ESSAYS ON HEALTHDETERMINANTS IN KENYA
Japheth Osotsi Awiti
A Thesis Submitted in Partial Fulfilment of the Requirements for the
Degree of Doctor of Philosophy in Economics in the University of Nairobi
2013
i
Declaration
ii
Acknowledgements
Generous financial assistance from the African Economic Research Consortium
(AERC), both towards the coursework component of my PhD studies under the Col-
laborative PhD Programme in Economics (CPP) and towards the research portion of
the PhD is acknowledged and appreciated. I am grateful for comments and guidance
from my supervisors, Professor Germano Mwabu and Professor Jane Kabubo–Mariara.
I would also like to recognize and appreciate the efforts of all the lecturers who taught
me in the coursework component of the PhD, both at the University of Dar es Salaam
and at the Joint Facility for Electives (JFE) in Nairobi. I would also like to thank my
classmates with whom I held very fruitful discussions. I am grateful for comments from
the resource persons and participants at AERC biannual research workshops where the
proposal and the work–in–progress for this thesis were presented. I am also grateful for
comments from participants at the School of Economics seminar where this work was
presented. Last but not least, I would like to appreciate the sacrifices made by my family
during the entire period of the PhD studies and especially during the thesis writing stage.
iii
Preface
This thesis is organized around three essays dealing with health determinants in
Kenya. The first essay investigates the effect of prenatal care use on the probability of
delivering a low birth weight infant. The second essay investigates the effect of preceding
birth interval length on the probability of a mother experiencing a miscarriage, stillbirth,
or an abortion. The third and final essay investigates the effect of smoking on self–rated
health status. The three essays give us an idea about the important health determinants
in Kenya.
The key contribution of the thesis to knowledge and policy is a demonstration that
failure to control for unobserved mother–specific effects, for instance via a multi–level
modelling framework, leads to an overstatement of the beneficial effects of prenatal
care use on birth outcomes. The second main contribution is a demonstration that
preceding birth intervals of length 36 to 59 months only improve maternal health after
the endogeneity of preceding birth interval has been controlled for. The third contribution
is showing that failure to control for sample selection bias, endogeneity of smoking and
unobserved heterogeneity leads to an understatement of the negative effects of smoking
on self–rated health status by about 50%.
The thesis is organized around 5 chapters. Chapter 1 introduces the thesis, Chapter
2 presents the first essay, Chapter 3 the second essay, Chapter 4 the third essay, while
Chapter 5 summarizes the thesis.
iv
Abstract
We investigate the effect of prenatal care use on infant health, the effect
of preceding birth interval length on maternal health, and the effect
of smoking on general health. We employ an estimation strategy that
controls for potential endogeneity of the key covariates, potential un-
observed heterogeneity, and potential sample selection bias. We obtain
three main results. First, models that do not control for unobserved
mother–specific effects overstate the beneficial effects of prenatal care on
infant health. We particularly find that after controlling for unobserved
mother–specific effects, adequate use of prenatal care decreases the prob-
ability of delivering a low–birth weight infant by 0.036, holding other
factors constant. Without such control, however, the corresponding
reduction would have been 0.26. Second, preceding birth interval is
an endogenous determinant of maternal health. In particular, we find
that preceding birth intervals of length 36 to 59 months can only be
shown to improve maternal health after we control for the endogeneity
of preceding birth interval. Third, failure to control for sample selection
bias, endogeneity of smoking and unobserved heterogeneity leads to an
understatement of the negative effects of smoking on self–rated health
status by about 50%. In particular, we find that after controlling for
sample selection bias, endogeneity of smoking and unobserved hetero-
geneity, the probability of individuals who smoke rating their own health
as “Poor” compared to their age–mates is higher than that of those who
do not smoke by 0.018, holding other factors constant. Without such
controls, however, the corresponding difference in probabilities is only
0.009. Our results imply that policies that promote adequate use of
prenatal care services, those that promote adequate spacing of births,
and those that discourage smoking should be pursued so as to improve
the health of the Kenyan people.
Contents
Declaration i
Acknowledgements ii
Preface iii
Abstract iv
List of figures vii
List of tables viii
1 Introduction 1
1.1 Definition and Importance of Health . . . . . . . . . . . . . . . . . . . . 1
1.2 Measurement of Health . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Health Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Key Health Indicators for Kenya . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Purpose and Objectives of the Thesis . . . . . . . . . . . . . . . . . . . . 7
1.6 General Modelling Framework for the Thesis . . . . . . . . . . . . . . . . 9
1.7 Contributions of the Thesis to Knowledge and Policy . . . . . . . . . . . 12
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Prenatal Care and Infant Health: A Multilevel Analysis 22
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Purpose and Objectives of the Study . . . . . . . . . . . . . . . . . . . . 29
2.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.7 Summary, Conclusions and Policy Implications . . . . . . . . . . . . . . . 60
v
Contents vi
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3 Preceding Birth Interval Length and Maternal Health 74
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.3 Purpose and Objectives of the Study . . . . . . . . . . . . . . . . . . . . 78
3.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.5 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.7 Summary, Conclusions and Policy Implications . . . . . . . . . . . . . . . 98
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4 Smoking and General Health 109
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.3 Purpose and Objectives of the Study . . . . . . . . . . . . . . . . . . . . 113
4.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.5 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.7 Summary, Conclusions and Policy Implications . . . . . . . . . . . . . . . 132
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5 Summary, Conclusions, and Policy Implications 144
5.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
5.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
5.3 Policy Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
List of figures
1.1 A General Framework for Analyzing Health Determinants . . . . . . . . 9
2.1 A Conceptual Framework for Analyzing the Effect of Prenatal Care Use
on Infant Health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 An Example of a Multilevel Data Structure . . . . . . . . . . . . . . . . . 42
3.1 A Conceptual Framework for Analyzing the Effect of Preceding Birth
Interval Length on Maternal Health . . . . . . . . . . . . . . . . . . . . . 80
4.1 A Conceptual Framework for Analyzing the Effect of Smoking on General
Health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
vii
List of tables
1.1 Indicators of Infant and Child Health at the Individual Level . . . . . . . 3
1.2 Indicators of Maternal Health at the Individual Level . . . . . . . . . . . 5
1.3 Key Health Indicators for Kenya for Selected Years . . . . . . . . . . . . 7
1.4 A Comparison of Kenya’s Key Health Indicators with Selected Regions . 8
2.1 Low Birth Weight Newborns (%) for Selected Countries and Regions . . 24
2.2 Indicators of Prenatal Care . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 WHO Recommended Minimum Prenatal Care Visits for Developing Countries 26
2.4 Maternal Risk Factors for Low Birth Weight . . . . . . . . . . . . . . . . 27
2.5 Summary of Studies Showing Positive Effects of Prenatal Care on Birth
Weight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.6 Factors Influencing Prenatal Care Use . . . . . . . . . . . . . . . . . . . 30
2.7 Variable Definitions for Prenatal Care and Infant Health Models . . . . . 47
2.8 Reporting of Birth Weight by Year of Survey . . . . . . . . . . . . . . . . 48
2.9 Descriptive Statistics for the Variables in Prenatal Care and Infant Health
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.10 Low Birth Weight Status by Year of Survey . . . . . . . . . . . . . . . . 50
2.11 Prenatal Care Use by Skill of Provider . . . . . . . . . . . . . . . . . . . 50
2.12 Adequacy of Prenatal Care Use, Analytic Sample for Prenatal Care Use . 51
2.13 Low Birth Weight and Adequacy of Prenatal Care Use . . . . . . . . . . 51
viii
List of tables ix
2.14 Average Marginal Effects for Sample Selection and Prenatal Care Probit
Models, Robust Z Statistics in Parentheses . . . . . . . . . . . . . . . . . 52
2.15 Average Marginal Effects from Single–Level Low Birth Weight Models,
Robust Z Statistics in Parentheses . . . . . . . . . . . . . . . . . . . . . 56
2.16 Average Marginal Effects from Multi–Level Low Birth Weight Models,
Robust Z Statistics in Parentheses (Number of Observations = 7331) . . 57
2.17 Average Marginal Effects from Our Chosen Low Birth Weight Models,
Robust Z Statistics in Parentheses . . . . . . . . . . . . . . . . . . . . . 58
3.1 Stillbirth rate for Selected Countries and WHO Regions, 2009 . . . . . . 75
3.2 Distribution of Preceding Birth Intervals in Kenya . . . . . . . . . . . . . 76
3.3 Determinants of Maternal Health . . . . . . . . . . . . . . . . . . . . . . 77
3.4 Risk Factors for Short Birth Intervals . . . . . . . . . . . . . . . . . . . . 79
3.5 Variable Definitions for Maternal Health Models . . . . . . . . . . . . . . 91
3.6 Descriptive Statistics for Maternal Health Models . . . . . . . . . . . . . 92
3.7 Distribution of Maternal Health Status by Survey Year, Percentages in
Parentheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.8 Distribution of Maternal Health Status by Preceding Birth Interval Length,
Percentages in Parentheses . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.9 Average Marginal Effects from Sample Selection and Preceding Birth
Interval Length Models, Robust Z Statistics in Parentheses . . . . . . . . 94
3.10 Average Marginal Effects for Maternal Health Status Models, Robust Z
Statistics in Parentheses . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.1 Some Facts about Tobacco Use and its Consequences in Kenya . . . . . . 111
4.2 Determinants of Smoking . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.3 Variable Definitions for Smoking and General Health Models . . . . . . . 122
4.4 Descriptive Statistics for Smoking and General Health Models . . . . . . 123
List of tables x
4.5 Distribution of Self–Rated Health Status in the Dataset . . . . . . . . . . 124
4.6 Distribution of Self–Rated Health Status by Smoking Status in the Dataset124
4.7 Average Marginal Effects for Sample Selection and Smoking Status Models,
Robust Z Statistics in Parentheses . . . . . . . . . . . . . . . . . . . . . 125
4.8 Average Marginal Effects for Probability of Rating Own Health as “Poor”,
Robust Z Statistics in Parentheses . . . . . . . . . . . . . . . . . . . . . 127
4.9 Average Marginal Effects for Probability of Rating Own Health as “Satis-
factory”, Robust Z Statistics in Parentheses . . . . . . . . . . . . . . . . 129
4.10 Average Marginal Effects for Probability of Rating Own Health as “Good”,
Robust Z Statistics in Parentheses . . . . . . . . . . . . . . . . . . . . . 130
4.11 Average Marginal Effects for Probability of Rating Own Health as “Very
Good”, Robust Z Statistics in Parentheses . . . . . . . . . . . . . . . . . 131
“And for man to look upon himself as a capital good, even if it did not impair his freedom,
may seem to debase him... by investing in themselves, people can enlarge the range of
choice available to them. It is one way free men can enhance their welfare. ”
— Theodore William Schultz (1902 – 1998)
Chapter 1
Introduction
1.1 Definition and Importance of Health
Health can be viewed as comprising physical, mental, and social well–being (Thomas and
Frankenberg, 2002; World Health Organization (WHO), 2006a; Mwabu, 2008; Strauss
and Thomas, 2008). Economists’ main interest in health is motivated by the fact that
good health is a component of human capital (Brock, 2002; Becker, 2007; Mwabu, 2008)
and, therefore, a key factor in wealth creation (Pritchett and Summers, 1996; Mwabu,
2001). Good health has also been shown to have positive effects on economic growth
(Arora, 2001; Bhargava, Jamison, Lau and Murray, 2001; Mayer, 2001; Bloom, Canning
and Sevilla, 2004; Qureshi and Mohyuddin, 2006).
We can look at health from a population point of view (in which case we talk about
population health) or from an individual point of view (in which case we talk about
individual health). From a policy perspective, sometimes we are specifically concerned
with the health of specific population groups (such as infants, children, adolescents, or
mothers) as opposed to general health. We can view infant health as referring to the
health of persons below the age of 1 year and child health as referring to the health of
persons aged 1 to 5 years. We can also view maternal health as the health of women
during the period when they are pregnant, at childbirth and the period immediately
following childbirth (Lule et al., 2005).
A policy maker’s interest in infant and child health is motivated by the fact that many
health problems that affect adults can be traced to early periods of one’s life (Hertzman
and Power, 2004; Currie, Stabile, Manivong and Roos, 2010). Research also shows that
infant health has effects on long–run adult health and other outcomes such as adult
1
Introduction 2
height, intelligence, education outcomes, and labour force outcomes (Black, Devereux
and Salvanes, 2007; Oreopoulos, Stabile, Walld and Roos, 2008; Al-Saleh and Renzo,
2009).
Policy makers have an interest in maternal health because research shows that good
maternal health increases labour supply, productive capacity, and economic well-being of
communities (Lule et al., 2005; Meyerhoefer and Sahn, 2010). Maternal health is also
important for infant and child health since neonatal and infant deaths can result from
poor maternal health (United Nations Population Fund (UNFPA), 2004).
1.2 Measurement of Health
The health status of a particular individual or population refers to all forms of the
individual’s or population’s health (Lopez et al., 2006). True health status is typically
unobservable but it can be approximated by using observable indicators (Stein, 2004),
which typically focus on the identification and assessment of negative deviations from what
would otherwise constitute normal life (Patrick and Bergner, 1990). The indicators help
improve health when they are applied in advocacy, accountability, system management,
quality improvement, and research (Etches, Frank, Di Ruggiero and Manuel, 2006).
Health status indicators can be constructed for individuals or entire populations. We
can broadly classify the indicators into three categories: indicators of infant and child
health, indicators of maternal health, and indicators of general health.
Indicators of Infant and Child Health
At the population level, the indicators of infant and child health status include neonatal
mortality rate, postneonatal mortality rate, infant mortality rate, birth weight distribution,
gestational age distribution (see, for example, Ruhm, 2000; Fayissa, 2001; Bhargava,
2003; Buitendijk et al., 2003; Alves and Belluzzo, 2004; Mwabu, 2009), child mortality
rate (see, for example, Fayissa, 2001), and under-five mortality rate. The most commonly
reported ones are, however, the birth weight distribution, the infant mortality rate, and
the under-five mortality rate.
For birth weight distribution, the emphasis is mainly on the percentage of children
with low birth weight. A child is considered to have low birth weight if his/her weight
Introduction 3
Table 1.1: Indicators of Infant and Child Health at the Individual Level
Indicator Literature source
Child survival Mosley and Chen, 1984; Schultz, 1984; Rutstein, 2000;
Kabubo–Mariara, Karienyeh and Kabubo, 2012.
Birth weight Rosenzweig and Schultz, 1982; Schultz, 1984;
Almond, Chay and Lee, 2005;
Oreopoulos, Stabile, Walld and Roos, 2008;
Mwabu, 2009;
Mbuya, Chideme, Chasekwa and Mishra, 2010.
Apgar score Almond, Chay and Lee, 2005;
Oreopoulos, Stabile, Walld and Roos, 2008;
Mwabu, 2009.
Gestation Oreopoulos, Stabile, Walld and Roos, 2008.
