13See Table 2.12 for a complete list of surveys used. Surveys before 1990 did not collect su -
ciently detailed data on recent births to be suitable for my analysis and surveys implemented after
1999 were excluded because the particular weather data used in this chapter relies on retroactive
reporting, which requires years of a time lag until data is cleaned and becomes available for use.
14DHS GPS Manual, 2004 (Macro International Inc. (1996)). While the surveys provide na-
tionally representative statistics, they are not suited for small area estimation. Comparing urban
and rural population was an important aim in many surveys and oversampling was conducted in
countries with a low share of urban population.
20
women?s month and year of birth, education level, residential characteristics (region
of residence, whether place of residence is urban or rural), children?s month and
year of birth, sex, birth order, and importantly, their survival status. For children
born within  ve years from the time of the survey, further detailed questions re-
garding the child?s health were asked, including for how long the child had been
breastfed. Given that the individual country samples pooled together are all from
surveys collected in the 1990s and that women between age 15 and 49 were sampled,
the distribution of observations across birth years is not even and some birth years
are over-represented, as shown in Figure 2.2, by country. In addition to that, the
sample for four countries was drawn from two separate surveys.
The most important sample restriction criterion concerns linking the correct
rainfall data to each household.15 Since interest lies in the impact of rainfall before
and after birth, it would be ideal to have geographic information on the cluster of
residence at the time of birth of each child.16 However, such data are not avail-
able and this information is inferred from the question on the number of years the
household has lived in the current residence.17 Combined with data on the child?s
age, it is possible to determine whether the child was born in the current sample
cluster. This way, children who were not born in the cluster of current residence
15The sample used to estimate the impact on mortality was restricted to \un nished cases". See
Section 2.5 for details on additional sample restriction criteria.
16Child fostering is a common practice in West Africa. For example, an estimated  fth of
non-orphaned children of age 7{14 were living away from their parents in C^ote d?Ivoire in 1985.
However, this percentage is very low for children less than two years old (Ainsworth (1992)). Child
fostering could be of concern since rainfall in the cluster of residence is merged to each child, but
it is unlikely to be a problem at least for the period of interest.
17 Any women who were visitors to the household at the time of the survey were also excluded
from the sample.
21
were dropped from the sample.18
Table 2.1 provides descriptive statistics on relevant child?s and mother?s char-
acteristics and Table 2.2 presents length of breastfeeding by percentile. In the ma-
jority of cases, the child was weaned because he or she entered weaning age (60 %),
followed by the death of child (17%) or because the mother became pregnant again
(12%). Slightly more than half of the children are male (51%), which is as expected,
3.4% of them were born in a multiple birth, and average birth order is 3.55.19 The
infant mortality rate, which measures the probability that a child survives until his
or her  rst birthday, is strikingly high at 11%. Only 18.6% of women have received
any formal education (13.5% have primary education and 5.1% have secondary ed-
ucation or more), and 27.8% live in an urban area. While an average child was
breastfed for about 13 months, 6.64% were not breastfed, 23.06% were breastfed
less than 6 months, and 40.75% were breastfed for less than a year.
Rainfall data was obtained from the Global Historical Climatology Network
(GHCN-Monthly, version 2) database, which collects weather station data from
around the world. This data was checked against mislocation of weather stations
and individual values were cleaned of discontinuities and inconsistencies before being
18Selective migration might be of concern if mothers with certain characteristics are more likely
to move to sample clusters with more than average rainfall. This would result in a downward bias
of estimates if mothers with healthier children were more likely to move away from the place of
residence and therefore be dropped from the sample. Additionally, children born before 1960 are
dropped from the sample because of the low number of observations per year.
19The average number of children per mother is slightly smaller than the total fertility rate for
this region reported elsewhere. This is not surprising, given that the sample of women surveyed
were between age 15 and 49 and many of them would not have completed their lifetime fertility.
The average total fertility for the nine countries ranges from 6 to 7.5 according to UN Population
Statistics for the period of 1950-2000 (United Nations Population Division (2008)).
22
made available for use.20 Rainfall was matched to households at the sample clus-
ter level using data from the nearest weather station. The median distance to the
nearest weather station is 62 km in the birth month of the children in the sample.
2.5 Results
2.5.1 Does rainfall shift the timing of births?
If rainfall a ects the opportunity cost of a woman?s time, it may a ect the tim-
ing of births. These results are presented in Table 2.3. The sample is constructed as
described in section 2.1 and consists of approximately 8.8 million woman-month ob-
servations. Column (1) presents baseline results with country dummies, and columns
(2) and (3) add region and mother  xed e ects, respectively. Mother  xed e ects
control for persistent, unobservable di erences across mothers, such as their fecun-
dity or their ability to control the timing of conception, both factors that are likely
to a ect their fertility outcome. In addition to reporting the coe cient estimates of
the regressors, test statistics for the impact of each variable in the rainy season are
also reported at the bottom of the table.21
The results on estimated coe cients on mother?s characteristics are not sur-
prising: women residing in urban areas or who have achieved any formal education
are less likely to give birth in any given month, and the age-fertility pro le is in-
creasing in a decreasing way. Estimated coe cients on long-run rainfall and rainfall
20The majority of rainfall data for Africa comes from the African Historical Precipitation Data,
which contains a total of 1,239 weather stations.
21The impact of rainfall shock before conception during the rainy season is the sum of the
coe cient on Rainfall Shock before Conception and its interaction with Rainy.
23
shocks show an interesting contrast across di erent birth seasons. Higher average
monthly rainfall before conception signi cantly increases the probability of giving
birth in the dry season, whereas it has a much smaller e ect on a birth in the rainy
season. However, the latter e ect is not statistically signi cant when mother  xed
e ects are included in the model. Rainfall during pregnancy may have an e ect
on how likely it is that a pregnancy results in a birth.22 Results suggest that hav-
ing a baby in gestation during months with higher historical rainfall is more likely
in the dry season, but the opposite e ect for the rainy season is not statistically
signi cant.23 Rainfall shocks before conception have a very small but statistically
signi cant negative impact on the probability of a birth during the rainy season,
but no e ect on a birth during the dry season. The magnitude of the coe cient
on long-run rainfall is much larger than the coe cient on the shock variable. This
is expected, as a woman does not know what the rainfall amount will turn out to
be when she plans her pregnancy. In terms of the magnitude of the e ects, a one
standard deviation change in long-run rainfall increases the probability that a birth
occurs by 0.7% points in the dry season, whereas a rainfall shock of the same mag-
nitude increases the probability of birth by 0.03% points in the rainy season. These
are not trivial e ects considering that the mean probability of birth in any given
month is 2%.
22Technically, a woman cannot give birth for nine months following a birth, since she cannot
conceive again once she has conceived. For comparison, I estimate the same speci cations excluding
those observations. Estimation results, not reported here, are qualitatively very similar. The sum
of the coe cients on the variables measuring rainfall shocks before conception and the interaction
with the rainy dummy is even larger and statistically signi cant at 2% with mother  xed e ects.
23If a child is born in the dry (rainy) season, a large part of the gestation months falls in the
rainy (dry) season.
24
It is conceivable that extreme weather events have a very di erent e ect on
pregnancy decisions because  oods and droughts might damage and kill the crops,
thereby limiting agricultural labor demand during those times. In other words, there
might be a nonlinear e ect of rainfall on opportunity costs. In order to account for
such extreme weather events, a dummy variable is  rst de ned to indicate whether
each month had a severe shortage or excess of rainfall: Speci cally, the dummy is
equal to one if the rainfall amount of that month fell in the extreme 5% tail of
the overall distribution of rainfall. Then, variables measuring extreme weather are
de ned as the number of months that had such an event, either in the three months
pre-conception or during the months of pregnancy. Table 2.4 reports regression
results accounting for such possibly nonlinear e ects. While the number of  ood
months before conception and the number of drought months in pregnancy seem
to matter even in the speci cation accounting for mother-level unobservables, the
coe cient estimates on long-run rainfall and rainfall shocks are remarkably sim-
ilar, which indicates that the probability of births is a ected very little by such
weather events. A one standard deviation change in long-run rainfall increases the
probability of birth by 0.7% in the dry season and the same magnitude of rainfall
shocks decrease the probability of birth by 0.07%. These estimates are statistically
signi cant at around 1%.
25
2.5.2 Reduced-form Impact on Infant Mortality
The sample for the infant mortality regression excludes any children that were
born less than one year from the time of survey, whether alive or not. This leaves
the  nal sample with 217,303 observations. The number of observations for each
survey is listed in the appendix.
Table 2.5 presents estimation results for the infant mortality model in equation
2.2. The signs of the coe cients on the mother?s and child?s control variables are
as expected: the child is less likely to reach his  rst birthday if the child is male, is
born in a multiple birth, is of higher birth order, and lives in a rural cluster. The
younger the mother, the lower the probability that the child dies before age one,
but the relationship bottoms out at age 35. Additionally, the more education the
mother has received, the more likely it is that the child survives.
The coe cient estimates on rainfall shocks one year before and after birth
are not signi cant in the baseline regression with country dummies (column(1)),
nor is long-run rainfall. Region dummies are added in column (2) to account for
unobservable factors that are time-invariant and persistent within the region. Fi-
nally, mother dummies are included in the results in the last column. I focus on
these results since they are the most robust ones. Models with mother dummies
capture mother-level unobservables to the degree that they are time-invariant: for
example, some mothers may have a better ability to shield their children from the
consequences of negative rainfall shocks if they have better health knowledge. These
estimates rely solely on the within-mother variation and thus the e ect of variables
26
that are constant within the mother (e.g., education) cannot be estimated anymore.
The results in column (3) show that while rainfall shocks before birth have a
statistically insigni cant e ect on survival, rainfall shocks during the  rst year after
birth negatively a ect survival of children born in the rainy season (as implied by
the statistically signi cant estimate of the sum of the variable and the interaction
with the \rainy" dummy: see F-stat at the bottom of the table). The coe cient
estimate suggests that conditional on other control variables, a positive one stan-
dard deviation shock in rainfall increases the probability of the child dying before
reaching age one by about 1.15% if the child was born in the rainy season. This
estimate is statistically signi cant at the 10% level in the speci cation with mother
 xed e ects.
The same sensitivity check that was performed in the previous section is done
again to account for potentially nonlinear e ects of rainfall. Table 2.6 reports these
results for infant mortality. If anything, the e ect of rainfall shocks on infant mor-
tality are slightly greater after accounting for extreme weather events.
The fact that rainfall shocks have adverse e ects on the survival of young
children according to these results may seem surprising at  rst, given that higher
rainfall is in general associated with more income, and therefore, one might expect
positive shocks to bene t the survival of children.24 However, more rainfall could
increase mortality due to the fact that parents have less time to take care of chil-
dren or because diseases such as malaria are more prevalent during the rainy season.
24Although not reported here, conditional on surviving the  rst year of life, rainfall shocks
actually decrease mortality of children born in the rainy season, but this result is likely driven by
selection.
27
While I am unable to eliminate the second channel as a possible cause, I explore
whether there is any evidence of the  rst in my data by looking at the impact of
rainfall shocks on breastfeeding decisions.
2.5.3 Rainfall shocks, Opportunity cost and Breastfeeding
The sample of children that reports length of breastfeeding has 51,912 obser-
vations, after dropping those that currently live in a residence di erent from the one
at the time of birth and those with missing breastfeeding information. Some moth-
ers report to have breastfed up to 5 years, and I remove those outliers. This leaves
a total of 49,161 observations. Estimation results on the length of breastfeeding are
shown in Table 2.7. The child is likely to be breastfed longer if the child is a girl,
born in a singleton birth, of lower birth order, and if the mother is younger, resides
in a rural area, and has no education.25 Rainfall shocks before birth are positively
associated with how long the child is breastfed for children born in the rainy sea-
son, which could possibly re ect an income e ect. The signs of the coe cients of
the rainfall variables are overall negative, and while not all of the coe cients on
individual interaction terms are signi cant, the sum of the variable with the inter-
action term is signi cant at the 1% level for rainfall shocks occurring after birth, as
reported at the bottom of the table. These results suggest that a positive deviation
of rainfall after birth consistently decreases the length of time the child is breastfed.
Regardless of season of birth, the e ect is greater for shocks occurring in the second
25In developing countries, it is common for less educated mothers to breastfeed longer (Cleland
and van Ginneken (1988)). There is no clear consensus on the urban-rural di erence in breastfeed-
ing practices, with some studies suggesting that rural women tend to breastfeed longer.
28
year of life relative to shocks in the  rst year. The e ect is also greater for children
born in the rainy season, which is as expected. This is true for shocks during the
second year but not during the  rst. A shock in the magnitude of one standard
deviation in the  rst year decreases the length of breastfeeding by a statistically
signi cant 30 days, regardless of the birth season.26
The consequences of reduced breastfeeding for mortality could be manifested
in several ways. It is possible that early weaning weakens immunity, or that sup-
plemental foods or liquids are not nutritious enough. On the other hand, if women
are restricted in their time they can devote to non-market household production,
they may not have enough time to prepare a separate meal for the newborn and
be provided with family food instead. The child might be exposed to food-borne
diseases during that process of premature transition from breast milk to solid food.
From a broader perspective, reduced breastfeeding during critical early childhood
might be representative of less time for \child care", which may a ect marginally
sick children.
It is well known that breastfeeding can be an e ective, although imperfect,
method of contraception. It appears that rainfall shocks before birth actually in-
crease the time children are breastfed if they are born in the rainy season. On the
other hand, rainfall shocks in the  rst year of life reduce the length of breastfeeding
for children born in the dry season, and reduce it even more for children born in
26The reason why the e ect of breastfeeding on mortality is not estimated directly is because
of obvious endogeneity issues. That is, there are mother-speci c unobservables that a ect both
the length of breastfeeding and mortality and therefore, I only examine the impact of exogenous
rainfall shocks on breastfeeding. Alternatively, to use rainfall as an instrument for breastfeeding
would likely violate the exclusion restriction in the second stage.
29
the rainy season. If rainfall induces women to wean children early, their fertility
may return sooner, resulting in another pregnancy, and this is consistent with what
the fertility results show. So whether this is a result of a shift in timing due to
fertility returning earlier because of weaning, or whether it is the direct consequence
of conscious timing decisions, the implication that opportunity costs in uence the
timing of births still stands and is consistent with both explanations.
2.5.4 Falsi cation Test with Neonatal Mortality
In the previous section, I argue that rainfall might adversely a ect the survival
of infant children through a post-birth channel, which is the length of time the child
is breastfed. To strengthen the case for post-birth channels leading to mortality,
I perform a type of falsi cation test by looking at how rainfall a ects neonatal
mortality (mortality before reaching one month of age). If rainfall has an e ect on
infant mortality through factors such as low birth weight or maternal malnutrition,
both of which are determined by conditions during pregnancy, then pre-birth factors
might also in uence infant mortality. Since detrimental in utero conditions are more
likely to lead to death within one month of life rather than thereafter, I examine
whether rainfall shocks right around birth a ect neonatal mortality. Results are
reported in Table 2.8, and across di erent speci cations, including one that controls
for mother  xed e ects, rainfall does not have a statistically signi cant e ect on
mortality in the  rst month of life when the child is born in the rainy season. For
children born in the dry season, there is a statistically signi cant e ect of rainfall
30
shock in the  rst month in columns (1) and (2) for speci cations with country and
region  xed e ects. However, the estimate turns insigni cant with the inclusion
of mother  xed e ects. This suggests that it is post-birth rather than pre-birth
channels that result in increased infant mortality.
2.5.5 Further results by urban{rural
Considering that variations in the opportunity cost of time of agricultural labor
created by  uctuations in rainfall is more relevant to rural households, it is natural
to expect households in rural areas to respond more strongly than households in
urban areas. To see if this is indeed true, the same regressions are run on subsamples
divided by urban or rural area of residence. While the majority of household income
in rural areas is generated by farm activities,27 there are at least two reasons why
urban areas may not be entirely free of seasonalities. First, women residing in urban
areas may engage in agriculture as a secondary source of income. Because of growing
food insecurity, urban agriculture has spread rapidly over the years.28 Moreover,
manufacturing industries that rely on the seasonal supply of raw material might
also experience  uctuations in opportunity cost.29 Therefore, women in urban areas
may still see variations in their opportunity cost in response to rainfall, although at
a smaller scale compared to women in rural areas.
Estimation results for the split sample are shown in Tables 2.9, 2.10, and 2.11,
for fertility, infant mortality, and breastfeeding, respectively. The speci cations are
27For example, in C^ote d?Ivoire, 75% of income came from farming (Kozel (1990)).
28Gardening vegetables or raising micro livestock has become more widespread in Ghana (Danso
et al. (2004)).
29The production of sorghum beer in Burkina Faso is one example.
31
the same as before. The impact on fertility is indeed greater in rural than urban
areas, although it seems to persist in both (Table 2.9). The e ect of rainfall shocks
on infant mortality is also mainly driven by rural children and the estimate is even
larger in magnitude than before. The results for breastfeeding are mixed: rainfall
shocks in the  rst year negatively impact the length of breastfeeding of rural but
not urban children who are born in the dry season. However, the opposite seems to
hold for children born in the rainy season, as the estimates are actually smaller for
rural children.