Nutritional status indicators Mosley and Chen, 1984; Rutstein, 2000;
Mukudi, 2003; Nakabo–Ssewanyana, 2003;
Alves and Belluzzo, 2004;
Kabubo-Mariara, Ndenge and Mwabu, 2009;
Kodjo, 2009; Case and Paxson, 2010;
Mbuya, Chideme, Chasekwa and Mishra, 2010;
Adewara and Visser, 2011.
at birth is less than 2,500 grams (Zegers–Hochschild et al., 2009; WHO, 2011a). The
infant mortality rate is the number of children who die aged under 1 year per 1,000 live
births (WHO, 2011a). The probability, subject to current age–specific mortality rates, of
a child born in a specific year dying before reaching age 5, expressed as a certain number
per 1,000 live births is referred to as the under–five mortality rate (WHO, 2005).
Table 1.1 summarizes the various indicators that have been used in the literature for
measuring infant and child health at the individual level.
The nutritional status indicators (such as height-for-age, weight-for-height, and weight-
for-age) are typically calculated using the World Health Organization (WHO)’s growth
standards (De Onis, Garza, Onyango and Martorell, 2006). Nutritional deficiencies in
childhood are associated with poor adult health outcomes, such as shorter adult height
(Victoria et al., 2008).
Introduction 4
The main indicators of malnutrition in children include underweight (having a weight
for age score that is less than −2 standard deviations of the WHO child growth standards
median), stunting (having a height for age score that is less than −2 standard deviations
of the WHO child growth standards median), wasting (having weight for height score
that is less than −2 standard deviations of the WHO child growth standards median) and
overweight (having a weight for height score that is greater than 2 standard deviations
of the WHO child growth standards median) (WHO, 2010). Underweight children are at
an increased risk of mortality, stunted and wasted children are at an increased risk of
illness and death, while obese children are at an increased risk of suffering from most
non–communicable diseases (Caulfield, De Onis, Blossner and Black, 2004; WHO, 2010).
Indicators of Maternal Health
At the population level, the typical indicators of maternal health status include maternal
mortality ratio (see, for example, McCarthy and Maine, 1992; Alexander et al., 2003;
WHO, 2006b) and maternal morbidity (Alexander et al., 2003; WHO, 2006b). Maternal
mortality ratio refers to the number of maternal deaths per 100,000 live births (WHO,
2011a). The death of a woman is said to be a maternal death if it results from causes
related to pregnancy or exacerbated by pregnancy or the management of the pregnancy
(WHO, 2011a). Maternal morbidity refers to illness and injury related to pregnancy and
childbirth (UNFPA, 2004).
Other indicators of maternal health at the population level include percentage of
women experiencing a live birth or a stillbirth following fertility treatment, distribution
of timing of first antenatal visit, and distribution of place of birth (Wildman et al., 2003).
Table 1.2 on page 5 shows the indicators of maternal health at the individual level
that have been used in the literature.
Other indicators of maternal health at the individual level include contraceptive use,
visit to obtain antenatal care in the first trimester, and safe delivery (Saikia and Singh,
2008).
Introduction 5
Table 1.2: Indicators of Maternal Health at the Individual Level
Indicator Literature source
Pregnancy outcome Kramer, 2003; Villamor and Cnattingius, 2006.
Pregnancy complications Koblinsky, 1995; Anderson et al., 2008;
Say, Souza and Pattinson, 2009.
Maternal morbidity Gandhi, Welz and Ronsmans, 2004.
Maternal mortality Mace and Sear, 1996.
Nutritional status indicators Koblinsky, 1995; Saikia and Singh, 2008.
Indicators of General Health
At the population level, there are several indicators of general health status which can be
broadly classified into two groups: health expectancies and health gaps (Murray, Salomon
and Mathers, 2002). Health status indicators that estimate the average time (in years)
that an individual could live in particular health states are generally referred to as health
expectancies (Mathers, 2002). Examples of health expectancies are disability–free life
expectancy and health–adjusted life expectancy (HALE) (Mathers, 2002).
Health status indicators that involve the quantification of differences between actual
health and some stated norm for health are referred to as health gaps (Murray, Mathers,
Salomon and Lopez, 2002). Examples of health gaps are disability–adjusted–life–years
(see, for example, Murray and Acharya, 1997; Mathers, Lopez and Murray, 2006) and
healthy–life–years (see, for example, Hyder, Rotllant and Morrow, 1998).
At the individual level, the main indicators of general health status include self-
reported general health status (where individuals rate own health in one of several
discrete categories compared to some reference group, such as age–mates); self-reported
morbidity, illness and normal activity (where individuals report whether or not they were
sick or injured during a reference period, such as four weeks; or the number of days the
individual failed to engage in some usual or “normal” activity due to sickness or injury);
self-reported physical functioning (where individuals report whether or not they had
difficulties in performing activities of daily living); nutrition–based indicators (such as
nutrient intakes, and anthropometrics which typically include height, weight, and body
mass index); haemoglobin levels (which can be used to compute the prevalence rates
of anaemia in both children and adults); and whether one is suffering from an acute or
chronic disease (Strauss and Thomas 1998, 2008; Mwabu, 2008).
Introduction 6
The most commonly used anthropometric measure for adults is the Body Mass Index
(BMI), typically computed by dividing an individual’s weight (in kilograms) by the square
of the individual’s height (in meters). The measure is used to classify an individual as
either underweight, normal, overweight or obese. An individual is considered underweight
if he/she has a BMI of less than 18.5, normal if he/she has a BMI of between 18.5 and
24.9, overweight if he/she has a BMI of between 25 and 29.9, and obese if he/she has a
BMI of 30 and above (National Institutes of Health (NIH), 1998; Hiza, Pratt, Mardis and
Anand, 2000). BMI values of above 25 are unhealthy and have been shown to increase
the risk of chronic diseases such as high blood pressure, diabetes, heart disease, stroke,
certain types of cancer, arthritis, and breathing problems (Hubert, Feinleib, McNamara
and Castelli, 1983; Colditz, Willet, Rotnitzky and Manson, 1995; Giovannucci et al.,
1995; Hochberg et al., 1995; Walker et al., 1996).
1.3 Health Determinants
Individual health outcomes are typically influenced by various health inputs such as
use of preventive and curative health services, health behaviours such as smoking and
alcohol consumption, and habits (Rosenzweig and Schultz, 1983; Mosley and Chen, 1984;
Schultz, 1984; McCarthy and Maine, 1992; Norris et al., 2003; Contoyannis and Jones,
2004; Maitra and Pal, 2007; Strauss and Thomas, 2008).
The actual mechanics for producing health at the individual level is likely to be
influenced by the individual’s personal socio-economic characteristics such as age, gender,
marital status, education, occupation, income, and ethnicity (Gilleskie and Harrison,
1998; Marmot, 2002; Fuchs, 2004; Strauss and Thomas, 2008; Cutler and Lleras-Muney,
2010); aspects of the individual’s family background such as the health of the individual’s
parents and his or her genetic endowment; and various environmental characteristics
such as the disease environment, accessibility of health facilities1 and the quality of care
(Mosley and Chen, 1984; Schultz, 1984; Strauss and Thomas, 2008).
1For a comprehensive discussion of access to health care, see Gulliford and Morgan (2003).
Introduction 7
Table 1.3: Key Health Indicators for Kenya for Selected Years
Indicator 1990 2000 2008/2009*
Life expectancy at birth (years) 61 54 60
Neonatal mortality rate (per 1,000 live births) 30 32 27
Infant mortality rate (per 1,000 live births) 64 66 55
Under – five mortality rate (per 1,000 live births) 99 105 84
Adult mortality rate (per 1,000 population) 253 421 319
Maternal mortality ratio (per 100,000 live births) 380 560 530
* Data for all other indicators is for 2009 while that for Maternal mortality is for 2008.
Source: WHO (2011b).
1.4 Key Health Indicators for Kenya
Table 1.3 summarizes the key health indicators for Kenya in the recent past. From
this table we can observe that Kenya’s health indicators are not very impressive. The
table shows, for example, that as of 2009, life expectancy at birth was 60 years, having
increased from 54 years in 2000. We can further see from the table that infant mortality
rate was 55 per 1,000 live births as of 2009, an improvement from a value of 66 in 2000.
The table also shows that as of 2008, maternal mortality ratio was 530 per 100,000 live
births, an improvement from a value of 560 in 2000.
Table 1.4 on page 8 compares Kenya’s health indicators with those for Africa, Europe
and the whole world. From the table, we can observe that although Kenya’s life expectancy
at birth (60 years) is higher than the average for the African region (54 years), it is still
lower than that for Europe (75 years) and the world average (68 years). This is also true
for all the other indicators shown in the table.
1.5 Purpose and Objectives of the Thesis
A careful look at the various health indicators for Kenya (such as those presented in
Table 1.3 and Table 1.4) leads to the conclusion that health issues in Kenya for infants,
children, mothers, and other adults still require urgent attention. People also generally
invest in various activities aimed at improving their health such as seeking adequate
Introduction 8
Table 1.4: A Comparison of Kenya’s Key Health Indicators with Selected Regions
Indicator Kenya Africa Europe World
Life expectancy at birth (years) 60 54 75 68
Neonatal mortality rate (per 1,000 live births) 27 36 7 24
Infant mortality rate (per 1,000 live births) 55 80 12 42
Under – five mortality rate (per 1,000 live births) 84 127 13 60
Adult mortality rate (per 1,000 population) 319 383 146 176
Maternal mortality ratio (per 100,000 live births)* 530 620 21 260
* Data for all other indicators is for 2009 while that for Maternal mortality is for 2008.
Source: WHO (2011b).
prenatal care when pregnant, spacing births optimally, quitting smoking, exercising, and
seeking health care from health facilities when sick (see, for example, the discussions in
Grossman, 1972 and Becker, 2007). People do not know a priori whether these activities
do actually improve their health. For us to be able to recommend actions on the part of
individuals and policy–makers that improve health (which could be indicated by better
health indicators), we need evidence that such actions do indeed improve health. It is
on this basis that this thesis investigates the effects of some specific health inputs and
behaviours on health outcomes for infants, mothers, and the general adult population in
Kenya.
Its specific objectives include:
1. Determining the effect of the adequacy of prenatal care use on infant health in
Kenya.
2. Determining the effect of preceding birth interval length on maternal health in
Kenya.
3. Determining the effect of smoking on general health in Kenya.
4. Drawing appropriate policy implications from the findings.
Introduction 9
Figure 1.1: A General Framework for Analyzing Health Determinants
Observed Variables Unobserved Variables
Preferences
Biological endowments/Genetics
Individual /Household Socio – economic Characteristics
Community Characteristics /Environmental factors
Health Inputs / Health Behaviours
Health Outcomes
Source: Author’s Own Construction Based on Schultz (1984).
1.6 General Modelling Framework for the Thesis
Conceptual Model
Following Schultz (1984), we can develop a general conceptual framework for analyzing the
effects of the various health inputs and behaviours on health outcomes. The framework
is illustrated in Figure 1.1. The framework shows how health outcomes are linked to
observable health inputs and behaviours, and unobservable biological endowments. It
further shows that the health inputs and behaviours are in turn influenced by unobservable
preferences and observable individual and household demographic and socio-economic
characteristics, community characteristics and environmental factors.
We use this framework in each of the essays only that in the essays we have a specific
health input/health behaviour and a specific health outcome.
Introduction 10
Theoretical Model
Our theoretical modelling framework for each of the essays follows Rosenzweig and Schultz
(1982, 1983) and Mwabu (2009). In this framework, individuals (or their representatives)
are assumed to maximize the utility, U , obtained from the consumption of goods and
services that have no direct effect on health, X, those goods that yield utility directly
but also affect individual health, Y , and the individual’s health status, H. For a typical
individual i, we can write this utility function as follows:
Ui = U(Xi, Yi, Hi). (1.1)
Since health cannot be purchased from the market, it has to be produced using
both marketed and non–marketed inputs (Grossman, 1972). We, therefore, assume that
the individual’s health status is influenced by purchased inputs into health, such as
health care, Z, other factors, Q, and unobservable biological endowments, µ. Thus, the
individual’s health status function (or health production function) is given by:
Hi = H(Zi, Yi, Qi, µi). (1.2)
The individual is assumed to maximize the utility function subject to the above health
production function and a budget constraint.
Estimation Issues
When estimating the models in each essay, we control for the following key issues:
potential sample selection bias, potential endogeneity of some of the covariates, and
potential unobserved heterogeneity.
Sample Selection Bias
In general, a selected sample is a non–random sample obtained from a restricted portion
of the population of interest (Wooldridge, 2002). Sample selection bias, in general, arises
whenever we only observe the variable of primary interest for a selected sample and there
is correlation between the unobservable factors determining inclusion in the sample and
the unobservable factors influencing the variable of primary interest (Vella 1992, 1998).
Introduction 11
The literature proposes several approaches to correcting for sample selection bias (see,
for example, Heckman, 1979; Olsen, 1980; Lee, 1982; Garen, 1984; Vella 1992, 1993). In
this thesis, we employ the approach suggested by Vella (1992) in the first essay found in
chapter 2, that suggested by Olsen (1980) in the second essay found in chapter 3, and
Heckman’s approach (Heckman, 1979) in the third essay found in chapter 42.
Endogeneity
An econometric model is said to suffer from the problem of endogeneity if there is a
correlation between the error term in the model and one or more of the regressors included
in the model (Stock and Watson, 2011). Endogeneity can result from the omission of
confounder variables from the model, reverse causality between the dependent variable
and the endogenous regressor, and measurement errors in the regressors (Cameron and
Trivedi, 2010). When a regression model suffers from the problem of endogeneity, the
estimated regression coefficients are inconsistent, and we can also not infer causality
between the dependent variable and the independent variables in such a model (Cameron
and Trivedi, 2010). Zohoori (1997) demonstrates that controlling for endoegeneity matters
in empirical studies. Terza, Basu and Rathouz (2008) discuss two Instrumental Variable
(IV)-based approaches to correcting for endogeneity bias in nonlinear models: two–stage
residual inclusion (2SRI) and two–stage predictor substitution (2SPS), and show that, in
non–linear models, the 2SRI approach is consistent while the 2SPS approach is not. We
consequently use the 2SRI approach in estimating our models since they are non–linear.
Unobserved Heterogeneity
Unobserved heterogeneity is said to occur in an econometric model if there is a non-
linear interaction between unobservable factors and the endogenous covariate which
causes the effect of the endogenous covariate on the variable of interest to differ among
population subjects (Zohoori and Savitz, 1997). The standard procedure for controlling
for unobserved heterogeneity is the control function approach (Florens, Heckman, Meghir
and Vytlacil, 2008; Mwabu, 2009; Wooldridge, 2010). We employ this approach in our
essays.
2The different approaches to correcting for sample selection bias are used to illustrate the differentmethods available in the literature for dealing with this problem.
Introduction 12
1.7 Contributions of the Thesis to Knowledge and
Policy
The thesis makes the following contributions to knowledge and policy using Kenyan data:
1. It shows that failure to control for unobserved mother–specific characteristics may
lead to an overstatement of the effects of adequate prenatal care use on reducing
the probability of low birth weight. This provides the rationale for using methods
that control for unobserved mother–specific effects in this kind of analysis, such as
multi–level modelling.
2. It shows that preceding birth interval length is an endogenous determinant of
maternal health. In particular, it shows that preceding birth intervals of length
36 to 59 months only improve maternal health after this endogeneity has been
controlled for. As such, failure to control for the endogeneity of preceding birth
interval may make preceding birth intervals of length 36 to 59 months to appear
to worsen maternal health. Models that link preceding birth intervals to maternal
health must, therefore, control for the endogeneity of preceding birth interval. The
finding that preceding birth interval improves maternal health provides evidence for
pursuing programmes aimed at increasing birth spacing.
3. It shows that failure to control for sample selection bias, endogeneity of smoking
and unobserved heterogeneity leads to an understatement of the negative effects
of smoking on self–rated health status by about 50%. For example, when we
do not control for sample selection bias, endogeneity of smoking and unobserved
heterogeneity, we find that smoking increases the probability of rating own health as
“Poor” compared to age–mates by 0.009, holding other factors constant. When we,
however, control for sample selection bias, endogeneity of smoking and unobserved
heterogeneity, we find that smoking increases the probability of rating own health
as “Poor” as compared to age–mates by 0.018, holding other factors constant.
Similarly, without controlling for sample selection bias, endogeneity of smoking and
unobserved heterogeneity, smoking reduces the probability of rating own health as
“Very Good” as compared to age–mates by 0.033, holding other factors constant.
However, after controlling for sample selection bias, endogeneity of smoking and
unobserved heterogeneity, we find that smoking reduces the probability of rating
own health as “Very Good” as compared to age–mates by 0.069, holding other
factors constant.
REFERENCES 13
References
[1] Adewara, S. O. and Visser, M., 2011. Use of anthropometric measures to analyzehow sources of water and sanitation affect children’s health in Nigeria.Environment for Development Discussion Paper No. Efd DP 11–02. [pdf]Environment for Development. Available at:http://www.rff.org/rff/documents/efd-dp-11-02.pdf [Accessed 2 January2012].
[2] Alexander, S., Wildman, K., Zhang, W., Langer, M., Vutuc, C. and Lindmark, G.,2003. Maternal health outcomes in Europe. European Journal of Obstretrics,Gynecology and Reproductive Biology 111 (Supplement 1), pp.S78-S87.
[3] Almond, D., Chay, K. Y. and Lee, D. S., 2005. The costs of low birth weight.Quarterly Journal of Economics 120 (3), pp.1031-1083.
[4] Al-Saleh, E. and Renzo, G. C. D., 2009. Actions needed to improve maternalhealth. International Journal of Gynecology & Obstetrics 106 (2), pp.115-119.
[5] Alves, D. and Belluzzo, W., 2004. Infant mortality and child health in Brazil.Economics and Human Biology 2 (3), pp.391-410.
[6] Anderson, F. W. J., Naik, S. I., Feresu, S. A., Gebrian, B., Karki, M. and Harlow,S., 2008. Perception of pregnancy complications in Haiti. International Journal ofGynecology & Obstetrics 100 (2), pp.116-123.
[7] Arora, S., 2001. Health, human productivity, and long–term economic growth.Journal of Economic History 61 (3), pp.699-749.
[8] Becker, G. S., 2007. Health as human capital: synthesis and extensions. OxfordEconomic Papers 59 (3), pp.379-410.
[9] Bhargava, A., 2003. Family planning, gender differences and infant mortality:evidence from Uttar Pradesh, India. Journal of Econometrics 112 (1), pp.225-240.
[10] Bhargava, A., Jamison, D. T., Lau, L. J. and Murray, C. J. L., 2001. Modeling theeffects of health on economic growth. Journal of Health Economics 20 (3),pp.423-440.
[11] Black, S. E., Devereux, P. J. and Salvanes, K. G., 2007. From the cradle to thelabour market? The effect of birth weight on adult outcomes. Quarterly Journal ofEconomics 122 (1), pp.409-439.
[12] Bloom, D. E., Canning, D. and Sevilla, J., 2004. The effect of health on economicgrowth: a production function approach. World Development 32 (1), pp.1-13.
[13] Brock, D. W., 2002. The separability of health and well-being. In: C. J. L. Murray,J. A. Salomon, C. D. Mathers and A. D. Lopez, eds. Summary measures ofpopulation health. Geneva: World Health Organization, pp.115–120.
REFERENCES 14
[14] Buitendijk, S., Zeitlin, J., Cuttini, M., Langhoff-Ross, J. and Bottu, J., 2003.Indicators of fetal and infant health outcomes. European Journal of Obstetrics,Gynecology, and Reproductive Biology 111 (Supplement 1), pp.S66-S77.
[15] Cameron, A. C. and Trivedi, P. K., 2010. Microeconometrics using stata. RevisedEdition. Stata Press.
[16] Case, A. and Paxson, C., 2010. Causes and consequences of early-life health.Demography 47 (Supplement 1), pp.s65-s85.
[17] Caulfield, L. E., De Onis, M., Blossner, M. and Black, R. E., 2004. Undernutritionas an underlying cause of child deaths associated with diarrhea, pneumonia,malaria, and measles. American Journal of Clinical Nutrition 80 (1), pp.193-198.
[18] Colditz, G. A., Willet, W. C., Rotnitzky, A. and Manson, J. E., 1995. Weight gainas a risk factor for clinical diabetes mellitus in women. Annals of InternalMedicine 122 (7), pp.481-486.
[19] Contoyannis, P. and Jones, A. M., 2004. Socio–economic status, health andlifestyle. Journal of Health Economics 23 (5), pp.965-995.
[20] Currie, J., Stabile, M., Manivong, P. and Roos, L., 2010. Child health and youngadult outcomes. Journal of Human Resources 45 (3), pp.517-548.
[21] Cutler, D. M. and Lleras-Muney, A., 2010. Understanding differences in healthbehaviours by education. Journal of Health Economics 29 (1), pp.1-28.
[22] De Onis, M., Garza, C., Onyango, A. W. and Martorell, R., 2006. WHO childgrowth standards. Acta Paediatricia, [pdf] 450 (Supplement 450), pp.3–101.Available at:http://www.who.int/childgrowth/standards/Acta_95_S450.pdf [Accessed 6August 2008].
[23] Etches, V., Frank, J., Di Ruggiero, E. and Manuel, D., 2006. Measuring populationhealth: a review of indicators. Annual Review of Public Health 27, pp.29–55.
[24] Fayissa, B., 2001. The determinants of infant and child mortality in developingcountries: the case of Sub-Saharan Africa. Review of Black Political Economy 29(2), pp.83-100.
[25] Florens, J. P., Heckman, J. J., Meghir, C. and Vytlacil, E., 2008. Identification oftreatment effects using control functions in models with continuous, endogenoustreatment and heterogeneous effects. Econometrica 76 (5), pp.1191-1206.
[26] Fuchs, V. R., 2004. Reflections on the socio-economic correlates of health. Journalof Health Economics 23 (4), pp.653-661.
[27] Gandhi, M. N., Welz, T. and Ronsmans, C., 2004. Severe acute maternalmorbidity in rural South Africa. International Journal of Gynecology & Obstetrics87 (2), pp.180-187.
REFERENCES 15
[28] Garen, J., 1984. The returns to schooling: a selectivity bias approach with acontinuous choice variable. Econometrica 52 (5), pp.1199-1218.
[29] Gilleskie, D. B. and Harrison, A., 1998. The effect of endogenous health inputs onthe relationship between health and education. Economics of Education Review 17(3), pp.279-297.
[30] Giovannucci, E., Ascherio, A., Rimm, E. B., Colditz, G. A., Stampfer, M. J. andWillet, W. C., 1995. Physical activity, obesity, and risk for colon cancer andadenoma in men. Annals of Internal Medicine 122 (5), pp.327-334.
[31] Grossman, M., 1972. On the concept of health capital and the demand for health.Journal of Political Economy 80 (2), pp.223–255.
[32] Gulliford, M. and Morgan, M., eds., 2003. Access to health care. Routledge.
[33] Heckman, J., 1979. Sample selection as a specification error. Econometrica 47 (1),pp.153-161.
[34] Hertzman, C. and Power, C., 2004. Child development as a determinant of healthacross the life course. Current Paediatrics 14 (5), pp.438-443.
[35] Hiza, A. H., Pratt, C., Mardis, A. L. and Anand, R., 2000. Body mass index andhealth. Nutrition Insights 16.
[36] Hochberg, M. C., Lethbridge–Cejku, M., Scott, W. W. W. (Jr.), Reichle, R.,Platto, C. C. and Tobin, J. D., 1995. The association of body weight, body fatnessand body fat distribution with osteoarthritis of the knee: data from the Baltimorelongitudinal study of aging. Journal of Rheumatology 22 (3), pp.488-493.
[37] Hubert, H. B., Feinleib, M., McNamara, P. M. and Castelli, W. P., 1983. Obesityas an independent risk factor for cardiovascular disease: a 26–year follow–up ofparticipants in the Framingham heart study. Circulation 67 (5), pp.968-977.
[38] Hyder, A. A., Rotllant, G. and Morrow, R. H., 1998. Measuring the burden ofdisease: healthy life-years. American Journal of Public Health 88 (2), pp.196-202.
[39] Kabubo–Mariara, J., Karienyeh, M. M. and Kabubo, F. M., 2012. Child survival,poverty and inequality in Kenya: does physical environment matter? AfricanJournal of Social Sciences 2 (1), pp.65–84.
[40] Kabubo-Mariara, J., Ndenge, G. K. and Mwabu, D. K., 2009. Determinants ofchildren’s nutritional status in Kenya: evidence from demographic and healthsurveys. Journal of African Economies 18 (3), pp.363-387.
[41] Koblinsky, M. A., 1995. Beyond maternal mortality - magnitude, interrelationship,and consequences of women’s health, pregnancy-related complications andnutritional status on pregnancy outcomes. International Journal of Gynecology &Obstetrics 48 (Supplement 1), pp.S21-S32.
REFERENCES 16
[42] Kodjo, A., 2009. Poverty and the anthropometric status of children: a comparativeanalysis of rural and urban households in Togo. AERC Research Paper 191. [pdf]Nairobi: African Economic Research Consortium. Available at:http://www.aercafrica.org/documents/rp191.pdf [Accessed 10 Agust 2011].
[43] Kramer, M. S., 2003. The epidemiology of adverse pregnancy outcomes: anoverview. Journal of Nutrition 133 (5), pp.1592S–1596S.
[44] Lee, L., 1982. Some approaches to the correction of selectivity bias. Review ofEconomic Studies 49 (3), pp.355-372.
[45] Lopez, A. D., Mathers, C. D., Ezzati, M., Jamison, D. T. and Murray, C. J. L.eds., 2006. Global burden of disease and risk factors. New York: Oxford UniversityPress and World Bank.
[46] Lule, E., Ramana, G. N. V., Oomman, N., Epp, J., Huntington, D. and Rosen, J.E., 2005. Achieving the millennium development goal of improving maternalhealth: determinants, interventions and challenges. Health, Nutrition andPopulation (HNP) Discussion Paper. World Bank.
[47] Mace, R. and Sear, R., 1996. Maternal mortality in a Kenyan pastoralistpopulation. International Journal of Gynecology & Obstetrics 54 (2), pp.137-141.
[48] Maitra, P. and Pal, S., 2007. Early childbirth, health inputs and child mortality:recent evidence from Bangladesh. IZA Discussion Paper No. 2841. [pdf] Availableat: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=998229[Accessed 7 July 2009].
[49] Marmot, M., 2002. The influence of income on health: views of an epidemiologist.Health Affairs 21 (2), pp.31-46.
[50] Mathers, C. D., 2002. Health expectancies: an overview and critical appraisal. In:C. J. L. Murray, J. A. Salomon, C. D. Mathers and A. D. Lopez, eds. Summarymeasures of population health. Geneva: World Health Organization, pp.177–204.
[51] Mathers, C. D., Lopez, A. D. and Murray, C. J. L., 2006. The burden of diseaseand mortality by condition: data, methods, and results for 2002. In: A. D. Lopez,C. D. Mathers, M. Ezzati, D. T. Jamison and C. J. L. Murray, eds. Global burdenof disease and risk factors. World Bank / Oxford University Press, pp.45–240.
[52] Mayer, D., 2001. The long-term impact of health on economic growth in LatinAmerica. World Development 29 (6), pp.1025-1033.
[53] Mbuya, M. N. N., Chideme, M., Chasekwa, B. and Mishra, V., 2010. Biological,social, and environmental determinants of low birth weight and stunting amonginfants and young children in Zimbabwe. Zimbabwe Working Papers, No. 7. [pdf]Calverton, Maryland, USA: ICF Macro. Available at:http://pdf.usaid.gov/pdf_docs/PNADR633.pdf [Accessed 10 August 2011].
REFERENCES 17
[54] McCarthy, J. and Maine, D., 1992. A framework for analyzing the determinants ofmaternal mortality. Studies in Family Planning 23 (1), pp.23-33.
[55] Meyerhoefer, C. D. and Sahn, D. E., 2010. The relationship between poverty andmaternal morbidity and mortality in Sub-Saharan Africa. In: O. Ajakaiye and G.Mwabu. eds. Reproductive health, economic growth and poverty in Africa:frameworks of analysis. Nairobi: University of Nairobi Press, pp.323–371.
[56] Mosley, W. H. and Chen, L. C., 1984. An analytical framework for the study ofchild survival in developing countries. Population and Development Review 10(Supplement), pp.25-45.
[57] Mukudi, E., 2003. Nutrition status, education participation, and schoolachievement among Kenyan middle-schooled children. Nutrition 19 (7/8),pp.612-616.
[58] Murray, C. J. L. and Acharya, A. K., 1997. Understanding DALYs. Journal ofHealth Economics 16 (6), pp.703-730.
[59] Murray, C. J. L., Mathers, C. D., Salomon, J. A. and Lopez, A. D., 2002. Healthgaps: an overview and critical appraisal. In: C. J. L. Murray, J. A. Salomon, C. D.Mathers and A. D. Lopez, eds. Summary measures of population health. Geneva:World Health Organization, pp.233–244.
[60] Murray, C. J. L., Salomon, J. A. and Mathers, C. D., 2002. A critical examinationof summary measures of population health. In: C. J. L. Murray, J. A. Salomon, C.D. Mathers and A. D. Lopez, eds. Summary measures of population health.Geneva: World Health Organization, pp.13–40.
[61] Mwabu, G., 2001. Health status in Africa: a regional profile. South AfricanJournal of Economics 69 (2), pp.319-335.
[62] Mwabu, G., 2008. Health economics for low-income countries. In: T. P. Schultzand J. Strauss, eds. Handbook of development economics, Volume 4. Amsterdam:Elsevier/North–Holland, pp.3305–3374.
[63] Mwabu, G., 2009. The production of child health in Kenya: a structural model ofbirth weight. Journal of African Economies 18 (2), pp.212-260.
[64] Nakabo–Ssewanyana, S., 2003. Food security and child nutrition status amongurban poor households in Uganda: implications for poverty alleviation. AERCResearch Paper 130. [pdf] Nairobi: African Economic Research Consortium.Available at: http://www.aercafrica.org/documents/rp130.pdf [Accessed 10August 2011].