2.6 Conclusion
Households in developing countries are a ected by various types of shocks and
it is important to understand how they respond to them. Rainfall shocks are of
particular interest given the high dependence on agriculture in many developing
countries. Rainfall in the development literature has often been associated with
positive outcomes given its positive e ect on income. This chapter suggests that
rainfall shocks can actually have adverse e ects on the survival of young children
through their impact on the opportunity cost of women across seasons and over
time. While this result may come as a surprise, it may be partly attributed to the
impact on the length of breastfeeding. More rainfall increases the mother?s oppor-
tunity cost of time, which reduces the time children are breastfed and may weaken
their immune system. The importance of opportunity costs is further substantiated
by exploring the e ect on the timing of birth: long-run rainfall increases the prob-
32
ability of giving birth in the dry season, whereas rainfall shocks have a small but
statistically signi cant negative e ect on the probability of giving birth in the rainy
season. This suggests that women are more likely in uenced by long-run variations
in rainfall when making their fertility decisions. Altogether, these results highlight
the importance of the opportunity cost of time. More work is warranted regarding
how changes in the opportunity costs of parents a ect the well-being of children.
33
Table 2.1: Descriptive Statistics (All households)
Mother?s characteristics Child?s characteristics
Age at birth 25.368 Male 0.51
(6.508) (0.5)
Primary education 0.135 Birth Order 3.547
(0.342) (2.30)
Secondary education or higher 0.051 Multiple birth 0.034
(0.220) (0.181)
Urban 0.278 Infant mortality rate 0.11
(0.448) (0.313)
Number of women 55,278 Number of children 217,293
Table 2.2: Breastfeeding in months
Mean (St. Dev) 13.06 (8.866)
Missing 1.42%
Zero months 6.64%
0 { 6 months 23.06%
0 { 12 months 40.75%
Observations 49,161
34
Table 2.3: Fertility
(1) (2) (3)
Mother?s age at birth 0.0039*** 0.0039*** 0.0024***
(0.0000) (0.0000) (0.0002)
Mother?s age at birth squared -0.0001*** -0.0001*** -0.0001***
(0.0000) (0.0000) (0.0000)
Urban -0.0028*** -0.0023***
(0.0001) (0.0002)
Primary education -0.0015*** -0.0012***
(0.0001) (0.0001)
Secondary education or higher -0.0051*** -0.0048***
(0.0002) (0.0002)
Rainfall shock before conception 0.0073 0.0063 0.0036
(0.0050) (0.0050) (0.0050)
Rainfall shock in pregnancy 0.0075** 0.0053 0.0041
(0.0036) (0.0036) (0.0037)
Long-run rainfall before conception 0.0835*** 0.0891*** 0.0805***
(0.0035) (0.0038) (0.0053)
Long-run rainfall in pregnancy -0.0200*** -0.0207*** -0.0251**
(0.0033) (0.0040) (0.0123)
Rainy 0.0009** 0.0010*** 0.0013***
(0.0004) (0.0004) (0.0004)
Interaction with Rainy
Rainfall shock before conception -0.0098* -0.0097* -0.0075
(0.0054) (0.0054) (0.0054)
Rainfall shock in pregnancy -0.0027 -0.0034 -0.0037
(0.0056) (0.0056) (0.0057)
Long-run rainfall before conception -0.0784*** -0.0845*** -0.0795***
(0.0038) (0.0039) (0.0039)
Long-run rainfall in pregnancy 0.0191*** 0.0209*** 0.0201***
(0.0034) (0.0034) (0.0035)
Constant -0.0262*** -0.0271*** -0.0335***
(0.0016) (0.0016) (0.0022)
Test for signi cance of variables
on births in the rainy season
Rainfall before conception
Sum of coe cients -0.0025 -0.0034 -0.0039
F-stat 1.93 3.48 4.19
Corresponding p-value 0.1648 0.0620 0.0407
Rainfall in pregnancy
Sum of coe cients 0.0048 0.0019 0.0004
F-stat 1.42 0.22 0.01
Corresponding p-value 0.2339 0.6389 0.9338
Long-run rainfall before conception
Sum of coe cients 0.0051 0.0046 0.001
35
F-stat 11.44 7.46 0.05
Corresponding p-value 0.0007 0.0063 0.8155
Long-run rainfall in pregnancy
Sum of coe cients -0.0009 0.0002 -0.005
F-stat 0.09 0.00 0.16
Corresponding p-value 0.7691 0.9499 0.6858
Country dummies x
Region dummies x
Mother dummies x
Observations 8831987 8831987 8831987
R2 0.0018 0.0018 0.0059
Note: This table reports linear probability model estimates for the e ect of rainfall shocks
on the probability of birth. Reported birth histories are used to construct a new sample
that consists of one observation for each woman for each month starting when she turns
15. The dependent variable is equal to one if the woman gave birth to a child in that
month and the unit of observation is a woman-month. Control variables include urban
(dummy equal to one if place of residence is in an urban area), education (dummies equal
to one if mother received either primary education or secondary education and above), and
mother?s age in each month. Sample is from 13 surveys from nine countries in West Africa.
Long-run rainfall is the 20-year average rainfall in the sample cluster and varies by month
and year. All rainfall shock variables are measured as an average relative to the month
of observation, t. Rainy is a dummy variable that takes on the value one if the month
of observation falls in the rainy season and zero otherwise, and varies by country and
month. All regressions include year and month dummies. The urban dummy is dropped
in columns (3) and (4) and mother?s education is dropped in column (4). Standard errors
are robust to heteroskedasticity and clustered at the sampling cluster level.
* signi cant at 10%; ** signi cant at 5%; *** signi cant at 1%.
36
Table 2.4: Fertility results with nonlinearity
(1) (2) (3)
Mother?s age at birth 0.0039*** 0.0039*** 0.0024***
(0.0000) (0.0000) (0.0002)
Mother?s age at birth squared -0.0001*** -0.0001*** -0.0001***
(0.0000) (0.0000) (0.0000)
Urban -0.0028*** -0.0022***
(0.0001) (0.0002)
Primary education -0.0015*** -0.0012***
(0.0001) (0.0001)
Secondary education or higher -0.0051*** -0.0048***
(0.0002) (0.0002)
Rainfall shock before conception 0.0035 0.0023 -0.0002
(0.0053) (0.0053) (0.0053)
Rainfall shock in pregnancy 0.0120*** 0.0095** 0.0084**
(0.0039) (0.0039) (0.0040)
Long-run mean rainfall before conception 0.0835*** 0.0892*** 0.0811***
(0.0035) (0.0038) (0.0053)
Long-run mean rainfall in pregnancy -0.0195*** -0.0200*** -0.0232*
(0.0033) (0.0040) (0.0123)
Rainy 0.0010** 0.0011*** 0.0014***
(0.0004) (0.0004) (0.0004)
# of  ood months before conception 0.0005** 0.0005** 0.0004*
(0.0002) (0.0002) (0.0002)
# of  ood months in pregnancy -0.0002** -0.0002 -0.0002
(0.0001) (0.0001) (0.0001)
# of drought months before conception -0.0001 -0.0001 -0.0002
(0.0002) (0.0002) (0.0002)
# of drought months in pregnancy 0.0003*** 0.0003*** 0.0003**
(0.0001) (0.0001) (0.0001)
Interaction with Rainy
Rainfall shock before conception -0.0078 -0.0076 -0.0056
(0.0055) (0.0054) (0.0055)
Rainfall shock in pregnancy -0.0027 -0.0034 -0.0037
(0.0056) (0.0056) (0.0057)
Long-run mean rainfall before conception -0.0785*** -0.0846*** -0.0797***
(0.0038) (0.0039) (0.0039)
Long-run mean rainfall in pregnancy 0.0188*** 0.0206*** 0.0197***
(0.0034) (0.0034) (0.0035)
Constant -0.0370*** -0.0377*** -0.0518***
(0.0027) (0.0027) (0.0030)
Test for signi cance of variablesn
on births in the rainy season
Rainfall before conception
F-stat 3.91 6.11 6.92
37
Corresponding p-value 0.0482 0.0135 0.0086
Rainfall in pregnancy
F-stat 4.31 1.79 0.98
Corresponding p-value 0.0379 0.1806 0.3227
Long-run rainfall before conception
F-stat 11.06 7.42 0.11
Corresponding p-value 0.0009 0.0065 0.7418
Long-run rainfall in pregnancy
F-stat 0.06 0.02 0.08
Corresponding p-value 0.8045 0.8819 0.7810
Country dummies x
Region dummies x
Mother dummies x
Observations 8842131 8842131 8842131
R2 0.0018 0.0018 0.0059
38
Table 2.5: Infant mortality
(1) (2) (3)
Male 0.0145*** 0.0146*** 0.0134***
(0.0015) (0.0015) (0.0020)
Multiple birth 0.2442*** 0.2454*** 0.2616***
(0.0073) (0.0073) (0.0093)
Birth Order 0.0134*** 0.0133*** 0.0232***
(0.0008) (0.0007) (0.0020)
Mother?s age at birth -0.0140*** -0.0139*** -0.0049
(0.0011) (0.0011) (0.0031)
Mother?s age at birth squared 0.0002*** 0.0002*** 0.0003***
(0.0000) (0.0000) (0.0000)
Urban -0.0309*** -0.0259***
(0.0022) (0.0026)
Primary education -0.0141*** -0.0100***
(0.0026) (0.0025)
Secondary education or higher -0.0219*** -0.0164***
(0.0033) (0.0033)
Rainfall shock before birth -0.0823 -0.0439 0.0012
(0.0698) (0.0710) (0.0954)
Rainfall shock in  rst year -0.0765 -0.0443 -0.0518
(0.0655) (0.0663) (0.0883)
Long-run rainfall 0.0005 0.0059 -0.0053
(0.0045) (0.0068) (0.0295)
Rainy 0.0039 0.0024 0.0008
(0.0037) (0.0037) (0.0049)
Interaction with Rainy
Rainfall shock before birth 0.1068 0.0830 0.0794
(0.0916) (0.0916) (0.1210)
Rainfall shock in  rst year 0.1256 0.1127 0.1857*
(0.0853) (0.0857) (0.1132)
Long-run rainfall -0.0002 0.0009 0.0007
(0.0033) (0.0034) (0.0044)
Constant 0.1502 0.1461 -0.1194
(0.1165) (0.1174) (0.1060)
Test for signi cance of variables
on mortality in the rainy season
Rainfall before birth
F-stat 0.1397 0.3498 0.9868
Corresponding p-value 0.7086 0.5543 0.3206
Rainfall in  rst year
F-stat 0.6336 1.1660 2.8402
Corresponding p-value 0.4261 0.2803 0.0920
Long-run rainfall
F-stat 0.0067 1.0005 0.0253
39
Corresponding p-value 0.9349 0.3172 0.8737
Country dummies x
Region dummies x
Mother dummies x
Observations 217293 217293 217293
R2 0.0370 0.0388 0.3104
Note: This table reports linear probability model estimates for infant mortality (mortality
before age one). The dependent variable is a dummy variable equal to one if the child died
before reaching its  rst birthday. Control variables include male (dummy equal to one
if child is male), a dummy for whether the child was born in a multiple birth, the birth
order of the child, urban (dummy equal to one if place of residence is in an urban area),
education (dummies equal to one if mother received either primary education or secondary
education and above), and mother?s age at birth. Sample is from 13 surveys from nine
countries in West Africa. Birth year ranges from 1950 to 1999. Long-run rainfall is the
20-year average rainfall in the sample cluster and varies by month and year. Rainfall
shock before birth is the deviation of average rainfall in the 12 months prior to birth
from long-run rainfall. Rainfall shock in  rst year is the deviation of rainfall during the
 rst year of life from the long-run average. Rainy is a dummy variable that takes on the
value one if the child was born in the rainy season and varies by country and month. All
regressions include year and month dummies. The urban dummy is dropped in columns
(3) and (4) and mother?s education is dropped in column (4). Standard errors are robust
to heteroskedasticity and clustered at the sampling cluster level.
* signi cant at 10%; ** signi cant at 5%; *** signi cant at 1%.
40
Table 2.6: Infant mortality results with nonlinearity
(1) (2) (3)
Male 0.0145*** 0.0146*** 0.0134***
(0.0015) (0.0015) (0.0020)
Multiple birth 0.2442*** 0.2454*** 0.2616***
(0.0073) (0.0073) (0.0093)
Birth Order 0.0134*** 0.0133*** 0.0233***
(0.0008) (0.0007) (0.0020)
Mother?s age at birth -0.0140*** -0.0139*** -0.0049
(0.0011) (0.0011) (0.0031)
Mother?s age at birth squared 0.0002*** 0.0002*** 0.0003***
(0.0000) (0.0000) (0.0000)
Urban -0.0310*** -0.0259***
(0.0022) (0.0026)
Primary education -0.0140*** -0.0100***
(0.0026) (0.0025)
Secondary education or higher -0.0219*** -0.0164***
(0.0033) (0.0033)
Rainfall shock before birth -0.1073 -0.0606 0.0235
(0.0778) (0.0793) (0.1045)
Rainfall shock in  rst year -0.0724 -0.0337 -0.0436
(0.0711) (0.0717) (0.0967)
Long-run rainfall 0.0002 0.0058 -0.0036
(0.0045) (0.0069) (0.0296)
Rainy 0.0039 0.0024 0.0009
(0.0037) (0.0037) (0.0049)
# months with excess rainfall before birth 0.0003 0.0005 -0.0012
(0.0016) (0.0016) (0.0020)
# of  ood months in dry season at age 1 -0.0005 -0.0003 -0.0002
(0.0016) (0.0016) (0.0021)
# months with rainfall shortage before birth -0.0021 -0.0007 0.0003
(0.0014) (0.0014) (0.0018)
# of drought months in rainy season at age 1 -0.0009 0.0005 0.0004
(0.0014) (0.0014) (0.0018)
Interaction with Rainy
Rainfall shock before birth 0.1075 0.0839 0.0782
(0.0916) (0.0917) (0.1211)
Rainfall shock in  rst year 0.1248 0.1119 0.1865*
(0.0853) (0.0858) (0.1133)
Long-run rainfall -0.0001 0.0009 0.0007
(0.0033) (0.0034) (0.0044)
Constant 0.1512 0.1461 -0.1207
(0.1165) (0.1174) (0.1060)
Test for signi cance of variables
on mortality in the rainy season
41
Rainfall before birth
F-stat 0.1397 0.3498 0.9868
Corresponding p-value 0.7086 0.5543 0.3206
Rainfall in  rst year
F-stat 0.5912 1.2417 2.5397
Corresponding p-value 0.4420 0.2652 0.1111
Long-run rainfall
F-stat 0.0004 0.9557 0.0097
Corresponding p-value 0.9845 0.3283 0.9214
Country dummies x
Region dummies x
Mother dummies x
Observations 217293 217293 217293
R2 0.0371 0.0388 0.3104
42
Table 2.7: Breastfeeding
(1) (2)
Male -0.1298* -0.1277*
(0.0665) (0.0663)
Multiple birth -1.4207*** -1.3796***
(0.3271) (0.3299)
Birth Order -1.8533*** -1.8789***
(0.1039) (0.1032)
Mother?s age at birth 0.1837*** 0.1986***
(0.0387) (0.0386)
Mother?s age at birth squared -0.0029*** -0.0031***
(0.0007) (0.0007)
Urban -0.6078*** -0.5512***
(0.0961) (0.1059)
Primary education -0.6305*** -0.5153***
(0.1028) (0.1061)
Secondary education or higher -1.0437*** -0.8355***
(0.1339) (0.1357)
Still breastfeeding -3.6512*** -3.6375***
(0.1432) (0.1427)
Died while breastfeeding -9.6772*** -9.6947***
(0.1586) (0.1609)
Rainfall shock before birth 1.4630 1.4199
(3.5185) (3.6820)
Rainfall shock in  rst year -11.0923*** -11.1762***
(3.7720) (3.8478)
Rainfall shock in second year -34.1548*** -35.3519***
(3.7914) (3.8383)
Long-run rainfall -7.7195*** -10.5555***
(2.2563) (2.9872)
Rainy 0.7030*** 0.5701***
(0.1607) (0.1583)
Interaction: Rainy
Rainfall shock before birth 5.7832 5.4669
(3.6932) (3.7128)
Rainfall shock in  rst year -0.5200 -0.5408
(4.3340) (4.3546)
Rainfall shock in second year -10.1392** -8.3804**
(4.2024) (4.1757)
Long-run rainfall -9.5172*** -7.9672***
(1.8349) (1.7904)
Constant 23.6542*** 25.1204***
(0.7517) (0.7662)
Country dummies x
Region dummies x
43
Observations 49161 49161
R2 0.3371 0.3447
Test for signi cance of variables
on breastfeeding in the rainy season
Rainfall shock before birth
F-stat 4.54 3.75
Corresponding p-value 0.0332 0.0529
Rainfall shock in  rst year
F-stat 11.63 10.82
Corresponding p-value 0.0007 0.0010
Rainfall shock in second year
F-stat 138.93 129.25
Corresponding p-value 0.0000 0.0000
Long-run rainfall
F-stat 81.78 47.59
Corresponding p-value 0.0000 0.0000
Note: This table reports ordinary least squares model estimates for the length of breast-
feeding of the child. The dependent variable is the number of months the child was
breastfed. Control variables include male (dummy equal to one if child is male), whether
the child was born in a multiple birth, the birth order of the child, urban (dummy equal to
one if place of residence is in an urban area), education (dummies equal to one if mother
received either primary education or secondary education and above), and mother?s age
at birth. Sample is from 13 surveys from nine countries in West Africa. Birth year ranges
from 1986 to 1999. Long-run rainfall is the 20-year average rainfall in the sample cluster
and varies by month and year. Rainfall shock before birth is the deviation of average
rainfall in the 12 months prior to birth from long-run rainfall. Rainfall shock in  rst year
is the deviation of rainfall during the  rst year of life from the long-run average. Rainy is
a dummy variable that takes on the value one if the child was born in the rainy season and
varies by country and month. All regressions include year and month dummies. Standard
errors are robust to heteroskedasticity and clustered at the sampling cluster level.