[65] National Institutes of Health (NIH), 1998. Clinical guidelines on the identification,evaluation, and treatment of overweight and obesity in adults: the evidence report.[pdf] National Institutes of Health. Available at:http://www.nhlbi.nih.gov/guidelines/obesity/ob_gdlns.pdf [Accessed 6
REFERENCES 18
August 2008].
[66] Norris, J. C., van der Laan, M. J., Lane, S., Anderson, J. N. and Block, G., 2003.Nonlinearity in demographic and behavioural determinants of morbidity. HealthServices Research 38 (6, Part II), pp.1791-1818.
[67] Olsen, R. J., 1980. A least squares correction for selectivity bias. Econometrica 48(7), pp.1815-1820.
[68] Oreopoulos, P., Stabile, M., Walld, R. and Roos, L. L., 2008. Short–, medium–,and long–term consequences of poor infant health: an analysis using siblings andtwins. Journal of Human Resources 43 (1), pp.88-138.
[69] Patrick, D. L. and Bergner, M., 1990. Measurement of health status in the 1990s.Annual Review of Public Health 11, pp.165-183.
[70] Pritchett, L. and Summers, L. H., 1996. Wealthier is healthier. Journal of HumanResources 30 (4), pp.841-868.
[71] Qureshi, H. A. and Mohyuddin, H. A., 2006. Health status, diseases, and economicdevelopment: a cross-country analysis. Journal of Developing Areas 39 (2),pp.121-128.
[72] Rosenzweig, M. R. and Schultz, T. P., 1982. The behaviour of mothers as inputsto child health: the determinants of birth weight, gestation, and rate of fetalgrowth. In: V. R. Fuchs, ed. Economic aspects of health. [pdf] Chicago: Universityof Chicago Press, pp.53–92. Available at:http://www.nber.org/chapters/c6545.pdf [Accessed 6 August 2008].
[73] Rosenzweig, M. R. and Schultz, T. P., 1983. Estimating a household productionfunction: heterogeneity, the demand for health inputs, and their effects on birthweight. Journal of Political Economy 91 (5), pp.723-746.
[74] Ruhm, C. J., 2000. Parental leave and child health. Journal of Health Economics19 (6), pp.931-960.
[75] Rutstein, S. O., 2000. Factors associated with trends in infant and child mortalityin developing countries during the 1990s. Bulletin of the World HealthOrganization 78 (10), pp.1256-1270.
[76] Saikia, N. and Singh, A., 2008. Does type of household affect maternal health?Evidence from India. Journal of Biosocial Science 41 (3), pp.329-353.
[77] Say, L., Souza, J. P. and Pattinson, R. C., 2009. Maternal near miss - towards astandard for monitoring quality of maternal health care. Best Practice & Research,Clinical Obstetrics & Gynaecology 23 (3), pp.287-296.
[78] Schultz, T. P., 1984. Studying the impact of household economic and communityvariables on child mortality. Population and Development Review 10 (Supplement),pp.215-235.
REFERENCES 19
[79] Stein, R. E. K., 2004. Measurement of children’s health. Ambulatory Pediatrics 4(4), pp.365-370.
[80] Stock, J. H. and Watson, M. W., 2011. Introduction to econometrics. ThirdEdition. Addison–Wesley.
[81] Strauss, J. and Thomas, D., 1998. Health, nutrition, and economic development.Journal of Economic Literature 36 (2), pp.766-817.
[82] Strauss, J. and Thomas, D., 2008. Health over the life course. In: T. P. Schultzand J. Strauss, eds. Handbook of development economics, Volume 4. Amsterdam:Elsevier/North–Holland, pp.3375–3474.
[83] Thomas, D. and Frankenberg, E., 2002. The measurement and interpretation ofhealth in social surveys. In: C. J. L. Murray, J. A. Salomon, C. D. Mathers and A.D. Lopez, eds. Summary measures of population health. Geneva: World HealthOrganization, pp.387–420.
[84] Terza, J. V., Basu, A. and Rathouz, P. J., 2008. Two-stage residual inclusionestimation: addressing endogeneity in health econometric modelling. Journal ofHealth Economics 27 (3), pp.531-543.
[85] United Nations Population Fund (UNFPA), 2004. State of world population 2004:the Cairo consensus at ten: population, reproductive health and the global effort toend poverty. [pdf] New York, NY: UNFPA. Avaialble at:http://www.unfpa.org/swp/2004/pdf/en_swp04.pdf [Accessed 6 August 2008].
[86] Vella, F., 1992. Simple tests for sample selection bias in censored and discretechoice models. Journal of Applied Econometrics 7 (4), pp.413-421.
[87] Vella, F., 1993. A simple estimator for models with censored endogenousregressors. International Economic Review 34 (2), pp.441-457.
[88] Vella, F., 1998. Estimating models with sample selection bias: a survey. Journal ofHuman Resources 33 (1), pp.127-169.
[89] Victoria, C. G., Adair, L., Fall, C., Hallal, P. C., Martorell, R., Richter, L. andSachdev, H. S., 2008. Maternal and child undernutrition: consequences for adulthealth and human capital. Lancet 371 (9609), pp.340–357.
[90] Villamor, E. and Cnattingius, S., 2006. Interpregnancy weight change and risk ofadverse pregnancy outcomes: a population–based study. Lancet 368 (9542),pp.1164–1170.
[91] Walker, S. P., Rimm, E. B., Ascheno, A., Kawachi, I., Stampfer, M. J. and Willet,W. C., 1996. Body size and fat distribution as predictors of stroke among US men.American Journal of Epidemiology 144 (12), pp.1143-1150.
[92] Wildman, K., Blondel, B., Nijhuis, J., Defoort, P. and Bakoula, C., 2003.European indicators of health care during pregnancy, delivery and the postpartum
REFERENCES 20
period. Journal of Obstetrics & Gynecology and Reproductive Biology 111(Supplement 10), pp.S53-S65.
[93] Wooldridge, J. M., 2002. Econometric analysis of cross section and panel data.MIT Press.
[94] Wooldridge, J. M., 2010. Econometric analysis of cross section and panel data.Second Edition. MIT Press.
[95] World Health Organization (WHO), 2005. World health report 2005: making everymother and child count. [pdf] Geneva: World Health Organization. Available at:http://www.who.int/whr/2005/whr2005_en.pdf [Accessed 6 August 2008].
[96] World Health Organization (WHO), 2006a. Constitution of the world healthorganization. [pdf] World Health Organization. Available at:http://www.who.int/governance/eb/who_constitution_en.pdf [Accessed 6August 2008].
[97] World Health Organization (WHO), 2006b. Report of a WHO technicalconsultation on birth spacing. [pdf] Geneva: World Health Organization. Availableat: www.who.int/making_pregnancy_safer/.../birth_spacing.pdf [Accessed6 August 2008].
[98] World Health Organization (WHO), 2010. Nutrition landscape information system(NLIS) country profile indicators: interpretation guide. [pdf] Geneva: WorldHealth Organization. Available at:http://whqlibdoc.who.int/publications/2010/9789241599955_eng.pdf
[Accessed 6 August 2011].
[99] World Health Organization (WHO), 2011a. International statistical classificationof diseases and related health problems, 10th Revision (ICD–10): volume 2instruction manual. 2010 edition. [pdf] World Health Organization. Available at:http://www.who.int/classifications/icd/ICD10Volume2_en_2010.pdf
[Accessed 12 October 2012].
[100] World Health Organization (WHO), 2011b. World health statistics 2011. [pdf]Geneva: World Health Organization. Available at:http://www.who.int/whosis/whostat/EN_WHS2011_Full.pdf [Accessed 3January 2012].
[101] Zegers–Hochschild, F., Adamson, G. D., de Mouzon, J., Ishihara, O., Mansour, R.,Nygren, K., Sullivan, E. and Vanderpoel, S., 2009. International committee formonitoring assisted reproductive technology (ICMART) and the World HealthOrganization revised glossary of ART terminology. Fertility and Sterility 92 (5),pp.1520–1524.
[102] Zohoori, N., 1997. Does endogeneity matter? A comparison of emprical analyseswith and without control for endogeneity. Annals of Epidemiology 7 (4),pp.258–266.
REFERENCES 21
[103] Zohoori, N. and Savitz, D. A., 1997. Econometric approaches to epidemiologicdata: relating endogeneity and unobserved heterogeneity to confounding. Annalsof Epidemiology 7 (4), pp.251–257.
Chapter 2
Prenatal Care and Infant Health: A
Multilevel Analysis
2.1 Introduction
As pointed out in Chapter 1, one of the indicators of infant and child health status
is birth weight (see, for example, Mwabu, 2009). Birth weight refers to the weight of
the newborn at birth, typically obtained within the first hour of birth for live births
(WHO, 2011). It is one of the most important predictors of infant health outcomes
(McCormick, 1985; Kramer, 1987; Pollack and Divon, 1992; Ohlsson and Shah, 2008).
It has been demonstrated to be a good summary measure of a child’s health at birth
(see, for example, Mwabu, 2009). The World Health Organization (WHO) defines low
birth weight as a weight at birth of less than 2,500 grams (Zegers–Hochschild et al., 2009;
WHO, 2011). In this chapter, we measure infant health using birth weight.
There are several undesirable health and other outcomes associated with low birth
weight such as an increased risk for illnesses and deaths among infants and neonates,
an increased risk for chronic non–communicable diseases in later life, and poor school
performance signalled by grade repetition (Institute of Medicine (IOM), 1985; Barker
and Osmond, 1986; Chaikind and Corman, 1991; Hack, Klein and Taylor, 1995; Paneth,
1995; Corman and Chaikind, 1998; UNICEF and WHO, 2004; Reyes and Manalich, 2005;
Ohlsson and Shah, 2008). Further, it is more expensive to care for low birth weight
infants as compared to caring for normal weight infants (Lewit, Baker, Corman and
Shiono, 1995; Ohlsson and Shah, 2008). Reducing the number of low birth weight babies
could have substantial economic benefits for low–income countries, such as increases
22
Prenatal Care and Infant Health: A Multilevel Analysis 23
in labour productivity and reduction in costs associated with infant illness and death
(Alderman and Behrman, 2006).
Low birth weight, being an indicator of inadequate foetal growth, may be a consequence
of either premature birth (birth before a gestation period of 37 weeks), intrauterine
growth retardation (IUGR), or both (IOM, 1985; Kramer, 1987; Kallan, 1993; Paneth,
1995; Ohlsson and Shah, 2008; UNICEF and WHO, 2004).
Table 2.1 on page 24 shows the percentage of newborns that are low birth weight
for Kenya as compared to other countries and the various World Health Organization
(WHO) regions for the periods 2000 – 2002, and 2005 – 2010. The table shows that the
percentage of low birth weight newborns in Kenya declined from 11% over the 2000 – 2002
period to 8% over the period 2005 – 2010. The table further shows that the percentage
of low birth weight newborns for Kenya were lower as compared to those of Uganda and
the United Republic of Tanzania both over the 2000 – 2002 period and over 2005 – 2010
period. Rwanda has, however, lower percentages compared to Kenya for both periods.
According to the table, the percentage of newborns who are low birth weight for Kenya
are also lower than the averages for Africa, South–East Asia, and Eastern Mediterranean
regions. Although this might appear impressive, it is still important to find out what
can be done to reduce the percentage of low birth weight newborns in Kenya to as low
levels as those in Rwanda or even lower.
Prenatal care, also called antenatal care, refers to the medical attention provided
to an expectant mother throughout the period of pregnancy but excluding labour and
delivery (Garrido, 2009). In the ideal scenario, prenatal care should involve the following:
provision of appropriate advice on health matters such as nutrition, hygiene, newborn
care and safer sex; identification of expectant women at risk of experiencing pregnancy
complications through appropriate screening and diagnosis; and either the treatment of
identified pre–existing illnesses and conditions or, where treatment is not possible at the
particular health facility, referral to an appropriate health facility that can deal with
the identified conditions (Berg, 1995). Prenatal care can further benefit both expectant
mothers and their unborn children through identification of expectant mothers at risk
of delivering low birth weight infants or experiencing complications during delivery and
reducing such risks by providing appropriate psychosocial, nutritional, and medical
interventions (Kramer, 1987; Alexander and Korenbrot, 1995).
Prenatal Care and Infant Health: A Multilevel Analysis 24
Table 2.1: Low Birth Weight Newborns (%) for Selected Countries and Regions
2000 – 2002 2005 – 2010
Kenya 11 8
Uganda 12 14
United Republic of Tanzania 13 10
Nigeria 14 12
Rwanda 9 6
Africa 14 13
Americas 9 8
South–East Asia 26 24
Europe 8 7
Eastern Mediterranean 17 21
Western Pacific 8 5
Source: WHO (2006, 2012).
Several indicators have been used in the literature to measure prenatal care use. Table
2.2 on page 25 gives examples of such indicators.
The World Health Organization (WHO) recommends a minimum of four prenatal
care visits at particular intervals, to skilled health personnel (doctors or nurses), for
expectant women in developing countries (Berg, 1995). The recommended timing for
each visit is shown in Table 2.3 on page 26. According to the table, for example, the first
prenatal care visit should occur within the first 16 weeks of pregnancy while the third
visit should occur at 32 weeks of pregnancy. There are further detailed recommendations
on what should be done at each visit (see Berg, 1995). Garrido (2009) shows that
the recommendations of WHO regarding prenatal care use in developing countries are
appropriate. In this chapter, we construct a prenatal care utilization index based on
WHO’s recommendations.
The rest of the chapter is structured as follows: Section 2.2 reviews the literature,
Section 2.3 discusses the purpose and objectives of the study, Section 2.4 discusses the
methodology of the study, descriptive statistics are discussed in Section 2.5, the results
are presented in Section 2.6, and Section 2.7 gives the summary, conclusions and policy
implications.
Prenatal Care and Infant Health: A Multilevel Analysis 25
Table
2.2:
Ind
icat
ors
ofP
renat
alC
are
Indic
ator
Exam
ple
sof
Stu
die
sU
sing
the
Indic
ator
Num
ber
ofpre
nat
alca
revis
its
LaV
eist
,K
eith
and
Guti
erre
z,19
95;
Rou
s,Jew
ell
and
Bro
wn,
2004
;
Eva
ns
and
Lie
n,
2005
;N
azim
and
Fan
,20
11.
Num
ber
ofpre
nat
alca
revis
its
War
ner
,19
95.
adju
sted
for
pre
gnan
cyle
ngt
h
Whet
her
pre
nat
alca
rew
asev
erin
itia
ted
Jew
ell,
2009
.
Auth
or-c
onst
ruct
edqual
ity
Naz
iman
dF
an,
2011
.
index
for
typ
eof
care
rece
ived
Tim
ing
offirs
tpre
nat
alca
revis
itF
rank,
Str
obin
o,Sal
keve
ran
dJac
kso
n,
1992
;
Gol
den
ber
g,P
atte
rson
and
Fre
ese,
1992
;
LaV
eist
,K
eith
and
Guti
erre
z,19
95;
War
ner
,19
95;
Joy
ce,
1999
;C
onw
ayan
dK
uti
nov
a,20
06;
Weh
by
etal
.,20
09a;
Weh
by
etal
.,20
09b.