* signi cant at 10%; ** signi cant at 5%; *** signi cant at 1%.
44
Table 2.8: Neonatal mortality
(1) (2) (3)
Male 0.0132*** 0.0132*** 0.0140***
(0.0011) (0.0011) (0.0015)
Multiple birth 0.1754*** 0.1755*** 0.1878***
(0.0068) (0.0067) (0.0080)
Birth Order 0.0068*** 0.0068*** 0.0151***
(0.0005) (0.0005) (0.0014)
Mother?s age at birth -0.0110*** -0.0110*** -0.0082***
(0.0008) (0.0008) (0.0025)
Mother?s age at birth squared 0.0002*** 0.0002*** 0.0002***
(0.0000) (0.0000) (0.0000)
Urban -0.0114*** -0.0088***
(0.0015) (0.0018)
Primary education -0.0061*** -0.0046**
(0.0018) (0.0018)
Secondary education or higher -0.0115*** -0.0096***
(0.0023) (0.0023)
Rainfall shock before birth -0.0721 -0.0475 -0.0590
(0.0450) (0.0458) (0.0619)
Rainfall shock in  rst month 0.0607* 0.0668* 0.0349
(0.0348) (0.0348) (0.0464)
Long-run rainfall -0.0005 0.0047 -0.0218
(0.0030) (0.0044) (0.0231)
Rainy 0.0012 0.0006 -0.0020
(0.0026) (0.0026) (0.0032)
Interaction with Rainy
Rainfall shock before birth 0.0380 0.0276 0.0614
(0.0597) (0.0600) (0.0802)
Rainfall shock in  rst month -0.0479 -0.0522 -0.0243
(0.0373) (0.0373) (0.0488)
Long-run rainfall 0.0003 0.0008 0.0030
(0.0023) (0.0023) (0.0029)
Constant 0.1686 0.1696 0.0641
(0.1151) (0.1159) (0.0923)
Test for signi cance of variables
on neonatal mortality in the rainy season
Rainfall before birth
F-stat 0.5905 0.1929 0.0017
Corresponding p-value 0.4423 0.6605 0.9670
Rainfall in  rst month
F-stat 1.3573 1.7514 0.5172
Corresponding p-value 0.2441 0.1858 0.4721
Long-run rainfall
F-stat 0.0037 1.5618 0.6642
45
Corresponding p-value 0.9512 0.2115 0.4152
Country dummies x
Region dummies x
Mother dummies x
Observations 232170 232170 232170
R2 0.0303 0.0312 0.3120
46
Table 2.9: Fertility by urban{rural with mother  xed e ects
(1) (2)
Urban Rural
Mother?s age at birth 0.0026*** 0.0023***
(0.0003) (0.0003)
Mother?s age at birth squared -0.0001*** -0.0001***
(0.0000) (0.0000)
Rainfall shock before conception 0.0011 0.0046
(0.0074) (0.0064)
Rainfall shock in pregnancy -0.0054 0.0065
(0.0058) (0.0045)
Long-run mean rainfall before conception 0.0543*** 0.0846***
(0.0083) (0.0063)
Long-run mean rainfall in pregnancy -0.0180 -0.0411***
(0.0218) (0.0144)
Rainy 0.0019*** 0.0011**
(0.0006) (0.0005)
Interaction with Rainy
Rainfall shock before conception -0.0076 -0.0077
(0.0078) (0.0069)
Rainfall shock in pregnancy 0.0020 -0.0055
(0.0084) (0.0072)
Long-run rainfall before conception -0.0567*** -0.0869***
(0.0060) (0.0048)
Long-run rainfall in pregnancy 0.0108** 0.0226***
(0.0047) (0.0042)
Constant -0.0592*** -0.0479***
(0.0049) (0.0035)
Test for signi cance of variables
on births in the rainy season
Rainfall before conception
F-stat 6.22 1.61
Corresponding p-value 0.0128 0.2052
Rainfall in pregnancy
F-stat 0.28 0.03
Corresponding p-value 0.5953 0.8587
Long-run rainfall before conception
F-stat 0.10 0.22
Corresponding p-value 0.75337 0.6385
Long-run rainfall in pregnancy
F-stat 0.11 1.62
Corresponding p-value 0.7428 0.2029
Observations 2776272 6065859
R2 0.0068 0.0056
47
Table 2.10: Infant mortality by urban{rural with mother
 xed e ects
(1) (2)
Urban Rural
Male 0.0133*** 0.0134***
(0.0035) (0.0023)
Multiple birth 0.2305*** 0.2685***
(0.0178) (0.0106)
Birth Order 0.0207*** 0.0237***
(0.0032) (0.0023)
Mother?s age at birth -0.0055* -0.0049
(0.0032) (0.0038)
Mother?s age at birth squared 0.0003*** 0.0003***
(0.0000) (0.0000)
Rainfall shock before birth 0.1912 -0.0470
(0.1752) (0.1114)
Rainfall shock in  rst year -0.0145 -0.0593
(0.1711) (0.1020)
Long-run rainfall -0.0146 0.0047
(0.0480) (0.0354)
Rainy 0.0014 0.0008
(0.0089) (0.0057)
Interaction with Rainy
Rainfall shock before birth -0.0568 0.1158
(0.1970) (0.1452)
Rainfall shock in  rst year 0.0390 0.2261*
(0.1970) (0.1339)
Long-run rainfall -0.0021 0.0013
(0.0075) (0.0052)
Constant -0.0949 -0.1287
(0.0983) (0.1310)
Test for signi cance of variables
on mortality in the rainy season
Rainfall before birth
F-stat 0.9907 0.4940
Corresponding p-value 0.3197 0.4822
Rainfall in  rst year
F-stat 0.0359 3.0547
Corresponding p-value 0.8498 0.0807
Long-run rainfall
F-stat 0.1233 0.0295
Corresponding p-value 0.7255 0.8637
Observations 60416 156877
R2 0.3353 0.3035
48
Table 2.11: Breastfeeding by urban{rural
(1) (2) (3) (4)
Urban Urban Rural Rural
Male -0.1558 -0.1459 -0.1204 -0.1153
(0.1221) (0.1198) (0.0773) (0.0773)
Multiple birth -0.5352 -0.4860 -1.7001*** -1.6865***
(0.4678) (0.4663) (0.3883) (0.3904)
Birth Order -1.4554*** -1.5206*** -1.9618*** -1.9648***
(0.1773) (0.1767) (0.1227) (0.1221)
Mother?s age at birth 0.1808*** 0.2072*** 0.1925*** 0.2036***
(0.0668) (0.0657) (0.0447) (0.0446)
Mother?s age at birth squared -0.0028** -0.0032*** -0.0030*** -0.0032***
(0.0012) (0.0012) (0.0008) (0.0008)
Primary education -0.9752*** -0.7923*** -0.5053*** -0.4150***
(0.1512) (0.1510) (0.1324) (0.1379)
Secondary education or higher -1.6711*** -1.3625*** -0.2684 -0.0833
(0.1688) (0.1724) (0.2136) (0.2163)
Still breastfeeding -4.5364*** -4.5617*** -3.3226*** -3.2910***
(0.1858) (0.1863) (0.1841) (0.1827)
Died while breastfeeding -9.6250*** -9.7105*** -9.5943*** -9.5870***
(0.3081) (0.3072) (0.1801) (0.1826)
Rainfall shock before birth 1.0973 3.8473 2.8905 2.5454
(5.6740) (5.9890) (4.1906) (4.3591)
Rainfall shock in  rst year -7.6245 -6.7117 -12.2300*** -12.2225***
(6.0057) (6.0913) (4.5167) (4.5777)
Rainfall shock in second year -22.8753*** -20.5897*** -37.7421*** -40.1386***
(5.4024) (5.3928) (4.5530) (4.6184)
Long-run rainfall -1.5222 -2.3587 -9.0577*** -13.0345***
(4.0671) (5.3582) (2.6576) (3.4080)
Rainy 1.3003*** 1.2594*** 0.5734*** 0.4107**
(0.2678) (0.2685) (0.1873) (0.1841)
Interaction: Rainy
Rainfall shock before birth 2.0624 0.3793 9.4954** 8.5722*
(6.5244) (6.4250) (4.3692) (4.4011)
Rainfall shock in  rst year -13.4866* -12.8076* 2.6124 2.0075
(7.3265) (7.2430) (5.1867) (5.2062)
Rainfall shock in second year -12.1393** -10.2075* -10.2766** -8.5823*
(6.0374) (5.9889) (5.1377) (5.1209)
Long-run rainfall -11.7794*** -11.0232*** -9.2975*** -7.4075***
(2.8664) (2.8661) (2.2015) (2.1339)
Constant 19.8516*** 20.6923*** 24.2904*** 26.0318***
(1.2084) (1.1508) (0.8852) (0.9091)
Observations 14765 14765 34396 34396
R2 0.3160 0.3273 0.3502 0.3580
Test for signi cance of variables
49
on breastfeeding in the rainy season
Rainfall before birth
F-stat 0.32 0.54 9.15 6.84
Corresponding p-value 0.5689 0.4608 0.0025 0.0090
Rainfall in  rst year
F-stat 17.08 14.09 5.25 5.52
Corresponding p-value 0.0000 0.0002 0.0221 0.0189
Rainfall in second year
F-stat 54.29 39.36 103.14 101.91
Corresponding p-value 0.0000 0.0000 0.0000 0.0000
Long-run rainfall
F-stat 15.25 7.26 68.38 47.20
Corresponding p-value 0.0001 0.0071 0.0000 0.0000
Country dummies x x
Region dummies x x
50
Table 2.12: Complete list of surveys
Country Year of survey Number of observations
Benin 1996 15,050
Burkina Faso 1992 16,376
1998 17,990
C^ote d?Ivoire 1994 16,786
1998 7,055
Ghana 1993 9,214
1998 9,799
Guinea 1999 22,138
Mali 1995 32,125
Niger 1992 19,142
1998 23,586
Nigeria 1990 24,614
Togo 1998 19,078
Total 217,303
51
52
Figure 2.1: Fraction of births by birth month, by country
Dark-colored bars correspond to rainy season months and light-colored
bars to dry season months.
53
54
Figure 2.2: Distribution of births by birth year, by country
55
Chapter 3
Health Shocks and Production Decisions of Agricultural Households
3.1 Introduction
There is an extensive literature linking health and work productivity, and thus
a farmer?s capacity to work (e.g., Strauss (1986)). However, as long as markets are
competitive, it can be shown that farm pro ts should not be impacted by supply
factors to the agricultural household, such as health shocks. Tests of this so-called
separation hypothesis have generally found support for it (see Lopez (1984) and Ben-
jamin (1992) for example). This chapter examines whether health shocks generated
by malaria infections a ect the production decisions of agricultural households.
Pitt and Rosenzweig (1986) extend the traditional consumer-cum-producer
agricultural household model to incorporate an explicit health production function
(as developed by Grossman (1972)). Their model essentially shows that with well-
functioning input and output markets, farm pro ts and the health of the farmer
are independent of each other. Using data from Indonesia, they test whether illness
episodes a ect farm pro ts and male labor supply. They  nd that while the labor
supply of males is reduced as a result of an illness, pro ts are not, which most
likely suggests that households were able to substitute family labor with hired labor,
con rming the separation between production and health.
Malaria is a disease that is recurring and thus anticipation of the disease may
involve adjustments in production decisions by the households a ected. I explore
the adjustment of households in response to malaria along two di erent margins:
First, whether households adjust their labor input, and second, whether households
choose to grow a less labor-intensive crop mix in anticipation of the disease.
56
The analysis in this chapter is complicated by the fact that the estimates may
be biased for several reasons. First, malaria incidence could be correlated with un-
observable factors, such as household-speci c managerial abilities, that are omitted
from the regression but also a ect the outcomes of interest. Second, the measure of
malaria being used may not be a good proxy for the true incidence of malaria, in
which case, measurement error would lead to an attenuation bias. Finally, there is
the possibility of reverse causality: agriculture and health are linked in a way that
causality might go in both directions. Agriculture is essential for the production of
health through the production of food, whereas bad health might reduce labor pro-
ductivity, which would further reduce income and result in further ill health (Hawkes
and Ruel (2006)). This last issue in particular is not present in some recent studies
that look at economic outcomes later in life, such as completed schooling or earn-
ings as an adult (as for example, Cutler et al. (2007) or Bleakley (2010)). While
the use of household  xed e ects regressions in some of the speci cations may deal
with the issue of household-level unobservables, it does not address the other issues
mentioned above, which call for the use of an instrumental variable. Moreover, the
attenuation bias due to measurement error is magni ed with the use of  xed e ects.
An ideal instrument in this case would be one that is a strong determinant of malaria
transmission (or in other words, that would result in a strong  rst stage regression)
and not be a direct determinant of the outcomes of interest, i.e., pro ts or labor
use. While the  rst condition is easier to satisfy, it proved to be very di cult to
 nd one that satis es the second condition as well.
The  rst candidate for an instrument for malaria incidence would be envi-
ronmental factors that govern the intensity of malaria transmission, i.e., rainfall,
temperature, humidity and altitude. Peak malaria transmission occurs during the
wet season because of its direct relationship with rainfall, which raises relative hu-
midity and promotes breeding activity of mosquitoes by increasing the breeding
57
surface. While rainfall is critical, excessive rainfall may interrupt transmission as
the breeding grounds might be  ushed away. The parasite develops best and trans-
mission is highest when mean temperature is between 20 C and 30 C. Low relative
humidity shortens the life of the mosquito and high relative humidity enhances
mosquito activity (Pampana (1963)). Finally, mosquitoes are inactive at altitudes
1500 meters and above. In Vietnam, most of the rice is grown in lowlands, while
some fruit crops are grown above certain altitudes (Food and Agriculture Organiza-
tion (1999b)). Due to their direct relationship with malaria transmission, di erent
combinations of these factors have been used as instruments to look at how early-
life exposure to malaria a ects education and wealth later in life (Barreca (2010),
Cutler et al. (2007) and Hong (2007)). Temperature is not used as an instrument
since Vietnam is geographically located in a tropical region close to the equator and
there is comparably little interannual variation in temperature. Thus transmission
is mainly in uenced by rainfall.1
Regardless of their impact on malaria transmission, these variables cannot
be used as instruments in this analysis. The outcomes of interest are agricultural
outcomes, which are directly a ected by all of these factors. Altitude is not a valid
instrument since it determines land cover and other geographical features (Hong
(2007)) and may therefore in uence which crops are grown in the high- and lowlands.
The exclusion restriction in the second stage of the instrumental variables regression
is not satis ed since these environmental factors strongly impact crop growth and
therefore, pro ts.
Another candidate, for which data is available, is province-speci c expenditure
on antimalarial drugs. These drugs were dispensed through the National Malaria
Control Program over the same period as the surveys, as part of a nationwide e ort
to control malaria. It is indeed the case that the central government allocated
1Therefore, exploiting di erences in ideal temperatures for malaria transmission and agriculture
is not an option (see Barreca (2010) who developed this approach).
58
more budget to provinces that had a higher prevalence of malaria (resulting in a
strong  rst-stage regression); however, program expenses are likely correlated with
the error term in the second stage, to the extent that government expenditures on
antimalarial drugs are correlated with other government expenditures.2
Acknowledging the di culties of  nding a suitable instrument that would al-
leviate potential endogeneity concerns of the malaria illness variable, I alternatively
exploit the panel nature of the survey by using a  xed-e ects estimation approach. I
 rst look at which characteristics are associated with having at least one individual
in the household reporting to have had a malaria episode. While certain household
and farm characteristics are signi cant determinants of malaria incidence, the re-
lationship largely disappears as household  xed e ects are added, suggesting that
the endogeneity of malaria illness is less of concern once these time-invariant unob-
servables are accounted for. I subsequently look at the relationship between malaria
incidence of household members and agricultural pro ts. I  nd a negative and sta-
tistically signi cant relationship between the household?s pro ts and the likelihood
that any household member experienced a malaria infection in the four weeks lead-
ing up to the household survey. While I am cautious in claiming that the e ect is
causal, there is reason to believe that at least some of the e ect might be, since
malaria infections would signi cantly impede the productivity of a ected household
members.
To explain why pro ts might have gone down, I next look at whether total
agricultural labor input is a ected. The e ect of the likelihood of illness incidence on
the number of total person-days employed in the household?s agricultural activities is
examined, where total labor input is the sum of all person-days worked by household
members, plus the total days of exchanged labor and hired labor. Surprisingly, total
2If the only bias associated with the malaria incidence variable was classical measurement error,
then household-level, reported malaria incidence could be instrumented using these o cial malaria
statistics. This is not the only issue however.
59
days of labor does not seem to be a ected even though pro ts are reduced. One
explanation could be that while farmers still go to the  eld regardless of their illness
and work when ill, they may not be able to work the same number of hours or work
as e ciently as when they are not. It is also possible that since the value of time
that household members work on farms is not explicitly accounted for in calculating
pro ts, that pro ts might be going down because households substitute hired labor
for their own labor. As will be discussed in more detail later, none of this appears to
be the case. When looking at hours worked by household members instead of days
worked, the results are qualitatively similar in that hours worked do not respond to
malaria episodes of household members. Moreover, disaggregating labor days into
household- and non-household labor does not make a di erence either.