Kes
sner
index
ofad
equac
yof
Kot
elch
uck
,19
94a;
Joy
ce,
1994
;
pre
nat
alca
rere
ceiv
edL
aVei
st,
Kei
than
dG
uti
erre
z,19
95;
Del
gado-
Rodri
guez
,Gom
ez–O
lmed
o,B
uen
o–C
avan
illa
san
dCalv
ez–V
arga
s,19
97;
Con
way
and
Kuti
nov
a,20
06;
Jew
ell,
2009
.
Adeq
uac
yof
pre
nat
alca
reuti
liza
tion
index
Kot
elch
uck
,19
94a;
Kot
elch
uck
,19
94b.
Index
esbas
edon
WH
OB
erg,
1995
;G
arri
do,
2009
.
reco
mm
endat
ions
for
dev
elop
ing
countr
ies
Prenatal Care and Infant Health: A Multilevel Analysis 26
Table 2.3: WHO Recommended Minimum Prenatal Care Visits for Developing Countries
Visit Recommended timing (Pregnancy weeks)
First ≤ 16
Second 24 – 28
Third 32
Fourth 36 – 38
Source: Berg (1995).
2.2 Literature Review
IOM (1985), Kramer (1987), Kallan (1993), Guyatt and Snow (2004), Mitgitti et al.
(2008), Ohlsson and Shah (2008), and Darling and Atav (2012) summarize the literature
on the determinants of low birth weight. They identify a number of maternal risk factors1
for low birth weight. The factors are summarized in Table 2.4 on page 27.
According to the table, for example, an adolescent mother is at an increased risk of
delivering a low birth weight infant. The table also shows that mothers with a previous
history of preterm/low birth weight births are at an increased risk of giving birth to a
low birth weight infant.
Since absent or inadequate use of prenatal care is a risk factor for low birth weight
(IOM, 1985), we would expect prenatal care use to improve birth weight. Although many
studies show that this is in fact the case (see, for example, the studies cited in Table 2.5
on page 28), Kramer (1987) cites studies that fail to find an effect of prenatal care on
reducing the risk of low birth weight. Other studies (see, for example, Grossman and
Joyce, 1990; Currie and Grogger, 2002) only find weak influences of prenatal care on the
health of infants. Why is this the case? One answer suggested by Conway and Deb (2005)
is that studies that find prenatal care to be ineffective in birth weight determination
typically combine “complicated” and “normal” pregnancies. Conway and Deb (2005)
proceed to demonstrate the effectiveness of prenatal care on “normal” pregnancies.
Table 2.5 on page 28 summarizes the key findings from some of the studies that find
prenatal care to be effective in lowering the risk of low birth weight.
1These are factors whose possession or presence is associated with an increased probability of givingbirth to a low birth weight infant (IOM, 1985; Lopez et al., 2006).
Prenatal Care and Infant Health: A Multilevel Analysis 27
Table 2.4: Maternal Risk Factors for Low Birth Weight
Main Factor Detailed Breakdown
Historical factors Short or long birth interval
Previous history of preterm/low birth weight births
Maternal history of being low birth weight
Demographic factors Adolescent mothers
Minority race
Acculturation
Unmarried/Cohabiting
Parity (0 or more than 4)
Nutritional factors Iron deficiency
Lack of fish oil in diet
Anthropometric factors Low Body Mass Index
Medical and pregnancy–related conditions Anatomical factors
Uterine factors
Placental factors
Infections
Malaria
Human Immuno–deficiency Virus (HIV)
Bacterial vaginosis
Trichomoniasis
Syphilis
Gonorrhea
Urinary tract infection
Periodontal infection
Others
Multiple pregnancy
Psychosocial factors Adverse psychosocial factors
Lifestyle–related factors Tobacco use
Heavy alcohol use
Cocaine use
Narcotic use
Environmental factors Environmental tobacco exposure
Violence/Maternal abuse Violence/abuse
Maternal trauma
Infertility and in vitro fertilization (IVF) treatment
Health care risks Absent or inadequate prenatal care
Source: IOM (1985, Table 1, p.7); Kramer (1987), Kallan (1993), Guyatt and Snow (2004), Mitgitti etal. (2008); Ohlsson and Shah (2008, Table 16.1, p.259).
Prenatal Care and Infant Health: A Multilevel Analysis 28
Table
2.5:
Su
mm
ary
ofS
tud
ies
Sh
owin
gP
osit
ive
Eff
ects
ofP
ren
atal
Car
eon
Bir
thW
eigh
t
Asp
ect
ofP
renat
alC
are
Eff
ect
onB
irth
Wei
ght
Exam
ple
sof
Stu
die
s
Num
ber
ofpre
nat
alca
revis
its
Incr
ease
Joy
ce,
1999
;K
orou
kia
nan
dR
imm
,20
02;
Rou
s,Jew
ell
and
Bow
n,
2004
;L
in,
2004
;
Jew
ell
and
Tri
unfo
,20
06;
Raa
tika
inen
,H
eisk
anen
and
Hei
non
en,
2007
;
Naz
iman
dF
an,
2011
.
Qual
ity
ofpre
nat
alca
reIn
crea
seM
wab
u,
2009
;N
azim
and
Fan
,20
11.
Del
ayin
seek
ing
pre
nat
alca
reD
ecre
ase
Weh
by
etal
.,20
09a.
Ear
ly(fi
rst
trim
este
r)in
itia
tion
ofpre
nat
alca
reIn
crea
ses
Fra
nk,
Str
obin
o,Sal
keve
ran
dJac
kso
n,
1992
;
War
ner
,19
95;
Weh
by
etal
.,20
09a.
Inad
equat
euse
ofpre
nat
alca
reD
ecre
ases
IOM
,19
85;
Bar
ros,
Tav
ares
and
Rodri
gues
,19
96;
Kot
elch
uck
,19
94b;
Gol
dan
i,B
arbie
ri,
Silva
and
Bet
tiol
,20
04;
Raa
tika
inen
,H
eisk
anen
and
Hei
non
en,
2007
.
Pre
nat
alca
revis
its
lost
earl
yin
the
pre
gnan
cyD
ecre
ase
Eva
ns
and
Lie
n,
2005
.
Prenatal Care and Infant Health: A Multilevel Analysis 29
The literature identifies several factors that influence prenatal care use. Table 2.6 on
page 30 summarizes some of these factors.
2.3 Purpose and Objectives of the Study
Looking at the literature, we can conclude that the issue concerning the effectiveness of
prenatal care in reducing adverse pregnancy outcomes such as low birth weight is not yet
settled. This is evidenced by the fact that although many studies have demonstrated the
effectiveness of prenatal care in reducing the occurrence of adverse pregnancy outcomes,
some studies (see, for example those cited in Kramer, 1987) find it to be ineffective. A
further closer look at the literature also reveals that there have been studies in Kenya
investigating the effect of particular types of prenatal care, such as immunization of
the mother against tetanus, on birth weight in Kenya (see, for example, Mwabu, 2009).
Other studies (see, for example, Simiyu, 2004; Were and Bwibo, 2009) simply analyze
illnesses and deaths of low birth weight infants. Yet other studies (see, for example,
Magadi, Diamond, Madise and Smith, 2004; Brown et al., 2008) simply examine the
effect of antenatal care on birth outcomes. To the best of our knowledge there are no
studies so far that link the adequacy of prenatal care use to low birth weight in Kenya.
Furthermore, most of the existing studies also use a single–level modelling framework.
It is important to link the adequacy of prenatal care received to birth outcomes
because some mothers may attend clinics for insufficient number of times or may not
get the prenatal care from qualified providers, making it difficult for us to assess the
effectiveness of prenatal care received from skilled providers.
It is for this reason that this chapter contributes to the literature by examining the
effect of the adequacy of prenatal care use on the probability of delivering a low birth
weight infant in Kenya. The chapter further adds to the literature by using a multilevel
modelling framework in its analysis.
Specifically, the chapter pursues the following objectives:
1. Determining the effect of the adequacy of prenatal care use by the mother on infant
health, where infant health is measured by whether or not the infant has low birth
weight.
2. Determining the factors that influence the adequacy of prenatal care use.
Prenatal Care and Infant Health: A Multilevel Analysis 30
Table
2.6:
Fac
tors
Infl
uen
cin
gP
ren
atal
Car
eU
se
Fac
tor
Exam
ple
sof
Stu
die
s
Typ
eof
faci
lity
vis
ited
Ob
erm
eyer
and
Pot
ter,
1991
.
(that
is,
whet
her
pri
vate
orpublic)
Mat
ernal
dem
ogra
phic
,G
olden
ber
g,
Pat
ters
onan
dF
rees
e,19
92;
soci
o–ec
onom
ic,
situ
atio
nal
,D
elga
do-
Rodri
guez
,Gom
ez–O
lmed
o,B
uen
o–C
avan
illa
san
dGa
lvez
–Var
gas,
1997
;
and
psy
chol
ogic
alfa
ctor
sC
elik
and
Hot
chkis
s,20
00;
Lin
,20
04;
Chan
dhio
k,
Dhillo
n,
Kam
bo
and
Sax
ena,
2006
;B
row
net
al.,
2008
;
Weh
by
etal
.,20
09a;
Jew
ell,
2009
.
Rac
eL
aVei
st,
Kei
than
dG
uti
erre
z,19
95.
Mat
ernal
hea
lth
stat
us
Weh
by
etal
.,20
09a.
Prenatal Care and Infant Health: A Multilevel Analysis 31
Figure 2.1: A Conceptual Framework for Analyzing the Effect of Prenatal Care Use on InfantHealth
Observed Variables Unobserved Variables
Maternal / Household
Preferences
Biology/Genetics/Maternal Health
Maternal /Household Demographic and Socio – economic Characteristics
Community Characteristics /Environmental factors
Prenatal Care
Infant Health
(Birth Weight)
Source: Author’s Own Construction Based on Figure 1.1 and Schultz (1984)
3. Drawing appropriate policy implications from the findings.
2.4 Methodology
Conceptual Framework
Based on the general framework presented in Figure 1.1 and Schultz (1984), we can
develop the conceptual framework shown in Figure 2.1 for the analysis of the effect of
prenatal care use on infant health.
According to the framework (also called the conceptual model), infant health (as
measured by birth weight) is influenced by prenatal care use and unobservable bio-
logical endowments of both the mother and the child, including true maternal health
status. Prenatal care use, in turn, is influenced by maternal/household demographic and
Prenatal Care and Infant Health: A Multilevel Analysis 32
socio–economic characteristics, community characteristics or environmental factors, and
unobservable maternal/household preferences. Prenatal care use is also influenced by
unobservable biological endowments of the mother and the unborn child.
Theoretical Framework
Following the general modelling framework discussed in Chapter 1, which is based on
Rosenzweig and Schultz (1982, 1983) and Mwabu (2009), we assume that an expectant
mother, j, maximizes the utility, Uj, obtained from her consumption of various goods
and services that have no impact on the health of her unborn child, Xj, and the health
status of her unborn child, Hj . We can represent the expectant mother’s utility function
as follows:
Uj = U(Xj, Hj). (2.1)
We assume that the health status of the unborn child, Hj, is in turn influenced by
the adequacy of prenatal care use, Zj, that affects health directly, other factors, Y , and
unobservable biological endowments, µj. The health production function of the unborn
child can, therefore, be represented by the following:
Hj = H (Zj, Y, µj) . (2.2)
The mother is assumed to maximize her utility function subject to the above health
production function and a budget constraint given by:
Ij = PxXj + PyY + PzZj (2.3)
where I is exogenous mother’s/household’s income, Px is the unit price of X, Py is a
vector of the unit prices for the respective components of Y , and Pz is the unit price of
Z.
Mwabu (2009) shows that manipulation of the above equations leads to the following
input demand equations:
Xj = X(Px, Py, Pz, Ij, µj) (2.4)
Zj = Z(Px, Py, Pz, Ij, µj) (2.5)
Prenatal Care and Infant Health: A Multilevel Analysis 33
As observed by Mwabu (2009), these equations show that commodity prices are
correlated with infant health.
Estimation Issues
As mentioned in Section 1.6, when estimating our models, we need to worry about
potential sample selection bias, potential endogeneity of some of the covariates, and
potential unobserved heterogeneity.
Sample Selection Bias
Our discussion of sample selection bias follows closely that by Vella (1998) on the same
issue. In this chapter, we only observe the birth weight of a child if it (the birth weight) is
reported in our dataset. Not all children in the dataset, however, have their birth weights
reported. We are concerned with determining the effect of prenatal care on the birth
weight of the children with reported birth weights so as to draw conclusions regarding
the effect of prenatal care on the birth weight of all children. Whether or not we have
sample selection bias depends on the difference between children whose birth weights
are reported and those for whom birth weights are not reported. If the sub–sample
of children with reported birth weights is randomly drawn from the population of all
children, selectivity bias will not arise if we were to just study this particular sub–sample.
If, however, non–reporting of birth weight is non–random, then the sample of children
with reported birth weights is also non–random. This implies that the children with
reported birth weights and those whose birth weights are not reported have different
characteristics. Sample selection bias will arise when some of the factors influencing the
reporting of birth weight also determine birth weight. If these factors are observable,
they can be included in the birth weight equation as conditioning variables. In such a
case there will be no sample selection bias.
If, however, the unobservable factors affecting the decision to report the birth weight
of the child are correlated with the unobservable factors affecting the birth weight itself,
then a relationship exists between reporting of birth weight and the process determining
the birth weight (Vella, 1998). In such a case, it is not enough to control for the observable
factors when explaining birth weight because the process determining whether or not
birth weight is reported also influences birth weight. If these unobservable factors (from
Prenatal Care and Infant Health: A Multilevel Analysis 34
the birth weight reporting equation) are correlated with the observable factors in the
birth weight equation, then failure to include an estimate of the unobservable factors in
the birth weight equation leads to sample selection bias.
As mentioned in Section 1.6, although several approaches to correcting for sample
selection bias have been proposed in the literature (see, for example, Heckman, 1979;
Olsen, 1980; Lee, 1982; Garen, 1984; Vella 1992, 1993), we use the approach suggested
by Vella (1992) since in our case the dependent variable is discrete.
This approach makes use of the generalized residuals from a probit regression (for
a discussion of generalized residuals, see Gourieroux, Monfort, Renault and Trognon,
1987). The approach typically involves identifying the selection factor and estimating a
probit model of the selection factor. Generalized residuals of the resulting model are then
included together with the selection factor in the equation of primary interest. Statistical
significance of the coefficient of the generalized residuals in the equation of primary
interest is then indicative of the presence of selection bias. If the estimated coefficient
of the generalized residuals is not statistically significant, however, there is no selection
bias.
In our case the selection factor is the reporting of birth weight. We, therefore,
estimate a model relating the decision to report birth weight to household and individual
characteristics, and some instruments. We then include in the equation of primary
interest (our infant health model) generalized residuals from this model together with a
variable indicating whether or not birth weight is reported.
Endogeneity
In our infant health model, we suspect that the covariate measuring the adequacy of
prenatal care use is endogenous. We suspect that this endogeneity is due to the presence
of unobservable factors in the infant health equation that are correlated with the adequacy
of prenatal care use chosen by the mother. If this endogeneity is not corrected for, the
resulting estimate of the coefficient of prenatal care use in our infant health model would
be inconsistent (Guevara and Ben–Akiva, 2008).