Yet another potential explanation could be that, if malaria transmission is
seasonal, farmers may come to expect productivity losses during certain periods
of the year and grow a di erent mix of crops in anticipation of the disease. Some
crops, notably rice and cash crops, are highly labor-intensive and need to be planted,
tended to and harvested at certain times of the year. Delays in the process may
harm the quality of the crop or result in crop loss. These are however also crops
that generate on average higher returns. On the other hand, crops such as cassava,
corn and sweet potatoes demand less labor input, are less sensitive to the timing
of harvest, but are also of low-return. I exploit these di erences across crops to
look at whether variations in the likelihood of illness is related to the crop choice
of households.3 More speci cally, I group the crops into highly labor-intensive,
high-return crops and less labor-intensive, low-return crops. Detailed information
on agricultural activities of households allows me to construct the share of land
allocated to each type of crops at the household level. I  nd that the share of
land devoted to rice and annual crops is smaller for households whose province
3Laxminarayan and Moeltner (2003) investigate whether malaria illness a ects crop yields.
Fisher and Datta (2007) examine whether malaria a ects how land is allocated to di erent crops.
60
experienced a greater prevalence of malaria illness in the past. This adjustment in
crop choice could have led to the decrease in pro ts. While not a direct cost of
malaria, these are indirect consequences of the disease that need to incorporated
into the calculation of the burden of the disease.
The remainder of this chapter proceeds as follows. Section 3.2 brie y sum-
marized related literature. Section 3.3 explains some background on malaria in
Vietnam. Section 3.4 explains the empirical strategy. Section 3.5 describes the data
sources. Section 3.6 presents the results and section 3.7 concludes.
3.2 Literature Review
The notion that health may a ect the productivity of farmers is not new. Basta
et al. (1979) implemented an iron and food supplementation treatment to a group of
plantation workers in Indonesia where anemia rates were high, to examine the e ect
of the treatment on labor productivity. They  nd that workers? physical performance
was improved, as measured by work output, physical capacity and morbidity. Antle
and Pingali (1994)  nd that pesticide use has a negative impact on farmer health and
that farmer health is positively related to productivity in two rice-growing regions in
the Philippines. This stream of literature has found consistent evidence that better
health status improves physical capacity and therefore improves labor productivity.
Some papers in this literature measure nutritional status by anthropometric
indicators such as height, weight, or body mass index (BMI). For example, Schultz
(2002)) estimates the e ect of height on hourly wages using local food prices and
parental education as instrumental variables. His estimates suggests that a one
centimeter increase in height results in an eight to ten percent increase in wages in
Ghana and Brazil. A handful of these studies speci cally focus on how diseases may
a ect farmers? health and their labor productivity. Baldwin and Weisbrod (1974)
and Weisbrod and Helminiak (1977) examine the impact of parasitic disease on
61
labor labor productivity but do not  nd a strong link to earnings. Kim et al. (1997),
using data from co ee plantation workers in Ethiopia,  nd that onchocerciasis (a
skin disease) reduces daily earnings of permanent workers by 10 to 15 percent.
The stream of literature that examines the relationship between health and
wages, income, and pro ts  nds more mixed results, which could be because the
majority of studies are not able to perfectly account for the endogeneity of health or
missing variables owing to data limitation. Deolalikar (1988)  nds that wages are
a ected by long-run nutritional status as measured by weight-for-height, but not by
short-run caloric intake, concluding that the human body cannot compensate for
chronic malnutrition. Sahn and Alderman (1988)  nd that calories a ect the wage
o er since better nutrition boosts labor productivity. Strauss (1986) uses local food
prices as an instrument and  nds that higher caloric intake raises family farm labor
productivity. Behrman et al. (1997)  nd that planting-stage calorie consumption has
productivity e ects that are realized in the harvest stage. All of the aforementioned
papers measure nutritional status by caloric consumption. Liu et al. (2008) use
self-reported health status and  nd a positive association with income, with returns
to health larger in rural areas where work is more likely to be manual.
While many studies are not able to address the issue of heterogeneity in health
and unobservable characteristics, some notable exceptions are Schultz and Tansel
(1997) who use instrumental variables to estimate how sickness a ects the wages
of individuals and  nd that disability days reduce daily wages and hours. Pitt and
Rosenzweig (1986) also use health program variables to instrument for illness and
 nd that while there is a reduction in labor supply, pro ts are una ected.
The main contribution of this chapter is that I explore the e ects on agricul-
tural outcomes of a disease which is endemic and recurs over time in a community.
While I am not able to overcome the methodological issues using an instrumental
variable, I instead exploit the panel nature of the survey to estimate a household-
62
 xed e ects model. The analysis by Audibert and Etard (2003) is similar in that
they test the link between health and agricultural productivity and labor input.
They  nd no e ect on either pro ts or labor use and suggest that the absence of an
e ect of improvement in health on hired labor could be because the additional time
available might be spent on leisure or the cultivation of crops other than rice, the
main cash crop.4
This chapter is also related to the literature that studies the economic bur-
den of malaria. Malaria, along with other tropical diseases, has been documented
as a factor constraining economic development (Conley (1972), Gallup and Sachs
(2001)). At the micro-level, researchers have attempted to gauge the economic bur-
den of malaria, by estimating the number of workdays or wages lost due to malaria
infections (see for example, Leighton and Foster (1994), Cropper et al. (1999), Chima
et al. (2003)). More recently, researchers have looked at broader and long-run im-
plications of the disease. Hong (2007)  nds that malarial risk leads to adverse long
run health outcomes, lower labor force participation, and lower wealth. Barecca
(2009)  nds that in utero and postnatal malaria exposure leads to lower educational
attainment. Bleakley (2010) studies the e ect of malaria eradication campaigns on
the income and education of native males in the United States, Brazil, Colombia,
and Mexico. Using malaria mortality rates and an ecology index to identify pre-
eradication disease prevalence, he  nds that childhood exposure to malaria lowers
labor productivity and leads to lower adult income. Cutler et al. (2007) use geo-
graphic variation in malaria prevalence in India prior to a nationwide eradication
program in the 1950s to study the e ect of malaria on educational attainment and
income. They  nd a positive e ect of the program on the income of men only,
which they attribute to di erences between men and women in their improvements
4Audibert et al. (2009) do not  nd an e ect of malaria prevalence on yields or the use of family
labor. On the other hand, Audibert et al. (2003)  nd that malaria had a large impact on cotton
production in C^ote d?Ivoire, which could be because cotton is a labor intensive crop.
63
in labor market productivity. I extend this literature by showing that there could
be further economic costs to the disease if households possibly adapt to the disease
environment by adjusting their production decisions ex-ante, thereby hurting their
pro ts.
3.3 Background on Malaria in Vietnam
Vivax malaria is the prevalent malaria strain in Vietnam. Although it is the
more benign type of malaria and has a low fatality rate, it is accompanied by poten-
tially severe symptoms.5 These include recurring fever, chills, severe headache, vom-
iting, and thirst, followed by a sweating stage, after which the symptoms subside.
These symptoms would leave the farmer bedridden for days, hurting his productiv-
ity and possibly pro ts (see Chima et al. (2003) for a review on Africa and Morel
et al. (2008) for a study in Vietnam).6 Even after individuals resume work, their
productivity may not immediately revert to pre-infection levels, especially for those
working in labor-intensive occupations (Hong (2007)). These attacks recur until the
infection is treated. Unlike many other diseases, immunity to malaria does not last
for a lifetime. Acquired immunity may develop and get stronger as the infection
proceeds, but even adults in hyperendemic areas do not develop full immunity to
the disease.
Malaria transmission in Vietnam is highly seasonal. Transmission peaks during
and after the wet season (from April through November), which coincides with the
period of high labor demand (Nam et al. (2005)). The  rst malaria eradication
e ort in Vietnam was started in 1958 and continued for two decades. There was
a dramatic reduction in malaria morbidity and mortality that was sustained until
5The prevalent malaria strain in Africa that carries a high fatality rate is Falciparum malaria.
6In the data used in this chapter, the average length of an episode is reported at approximately
7.5 days, out of which 6.3 days individuals reported to have been unable to carry on their usual
activities.
64
the early 1980s. However, the situation reversed due to increasing lack of resources,
post-war migration and insu cient health infrastructure. The number of malaria
cases peaked in 1991, with more than 1.6 million cases and nearly  ve thousand
deaths resulting from about 150 outbreaks (Ettling (2002), Hung et al. (2002)).
In 1991, the central government made malaria a top health priority. A na-
tional malaria control program was begun, which provided clear control guidelines
for intervention strategies based on the level of endemicity. The program focused on
prevention, early diagnosis, and e cient treatment. Primary health Care (PHC) net-
works were extended, and surveillance and response to outbreaks was strengthened.
Insecticide-treated bednets and antimalarial drugs were distributed and extensive
indoor residual spraying was carried out. A stronger economy facilitated access to
treatment options and contributed to an increase in the government?s health bud-
get. This concerted e ort resulted in a 92% reduction in the number of epidemic
outbreaks per year, near elimination of malaria fatalities and a large decline in the
number of cases over the following 5-6 years. However, Vietnam has not succeeded
in eliminating malaria until today and downward trends in number of cases started
to reverse in the late 1990s.
3.4 Empirical Strategy
I estimate the impact of malaria on farm pro ts using data from the 1992/93
and 1997/98 Living Standard Measurement Surveys. In the health section of the
survey, each household member is asked about any illness episodes in the four weeks
prior to the survey. I construct an illness shock variable that measures the number
of self-reported malaria cases in the household and divide it by household size.7
7One might argue that the e ect on pro ts might depend on who in the household experiences
malaria. In other words, the e ect on pro ts might di er whether an adult, child or elderly person
had malaria. To explore this, I replace the original variable, which measures the share of household
members who experienced a malaria illness, with the same variable but measured by demographic
group: i.e., I include in the regression the share of adults, elderly and children in the household
65
Unfortunately, the survey underwent a considerable change in the format of the
questionnaires between the two waves. In the 1997/98 wave, instead of asking for
the speci c diseases experienced, households were asked which symptoms they had
experienced in the prior four weeks. In the second wave of the survey, I use any
responses that were coded as \fever" to proxy for malaria illness. Fever is one of
the main symptoms of malaria, but could be an indication of an illness other than
malaria (e.g., dengue fever). This variable therefore su ers from measurement error.
I also de ne a narrower proxy for malaria, which is a combination of headache, fever,
vomiting and cough. The results are qualitatively the same when this de nition of
malaria is used but are shown as an alternative nevertheless.
To study the impact of malaria on agricultural pro ts and labor use, the
following equation is estimated:
Yi =  + Xi + Ai + Malariai + Landtitlei +Yeari +Provincei + i (3.1)
where the subscript i notates the household. Yi is either the logarithm of pro ts or
the logarithm of the total number of person-days used by each household. Total la-
bor use of the household is constructed as the sum of any household labor employed
in the last 12 months, plus unpaid or exchanged labor, and hired labor. Xi includes
household composition variables (share of adult females, elderly people and chil-
dren,8 logarithm of household size, age, sex and education of household head) and
Ai initially includes the logarithm of total land area cultivated by the household and
the number of pieces of farming equipment (such as a tractor or insecticide pumps)
owned by the household.9 LandTitlei measures the number of land use certi cates
who experienced malaria in the prior four weeks. These variables are set to zero if there are zero
members of the corresponding demographic group in the household. Results appear strongest for
children. The negative relationship between pro ts and malaria illness of adults is present in the
baseline model but disappears with the inclusion of additional variables and  xed e ects.
8Share of adult males is the omitted category.
9Farming equipment for livestock food processing and aquaculture are excluded.
66
distributed in the province by 1998, divided by the total number of households.
Yeari and Provincei control for year- and province-speci c trends, respectively,
and  i is the idiosyncratic error term. To the baseline regression, I subsequently add
whether the household has access to irrigation by looking at whether it cultivated
any irrigated land, and variables indicating the share of land of good and medium
quality (with share of land of bad quality being the omitted category). Province-
and year-  xed e ects are then replaced by Province x Year  xed e ects, to see
whether the results are robust to time-varying, province-speci c unobservables.
I consider only households that report to be \farm" households, which reduces
the sample to a little less than two thirds of the original sample. The relationship
between pro ts and the likelihood of a malaria infection of any household member
is examined  rst. Pro ts are calculated as the sum of crop revenue from food
crops (including rice, corn and sweet potatoes), other annual crops (e.g., soy beans,
peanuts, sesame), perennial crops (e.g., tea, co ee, rubber) and fruit crops (e.g.,
oranges, pineapples, mangoes) minus any expenses. Revenue from crop byproducts
are also included, as well as revenue from agricultural sideline activities such as
livestock, agro-forestry, and aquaculture. Expenses from the purchase of seeds,
fertilizers, insecticides, transport, storage and hired labor are deducted from total
revenue. The value of any output that was given to laborers or landlords and any
payments for use of land are also excluded. These payments include cash or in-kind
payments on allocated, auctioned or rented-in annual land. If payment was for land
use over multiple years, the amount was rescaled to re ect annual payments. Finally,
pro ts are adjusted for in ation between the years using an overall price index.10
10As a caveat, \priced" own labor is not accounted for in the measure of pro ts. Therefore,
the impact of malaria on pro ts, when they do not include the cost of the household?s own labor,
would be expected to be automatically negative as the farm would replace the labor of their own
(which costs nothing) with that of outside workers (which would be subtracted from the pro ts).
Attempts to subtract the value of the time each household member works on the farm proved to
be di cult as these attempts led pro ts to be negative for a large fraction of households in the
data.
67
The resulting measure of pro ts has some extreme outliers, which is likely due
to the many pieces of information that were put together (also reported in McCaig
et al. (2009)). In order for those outliers to not bias my estimates, observations in
the extreme 1% of each tail are dropped from the sample used in the regressions. In
the labor use model, observations are dropped if the household reported zero days
of total labor input.
I am interested in whether the impact of malaria on pro ts could possibly be
explained by its impact on crop choice. Rice is by far the most important crop in
Vietnam, most of which is planted on irrigated lowland. Other major crops include
corn, sweet potato, cassava, legume, soybean, peanut, sugar cane, co ee and black
pepper. The planting and harvesting of rice and other cash crops is extremely labor
intensive. Moreover, those activities need to be carried out in a timely manner.
Despite advances in modern technology, many tasks are still performed manually.
On the other hand, some food crops such as corn, sweet potato and cassava are less
labor intensive in general, but they also yield lower pro ts.11.
My hypothesis is that households, based on past illness experiences, may an-
ticipate malaria and choose to grow a less labor-intensive crop mix in an e ort to
minimize loss in pro ts. Since those crops also tend to be less pro table in general,
this would reduce farm pro ts. The following equation is estimated to test this
hypothesis:
Land highi;98 =  + Xi;98 + Ai;98 + 1Malariai;92 + 2Malariap;92
+ Landtitlep;98 +Yeari;98 +Provincei;98 + i
(3.2)
where i denotes the household and p the province of residence. The dependent
11There are three main cropping seasons in Vietnam: The main season, with planting occurring
from May to August and harvesting from September to December, the winter-spring season and
the summer-autumn season.12 The wet season occurs from April to November. Temperatures are
higher year-round in central and southern Vietnam, and more temperate in the North (Food and
Agriculture Organization (1999a)
68
variable Land highi;98 is the share of land allocated to high-return, highly labor
intensive crops by household i. All regressors are measured in 1998, except for
Malariai;92, which measures the number of malaria episodes in the household in
1992, weighted by household size.13 Since the disease is endemic at the community
level, I use province-level malaria prevalence data to proxy for the intensity of the
disease that households may expect to experience on average. This is measured
by Malariap;92. Thus I look at how malaria illness in the household and at the
province-level in 1992 a ects the household?s own crop choice decisions 5-6 years
later, conditional on other household characteristics which might also in uence how
agricultural land is allocated to crops of di erent labor intensity.
While the speci cation to be estimated is very similar to equation 3.1, there
are a few di erences. First, I use only the sample of households that was surveyed
in both years, to look at whether their own or province-level past malaria illness
a ects subsequent crop choice of the household. Of the 4800 households originally
interviewed in 1992/93, 495 households were not reinterviewed either because they
were not available or because they refused to be surveyed. These households were
replaced with a randomly selected household in the same village.14 The households
that were interviewed again for the second wave of the survey were either still living
in the same residence or they moved to a di erent residence within the same village.
Second, since the outcome is measured as \share of land", perennial and fruit crops
are excluded since many households report how much perennial and fruit crops they
grew in number of trees instead of land area cultivated, which would have created
an aggregation issue.
There are two reasons why I look at later (instead of contemporaneous) crop
13The correlation between malaria illness in 1992 and 1998 is rather low, indicating that malaria
incidence does not necessarily repeat itself in the same household and providing another argument
against the endogeneity of the disease.
14The households that dropped out of the sample from 1992 were on average smaller in size,
cultivated less land, and the household head was younger, less likely to be male and more likely to
have received formal education.
69
choice. One reason is that information on malaria illness is available only for the
same year as the household reports its agricultural activities, creating a timing
issue: planting decisions have already been made by the time households experience
malaria. Another is that households may not change their crop mix from one year
to another: in other words, crop choice might be \sticky" and it may take time to
adjust production decisions. This is con rmed by the fact that the share of land
allocated to high-return, highly labor intensive crops changes little between the two
waves.15
3.5 Data Sources
3.5.1 Vietnam Living Standard Measurement Survey (LSMS)
The household surveys used in this study are the Vietnam Living Standard
Measurement Surveys (VLSS) from 1992/93 and 1997/98. Data collection was
funded by the UNDP and the Swedish International Development Authority (SIDA),
and implemented with technical assistance from the World Bank as part of the Living
Standard Measurement Study (LSMS) Household Surveys. Each survey consisted
of three parts: a household survey, price survey and community survey. The house-
hold survey contains information on the demographic characteristics of all household
members, as well as information on education, health, migration, housing, fertility,
and most importantly, agricultural activities. The price survey includes observed
prices of important food and non-food commodities. The commune questionnaires
were administered in rural areas only and contain information on demographics,
education and health infrastructure, and main employment activities. Sample selec-
tion was based on the 1989 Census. There were 4800 households originally sampled
in 1992-93 and 6002 households in 1997-98, representing 7 regions,53 provinces (61
15The mean change in the share of land devoted to high-return, highly labor intensive crops is
just over 3%.
70
provinces in 1998, after a reorganization of provinces) and 150 communes, of which
80% were rural.16 Households were interviewed from October 1992 to October 1993
in the  rst wave, and from December 1997 to December 1998 in the second. Approx-
imately 4300 of the households surveyed in 1992 were reinterviewed in 1998, allowing
the data set to be used as a panel. The  nal sample includes households from 51
provinces: 2 provinces, Kon Tum and Lai Chau, did not have any households who
grew crops in either of the years of the survey.
3.5.2 Economic Reform and Land Title Data
Vietnam went through a series of structural economic reforms (called \Doi
Moi") starting around 1986. These reforms were aimed at transforming the economy
from a centrally planned to a market-oriented system and encompassed liberalization
of product and input markets, agricultural diversi cation, assignment of individual
land rights and trade liberalization (Benjamin and Brandt (2002)), Kirk and Tuan
(2009)). These economic reforms continued during the period 1992-1998. Trade
restrictions on rice and fertilizers were relaxed in 1996 and the rice export quota
was increased, which led to a signi cant increase in the price of rice. While land was
o cially still the property of the State, non-tradeable land use rights were assigned
to individuals for up to  fteen years.
The Land Law of 1993  nally granted households the permission to transfer,
inherit, exchange, lease or mortgage their land-use rights by distributing Land Use
Certi cates (LUC, hereafter), which guaranteed e ective land ownership. The law
16The 7 regions are Northern Uplands, Red River Delta, North Central Coast, Central Coast,
Central Highlands, Southeast and Mekong Delta. There were 53 provinces in 1992 and 61 provinces
in 1998. The di erence in the number of provinces is due to the fact that 8 provinces were split into
two provinces in 1996. Bac Thai province split into Bac Can and Thai Nguyen, Ha Bac province
split into Bac Giang and Bac Ninh, Hai Hung province split into Hai Duong and Hung Yen, Minh
Hai province split into Bac Lieu and Ca Mau, Nam Ha province split into Ha Nam and Nam Dinh,
Quang Nam-Da Nang province split into Quang Nam province and Da Nang municipality, Song
Be province split into Binh Duong and Binh Phuoc, and Vinh Phu province split into Phu Tuo
and Vinh Phuc.
71
increased tenure over the land that farms had been allocated and was seen as setting
the grounds for a formal market for land (Do and Iyer (2007)). The  rst two reforms
were implemented at the national level and are therefore less likely to a ect the
results. However, the distribution of LUCs was carried out by communal authorities
and thus did not occur in a uniform way across communes. By the end of 1998, there
was considerable variation across provinces in how far the process had proceeded.
Land tenure is of concern as far as it might a ect factors that are related to
farm pro ts, labor input and crop choice decisions. Studies have shown that land
tenure or property rights increase labor supply (Field (2007)) and investment in the
land (Besley (1995), Goldstein and Udry (2008)). Although it would be ideal to
have information on whether the household itself held a registered LUC, this data
is not available, and instead data on the proportion of households in a province who
had a registered land title is used to control for the e ect of the reform. The data
comes from the records of the General Department of Land Administration (GDLA)
in Hanoi.17 Using this data, Do and Iyer (2007)  nd that provinces that made
greater progress in land titling issuance had a greater increase in the proportion of
cultivated area devoted to multi-year crops in Vietnam. Following their approach, I
take province-level di erences in registration level of LUCs as plausibly exogenous.18
17Data was obtained from Quy-Toan Do at the World Bank who used the data to examine the
impact of land titles on land use and crop investment (see Do and Iyer (2007)).
18Heterogeneity across provinces in registration levels could result from a variety of factors, in-
cluding some that might a ect the agricultural outcomes of interest. Do and Iyer (2007) argue that
province- level di erences are due to bureaucratic and other reasons exogenous to the household.
They run alternative speci cations controlling for household expenditure that might be related to
registration but  nd similar results.
72
3.6 Results
3.6.1 Description of the Data
The means and standard deviations of all variables are presented by survey
year in Tables 3.1 and 3.2, for the pooled sample and the panel sample, respectively.
While pro ts on average increased over the years, total days worked on the farm
went down, as did total hours worked by household members. This might be due
to substitution to capital (ownership of farm equipment went up). Nearly 80% of
households were headed by a male, who is older in 1997/98 compared to 1992/93.
This is not surprising given that a large share of households in 1998 were reselected
from 1992. The size of total land cultivated by households slightly increased, as
did the share of households with access to irrigation. Somewhat surprisingly, land
quality seems to have deteriorated over the years: however, this may be largely
attributed to measurement error, in particular stemming from the format change in
the questionnaires.
The last two rows in the tables report means of the two malaria illness mea-
sures, as calculated by the number of cases per household, divided by household
size. The mean of the  rst proxy, which uses fever as a proxy for malaria in 1998,
actually increases by about two-fold despite the national downward trend in malaria
cases over this period. As noted above, the use of fever to proxy for malaria likely
overestimates the incidence of the disease. Malaria is not the only disease that is
accompanied by fever: in uenza and dengue fever are also characterized by fever
and are not uncommon in Vietnam, although the incidence of dengue fever is lower
than malaria. I therefore construct construct an alternative measure of malaria
which combines four symptoms reported in the survey: headache, fever, vomiting
and cough. Some of the symptoms may overlap with other diseases, but not all
73
of them.19 Since individuals who experienced malaria may report some, but not
all of these symptoms, this proxy may lead to an underestimate of the actual inci-
dence. Thus, it might be informative to look at the association of both measures of
malaria and agricultural outcomes. The share of land allocated to labor-intensive,
high-return crops is greater than 76%. Rice is the single most important crop in
Vietnam and therefore it usually takes up a large proportion of the land for most
households, which explains why the means are all in the higher range.
Descriptive statistics for panel households are overall comparable, although
due to possible changes in the household composition, the change in the share of in-
dividuals in each demographic group is slightly larger compared to the corresponding
change in the variable from the cross-sectional sample. Panel households are slightly
larger in size, and the share of good quality land is higher in 1998 compared to non-
panel households. They also own more farm equipment and are more likely to hold
a land use certi cate in 1998.20
Table 3.3 shows the geographic distribution of households across seven major
regions in Vietnam and Table 3.4 presents, by province, the percentage of households
that held a land use certi cate as of 1998. The number ranges widely from 11.89%
to 100%.21 Since the registration of land use certi cates started to roll out after the
Land Reform Law of 1993, the equivalent variable for households in 1992 is set to
zero.
3.6.2 Models of Malaria Incidence
I  rst examine which households are more likely to report anyone in the house-
hold having had malaria in the four weeks prior to the survey. Speci cally, I estimate
19Using this measure, illness incidence goes down to about 0.002 in 1998.
20The number of pieces of farm equipment measures how many pieces of tractors, water pumps,
threshing machines, etc were owned by the household.
21Two provinces are not shown since there were no households in the  nal regression sample
residing in those provinces.
74
a logit model to analyze which characteristics of the household are associated with
malaria incidence. Table 3.5 reports marginal e ects from this logit model. Columns
(1) and (2) report results for the malaria measure that proxies malaria illness in 1998
using only fever, whereas columns (3) and (4) report results using the alternative
malaria proxy which combines multiple symptoms to measure malaria illness in the
1998 survey. Further, speci cations in columns (1) and (3) include province and
year  xed e ects, and columns (2) and (4) include province x year  xed e ects. To
account for the possibility that the stochastic error terms are correlated within the
household?s province of residence, standard errors are clustered by province.
Interview month dummies are included to account for the seasonality of the
disease since households were interviewed during di erent months of the year and
the health questions were asked in reference to the four weeks prior to the survey.
Not surprisingly, the coe cients on some of the wet season months such as July
and August are positive and statistically signi cant, indicating that households are
more likely to experience malaria during this period relative to January (the omitted
month). In particular, households interviewed in July are more likely to report
having had an illness even after controlling for province-year-speci c  xed e ects.
This relationship, however, is not present when using the narrow proxy for malaria.
Across all speci cations, there are statistically signi cant associations between
the likelihood that there was a malaria episode in the household and household size
and education of the household head. While the coe cient on whether the household
had any irrigated land iss insigni cant in columns (2) and (4) with province-year
 xed e ects, the coe cients on some household characteristics remain large and
statistically signi cant. Moreover, the results using the alternative malaria proxy
in columns (3) and (4) suggest that land quality may matter as well. Along with
household head?s education, these are all factors that are likely to be positively
correlated with income, which raises the possibility of reverse causality. The number
75
of observations in column (4) is lower by about a quarter of the sample because of
the large number of households who report no malaria incidence using the narrow
measure in 1998. This causes some provinces to drop out of the sample entirely when
province-year  xed e ects are included and if they have no households reporting a
malaria case.
Columns (1) and (2) in Table 3.6 report estimates of the same model as in Table
3.5, but with household  xed e ects. This is to see whether controlling for unob-
servables at the household-level decreases the correlation between certain household
characteristics and malaria incidence, and thus the potential for reverse causality.
Here the within-household variation in malaria incidence and regressors is exploited
to purge any unobservable household-level heterogeneity. The dependent variable
is the malaria proxy using fever in column (1), and the narrow malaria proxy in
column (2). Many households had no variation in their malaria responses between
1992 and 1998. Therefore, the number of observations used in the panel regression
drops substantially.
Compared to estimates in Table 3.5, the statistically signi cant relationship
between malaria incidence and some of the variables (e.g., share of good quality land)
disappears, partly alleviating concerns that malaria incidence is endogenous at the
household-level. The sign on the age of household head switches and is now positive
and signi cant at the 10% level. Moreover, the number of pieces of farm equipment
owned has a positive and statistically signi cant association with malaria illness. In
column (2), where malaria incidence is proxied using the alternative measure, the
only factor that still matters is the share of elderly people in the household.
3.6.3 Malaria Illness and Farm Pro ts
Table 3.7 reports estimates of equation 3.1 where the dependent variable is
total agricultural pro ts. The relationship between pro ts and malaria illness is
76
negative and statistically signi cant at the 1% level in the baseline results in col-
umn (1). Lower farm pro ts are associated with more members of a household
su ering from malaria, although the relationship may not be entirely causal. Re-
verse causality could be present if households who earn more farming pro ts can
a ord better treatment and thus su er from fewer illness episodes. The magnitude
of the malaria coe cient indicates that in a household consisting of six people, if one
additional individual becomes sick from malaria, pro ts are lower by roughly 4%.
This statistically signi cant, negative relationship largely holds up to the inclusion
of province x year  xed e ects in column (3) that control for province-time-speci c
trends although estimates are smaller.
Other variables also explain variation in farm pro ts. Across columns, total
land area and the number of pieces of farm equipment owned by the household are
positively associated with farm pro ts, although the latter is measured very impre-
cisely in column (4). Pro ts are higher if the household head is older, male and
received formal education. Column (2) adds land quality variables and an irriga-
tion dummy to the speci cation, where the latter measures whether the household
cultivated any irrigated land. Land quality variables are shown to have a positive
association with pro ts. The proportion of households in a province that received
a formal land use certi cate has a positive sign, but the coe cient is not statisti-
cally signi cant in any of the speci cations. Province x year  xed e ects are further
added in column (3) to account for province-year speci c, unobservable trends. The
land title variable drops out because it varies only at the province-year level. These
results are overall similar to the ones in column (2). Again, standard errors are
clustered by province and are robust to heteroskedasticity.
While I am not able to deal with the measurement error problem associated
with malaria and potential reverse causality in a rigorous way, I try to address the
issue of omitted variables using household  xed e ects, which will capture household-
77
speci c omitted variables that are time-invariant. These results are shown in column
(4). The drop in the number of observations is obviously due to the fact that not all
households were reinterviewed in 1997/98. The coe cient measuring the intensity
of malaria illness in the household is still negative and statistically signi cant at
the 10% level. Although there is a loss of precision, the size of the coe cient is
comparable to column (2). Except for the household head?s education and the share
of female adults, most household characteristics are now insigni cant.
The 4-week malaria illness variable is an imperfect measure of the extent that
the household is exposed to the disease over the year, due to the length of the
reporting period and the fact that transmission is seasonal. To capture seasonality
in transmission, I include dummies for the month the household was interviewed, and
interact those dummies with the household?s malaria illness variable. The interview
month dummies may also capture any other sort of measurement error that is related
to which month of the year the household is interviewed: households may remember
more or less precisely how much they earned in yields and pro ts depending on how
much time has passed since the main harvest, even though all households were asked
about agricultural activities over the entire past 12 months.
These results with interview month dummies are presented in Table 3.8. Prof-
its are persistently higher for households that were interviewed in the main agri-
cultural season, from May through October. Testing for joint signi cance of the
malaria illness variable and its interaction terms with interview months, it does
appear to matter when the household was interviewed: The negative relationship
between pro ts and malaria illness is statistically signi cant when households were
interviewed in May through August and December in column (1), and additionally
in February in column (2). In columns (3) and (4), only the interaction terms with
June and August, and February and July are jointly signi cant, respectively. All of
these months fall in the wet season, when malaria transmission is higher. The size
78
of the e ect is also greater by at least a quarter.
As previously mentioned, households were not speci cally asked in 1997/98
which illness they su ered from, but which symptoms they experienced (and recog-
nized). To see the impact of using symptoms other than fever to proxy for malaria,
the same speci cations as in Table 3.7 are estimated with this alternative proxy,
which combines several malaria symptoms. These are presented in Table 3.9. Co-
e cient estimates on malaria illness are larger in column (4) compared to those in
Table 3.7, which is to be expected if fever was indeed capturing measurement error
by including other diseases. According to estimates with household  xed e ects in
column (4), one more case of malaria illness in a household of six is associated with
around 10% lower pro ts.
3.6.4 Malaria Illness and Use of Agricultural Labor
Given the negative relationship between pro ts and malaria illness, the next
question is whether any of the e ect could result because farmers fall ill and are not
able to provide su cient labor. In order to investigate this, I examine how the total
number of person-days employed is associated with malaria illness. These estimates
are presented in Table 3.10. Across all speci cations, the total number of person-
days employed by the household in the past 12 months is positively associated with
total land area, household size and number of farm equipment owned. The larger
the share of children and elderly in the household, the lower the number of total
labor days used. This may not be surprising given that households largely rely on
their own labor input rather than exchanged or paid labor.
The important result is that malaria illness shocks do not appear to impact
total labor use of the household. This result persists throughout di erent speci ca-
tions: with the inclusion of land quality and irrigation variables, with the addition
of province x year  xed e ects and even with household  xed e ects. Since total
79
labor is measured as the sum of household and non-household labor, the question
is whether there are no e ects overall because households reduce their labor supply
and substitute paid labor for their own. In order to address that concern, I run the
same model on total labor days worked by members of the household. These results
are reported in Table 3.11, which indicate that there is no evidence of such substi-
tution. Using the alternative proxy for malaria illness does not change the results,
as reported in Table 3.12. Another possibility is that while household members still
opt to work the same number of days on the  eld, they may reduce the number
of hours spent working per day. I examine whether illness episodes due to malaria
a ects number of total hours worked by household members, but do not  nd any
evidence of this (results not reported here). 22
3.6.5 The Impact of Malaria on Crop Choice
If malaria does not a ect labor employed in agriculture, what explains the
impact on pro ts? An alternative explanation could be that households, in antici-
pation of the seasonal disease, choose to grow less rice, soy beans and peanuts, which
are crops that are highly labor-intensive and o er higher returns, and instead grow
more corn, cassava and sweet potatoes, which are crops that are less labor-intensive,
but also less pro table. This might be especially true in a context where the disease
is endemic and occurs during seasons when labor demand peaks, as in Vietnam.
In order to test whether households, in making their crop choice decisions, take
into account how likely it is that they will contract malaria, which might weaken
their productivity and ultimately hurt pro ts, I estimate equation 3.2 to see whether
previous malaria experience at the household- or province-level a ects households?
crop choice decisions 5-6 years later. While previous household-level malaria illness
measures the household?s own propensity to contract the illness (possibly related to
22This result also suggests that the negative relationship between malaria and pro ts is not
driven by the fact that own-labor is not factored in.
80
household-speci c factors), province-level malaria prevalence measures how likely it
is for a household on average to become ill.
The estimation results for models of crop choice are shown in Table 3.13. In
order to look at how households? health shocks a ected their own crop choice later
on, here I consider only panel households who were surveyed in both years. The
larger the area of the land cultivated and the better the quality of the land, the
more land the household allocated to such pro table crops. While household-level
malaria illness is not signi cant in columns (1) and (2), province-level malaria is
statistically signi cant once it is added in column (3). These  ndings suggest that
households respond to previously experienced province-level malaria prevalence, but
that they do not respond to their own experience of malaria incidence. Interestingly,
these results suggest that households do not necessarily take into account their own
illness experience, but rather, how likely it was for a household in their province
to have su ered from malaria. Past experience may serve as a proxy for how likely
members of a household think they might become ill in any given year. This result
is also in line with the fact that the correlation of malaria illness variables in 1992
and 1998 is not very high, suggesting that malaria illness may not be as recurring
as often suggested.