We employ the Two–Stage Residual Inclusion (2SRI) method (Terza, Basu and
Rathouz, 2008) in an attempt to correct for this endogeneity. For simplicity, we assume
that this is the only endogenous covariate in our model. Following Bollen, Guilkey
and Mroz (1995), we test for the endogeneity of the adequacy of prenatal care use in
Prenatal Care and Infant Health: A Multilevel Analysis 35
the infant health production equation by testing for the statistical significance of the
generalized residuals from the prenatal care equation. If the coefficient of the residuals
is significantly different from zero, then the adequacy of prenatal care use variable is
endogenous; otherwise, it is exogenous.
Unobserved Heterogeneity
Unobserved heterogeneity will exist in our infant health model if there are some unobserv-
able factors that interact non-linearly with the adequacy of prenatal care use causing the
effect of prenatal care use on infant health to differ amongst children in the population
(Zohoori and Savitz, 1997). As mentioned in Section 1.6, the standard procedure for
correcting for unobserved heterogeneity is the control function approach (see, for example,
Florens, Heckman, Meghir and Vytlacil, 2008; Mwabu, 2009; Wooldridge, 2010). We
adopt this approach for our model.
This approach involves including in our equation of primary interest (in our case, the
infant health equation) interactions between the residuals and the endogenous explanatory
variable. If the coefficient of the resulting interaction term is statistically significant,
there is unobserved heterogeneity in our infant health model. If the coefficient is not
statistically significant, there is no unobserved heterogeneity in our infant health model.
Model Identification
Our discussion in this subsection follows that in Mwabu (2009) on the issue of model
identification. So as to properly interpret the estimated parameters of our infant health
model, it is important that infant health effects of the endogenous covariate (in our
case, the adequacy of prenatal care use) and of the sample selection rule be identified.
Because we have one endogenous variable in our model, identification requires at least two
exclusion restrictions since we have a situation that requires the simultaneous solution of
two equations. A minimum of one instrument is required for the endogenous covariate
and another minimum of one exogenous variable is also required for the determination of
selection of observations into the sub–sample to be used for estimation purposes (Mwabu,
2009).
The variables chosen as instruments should be uncorrelated with the stochastic error
term in the infant health equation (i.e. they should be valid or exogenous), should be
Prenatal Care and Infant Health: A Multilevel Analysis 36
correlated with the endogenous variable in the infant health equation (i.e. they should
be relevant, or rather, their effects on the endogenous explanatory variable in the infant
health equation should be statistically significant), and should be excluded from the infant
health equation (Wooldridge, 2002; Murray, 2006; Mwabu, 2009; Brookhart, Rassen and
Schneeweiss, 2010). Guevara–Cue (2010) discusses various tests that can be used to
check the validity of instrumental variables in discrete choice models.
In our case, therefore, the variables we use as instruments for prenatal care use should
first, affect prenatal care use or be associated with prenatal care use; second, they should
be unrelated to mother or household characteristics; and third, they should be related to
birth weight only through their association with prenatal care (Brookhart, Rassen and
Schneeweiss, 2010).
Examples of variables that have been used as instruments for prenatal care in the
literature include number of prenatal care clinics or providers per capita (see, for example,
Wehby et al., 2009a), distance from residence to prenatal care clinics (see, for example,
Wehby et al., 2009a; Wehby et al., 2009c), population per hospital bed (see, for example,
Wehby et al., 2009a), unemployment rate (see, for example, Wehby et al., 2009a), rate of
uninsured females (see, for example, Wehby et al., 2009a; Wehby et al., 2009c), price of
prenatal care (see, for example, Wehby et al., 2009c), bus strikes (see, for example, Evans
and Lien, 2005), whether mother cohabits with father of child (see, for example, Conway
and Deb, 2005), and mother’s income (see, for example, Conway and Deb, 2005).
We use the “average distance to the nearest health facility” and the “health facilities
per 100,000 of population” as instruments in our models. Our models are, therefore,
exactly identified (Murray, 2006). The choice of distance to the nearest health facility as
an instrument is based on the assumption that distances to health facilities are correlated
with prenatal care use. Since mothers have other uses for their time (such as engaging in
paid work, housework, and child care), they must optimally allocate the time available
to them amongst the various uses (Becker, 1965; Agenor and Canuto, 2012). The longer
the distance to the nearest health facility, the higher the opportunity cost to the mother
of visiting the facility for prenatal care. Research (see, for example, Arcury et al., 2005;
Fortney et al., 2005; Kosimbei, 2005; Qian et al., 2009) actually shows that distance
to the health facility significantly influences the utilization of health care services. We
would, therefore, expect mothers’ utilization of prenatal care to be limited the longer the
distance to the nearest health facility. Consequently, we expect mother’s utilization of
prenatal care to be inadequate the longer the distance to the nearest health facility.
Prenatal Care and Infant Health: A Multilevel Analysis 37
One argument in the literature against the use of distance to the nearest health
facility as an instrumental variable is that mothers can choose to live near health facilities
because of their health–related problems or preferences (Rosenzweig and Schultz, 1982;
Garrido, 2009). This then undermines the argument that the distances are exogenous.
To overcome this possibility, we use provincial level averages for the distance to the
nearest facility. This is because, even though an individual mother may choose to live
near a health facility because of health issues or preferences, all the women in a province
are unlikely to make this decision simultaneously every time they are pregnant. As such,
an individual woman’s decision may not immediately affect the average distance to the
nearest health facility in a province. Furthermore, if the relocation of a mother is from
one area of the province to another area of the province, this does not change the average
distance to the nearest health facility in the province.
The health facilities per 100,000 of population is aimed at indicating the overall
accessibility and availability of health care in a particular region. We expect that the
higher the number of health facilities per 100,000 of population, the more the health
care, including prenatal care, is accessible and available for use. Consequently, we expect
that the higher the number of health facilities per 100,000 of population, the higher the
probability of adequate prenatal care use.
We test the validity of our instruments using a procedure due to Guevara–Cue (2010).
This procedure involves four steps. In the first step, we estimate a model of prenatal
care in which we include both instruments. We then obtain the generalized residuals
of this model. In the second step, we estimate a model of infant health in which the
residuals from the prenatal care model are included as an additional regressor. Call this
“Model 1”. The third step involves a re–estimation of the infant health model estimated
in the second step but including one of the instruments as an additional regressor. Call
this “Model 2”. In the final step, a likelihood ratio test is conducted comparing the two
models. A failure to reject the null hypothesis is indicative of instrument validity.
Empirical Model
We need to formulate an empirical version of the infant health model. Since we are
measuring the infant’s health status by considering whether or not the infant has low
birth weight, our dependent variable is binary and the appropriate model in our case is the
binary regression model (Long, 1997; Long and Freese, 2006). A comprehensive discussion
of the binary regression model can be found in Long (1997), Long and Freese (2006),
Prenatal Care and Infant Health: A Multilevel Analysis 38
Wooldridge (2002, 2010), and Greene (2008, 2012). We formulate both a single–level
model and a multilevel model of infant health.
Single–level Model
Long and Freese (2006, p.132) discuss the three methods that can be used to derive the
binary regression model: assuming that there is an unobserved variable that is linked
to the observed outcome through a measurement equation, constructing the model as a
probability model, and generating the model as a random utility model. We adopt the
latent variable method because of its appeal to intuition.
To formulate the single–level model, we let Hi be the observed health status of the
ith infant which is defined as follows:
Hi =
1 if infant i has low birth weight,
0 otherwise.(2.6)
We then assume that there is an unobserved or latent continuous variable H∗i (repre-
senting, for example, a true infant health status), which is linked to the observed infant
health status via the following equation:
Hi =
1 if H∗i > 0,
0 otherwise.(2.7)
The latent variable is, in turn, linked to covariates by the following equation:
H∗i = β1 + β2Zi + β3Y + ε1i (2.8)
where Z is an indicator of the adequacy of prenatal care use, Y is a vector of controls,
and ε1 is a stochastic error term.
For given values of Z and Y , it can be shown (see, for example, Long and Freese,
2006; Cameron and Trivedi, 2010) that
Prenatal Care and Infant Health: A Multilevel Analysis 39
Pr (Hi = 1 | Z, Y ) = Pr (H∗i > 0 | Z, Y ) (2.9)
where Pr stands for “probability”.
From Equations (2.8) and (2.9), it can be shown that
Pr (Hi = 1 | Z, Y ) = Pr (−ε1i ≤ β1 + β2Zi + β3Y | Z, Y ) = F (β1 + β2Zi + β3Y )
(2.10)
where F (·) is the cumulative density function of −ε1i, which is the same as that of ε1i if
ε1i is symmetric.
If we assume that ε1i follows a logistic distribution, we end up with the logit model;
while if we assume that ε1i follows the standard normal distribution, we end up with the
probit model (Rabe–Hesketh and Skrondal, 2008). We adopt the probit model because
of its popularity in the literature.
In our case, the relevant probit model is given by:
Pr (Hi = 1 | Z, Y ) = F (β1 + β2Zi + β3Y ) = Φ (β1 + β2Zi + β3Y ) (2.11)
where Φ(·) is the standard normal cumulative distribution function.
Because Z is potentially endogenous in Equation (2.8), we have to control for this
potential endogeneity. To use the Two–Stage–Residual–Inclusion method to control for
this potential endogeneity, we estimate a model for the adequacy of prenatal care use,
obtain generalized residuals from the estimated model (Gourieroux, Monfort, Renault
and Trognon, 1987), and then include these generalized residuals together with the
adequacy of prenatal care variable in our structural equation of interest.
The adequacy of prenatal care use variable is constructed based on the WHO recom-
mendations (Berg, 1995). In particular, we consider prenatal care use to be adequate
if it is obtained from a skilled provider, the total number of visits is at least four, and
the first visit occurs within four months of pregnancy. The adequacy of prenatal care
Prenatal Care and Infant Health: A Multilevel Analysis 40
variable is defined as follows:
Zi =
1 if mother sought adequate prenatal care while pregnant with infant i,
0 otherwise.
(2.12)
Using the latent variable formulation, we can define a latent variable Z∗i that is related
to Zi via the following equation:
Zi =
1 if Z∗i > 0,
0 otherwise.(2.13)
This latent variable is linked to the covariates using the equation
Z∗i = α1 + α2Y + α3Q+ ε2i (2.14)
where Y is a vector of controls, Q is a vector of instruments, and ε2 is a stochastic error
term.
Assuming a standard normal distribution for ε2 leads to a probit model given by:
Pr (Zi = 1) = Φ (α1 + α2Y + α3Q) (2.15)
We estimate this model, obtain its generalized residuals, and include the generalized
residuals as an additional variable in the structural equation of interest.
To control for possible non–random selection of the estimation sample, we also estimate
a sample selection equation. Let selection into the sample be given by the following
Ii =
1 if infant’s birth weight is reported,
0 otherwise.(2.16)
Prenatal Care and Infant Health: A Multilevel Analysis 41
We assume that there is a latent variable linked to the reporting of birth weight and
related to the covariates via the following equation:
I∗i = γ1 + γ2Y + γ3Q+ ε3i (2.17)
where Y is a vector of controls, Q is a vector of instruments, and ε3 is a stochastic error
term.
Assuming a standard normal distribution for ε3 leads to a probit model given by:
Pr (Ii = 1) = Φ (γ1 + γ2Y + γ3Q) (2.18)
We also include the generalized residuals from the estimation of the above model in
our structural equation of interest.
To control for potential unobserved heterogeneity, we include the interaction of the
adequacy of prenatal care with the generalized residuals from the prenatal care equation.
Equation (2.8) is, therefore, extended as follows:
H∗i = β1 + β2Zi + β3Y + β4Ii + β5ε2i + β6ε3i + β7Zε2i + ε1i (2.19)
where Z is an indicator of the adequacy of prenatal care use, Y is a vector of controls, I
is an indicator of whether or not birth weight is reported, ε2 are generalized residuals
from the prenatal care model, ε3i are generalized residuals from the sample selection
model, and ε1 is a stochastic error term.
Consequently, we end up with the following single–level probit model
Pr (Hi = 1) = Φ (β1 + β2Zi + β3Y + β4Ii + β5ε2i + β6ε3i + β7Zε2i) . (2.20)
Multilevel Analysis
Sometimes economists use data collected at different levels of analysis, for example, most
household surveys collect data at the individual, household and even community levels.
Another example is where we may have data collected on the characteristics of children,
their mothers, and the regions they live in. Such data could easily be organized in a
hierarchical manner where, for instance, an individual is linked to a particular household
Prenatal Care and Infant Health: A Multilevel Analysis 42
Figure 2.2: An Example of a Multilevel Data Structure
Level 2: Mother Mother 1 Mother 2
Child 1 Child 2 Child 3 Child 4 Level 1: Child
Mother 3
which is in turn linked to a particular community. Data organized in such a manner
where units at one level are linked to more larger units at a higher level is referred to as
multilevel data (Steenbergen and Jones, 2002)2. The analysis of multilevel data is referred
to as multilevel analysis (Snijders and Bosker, 1999). For a discussion of substantive and
statistical motivation for using multilevel analysis, see Steenbergen and Jones (2002).
Figure 2.2 shows an example of a multilevel data structure. In the figure, children
are the level 1 units while mothers are level 2 units.3 Children are nested in mothers.
Notice from the figure that it is possible for a mother to have more than one child, as is
the case with mother 2 in the figure, but it is not possible for a child to have more than
one biological mother.
In multilevel analysis, we aim at explaining the variability in a predicted variable
obtained at the lowest level in the data structure based on data gathered at all the levels
of nesting (Snijders and Bosker, 1999; Steenbergen and Jones, 2002). For example, for
the data structure shown in Figure 2.2, multilevel analysis would aim at explaining the
2In general, multilevel data may consist of several levels of analysis that can be related to one anotherin some hierarchical order (Steenbergen and Jones, 2002).
3It is conventional in multilevel analysis to refer to the lowest level of analysis as level 1, the nexthighest level of analysis as level 2, and so on (see, for example, Steenbergen and Jones, 2002).
Prenatal Care and Infant Health: A Multilevel Analysis 43
variability in a predicted variable measured at the child level, such as birth weight, based
on data gathered from both children and mothers.
Multi–level Model
The multilevel model is obtained by breaking the stochastic error terms in our single–
level models into two parts, a mother–specific component, ζ, and an infant–specific
component, ε. The mother–specific component, ζ, controls for unobservable mother–
specific characteristics that affect the dependent variable of interest (e.g. infant health,
adequacy of prenatal care use, reporting of birth weight), and is assumed to remain
unchanged across infants born to the same mother but independent across mothers (Rabe–
Hesketh and Skrondal, 2008). The infant–specific component, ε, varies between infants
as well as mothers but is assumed to be independent across both infants and mothers
(Rabe–Hesketh and Skrondal, 2008). It is also further assumed that ζ is independent of
ε (Rabe–Hesketh and Skrondal, 2008)
The multilevel counterparts of our binary outcomes are as follows:
Hij =
1 if infant i from mother j has low birth weight,
0 otherwise.(2.21)
Zij =
1 if mother j sought adequate prenatal care when pregnant with infant i,
0 otherwise.