3.7 Conclusion
Malaria is still endemic in a number of developing countries and an estimated
three billion people are said to be at risk of malaria each year (Boschi-Pinto and
Shibuya (2008)). In this chapter, I show that malaria illness measured at the house-
hold level is negatively correlated with agricultural pro ts. However, this does not
appear to be due to a decrease in the total number of labor days the household em-
ployed. As an alternative explanation for the decrease in pro ts, I present evidence
that the household?s land allocation decisions might be a ected. Since malaria is
81
a seasonal and endemic disease in Vietnam, households living in an area where the
likelihood of contracting malaria is higher may choose to allocate less land to highly
labor-intensive, but also more pro table crops in anticipation of the disease. Results
seem to con rm this: more malaria episodes experienced on average in a province
are associated with a smaller share of land devoted to rice and cash crops  ve years
later.
As Hong (2007) notes, most papers in the literature that aim at calculating
the impact of malaria on losses in economic activities fail to capture long-term con-
sequences of the disease and therefore the impact on economic activities could likely
be underestimated. Along the same line, I argue that the welfare of agricultural
households can still be negatively impacted even if no e ect on labor supply is
found, as households may adjust their production decisions ex-ante.
82
T able 3.1: Descriptiv e Statistics (All households)
1992 1998
V ariable Mean Std. Dev. N Mean Std. Dev. N
Ln(pro ts) 7.919 1.06 2802 8. 559 0.945 3229
Ln(T otal lab or da ys) 6.416 0.771 2796 6. 163 0.873 3098
Ln(T otal lab or da ys w ork ed b y HH mem b er s ) 6.219 0.799 2791 6.191 0.737 3020
Ln(T otal hours w ork ed b y HH mem b ers) 7.966 0.912 2791 7.859 0.812 3020
Share of adult female in h ousehold 0.301 0.17 2 2802 0.307 0.183 3229
Share of elderly in household 0.066 0.175 2802 0.089 0.214 3229
Share of c hildren in household 0.376 0.231 2802 0. 333 0.23 3229
Ln(household size) 1.541 0.45 6 2802 1.486 0.469 3229
Age of household head 45.277 14.775 2802 47.486 13.752 3229
Whether household head is male 0.787 0.4 1 2802 0.787 0.41 3229
Whether household head has education 0.873 0.33 3 2802 0.897 0.305 3229
Ln(T otal land area) 8.346 0.864 2802 8.544 1.003 3229
Whether household has access to irr igation 0.971 0.16 8 2802 0.983 0.131 3229
Share of go o d qualit y land 0.276 0.367 2802 0.241 0.341 3229
Share of mediu m qualit y land 0.509 0.387 2802 0.37 0.362 3229
# of farm equipmen t 0.926 1.021 2802 1.478 1.304 3229
% households in pro vince with LUC 0 0 2082 0.714 0.249 3229
Share of land gro wing high-return crops 0.767 0.236 2790 0. 769 0.261 3090
Malaria incidence of HH mem b ers 0.048 0.143 2802 0.097 0.199 3229
Malaria incidence of HH mem b ers (narro w) 0.048 0.143 2802 0.002 0.026 3229
83
T able 3.2: Descriptiv e Statistics (P anel households)
V ariable Mean Std. Dev. Mean Std. Dev. N
Ln(Pro ts) 7.982 0.989 8.56 6 0.909 2141
Ln(T otal lab or da ys) 6.478 0.729 6.222 0.830 2066
Ln(T otal lab or da ys w ork ed b y HH mem b e rs) 6.308 0.753 6.23 1 0.741 2029
Share of adult female in household 0.297 0.167 0.311 0.187 2141
Share of elderl y in household 0.064 0.167 0.09 5 0.219 2141
Share of c hildren in household 0.383 0.23 0.327 0.232 2141
Ln(household size) 1.56 0.448 1.49 0.473 2141
Age of household head 45.412 1 4.553 47.8 94 13.612 2141
Whether household head is male 0.8 0.4 0.78 5 0.411 2141
Whether household head has education 0.87 0.337 0.89 3 0.309 2141
Ln(T otal land area) 8.385 0 .821 8.505 0.956 2141
Whether household has access to ir rigation 0.972 0.164 0.984 0.127 2141
Share of go o d qualit y land 0.27 0.361 0.271 0.345 2141
Share of mediu m qualit y land 0.5 0.379 0.35 2 0.343 2141
# of farm equipmen t 0.998 1 .034 1.56 1.312 2141
% households in pro vince with LUC 0 0 0.735 0.244 2141
Share of land gro wing high-return crops .761 .225 .802 .226 2110
Malaria incidence of HH mem b ers 0.05 0.147 0.087 0.185 2141
Malaria incidence of HH mem b ers (alternativ e) 0.05 0.147 0.002 0.023 2141
84
Table 3.3: Geographic distribution of households
Region Number of households (% of total)
Northern Uplands 1176 (19.5)
Red River Delta 1366 (22.65)
North Central Coast 965 (16)
South Central Coast 652 (10.81)
Central Highlands 375 (6.22)
Southeast 388 (6.43)
Mekong River Delta 1109 (18.39)
Table 3.4: Means by province, grouped by region
Province % HH with land title
in 1998
Northern Uplands 65.9
Bac Can 93.91
Bac Giang 31.25
Cao Bang 77.14
Ha Giang 75.72
Hoa Binh 100.00
Lang Son 87.20
Lao Cai 85.88
Phu Tho 35.84
Quang Ninh 91.88
Son La 31.13
Tuyen Quang 95.62
Yen Bai 50.49
Red River Delta 81.72
Ha Nam 74.04
Ha Tay 11.89
Hai Phong 97.34
Hanoi 35.02
Hung Yen 92.48
Ninh Binh 99.48
Thai Binh 50.12
North Central Coast 74.39
Ha Tinh 90.98
Nghe An 91.45
Quang Binh 99.57
Quang Tri 40.54
Thanh Hoa 98.87
85
Thua Thien Hue 49.15
Central Coast 77.26
Binh Dinh 93.95
Binh Thuan 76.67
Khanh Hoa 84.78
Ninh Thuan 77.71
Phu Yen 88.02
Quang Nam 70.69
Quang Ngai 37.37
Central Highlands 60.09
Dac Lac 83.12
Gia Lai 29.91
Lam Dong 89.04
Southeast 78.58
Ba Ria Vung Tau 65.87
Binh Duong 64.42
Dong Nai 57.99
Ho Chi Minh City 68.74
Tay Ninh 81.42
Mekong Delta 88.97
An Giang 91.72
Ben Tre 89.59
Ca Mau 89.82
Can Tho 86.32
Dong Thap 85.94
Kien Giang 31.95
Long An 90.53
Soc Trang 74.77
Tien Giang 75.52
Tra Vinh 80.09
Vinh Long 80.01
86
Table 3.5: Determinants of malaria illness
(1) (2) (3) (4)
Dependent variable Malaria Malaria Alt. malaria Alt. malaria
Ln(Total land area) -0.0046 -0.0148 0.0027 0.0024
(0.0815) (0.0682) (0.0864) (0.0897)
Whether household has access to irrigation -0.6887* 0.0991 -0.5602* -0.1511
(0.3993) (0.1957) (0.3125) (0.3245)
Share of good quality land -0.3576 -0.1231 -1.0208*** -0.9046**
(0.2535) (0.2200) (0.3838) (0.3778)
Share of medium quality land -0.3155 0.0142 -0.6502** -0.5056*
(0.2507) (0.1843) (0.2875) (0.2883)
Share of adult female in household -0.0741 -0.1872 -0.7464 -0.8969*
(0.2607) (0.2831) (0.4995) (0.5329)
Share of elderly in household -0.4305 -0.4371 -1.3913*** -1.4841***
(0.3102) (0.2997) (0.5196) (0.5493)
Share of children in household 0.3219 0.2326 -0.4008 -0.5210
(0.2703) (0.2736) (0.4721) (0.5067)
Ln(household size) 0.5295*** 0.5996*** 0.3829** 0.4589**
(0.1317) (0.1242) (0.1691) (0.1836)
Age of household head -0.0062** -0.0071*** -0.0073 -0.0082
(0.0028) (0.0025) (0.0061) (0.0064)
Whether household head is male 0.0070 -0.0523 0.2747* 0.2547
(0.0761) (0.0853) (0.1475) (0.1580)
Whether household head has education -0.5842*** -0.5232*** -0.5929*** -0.6160***
(0.1501) (0.1545) (0.1884) (0.2256)
# of farm equipment -0.0472 -0.0686 -0.0635 -0.1031*
(0.0420) (0.0457) (0.0576) (0.0549)
Year = 1998 1.3288* 0.0481 -2.8776* -16.9849***
(0.7177) (0.2869) (1.7369) (0.3017)
Interview month = Feb -0.1218 0.2598 -0.0689 1.2923
(0.3356) (0.3704) (1.2433) (1.0958)
Interview month = Mar 0.2355 0.4944 -0.4943 0.7265
(0.3839) (0.3974) (1.3453) (1.1232)
Interview month = Apr 0.2951 0.5606* -0.2135 1.2458
(0.3150) (0.3349) (1.3272) (1.0668)
Interview month = May 0.3229 0.5060 -0.4877 0.6325
(0.2796) (0.3231) (1.2898) (1.1295)
Interview month = Jun 0.4580* 0.4534 -0.4675 0.4529
(0.2749) (0.2853) (1.3503) (1.1977)
Interview month = Jul 0.6707** 0.9588*** 0.2449 1.4264
(0.2844) (0.2603) (1.2495) (1.1323)
Interview month = Aug 0.3353 0.6005** -0.7194 0.3152
(0.2830) (0.2909) (1.3496) (1.2323)
Interview month = Sep 0.0954 0.4186 -1.2384 -0.0942
(0.2714) (0.3164) (1.3306) (1.2044)
Interview month = Oct 0.2911 0.6010* -0.2570 1.0969
87
(0.3275) (0.3490) (1.3284) (1.1688)
Interview month = Nov -0.0248 0.3036 -0.8926 0.4245
(0.3143) (0.3165) (1.3052) (1.1030)
Interview month = Dec 0.0614 0.4142* -0.3569 0.8658
(0.2808) (0.2483) (1.2169) (1.1010)
% households with land titles in province -0.5666 -0.8546
(0.9110) (2.3556)
Province FE, Year FE x x
Province x Year FE x x
Observations 6336 6336 6276 4579
Pseudo R2 0.1172 0.1767 0.3212 0.2700
This table reports marginal e ects from logit regressions on the determinants of malaria
illness in the household. The dependent variable in columns (1) and (2) is a dummy equal
to one if the household reported at least one household member to have had a malaria
episode in the four weeks prior to the survey. The dependent variable in columns (3) and
(4) is the alternative malaria measure, which in 1998, proxies malaria using a combination
of several symptoms (fever, cough, chill, and headache). Standard errors are robust to
heteroskedasticity and clustered at the province level.
* signi cant at 10%; ** signi cant at 5%; *** signi cant at 1%.
88
Table 3.6: Determinants of malaria illness (Households  xed
e ects)
(1) (2)
Dependent variable Malaria Alt. malaria
Ln(Total land area) -0.0701 0.1091
(0.2807) (0.2768)
Whether household has access to irrigation -1.5543** -0.2623
(0.6717) (0.6713)
Share of good quality land -0.3768 -0.4998
(0.4270) (0.8661)
Share of medium quality land -0.0815 -0.2270
(0.4169) (0.4177)
Share of adult female in household 0.7118 -1.2741
(0.5878) (2.2654)
Share of elderly in household -0.2771 -1.9420*
(0.6912) (1.1229)
Share of children in household 0.8040 0.9777
(0.5361) (1.8000)
Ln(household size) 0.4554 -0.0084
(0.2796) (0.8096)
Age of household head 0.0136* 0.0294
(0.0077) (0.0196)
Whether household head is male 0.3882 -0.3324
(0.2529) (0.6654)
Whether household head has education -0.0205 0.5282
(0.3353) (0.4962)
# of farm equipment 0.1223** 0.0902
(0.0484) (0.2664)
% households with land titles in province -0.3489 -0.3847
(0.8384) (1.7112)
Interview month = Feb -0.5413 -2.5238
(0.6486) (1.8827)
Interview month = Mar 0.0374 -0.7373
(0.6793) (1.7489)
Interview month = Apr 0.3863 -1.9223
(0.6936) (1.8309)
Interview month = May 0.1060 -2.1639
(0.5594) (1.8293)
Interview month = Jun 0.4608 -0.8686
(0.4857) (1.7979)
Interview month = Jul 0.2886 -1.8095
(0.6345) (1.6362)
Interview month = Aug 0.0033 -2.3881
(0.6326) (1.8430)
Interview month = Sep -0.1339 -0.6437
89
(0.5684) (1.4236)
Interview month = Oct 0.2891 -1.9171
(0.5226) (1.6909)
Interview month = Nov -0.4036 -2.1263
(0.6011) (1.5362)
Interview month = Dec -0.4697 -0.3739
(0.5402) (1.2073)
Year = 1998 1.1425 -2.9920**
(0.6963) (1.2320)
Observations 1904 864
Pseudo R2 0.1940 0.7317
This table reports marginal e ects from conditional logit regressions on the determinants
of malaria illness in the household using household  xed e ects. The dependent variable
in column (1) is a dummy equal to one if the household reported at least one household
member to have had a malaria episode in the four weeks prior to the survey. The dependent
variable in column (2) is the alternative narrow malaria measure, which in 1998, proxies
malaria using a combination of several symptoms (fever, cough, chill, and headache).
Standard errors are robust to heteroskedasticity and clustered at the province level.
* signi cant at 10%; ** signi cant at 5%; *** signi cant at 1%.
90
Table 3.7: Relationship between malaria illness and farm
pro ts
(1) (2) (3) (4)
Malaria illness of household members -0.3342*** -0.2881*** -0.1609** -0.2652*
(0.0714) (0.0608) (0.0745) (0.1461)
Ln(Total land area) 0.4987*** 0.5271*** 0.5433*** 0.3307***
(0.0440) (0.0471) (0.0444) (0.1030)
Whether household has access to irrigation 0.0848 0.0649 -0.0616
(0.0926) (0.0950) (0.2498)
Share of good quality land 0.4905*** 0.4795*** -0.3728
(0.0845) (0.0799) (0.2775)
Share of medium quality land 0.2766*** 0.2562*** -0.1120
(0.0731) (0.0688) (0.1985)
Share of adult female in household -0.0084 -0.0511 -0.0780 0.3667***
(0.1062) (0.1064) (0.1019) (0.1295)
Share of elderly in household -0.2378*** -0.2789*** -0.2867*** 0.0039
(0.0795) (0.0783) (0.0718) (0.0027)
Share of children in household -0.1253 -0.1109 -0.0950 -0.0915
(0.0859) (0.0865) (0.0867) (0.0989)
Ln(household size) 0.3477*** 0.3045*** 0.2907*** 0.0243
(0.0510) (0.0551) (0.0510) (0.1373)
Age of household head 0.0018* 0.0021* 0.0022** 0.0462
(0.0011) (0.0011) (0.0010) (0.0345)
Whether household head is male 0.0533* 0.0612* 0.0612* 0.1627
(0.0313) (0.0359) (0.0350) (0.2975)
Whether household head has education 0.1745*** 0.1488*** 0.1317*** 0.4420*
(0.0448) (0.0461) (0.0442) (0.2303)
# of farm equipment 0.1526*** 0.1420*** 0.1455*** 0.1500
(0.0193) (0.0179) (0.0181) (0.1924)
% households with land titles in province 0.0566 0.0274 0.1696
(0.2334) (0.2195) (0.1777)
Year = 1998 0.4577** 0.5281*** 0.7972*** 0.1056
(0.1838) (0.1761) (0.0165) (0.1268)
Constant 2.4667*** 2.1696*** 1.9551*** 4.3186***
(0.3555) (0.3869) (0.3647) (0.8664)
Province FE, Year FE x x
Province x Year FE x
Household FE x
Observations 6185 6031 6031 4282
Adjusted R2 0.4776 0.4903 0.5161 0.7826
This table reports results from ordinary least square regressions estimating the relationship
between farm pro ts and malaria illness in the household. Malaria incidence is measured as
the number of episodes reported by the household, weighted by household size. Standard
errors are robust to heteroskedasticity and clustered at the province level.
* signi cant at 10%; ** signi cant at 5%; *** signi cant at 1%.
91
Table 3.8: Relationship between malaria illness and farm
pro ts, controlling for interview month
(1) (2) (3) (4)
Malaria illness of household members 0.2635 0.3129 0.3993 0.4126
(0.3398) (0.3065) (0.2936) (0.5027)
Ln(Total land area) 0.4948*** 0.5234*** 0.5380*** 0.3265***
(0.0432) (0.0450) (0.0424) (0.0984)
Whether household has access to irrigation 0.1081 0.0684 0.1159
(0.0909) (0.0948) (0.1500)
Share of good quality land 0.5013*** 0.4879*** 0.2020
(0.0835) (0.0743) (0.1605)
Share of medium quality land 0.2684*** 0.2444*** 0.0781
(0.0750) (0.0687) (0.1203)
Share of adult female in household -0.0033 -0.0511 -0.0839 -0.0394
(0.1051) (0.1040) (0.0998) (0.2502)
Share of elderly in household -0.2351*** -0.2804*** -0.2895*** -0.3325
(0.0802) (0.0776) (0.0712) (0.2799)
Share of children in household -0.1224 -0.1053 -0.0992 -0.0648
(0.0868) (0.0870) (0.0890) (0.2055)
Ln(household size) 0.3524*** 0.3074*** 0.2967*** 0.3586***
(0.0508) (0.0543) (0.0501) (0.1232)
Age of household head 0.0018* 0.0022** 0.0023** 0.0042
(0.0010) (0.0010) (0.0010) (0.0027)
Whether household head is male 0.0504 0.0576 0.0561 -0.0719
(0.0323) (0.0363) (0.0366) (0.0983)
Whether household head has education 0.1736*** 0.1521*** 0.1365*** 0.0195
(0.0449) (0.0456) (0.0453) (0.1403)
# of farm equipment 0.1544*** 0.1440*** 0.1486*** 0.0483
(0.0161) (0.0149) (0.0153) (0.0313)
% households with land titles in province 0.0454 0.0242 0.0789
(0.2339) (0.2246) (0.2788)
Interview month = Feb 0.1440 0.1333 0.0760 0.2922
(0.1228) (0.1154) (0.1095) (0.2175)
Interview month = Mar -0.0245 0.0112 -0.0690 0.1720
(0.1101) (0.1121) (0.0994) (0.1814)
Interview month = Apr 0.1874* 0.1994* 0.1443 0.2356
(0.1106) (0.1056) (0.1173) (0.1494)
Interview month = May 0.1967 0.1950* 0.1328 0.2964**
(0.1228) (0.1087) (0.1307) (0.1374)
Interview month = Jun 0.1900 0.2166 0.1796 0.4234**
(0.1334) (0.1337) (0.1587) (0.1597)
Interview month = Jul 0.2840*** 0.3142*** 0.2542** 0.4240***
(0.0932) (0.0901) (0.1234) (0.1484)
Interview month = Aug 0.1259 0.1696 0.0962 0.4166**
(0.1227) (0.1253) (0.1271) (0.1691)
Interview month = Sep 0.3084** 0.3501*** 0.3008** 0.3793**
92
(0.1221) (0.1232) (0.1232) (0.1539)
Interview month = Oct 0.1371 0.1775* 0.1017 0.3754**
(0.1148) (0.1021) (0.1143) (0.1572)
Interview month = Nov 0.1086 0.1208 0.1074 0.2603
(0.1209) (0.1165) (0.1158) (0.1605)
Interview month = Dec 0.0675 0.0638 -0.0272 0.2261
(0.0682) (0.0700) (0.0605) (0.1454)
Interaction with HH malaria illness
Interview month = Feb -0.7102 -0.8274* -0.7589* -1.2945**
(0.4739) (0.4450) (0.4382) (0.5207)
Interview month = Mar -0.4951 -0.5037 -0.4443 -0.8498
(0.3366) (0.3249) (0.3029) (0.5352)
Interview month = Apr -0.4080 -0.4420 -0.3135 -0.2467
(0.3429) (0.2963) (0.2927) (0.5210)
Interview month = May -0.8039** -0.8043** -0.6769** -0.6994
(0.3999) (0.3369) (0.3218) (0.7746)
Interview month = Jun -0.7196* -0.7690** -0.7801** -0.9195
(0.4102) (0.3738) (0.3687) (0.6586)
Interview month = Jul -0.8112** -0.6702** -0.3965 -1.0860*
(0.3643) (0.3190) (0.4359) (0.6010)
Interview month = Aug -0.6669* -0.6389* -0.8736** -0.2529
(0.3782) (0.3442) (0.3342) (0.6964)
Interview month = Sep -0.1663 -0.3046 -0.3791 -0.7023
(0.3942) (0.3848) (0.3694) (0.6229)
Interview month = Oct -0.5311 -0.5423 -0.5260 -0.5149
(0.4657) (0.4088) (0.3914) (0.8859)
Interview month = Nov -0.5287 -0.6485* -0.5506* -0.6700
(0.3858) (0.3750) (0.3257) (0.8055)
Interview month = Dec -0.9086** -0.9008** -0.7494** -0.8687
(0.3622) (0.3598) (0.3318) (0.5530)
Year = 1998 0.4584** 0.5202*** 0.5455*** 0.4871**
(0.1821) (0.1818) (0.0889) (0.2260)
Constant 2.3575*** 2.0014*** 2.0016*** 4.0480***
(0.3592) (0.3925) (0.3770) (0.8533)
Test for signi cance of variables
by interview month
Interview month = Feb
Sum of coe cients -0.4467 -0.5145 -0.3596 -0.8820
p-value 0.1450 0.0901 0.1909 0.0218
Interview month = Mar
Sum of coe cients -0.2316 -0.1908 -0.0450 -0.4373
p-value 0.2841 0.4865 0.8670 0.4071
Interview month = Apr
Sum of coe cients -0.1445 -0.1291 0.0858 0.1659
p-value 0.4193 0.3917 0.6136 0.5818
Interview month = May
Sum of coe cients -0.54048 -0.4914 -0.2776 -0.2869
93
p-value 0.0407 0.0137 0.1281 0.6357
Interview month = Jun
Sum of coe cients -0.4561 -0.4561 -0.3808 -0.5069
p-value 0.0383 0.0260 0.0895 0.2252
Interview month = Jul
Sum of coe cients -0.5477 -0.3573 0.0028 -0.6735
p-value 0.0001 0.0019 0.9932 0.0477
Interview month = Aug
Sum of coe cients -0.4034 -0.3260 -0.4744 0.1597
p-value 0.0055 0.0305 0.0085 0.7358
Interview month = Sep
Sum of coe cients 0.0972 0.0083 0.0201 -0.2897
p-value 0.5924 0.9684 0.9208 0.3428
Interview month = Oct
Sum of coe cients -0.2676 -0.2294 -0.1267 -0.1024
p-value 0.3103 0.3191 0.5893 0.8558
Interview month = Nov
Sum of coe cients -0.2652 -0.3356 -0.1513 -0.2575
p-value 0.1995 0.1061 0.3924 0.5085
Interview month = Dec
Sum of coe cients -0.6451 -0.5879 -0.3502 -0.4562
p-value 0.0053 0.0204 0.1215 0.4583
Province FE, Year FE x x
Province x Year FE x
Household FE x
Observations 6185 6031 6031 4282
Adjusted R2 0.4830 0.4959 0.5218 0.5693
This table reports results from ordinary least square regressions estimating the relationship
between farm pro ts and malaria illness in the household. Malaria incidence is measured as
the number of episodes reported by the household, weighted by household size. Compared
to Table 4, speci cations additionally control for the month the survey was administered on
the household, to account for the seasonality of the disease and the fact that households
reported on illness episodes in the four weeks prior to the survey. Standard errors are
robust to heteroskedasticity and clustered at the province level. * signi cant at 10%; **
signi cant at 5%; *** signi cant at 1%.
94
Table 3.9: Relationship between pro ts and malaria illness
using alternative malaria measure
(1) (2) (3) (4)
Malaria illness of household members -0.5656*** -0.4911*** -0.1390 -0.6651***
(0.1259) (0.1153) (0.2179) (0.2279)
Ln(Total land area) 0.4984*** 0.5266*** 0.5428*** 0.3295***
(0.0443) (0.0472) (0.0446) (0.0990)
Whether household has access to irrigation 0.0637 0.0651 0.0842
(0.0951) (0.0934) (0.1616)
Share of good quality land 0.4901*** 0.4819*** 0.1528
(0.0845) (0.0804) (0.1754)
Share of medium quality land 0.2695*** 0.2563*** 0.0764
(0.0737) (0.0697) (0.1221)
Share of adult female in household -0.0171 -0.0596 -0.0824 -0.0499
(0.1066) (0.1068) (0.1010) (0.2541)
Share of elderly in household -0.2439*** -0.2846*** -0.2891*** -0.3564
(0.0791) (0.0780) (0.0708) (0.2805)
Share of children in household -0.1447* -0.1277 -0.1039 -0.1270
(0.0854) (0.0866) (0.0854) (0.1970)
Ln(household size) 0.3504*** 0.3071*** 0.2926*** 0.3617***
(0.0513) (0.0554) (0.0513) (0.1258)
Age of household head 0.0018 0.0021* 0.0022** 0.0041
(0.0011) (0.0011) (0.0010) (0.0027)
Whether household head is male 0.0562* 0.0637* 0.0619* -0.0834
(0.0316) (0.0359) (0.0351) (0.0996)
Whether household head has education 0.1733*** 0.1478*** 0.1339*** 0.0174
(0.0457) (0.0463) (0.0444) (0.1353)
# of farm equipment 0.1526*** 0.1424*** 0.1461*** 0.0442
(0.0192) (0.0179) (0.0181) (0.0337)
% households with land titles in province 0.0858 0.0503 0.1781
(0.2289) (0.2164) (0.2857)
Year = 1998 0.3935** 0.4736*** 0.7901*** 0.3847*
(0.1784) (0.1736) (0.0194) (0.2232)
Constant 2.4912*** 2.2144*** 1.9556*** 4.4337***
(0.3555) (0.3836) (0.3608) (0.8578)
Province FE, Year FE x x
Province x Year FE x
Household FE x
Observations 6185 6031 6031 4282
Adjusted R2 0.4773 0.4900 0.5156 0.5647
This table is analogous to Table 4 except that the alternative malaria proxy is used,
which proxies malaria using a combination of several symptoms (fever, cough, chill, and
headache). Malaria incidence is measured as the number of episodes reported by the
household, weighted by household size. Standard errors are robust to heteroskedastic-
ity and clustered at the province level. * signi cant at 10%; ** signi cant at 5%; ***
signi cant at 1%.
95
Table 3.10: Malara incidence and total agricultural labor use
of household
(1) (2) (3) (4)
Malaria illness of household members 0.0083 0.0389 -0.0164 0.1417
(0.0704) (0.0760) (0.0708) (0.1371)
Ln(Total land area) 0.2195*** 0.2300*** 0.2360*** 0.1355*
(0.0322) (0.0351) (0.0367) (0.0709)
Whether household has access to irrigation 0.1788* 0.1915* 0.1202
(0.0918) (0.1089) (0.1102)
Share of good quality land 0.0965 0.0686 0.1905
(0.0803) (0.0825) (0.1527)
Share of medium quality land 0.1011 0.0603 0.0772
(0.0738) (0.0696) (0.1143)
Share of adult female in household 0.0339 0.0365 0.0065 0.0153
(0.0749) (0.0749) (0.0714) (0.2495)
Share of elderly in household -0.6043*** -0.6096*** -0.6273*** -0.5408**
(0.0892) (0.0909) (0.0842) (0.2505)
Share of children in household -0.6169*** -0.6044*** -0.6564*** -0.3414**
(0.0725) (0.0754) (0.0675) (0.1699)
Ln(household size) 0.6660*** 0.6508*** 0.6420*** 0.6469***
(0.0411) (0.0428) (0.0417) (0.1027)
Age of household head 0.0023*** 0.0021*** 0.0014 0.0042
(0.0008) (0.0008) (0.0009) (0.0030)
Whether household head is male 0.0389* 0.0452* 0.0309 0.1149
(0.0224) (0.0230) (0.0235) (0.0953)
Whether household head has education 0.0389 0.0231 0.0359 -0.0027
(0.0341) (0.0347) (0.0345) (0.1237)
# of farm equipment 0.0840*** 0.0790*** 0.0778*** 0.0792**
(0.0143) (0.0151) (0.0117) (0.0390)
% households with land titles in province -0.1310 -0.1241 -0.0764
(0.2515) (0.2590) (0.3703)
Year = 1998 -0.2535 -0.2521 -0.0406** -0.2347
(0.1856) (0.1901) (0.0171) (0.2861)
Constant 3.5823*** 3.3346*** 3.1952*** 3.9071***
(0.2501) (0.2962) (0.3017) (0.5752)
Province FE, Year FE x x
Province x Year FE x
Household FE x
Observations 6045 5894 5894 4132
Adjusted R2 0.3926 0.3944 0.4626 0.3817
This table reports results from ordinary least square regressions estimating the relation-
ship between total number of household and non-household (hired and exchanged) labor
days employed by the household, and malaria illness experienced by household mem-
bers. Malaria incidence is measured as the number of episodes reported by the household,
weighted by household size. Standard errors are robust to heteroskedasticity and clustered
at the province level. * signi cant at 10%; ** signi cant at 5%; *** signi cant at 1%.
96
Table 3.11: Malaria incidence and total labor use of house-
hold members only (excluding non-household labor)
(1) (2) (3) (4)
Malaria illness of household members 0.0457 0.0630 0.0839* 0.0743
(0.0557) (0.0561) (0.0476) (0.1121)
Ln(Total land area) 0.1609*** 0.1641*** 0.1723*** 0.0645
(0.0243) (0.0279) (0.0277) (0.0724)
Whether household has access to irrigation 0.1946*** 0.1559* 0.1844
(0.0683) (0.0808) (0.1172)
Share of good quality land -0.0193 -0.0713 0.1582
(0.0737) (0.0723) (0.1479)
Share of medium quality land 0.0014 -0.0333 0.0318
(0.0512) (0.0538) (0.1008)
Share of adult female in household 0.0047 0.0094 -0.0125 -0.0523
(0.0662) (0.0671) (0.0667) (0.2193)
Share of elderly in household -0.6723*** -0.6787*** -0.6870*** -0.7245***
(0.0887) (0.0898) (0.0876) (0.2567)
Share of children in household -0.7820*** -0.7738*** -0.7775*** -0.5257***
(0.0597) (0.0600) (0.0629) (0.1352)
Ln(household size) 0.7767*** 0.7702*** 0.7583*** 0.8373***
(0.0314) (0.0322) (0.0316) (0.0931)
Age of household head 0.0021*** 0.0019*** 0.0019** 0.0010
(0.0007) (0.0007) (0.0008) (0.0024)
Whether household head is male 0.0376* 0.0407* 0.0292 0.0839
(0.0224) (0.0222) (0.0219) (0.0865)
Whether household head has education 0.0443 0.0308 0.0371 0.0653
(0.0356) (0.0361) (0.0356) (0.0960)
# of farm equipment 0.0652*** 0.0628*** 0.0646*** 0.0397
(0.0135) (0.0139) (0.0117) (0.0349)
% households with land titles in province -0.0275 -0.0520 -0.0482
(0.1937) (0.1924) (0.3045)
Year = 1998 -0.0903 -0.0757 -0.0437*** -0.0348
(0.1496) (0.1461) (0.0140) (0.2383)
Constant 3.8281*** 3.6306*** 3.5947*** 4.2544***
(0.1981) (0.2384) (0.2315) (0.5967)
Province FE, Year FE x x
Province x Year FE x
Household FE x
Observations 5961 5811 5811 4058
Adjusted R2 0.4469 0.4454 0.4826 0.4738
This table reports results from ordinary least square regressions estimating the relationship
between total number of labor days worked by household members only, and malaria illness
among household members. Malaria incidence is measured as the number of episodes
reported by the household, weighted by household size. Standard errors are robust to
heteroskedasticity and clustered at the province level. * signi cant at 10%; ** signi cant
at 5%; *** signi cant at 1%.
97
Table 3.12: Total labor use regressions with alternative
malaria measure
(1) (2) (3) (4)
Malaria illness of household members 0.0069 0.1103 0.0835 0.1063
(0.1538) (0.1538) (0.0915) (0.3602)
Ln(Total land area) 0.2195*** 0.2300*** 0.2357*** 0.1343*
(0.0325) (0.0353) (0.0370) (0.0697)
Whether household has access to irrigation 0.1866* 0.1926 0.1130
(0.0854) (0.1089) (0.1202)
Share of good quality land 0.0981 0.0700 0.1882
(0.0791) (0.0821) (0.1502)
Share of medium quality land 0.1039 0.0612 0.0782
(0.0729) (0.0694) (0.1176)
Share of adult female in household 0.0340 0.0375 0.0077 0.0182
(0.0749) (0.0748) (0.0712) (0.2475)
Share of elderly in household -0.6043*** -0.6083*** -0.6260*** -0.5463**
(0.0896) (0.0912) (0.0842) (0.2507)
Share of children in household -0.6166*** -0.6023*** -0.6558*** -0.3308*
(0.0729) (0.0762) (0.0678) (0.1679)
Ln(household size) 0.6659*** 0.6507*** 0.6427*** 0.6482***
(0.0412) (0.0429) (0.0419) (0.1037)
Age of household head 0.0023** 0.0021* 0.0014 0.0042
(0.0008) (0.0008) (0.0009) (0.0030)
Whether household head is male 0.0389 0.0446 0.0309 0.1163
(0.0225) (0.0229) (0.0235) (0.0944)
Whether household head has education 0.0387 0.0244 0.0377 -0.0020
(0.0359) (0.0363) (0.0355) (0.1255)
# of farm equipment 0.0839*** 0.0790*** 0.0781*** 0.0800**
(0.0143) (0.0151) (0.0118) (0.0390)
% households with land titles in province -0.1315 -0.1277 -0.0836
(0.2498) (0.2578) (0.3725)
Year = 1998 -0.2523 -0.2421 -0.0380* -0.2181
(0.1834) (0.1881) (0.0182) (0.2850)
Constant 3.5822*** 3.3208*** 3.1881*** 3.9155***
(0.2496) (0.2926) (0.3002) (0.5554)
Province FE, Year FE x x
Province x Year FE x
Household FE x
Observations 6045 5894 5894 4132
Adjusted R2 0.3926 0.3945 0.4627 0.3811
This table reports results from ordinary least square regressions estimating the relation-
ship between total number of labor days employed by the household and an alternative
malaria measure, which proxies malaria using a combination of several symptoms (fever,
cough, chill, and headache). Malaria incidence is measured as the number of episodes
reported by the household, weighted by household size. Standard errors are robust to
heteroskedasticity and clustered at the province level.