(2.22)
Iij =
1 if the birth weight for infant i from mother j is reported,
0 otherwise.(2.23)
For the multilevel case, the binary outcomes are related to the respective latent
continuous responses via the following equations:
Hij =
1 if H∗ij > 0,
0 otherwise.(2.24)
Prenatal Care and Infant Health: A Multilevel Analysis 44
Zij =
1 if Z∗ij > 0,
0 otherwise.(2.25)
Iij =
1 if I∗ij > 0,
0 otherwise.(2.26)
The multilevel latent responses are given by:
H∗ij = β1 + β2Zij + β3Y + ζ1j + ε1ij (2.27)
Z∗ij = α1 + α2Y + α3Q+ ζ2j + ε2ij (2.28)
I∗ij = γ1 + γ2Y + γ3Q+ ζ3j + ε3ij (2.29)
where
Y is a vector of controls,
Q is a vector of instruments,
ζ1j, ζ2j, ζ3j are random intercepts that control for unobservable mother–specific char-
acteristics,
ε1ij, ε2ij, ε3ij are infant–specific stochastic error terms.
We assume that ζ1j ∼ N (0, ψ1), ζ2j ∼ N (0, ψ2), and ζ3j ∼ N (0, ψ3). ε1ij, ε2ij, ε3ij are
all assumed to follow the standard normal distribution.
The corresponding multilevel probit models are given by:
Pr (Hij = 1) = Φ (β1 + β2Zij + β3Y + ζ1j) (2.30)
Pr (Zij = 1) = Φ (α1 + α2Y + α3Q+ ζ2j) (2.31)
Pr (Iij = 1) = Φ (γ1 + γ2Y + γ3Q+ ζ3j) (2.32)
Prenatal Care and Infant Health: A Multilevel Analysis 45
To control for potential endogeneity of prenatal care, potential sample selection bias
and potential unobserved heterogeneity, we extend Equation (2.27) as follows:
H∗ij = β1 + β2Zij + β3Y + β4Iij + β5ε2ij + β6Zij ε2ij + β7ε3ij + ζ1j + ε1ij (2.33)
where ε2ij are generalized residuals from the multilevel prenatal care model, and ε3ij
are the generalized residuals from the multilevel sample selection model.
The resulting multilevel probit model is given by:
Pr (Hij = 1) = Φ (β1 + β2Zij + β3Y + β4Iij + β5ε2ij + β6Zij ε2ij + β7ε3ij + ζ1j) . (2.34)
We can quantify the dependence among the various dichotomous responses for the
same mother by using the conditional intraclass correlation, ρ, of the latent responses
H∗ij (Rabe–Hesketh and Skrondal, 2008). For our models, this is given by:
ρ =ψ
ψ + 1. (2.35)
To check whether multilevel modelling is appropriate for our study, we perform a
Likelihood Ratio test for ρ = 0, which is equivalent to testing whether ψ = 0 (Rabe–
Hesketh and Skrondal, 2008). If we reject this hypothesis then the implication is that a
multilevel model is appropriate; a failure to reject the hypothesis would indicate that the
multilevel model is not appropriate (Rabe–Hesketh and Skrondal, 2008).
We estimate our models using Stata 11 software (StataCorp, 2009). The multilevel
models are estimated using the gllamm command (Rabe–Hesketh, Skrondal and Pickles,
2002; Rabe–Hesketh, Skrondal and Pickes, 2004; Rabe–Hesketh, Skrondal and Pickles,
2005; Rabe–Hesketh and Skrondal, 2008) that runs in Stata 11 (StataCorp, 2009).
Data
The main data set we use are the Demographic and Health Survey (DHS) data sets
for Kenya collected in 1993, 1998, 2003, and 20084. A good guide to Demographic and
Health Survey (DHS) data sets can be found in Rutstein and Rojas (2006). Demographic
4More information on Demographic and Health Surveys can be obtained by visiting http://www.
measuredhs.com/What-We-Do/Survey-Types/DHS.cfm
Prenatal Care and Infant Health: A Multilevel Analysis 46
and Health Surveys are nationally representative household surveys that provide a wide
range of household level data on child and maternal health.
Data on distances to the nearest health facilities comes from the 2005 Kenya Integrated
Household and Budget Survey (KIHBS), conducted by the Kenya National Bureau of
Statistics. Data on health facilities per 100,000 of population was obtained from the
Ministry of Health.
Table 2.7 on page 47 shows the variable definitions for the various variables found in
our models.
2.5 Descriptive Statistics
The overall sample pools together observations for 1993, 1998, 2003, and 2008. A key
observation of our dataset is that some children in the dataset do not have their birth
weights reported. Table 2.8 on page 48 summarizes, for our pooled sample, information
about reporting of birth weight by year of survey.
From the table, we can observe that about 55% of the children in the pooled sample
do not have information on birth weight. As explained in Section 2.4 above, this is likely
to lead to sample selection bias and is dealt with using the approach suggested by Vella
(1992).
Our analytic samples differ depending on the model being estimated due to missing
values on some of the other variables. For instance, our sample selection model uses
18,974 observations, our adequacy of prenatal care use model (the demand model for
adequate prenatal care) uses 15,713 observations while our birth weight model uses 7,331
observations.
Table 2.9 on page 49 presents the descriptive statistics for some of the variables in
our models, based on the pooled sample.
Table 2.10 on page 50 shows that only 533 (or 7.3%) of the children in the analytic
sample are low birth weight children. We can note from the table, however, that the
proportion of children who are low birth weight declined steadily from about 8.5% in
1993 to about 5.4% in 2008.
Prenatal Care and Infant Health: A Multilevel Analysis 47
Table 2.7: Variable Definitions for Prenatal Care and Infant Health Models
Variable Definition
Low birth weight 1 if child’s weight at birth is less
than 2,500 grams, 0 otherwise.
Birth weight reported 1 if child’s birth weight is reported,
0 otherwise.
Adequate prenatal care 1 if prenatal care use is sought from a
skilled provider (doctor or nurse), the total number of visits is
at least four, and the first prenatal care
visit occurs within four
months of pregnancy; 0 otherwise.
Mother’s age at birth of child mother’s age at time of birth of child in years.
Mother’s education Mother’s years of education.
Parity Number of previous births to the mother.
Number of living children Number of living children born to mother .
Mother wanted pregnancy 1 if Yes, 0 otherwise.
Urban residence 1 if area of residence is urban, 0 otherwise.
Male child 1 if sex of child is male, 0 otherwise.
Twin or multiple birth child 1 if child is twin or from a multiple birth, 0 otherwise.
First born child 1 if child is first born, 0 otherwise.
Asset index Household’s asset index, ranges from 0 to 10.
1993 1 if survey year is 1993.
1998 1 if survey year is 1998.
2003 1 if survey year is 2003.
2008 1 if survey year is 2008.
Average distance to nearest health facility Provincial level average distance to nearest
health facility in kilometres.
Health facilities per 100,000 Number of health facilities
per 100,000 of population, measured at provincial level.
Health facilities per 100,000 squared Square of number of health facilities
per 100,000 of population.
Selection residual Generalized residuals from the selection model
Prenatal care residual Generalized residuals from
the prenatal care model.
Interaction of prenatal care with residual Interaction of adequacy of
prenatal care with its residual.
Table 2.11 on page 50 shows prenatal care use by skill of provider for the analytic
sample for the adequacy of prenatal care model. According to the table, in about 92% of
the cases, prenatal care was sought from a skilled provider (that is, either a doctor or a
Prenatal Care and Infant Health: A Multilevel Analysis 48
Table 2.8: Reporting of Birth Weight by Year of Survey
Year of Survey Reported Birth Weight Total
No Yes
1993 3,450 2,665 6,115
(56.4%) (43.6%) (100%)
1998 2,008 1,523 3,531
(56.9%) (43.1%) (100%)
2003 3,271 2,678 5,949
(55%) (45%) (100%)
2008 3,199 2,880 6,079
(52.6%) (47.4%) (100%)
Total 11,928 9,746 21,674
(55%) (45%) (100%)
Note: Percentages are shown in parentheses.
nurse). Table 2.12 on page 51 shows the adequacy of prenatal care use by year of survey.
According to the table, about 74% of the observations are associated with inadequate
use of prenatal care. Table 2.13 on page 51 shows the distribution of low–birth weight
status by adequacy of prenatal care use for our analytic sample for the birth weight
model. From the table, about 68% of low birth weight observations also had inadequate
use of prenatal care.
2.6 Results
First–Stage Models
We estimate our models in two stages. In the first stage, we estimate sample selection
models and prenatal care models. In the second stage, we estimate the low birth weight
models.
One important question we may want to answer after the estimation of our models is
how changes in the explanatory variables affect the probabilities of a positive outcome.
Prenatal Care and Infant Health: A Multilevel Analysis 49
Table 2.9: Descriptive Statistics for the Variables in Prenatal Care and Infant Health Models
Variable Number of Mean Standard Minimum Maximum
Observations Deviation
Low birth weight 9,746 0.076 0.265 0 1
Birth weight reported 21,674 0.450 0.497 0 1
Adequate Prenatal Care 17,617 0.262 0.440 0 1
Mother’s age at birth of child 20,070 26.305 6.50 12 49
Mother’s education 21,672 6.212 3.99 0 23
Parity 21,674 0.355 0.577 0 5
Number of living children 21,674 3.563 2.266 0 17
Mother wanted pregnancy 21,674 0.555 0.497 0 1
Urban residence 21,674 0.194 0.395 0 1
Male child 21,674 0.507 0.50 0 1
Twin or multiple birth child 21,674 0.031 0.173 0 1
First born child 21,674 0.230 0.421 0 1
Asset index 20,500 3.476 1.905 0 10
Average distance to nearest health facility 21,674 8.043 4.422 3.11 22.64
Health facilities per 100,000 21,674 17.900 4.736 12 26
Health facilities per 100,000 squared 21,674 342.855 175.354 144 676
This question can be answered by reporting the marginal effects of the respective covariates
(Long, 1997). The marginal effect is computed by taking the partial derivative of the
estimated probability model with respect to the variable of interest (Long, 1997). As
shown by Long (1997, p.72), the resulting partial derivative is a function of all the
variables and the estimated parameters in the model and assumes the same sign as the
estimated coefficient.
This partial derivative can either be evaluated at the means of the various variables,
leading to what is called the marginal effect at the means, or it can be computed for
each observation and then averaged over all observations, leading to what is referred to
as the average marginal effect (Long, 1997; Bartus, 2005; Greene, 2008; Cornelissen and
Sonderhof, 2009). The average marginal effects are preferable to the marginal effects
at means (Bartus, 2005; Greene, 2008). We, therefore, compute and report the average
marginal effects for the variables in our models.
For a dummy explanatory variable, the marginal effect is given by the differences in
the probabilities when the variable assumes the value of 1 and when it assumes the value
of 0 (Long, 1997).
Prenatal Care and Infant Health: A Multilevel Analysis 50
Table 2.10: Low Birth Weight Status by Year of Survey
Year of Survey Child Low–Birth Weight Total
No Yes
1993 2,260 210 2,470
(91.5%) (8.5%) (100%)
1998 1,312 115 1,427
(91.9%) (8.1%) (100%)
2003 1,495 110 1,605
(93.1%) (6.9%) (100%)
2008 1,731 98 1,829
(94.6%) (5.4%) (100%)
Total 6,798 533 7,331
(92.7%) (7.3%) (100%)
Note: Percentages are indicated in parentheses.
Table 2.11: Prenatal Care Use by Skill of Provider
Year of Survey Provider of Prenatal Care Total
Unskilled Skilled
1993 262 5,808 6,070
(4.3%) (95.7%) (100%)
1998 253 3,256 3,509
(7.2%) (92.8%) (100%)
2003 547 3,404 3,951
(13.8%) (86.2%) (100%)
2008 425 3,648 4,073
(10.4%) (89.6%) (100%)
Total 1,487 16,116 17,603
(8.4%) (91.6%) (100%)
Note: Percentages are shown in parentheses.
Table 2.14 on page 52 shows the results for the sample selection model and the
adequacy of prenatal care model. We show the results for the multilevel model and those
for the single level model, for comparison purposes. The single level model results are
Prenatal Care and Infant Health: A Multilevel Analysis 51
Table 2.12: Adequacy of Prenatal Care Use, Analytic Sample for Prenatal Care Use
Year of Survey Adequacy of Prenatal Care Use Total
Inadequate Adequate
1993 4,272 1,793 6,065
(70.4%) (29.6%) (100%)
1998 2,585 928 3,513
(73.6%) (26.4%) (100%)
2003 3,081 877 3,958
(77.8%) (22.2%) (100%)
2008 3,068 1,013 4,081
(75.2%) (24.8%) (100%)
Total 13,006 4,611 17,617
(73.8%) (26.2%) (100%)
Note: Percentages are shown in parentheses.
Table 2.13: Low Birth Weight and Adequacy of Prenatal Care Use
Low Birth Weight Adequacy of Prenatal Care Use Total
Inadequate Adequate
No 4,701 2,097 6,798
(69.2%) (30.8%) (100%)
Yes 361 172 533
(67.7%) (32.3%) (100%)
Total 5,062 2,269 7,331
(69%) (31%) (100%)
Note: Percentages are shown in parentheses.
Source: Author’s Own Computation.
shown in columns (1) and (3) of the table while the multilevel model results are shown
in columns (2) and (4) of the table.
Columns (1) and (2) in the table show the results for the sample selection model
while columns (3) and (4) show the results for the adequacy of prenatal care model. A
careful look at columns (1) and (2) indicates that mothers who are older, more educated,
have fewer living children, wanted their pregnancies, reside in urban areas, have twin
Prenatal Care and Infant Health: A Multilevel Analysis 52
Table 2.14: Average Marginal Effects for Sample Selection and Prenatal Care Probit Models,Robust Z Statistics in Parentheses
Variable Sample Selection Model Prenatal Care Model
(Report birth weight =1) (Adequate prenatal care =1)
(1) (2) (3) (4)
Mother’s age at birth of child 0.011 0.025 0.009 0.012
(2.54) (2.88) (2.09) (2.25)
Square of mother’s age at birth of child -0.000048 -0.00010 -0.000097 -0.00013
(-0.65) (-0.70) (-1.26) (-1.42)
Mother’s education 0.026 0.058 0.005 0.006
(24.41) (18.29) (4.47) (4.74)
Parity -0.013 0.008 0.033 0.038
(-1.97) (0.72) (3.57) (3.94)
Number of living children -0.027 -0.062 -0.014 -0.016
(-9.64) (-9.86) (-4.61) (-4.54)
Mother wanted pregnancy 0.039 0.078 0.056 0.064
(5.84) (6.00) (7.91) (8.05)
Urban residence 0.163 0.347 0.031 0.036
(16.50) (15.88) (2.98) (2.90)
Male child 0.013 0.019 0.005 0.007
(2.06) (1.63) (0.72) (0.98)
Twin or multiple birth child 0.095 0.182 0.037 0.038
(4.56) (4.19) (1.49) (1.16)
First born child 0.097 0.201 0.008 0.008
(9.04) (9.86) (0.69) (0.68)
Asset index 0.050 0.112 0.015 0.016
(23.56) (18.03) (6.43) (6.12)
1998 -0.070 -0.140 -0.045 -0.048
(-7.09) (-7.45) (-4.59) (-4.80)
2003 -0.087 -0.184 -0.106 -0.105
(-9.98) (-10.14) (-10.35) (-12.09)
2008 -0.057 -0.124 -0.080 -0.081
(-6.55) (-6.82) (-7.92) (-8.72)
Average distance to nearest health facility 0.010 0.022 -0.003 -0.004
(9.53) (8.77) (-2.78) (-2.73)
Health facilities per 100,000 -0.057 -0.124 0.032 0.033
(-8.26) (-7.91) (4.23) (3.77)
Health facilities per 100,000 squared 0.002 0.004 -0.001 -0.001
(10.35) (9.69) (-4.92) (-4.45)
ψ 2.324 1.166
ρ 0.70 0.54
Likelihood Ratio Test for ρ = 0
χ21(P − value) 1221.84 (0.00) 296.78 (0.00)
Number of observations 18974 18974 15713 15713
Prenatal Care and Infant Health: A Multilevel Analysis 53
or multiple births, or are members of wealthy households are more likely to report the
infant’s birth weight, holding other factors constant. Further, first born children are
more likely to have their birth weights reported, holding other factors constant.