98
Table 3.13: E ect of malaria incidence at the household- and
province-level in 1992 on crop choice in 1998
(1) (2) (3)
Malaria illness of household members in 1992 -0.0730 -0.0560 -0.1225
(0.0739) (0.0691) (0.0803)
Malaria cases per 1000 in population in 1992 -0.0028***
(0.0009)
Ln(Total land area) 0.0422** 0.0562*** 0.0475***
(0.0170) (0.0143) (0.0157)
Whether household has access to irrigation 0.0161 0.0121
(0.0403) (0.0452)
Share of good quality land 0.1792*** 0.1599***
(0.0530) (0.0493)
Share of medium quality land 0.1351*** 0.1153***
(0.0309) (0.0333)
Share of adult female in household 0.0494 0.0440 0.0503
(0.0322) (0.0338) (0.0396)
Share of elderly in household 0.0009 0.0029 0.0165
(0.0314) (0.0314) (0.0371)
Share of children in household 0.0359 0.0478* 0.0764**
(0.0287) (0.0283) (0.0365)
Ln(Household size) 0.0334** 0.0151 0.0003
(0.0166) (0.0175) (0.0222)
Age of household head -0.0003 -0.0002 0.0005
(0.0005) (0.0005) (0.0006)
Whether household head is male 0.0082 0.0100 0.0142
(0.0135) (0.0128) (0.0117)
Whether household head has education -0.0606* -0.0527 0.0151
(0.0333) (0.0319) (0.0463)
# of farm equipment -0.0031 -0.0063 -0.0044
(0.0093) (0.0091) (0.0091)
% households with land titles in province 0.0590
(0.0452)
Constant 0.1036 -0.0459 0.2150
(0.1515) (0.1211) (0.1634)
Observations 2571 2571 2571
Adjusted R2 0.3005 0.3289 0.1205
This table reports ordinary least sqaure results on how malaria incidence reported by the
household or at the province-level a ects households decision to allocate land to crops
of di erent labor intensity 5-6 years later. The dependent variable measures the share
of land each household devoted to highly labor- intensive, high-return crops such as rice
and industrial crops such as peanut and soybean. Malaria incidence of the household is
measured as the number of episodes reported by the household, weighted by household
size. Province malaria measures the number of reported, slide-positive malaria cases per
1000 people in the population. Speci cations in columns (1) and (2) include province  xed
e ects. Standard errors are robust to heteroskedasticity and clustered at the province level.
99
Bibliography
Adepoju, A. and C. Oppong (1994). Gender, Work and Population in Sub-Saharan
Africa. International Labor Organization.
Ainsworth, M. (1992). Economic Aspects of Child Fostering in C^ote d?Ivoire. World
Bank, LSMS Working Paper No. 92.
Alderman, H., J. Hoddinott, and B. Kinsey (2006). Long Term Consequences of
Early Child Malnutrition. Oxford Economic Papers 58, 450{474.
Alderman, H. and C. Paxson (1992). Do the Poor Insure? A Synthesis of the
Literature on Risk and Consumption in Developing Countries. World Bank Policy
Research Working Paper No. 1008.
Antle, J. and P. Pingali (1994). Pesticides, Productivity, and Farmer Health: A
Philippine Case Study. American Journal of Agricultural Economics 76, 418{
430.
Artadi, E. V. (2005). Going into Labor: Earnings vs. Infant Survival in Rural Africa.
Working paper.
Audibert, M., J.-F. Brun, J. Mathonnat, and M. Henry (2009). Malaria, Production
and Income of the Producers of Co ee and Cocoa: an Analysis from Survey Data
in C^ote d?Ivoire. CERDI, Working Paper Series, E 2006.31.
Audibert, M. and J.-F. Etard (2003). Productive Bene ts after Investment in Health
in Mali. Economic Development and Cultural Change 51, 769{782.
Audibert, M., J. Mathonnat, and M.-C. Henry (2003). Social and Health Deter-
minants of the E ciency of Cotton Farmers in Northern C^ote d?Ivoire. Social
Science & Medicine 56, 1705{1717.
Baker, M. and K. Milligan (2008). Maternal Employment, Breastfeeding and Health:
Evidence from Maternity Leave Mandates. Journal of Health Economics 27(4),
871{887.
Baldwin, R. and B. Weisbrod (1974). Disease and Labor Productivity. Economic
Development and Cultural Change 22, 414{435.
Barreca, A. (2010). The Long Term Impact of In Utero and Postnatal Exposure to
Malaria. Journal of Human Resources Forthcoming.
Basta, S. S., M. Soekirman, D. Karyadi, and N. Scrimshaw (1979). Iron De ciency
Anemia and the Productivity of Adult males in Indonesia. American Journal of
Clinical Nutrition 32, 916{25.
Behrman, J., A. Foster, and M. Rosenzweig (1997). The Dynamics of Agricultural
Production and the Calorie-Income Relationship: Evidence from Pakistan. Jour-
nal of Econometrics 77, 187{207.
100
Benjamin, D. (1992). Household Composition, Labor Markets, and Labor Demand:
Testing for Separation in Agricultural Household Models. Econometrica 60(2),
287{322.
Benjamin, D. and L. Brandt (2002). Agriculture and Income Distribution in Rural
Vietnam under Economic Reforms: A Tale of Two Regions. University of Toronto
Working Paper.
Besley, T. (1995). Property Rights and Investment Incentives: Theory and Evidence
from Ghana. Journal of Political Economy 103, 903{937.
Betran, A., M. de Onis, J. A. Lauer, and J. Villar (2003). Ecological Study of
E ect of Breastfeeding on Infant Mortality in Latin America. British Medical
Journal 323, 1{5.
Bhandari, N., R. Bahl, S. Mazumder, J. Martines, R. Black, M. Bhan, and the
other members of the Infant Feeding Study Group (2003). E ect of Community-
based Promotion of Exclusive Breastfeeding on Diarrhoeal Illness and Growth: A
Cluster Randomised Controlled Trial. The Lancet 361, 1418{23.
Black, R., S. Morris, and J. Bryce (2003). Where and Why are 10 Million Children
Dying Every Year? The Lancet 361, 2226{34.
Bleakley, H. (2010). Malaria Eradication in the Americas: A Retrospective Analysis
of Childhood Exposure. American Economic Journal: Applied Economics 2, 1{45.
Boschi-Pinto, L. V. and K. Shibuya (2008). Estimating Child Mortality due to
Diarrhoea in Developing Countries. Bulletin of the World Health Organization 86,
710{717.
Briend, A., B. Wojtyniak, and M. Rowland (1988). Breast feeding, Nutritional State,
and Child Survival in Rural Bangladesh. British Medical Journal 296, 879{882.
Chima, R. I., C. Goodman, and A. Mills (2003). The Economic Impact of Malaria
in Africa: a Critical Review of the Evidence. Health Policy 63, 17{36.
Cleland, J. G. and J. K. van Ginneken (1988). Maternal Education and Child
Survival in Developing Countries: The Search for Pathways of In uence. Social
Science Medicine 27(12), 1357{1368.
Clemens, J. D., M. Rao, J. Chakraborty, M. Yunus, M. Ali, B. Kay, F. P. L. van
Loon, A. Na cy, , and D. Sack (1997). Breastfeeding and the Risk of Life-
threatening Enterotoxigenic coli Diarrhea in Bangladeshi Infants and Children.
Pediatrics 100(6), e2.
Conley, G. N. (1972). The Impact of Malaria on Economic Development: A Case
Study. The American Journal of Tropical Medicine and Hygiene 21, 668{674.
101
Cropper, M., M. Haile, J. Lampietti, C. Poulos, and D. Whittington (1999). The
Value of Preventing Malaria in Tembien, Ethiopia. Washington DC: The World
Bank.
Cutler, D., W. Fung, M. Kremer, M. Singhal, and T. Vogl (2007). Mosquitoes: The
Long-term E ects of Malaria Eradication in India. NBER Working paper No.
13539.
Danso, G., O. Co e, L. Annang, E. Obuobie, and B. Keraita (2004). Gender and
Urban Agriculture: The Case of Accra, Ghana. International Water Management
Institute.
Deaton, A. (1992). Household Saving in LDC?s Credit Markets, Insurance, and
Welfare. Scandinavian Journal of Economics 94.
Deolalikar, A. (1988). Nutrition and Labor Productivity in Agriculture: Estimates
for Rural South India. Review of Economics and Statistics 70, 406{413.
Dercon, S. (1996). Risk, Crop Choice, and Savings: Evidence from Tanzania. Eco-
nomic Development and Cultural Change 44, 485{513.
Dercon, S. (2004). Growth and Shocks: Evidence from Rural Ethiopia. Journal of
Development Economics 74, 309{329.
Dercon, S. and J. Hoddinott (2003). Health, Shocks and Poverty Persistence. World
Institute for Development Economics Research Discussion Paper No. 2003/08.
Dercon, S. and P. Krishnan (2000). In Sickness and in Health: Risk Sharing within
Households in Rural Ethiopia. Journal of Political Economy 108(4), 688{727.
Do, Q.-T. and L. Iyer (2007). Land Titling and Rural Transition in Vietnam. World
Bank Policy Research Working Paper.
Du o, E. and C. Udry (2004). Intrahousehold Resource Allocation in C^ote d?Ivoire:
Social Norms, Separate Accounts, and Consumption Choices. NBER Working
paper, No.10498.
Ettling, M. (2002). The Control of Malaria in Vietnam from 1980 to 2000: What
went right? WHO Regional O ce for the Western Paci c.
Field, E. (2007). Entitled to Work: Urban Property Rights and the Labor Supply
in Peru. Quarterly Journal of Economics 122, 1561{1602.
Fisher, M. and S. Datta (2007). Malaria Risk, Crop Choice, and Agricultural Income
in Northern Tanzania. Abstract accessed at http://www.aaea.org/2007am/
abstracts/default.asp?mode=detail&abstractid=174603 on March 13, 2008.
Food and Agriculture Organization (1999a). Aquastat: Viet Nam. http://www.
fao.org/nr/water/aquastat/countries/viet_nam/index.stm.
102
Food and Agriculture Organization (1999b). Deciduous Fruit Production in Asia
and the Paci c. FAO/RAP Publication: 1998/10.
Foster, A. (1995). Prices, Credit Markets and Child Growth in Low-Income Rural
Areas. Economic Journal 105, 551{70.
Gallup, J. L. and J. D. Sachs (2001). The Economic Burden of Malaria. American
Journal of Tropical and Medical Hygiene 64, 85{96.
Gertler, P. and J. Gruber (2002). Insuring Consumption against Illness. American
Economic Review 92, 51{70.
Goldstein, M. and C. Udry (2008). The Pro ts of Power: Land Rights and Agricul-
tural Investment in Ghana. Journal of Political Economy 116, 981{1022.
Grimard, F. (1997). Household Consumption Smoothing through Ethnic Ties: Ev-
idence from C^ote d?Ivoire. Journal of Development Economics 53, 391{422.
Grossman, M. (1972). On the Concept of Health Capital and the Demand for Health.
The Journal of Political Economy 80, 223{255.
Hawkes, C. and M. Ruel (2006). The Links between Agriculture and Health: an
Intersectoral Opportunity to improve the Health and Livelihoods of the Poor.
Bulletin of the World Health Organization 84.
Hoddinott, J. and B. Kinsey (2001). Child Health in the Time of Drought. Oxford
Bulletin of Economics and Statistics 63, 409{436.
Hong, S. C. (2007). Malaria and Economic Productivity: The American Case.
University of Chicago mimeo.
Hung, L. Q., peter J. de Vries, P. T. Giao, N. V. Nam, T. Q. Binh, M. Chong,
N. Quoc, T. Thanh, L. Hung, and P. Kager (2002). Control of Malaria: a Suc-
cessful Experience from Viet Nam. Bulletin of the World Health Organization 80,
660{666.
International Rice Research Institute, Rice statistics, Info by country (n.d.). http:
//www.irri.org/science/cnyinfo/vietnam.asp.
Jacoby, H. and E. Skou as (1997). Risk, Financial Markets, and Human Capital in
a Developing Country. Review of Economic Studies 64.
Jensen, R. (2000). Agricultural Volatility and Investments in Children. American
Economic Review, Papers and Proceedings 90, 399{404.
Kim, A., A. Tandon, A. Hailu, and Others (1997). Health and Labor Productivity:
The Economic Impact of Onchocercal Skin Disease. World Bank Policy Research
Working Paper No. 1836.
103
Kirk, M. and N. D. A. Tuan (2009). Land-Tenure Policy Reforms: Decollectivization
and the Doi Moi System in Vietnam. International Food Policy Research Institute
Discussion Paper 00937.
Kozel, V. (1990). The Composition and Distribution of Income in C^ote d?Ivoire.
Technical Report, World Bank LSMS Working Paper No. 68.
Kramer, M., B. Chalmers, E. Hodnett, and et al (2001). Promotion of Breastfeeding
Intervention Trial (PROBIT): A Randomized Trial in the Republic of Belarus.
Journal of the American Medical Association 285(4), 413{420.
Laxminarayan, R. and K. Moeltner (2003). Malaria, Adaptation and Crop Choice.
Working paper.
Leighton, C. and R. Foster (1994). Economic Impacts of Malaria in Kenya and
Nigeria. Abt Associates, Health Financing and Sustainability Project, Bethesda,
Maryland.
Levine, D. and D. Yang (2006). A Note on the Impact of Local Rainfall on Rice
Output in Indonesian Districts. mimeo.
Lindberg, L. D. (1996). Women?s Decisions about Breastfeeding and Maternal Em-
ployment. Journal of Marriage and Family 58(1), 239{251.
Liu, G. G., W. H. Dow, A. Z. Fu, J. Akin, and P. Lance (2008). Income Productivity
in China: On the Role of Health. Journal of Health Economics 27, 27{44.
Lopez, R. (1984). Estimating Labour Supply and Production Decisions of Self-
employed Farm Producers. European Economic Review 24, 61{82.
Maccini, S. and D. Yang (2009). Under the Weather: Health, Schooling and Eco-
nomic Consequences of Early-Life Rainfall. American Economic Review 99(3),
1006{26.
Macro International Inc. (1996). Sampling Manual. DHS-III Basic Documentation
No. 6 Calverton, Maryland.
McCaig, B., D. Benjamin, and L. Brandt (2009). The Evolution of Income Inequality
in Vietnam, 1993-2006 (data appendix). Working Paper.
Morel, C. M., N. D. Thang, N. X. Xa, L. X. Hung, L. K. Thuan, P. V. Ky, A. Erhart,
A. J. Mills, and U. D?Alessandro (2008). The Economic Burden of Malaria on the
household in South-Central Vietnam. Malaria Journal 7.
Morrow, A., M. L. Guerrero, J. Shults, J. Calva, C. Lutter, J. Bravo, G. Ruiz-
Palacios, R. Morrow, and F. Butterfoss (1999). E cacy of Home-based Peer
Counselling to Promote Exclusive Breastfeeding: A Randomised Controlled Trial.
Lancet 353, 1226{1231.
104
Nam, N. V., P. de Vries, L. V. Toi, and N. Nagelkerke (2005). Malaria Control
in Vietnam: the Binh Thuan Experience. Tropical Medicine and International
Health 10, 357{365.
Noble, S. and The ALSPAC Study Team (2001). Maternal Employment and the
Initiation of Breastfeeding. Acta Paediatrica 90(4), 423{428.
Pampana, E. (1963). A Textbook of Malaria Eradication. London: Oxford University
Press.
Paxson, C. H. (1992). Using Weather Variability to Estimate the Response of
Savings to Transitory Income in Thailand. American Economic Review 82(1),
15{33.
Pitt, M. and M. Rosenzweig (1986). Agricultural Prices, Food Consumption and
the Health and Productivity of Indonesian Farmers. In Agricultural Household
Models: extensions, Applications, and Policy. Johns Hopkins University Press,
Baltimore.
Pitt, M. and W. Sigle (1997). Seasonality, Weather Shocks and the Timing of Births
and Child Mortality in Senegal. Working Paper.
Sahn, D. and H. Alderman (1988). The E ect of Human Capital on Wages and the
Determinants of Labor Supply in a Developing Country. Journal of Development
Economics 29, 157{183.
Schultz, T. (2002). Wage Gains associated with Height as a Form of Health Human
Capital. Economic Center Discussion Paper, Yale University.
Schultz, T. and A. Tansel (1997). Wage and Labor Supply E ects of Illness in C^ote
d?Ivoire and Ghana: Instrumental Variable Estimates for Days Disabled. Journal
of Development Economics 53, 251{286.
Strauss, J. (1986). Does Better Nutrition raise Farm Productivity? Journal of
Political Economy 94, 297{320.
Townsend, R. (1994). Risk and Insurance in Village India. Econometrica 62, 539{91.
United Nations Population Division (2008). World Population Prospects.
http://esa.un.org/unpp/index.asp.
Victora, C., P. Vaughan, C. Lombardi, S. Fuchs, L. Gigante, P. Smith, L. Nobre,
A. M. Texiera, L. Moreira, and F. Barros (1987). Evidence for a Strong Protective
E ect of Breastfeeding against Infant Death due to Infectious Diseases in Brazil.
Lancet 330(8554), 318{322.
Weisbrod, B. and T. Helminiak (1977). Parasitic Diseases and Agricultural Labor
Productivity. Economic Development and Cultural Change 25, 505{522.
105
WHO (2000). Who Collaborative Study Team on the Role of Breastfeeding in
the Prevention of Infant and Child Mortality due to Infectious Diseases in Less
Developed Countries: A Pooled Analysis. Lancet 355, 451{455.
WHO (2008). The Global Burden of Disease: 2004 update. World Health Organi-
zation.
World Bank (2009). Gender in Agriculture Sourcebook. World Bank: Washington
DC.
Yoon, P. W., R. E. Black, L. H. Moulton, and S. Becker (1996). E ect of Not
Breastfeeding on the Risk of Diarrheal and Respiratory Mortality in Children
under 2 Years of Age in Metro Cebu in the Philippines. American Journal of
Epidemiology 143(11), 1142{1148.
106