Columns (3) and (4) present the prenatal care use model. From the two columns, we
can observe that, holding other factors constant, the longer the average distance to the
nearest health facility, the lower the probability of seeking adequate prenatal care. This
is in line with our expectations. As expected, also, more health facilities per 100,000 of
population increase the probability of seeking adequate prenatal care, if other factors are
held constant.
We can also observe from Columns (3) and (4) in the table that the older the mother
at the time of birth of the child, the higher the probability of seeking adequate prenatal
care, holding other factors constant. This result is consistent with the findings from
the literature (see, for example, McDonald and Coburn, 1988; Ribeiro et al., 2009).
Consistent with the literature (see, for example, Raghupathy, 1996; Celik and Hotchkiss,
2000), we find that holding other factors constant, highly educated mothers are more
likely to seek adequate prenatal care.
The results also indicate that, holding other factors constant, the probability of using
adequate prenatal care is higher if the mother wanted the pregnancy as compared to a
case where the pregnancy is unwanted by the mother. This finding is consistent with the
finding in the literature (see, for example, Delgado–Rodriguez, Gomez–Olmedo, Bueno–
Cavanillas and Galvez–Vargas, 1997; Eggleston, 2000; Magadi, Madise and Rodrigues,
2000).
Our results also indicate that urban residence is associated with a higher probability
of seeking adequate prenatal care as compared to rural residence. The results in the
literature are mixed. Some studies (see, for example, Celik and Hotchkiss, 2000) do not
find such a relationship while others (see, for example, Baldwin et al., 2002) support our
finding.
In the table, the likelihood ratio test for ρ = 0 is a test of the null hypothesis that
the variance of the random intercept is zero. From Table 2.14 on page 52, we can see
that this hypothesis is rejected in both the sample selection and prenatal care models.
We, therefore, conclude that the multilevel versions of our models are appropriate.
A closer look at the results shows that, in most cases (but definately in not all cases),
the average marginal effects are higher for the multilevel model than for the single–level
Prenatal Care and Infant Health: A Multilevel Analysis 54
model, in absolute terms. This implies that, in general, the single–level model understates
the effect of the respective covariate on the variable of interest.
Infant Health Models
Table 2.15 on page 56 shows the results for the single–level low birth weight models.
The columns showing the results have been labelled (1), (2), (3) and (4). Column (1) of
Table 2.15 shows the basic model; column (2) shows the model controlling for sample
selection bias; column (3) shows the model controlling for both sample selection bias
and endogeneity of prenatal care use; while column (4) shows the model controlling for
sample selection bias, endogeneity of prenatal care use and unobserved heterogeneity.
Since the selection residual is statistically significant at the 5% level of significance
in column (2), we conclude that the model in column (1) does suffer from selection
bias. Looking at the model in column (3), we can conclude that prenatal care is an
endogenous determinant of infant health since the coefficient of the prenatal care residual
is statistically significant at the 5% level of significance. Notice also that the average
marginal effect of the adequate prenatal care variable is now negative unlike the case
of the models in columns (1) and (2). The model in column (4) shows that there is no
unobserved heterogenity in our model since the coefficient of the interaction of prenatal
care with its residual is not statistically significant. In the single–level case, therefore,
the model in column (3) is the most appropriate. Comparing the model in column (3) to
the basic model in column (1), we note the following observations which point to the
importance of controlling for the endogeneity of prenatal care use in models of infant
health:
i. In column (1) of Table 2.15, the average marginal effect associated with adequate
prenatal care use is positive, implying that adequate use of prenatal care is likely to
worsen infant health. The result in column (3), however, indicates that adequate use
of prenatal care improves infant health. This means that if we had not controlled
for the endogeneity of prenatal care use, we would have concluded that adequate
use of prenatal care worsens infant health. Controlling for endogeneity, therefore,
helps us arrive at the conclusion that adequate use of prenatal care improves infant
health.
ii. In column (1) of Table 2.15, the average marginal effect of adequate use of prenatal
care is statistically insignificant but in column (3) it is statistically significant at
Prenatal Care and Infant Health: A Multilevel Analysis 55
the 5% level of significance. Thus, failure to control for endogeneity of prenatal
care use makes it (prenatal care) appear to have no effect on infant health. Once
we control for this endogeneity, however, we find that adequate use of prenatal care
improves infant health.
Table 2.16 on page 57 shows the results for the multi–level birth weight models. The
column of results are also labelled as (1), (2), (3) and (4). Column (1) of the table shows
the basic model; column (2) shows the model that controls for sample selection bias;
column (3) shows the model that controls for sample selection bias and endogeneity of
prenatal care use; while column (4) shows the model that controls for sample selection
bias, endogeneity of prenatal care use and unobserved heterogeneity. The multi–level
results shown in column (2) of Table 2.16 imply that there is no sample selection bias since
the coefficient of the selection residual is not statistically significant. From column (3) of
Table 2.16, we can conclude that prenatal care use is endogenous since the coefficient
for prenatal care residual is statistically significant. Based on the results in column (4),
we can conclude that there is no unobserved heterogeneity because the coefficient of the
interaction of prenatal care with its residual is not statistically significant. The best
multi–level model, therefore, appears to be the model in column (3) since the model in
column (4) does not improve the results obtained in column (3).
We conclude from column (3) of Table 2.16 that, just as was the case in the single–level
model, prenatal care use is endogenous even in the multilevel model. We can, therefore,
conclude that prenatal care use is an endogenous determinant of infant health measured
by birth weight. The result that prenatal care use is endogenous is consistent with the
findings in the literature (see, for example, Rous, Jewell and Brown, 2004; Conway and
Kutinova, 2006). Based on a Likelihood Ratio test for instrument validity for the model
in column (3) of Table 2.16, not explicitly shown in the table, we conclude that the
instruments are valid (χ21 = 0.07, p− value = 0.798).
We show in Table 2.17 on page 58 the single–level results based on the model in
column (3) of Table 2.15 on page 56 and the multi–level results based on the model in
column (3) of Table 2.16 on page 57, for comparison purposes. We focus our attention
on the results from the multilevel model, since we have already demonstrated that the
multilevel model is an appropriate model for studying the effect of prenatal care on low
birth weight.
Prenatal Care and Infant Health: A Multilevel Analysis 56
Table 2.15: Average Marginal Effects from Single–Level Low Birth Weight Models, Robust ZStatistics in Parentheses
Variable Low Birth Weight = 1
(1) (2) (3) (4)
Adequate prenatal care 0.003 0.004 -0.260 -0.298
(0.44) (0.66) (-2.40) (-2.33)
Mother’s age at birth of child -0.008 -0.008 -0.006 -0.005
(-2.16) (-2.30) (-1.44) (-1.24)
Square of mother’s age at birth of child 0.00016 0.00016 0.00014 0.0001
(2.38) (2.39) (2.05) (1.94)
Mother’s education -0.003 -0.004 -0.006 0.0001
(-2.88) (-3.41) (-0.28) (0.05)
Parity -0.005 -0.004 0.004 0.006
(-0.61) (-0.51) (0.46) (0.60)
Number of living children -0.002 -0.000001 -0.006 -0.008
(-0.82) (-0.00) (-1.65) (-1.70)
Mother wanted pregnancy -0.003 -0.006 0.013 0.017
(-0.50) (-0.90) (1.29) (1.38)
Urban residence 0.014 0.007 0.031 0.036
(1.80) (0.83) (2.22) (2.15)
Male child -0.012 -0.013 -0.009 -0.009
(-1.98) (-2.11) (-1.58) (-1.48)
Twin or multiple birth child 0.325 0.309 0.363 0.374
(8.08) (7.58) (7.58) (7.17)
First born child 0.035 0.029 0.039 0.041
(3.43) (2.81) (3.34) (3.31)
Asset index -0.003 -0.007 0.002 0.004
(-1.67) (-2.51) (0.45) (0.68)
1998 -0.005 -0.0009 -0.019 -0.023
(-0.65) (-0.11) (-1.68) (-1.73)
2003 -0.017 -0.012 -0.049 -0.057
(-1.99) (-1.34) (-2.79) (-2.55)
2008 -0.031 -0.028 -0.055 -0.061
(-3.60) (-3.11) (-3.68) (-3.37)
Selection residual -0.038 0.013 0.022
(-1.98) (0.45) (0.67)
Prenatal care residual 0.157 0.199
(2.44) (1.99)
Interaction of prenatal care with residual -0.027
(-0.55)
LR Test for Instrument Validity
χ21(P − value) 0.15 (0.696)
Number of observations 7331 7331 7331 7331
Note: An explanatory variable indicating whether or not birth weight was reported was included in
models controlling for selection bias but it was automatically dropped by the software during estimation.
Prenatal Care and Infant Health: A Multilevel Analysis 57
Table 2.16: Average Marginal Effects from Multi–Level Low Birth Weight Models, Robust ZStatistics in Parentheses (Number of Observations = 7331)
Variable Low Birth Weight = 1
(1) (2) (3) (4)
Adequate prenatal care 0.002 0.002 -0.036 -0.035
(0.56) (0.55) (-2.22) (-2.14)
Mother’s age at birth of child -0.005 -0.005 -0.005 -0.005
(-2.14) (-2.12) (-1.97) (-1.96)
Square of mother’s age at birth of child 0.0001 0.0001 0.0001 0.0001
(2.30) (2.30) (2.22) (2.21)
Mother’s education -0.002 -0.002 -0.001 -0.001
(-2.68) (-2.40) (-1.97) (-1.96)
Parity -0.003 -0.003 -0.002 -0.002
(-0.53) (-0.63) (-0.32) (-0.32)
Number of living children -0.001 -0.001 -0.002 -0.002
(-0.70) (-0.68) (-1.01) (-1.02)
Mother wanted pregnancy -0.002 -0.002 0.0002 0.0002
(-0.65) (-0.42) (0.05) (0.06)
Urban residence 0.009 0.010 0.012 0.012
(1.92) (2.02) (2.36) (2.36)
Male child -0.007 -0.007 -0.007 -0.007
(-1.87) (-1.86) (-1.78) (-1.78)
Twin or multiple birth child 0.100 0.102 0.104 0.104
(6.51) (6.66) (6.57) (6.58)
First born child 0.019 0.020 0.020 0.020
(3.19) (3.26) (3.29) (3.29)
Asset index -0.002 -0.002 -0.001 -0.001
(-1.52) (-1.21) (-0.72) (-0.71)
1998 -0.003 -0.003 -0.006 -0.006
(-0.58) (-0.65) (-1.04) (-1.05)
2003 -0.010 -0.011 -0.016 -0.016
(-1.86) (-1.92) (-2.58) (-2.58)
2008 -0.019 -0.019 -0.023 -0.023
(-3.18) (-3.23) (-3.63) (-3.62)
Selection residual 0.007 0.004 0.005
(0.74) (0.50) (0.51)
Prenatal care residual 0.039 0.041
(2.42) (1.78)
Interaction of prenatal care with residual -0.004
(-0.14)
ψ 0.835 0.818 0.832 0.831
ρ 0.46 0.45 0.45 0.45
LR Test for ρ = 0: χ21(P − value) 18.09 (0.00) 17.80 (0.00) 18.01 (0.00) 17.94 (0.00)
Note: An explanatory variable indicating whether or not birth weight was reported was included in
models controlling for selection bias but it was automatically dropped by the software during estimation.
Prenatal Care and Infant Health: A Multilevel Analysis 58
Table 2.17: Average Marginal Effects from Our Chosen Low Birth Weight Models, Robust ZStatistics in Parentheses
Variable Low Birth Weight = 1
Single Level Model Multi–Level Model
Adequate prenatal care -0.260 -0.036
(-2.40) (-2.22)
Mother’s age at birth of child -0.006 -0.005
(-1.44) (-1.97)
Square of mother’s age at birth of child 0.0001 0.0001
(2.05) (2.22)
Mother’s education -0.0006 -0.001
(-0.28) (-1.97)
Parity 0.004 -0.002
(0.46) (-0.32)
Number of living children -0.006 -0.002
(-1.65) (-1.01)
Mother wanted pregnancy 0.013 0.0002
(1.29) (0.05)
Urban residence 0.031 0.011
(2.22) (2.36)
Male child -0.009 -0.007
(-1.58) (-1.78)
Twin or multiple birth child 0.363 0.104
(7.58) (6.57)
First born child 0.039 0.020
(3.34) (3.29)
Asset index 0.002 -0.001
(0.45) (-0.72)
1998 -0.019 -0.006
(-1.68) (-1.04)
2003 -0.049 -0.016
(-2.79) (-2.58)
2008 -0.055 -0.023
(-3.68) (-3.63)
Selection residual 0.013 0.004
(0.45) (0.50)
Prenatal care residual 0.157 0.039
(2.44) (2.42)
ψ 0.832
ρ 0.45
LR Test for ρ = 0: χ21(P − value) 18.01 (0.00)
LR Test for Instrument Validity: χ21(P − value) 0.15 (0.696) 0.07 (0.798)
Number of observations 7331 7331
Prenatal Care and Infant Health: A Multilevel Analysis 59
The results show that adequate use of prenatal care reduces the probability of having
a low birth weight infant, holding other factors constant. This actually implies that
prenatal care use is effective in preventing low birth weight. This finding is consistent
with the findings from the studies in the literature that find prenatal care use to be an
effective determinant of birth weight (see, for example, Conway and Deb, 2005; Jewell
and Triunfo, 2006).
A comparison of the results from the multi–level model and the single–level model
shows that the single–level model overstates the effect of adequate use of prenatal care
on reducing the probability of low birth weight. In the single–level model, holding other
factors constant, adequate use of prenatal care reduces the probability of delivering a low
birth weight infant by 0.26. In the multilevel model, however, the corresponding reduction
is only 0.036. Consequently, when we do not control for unobserved mother–specific
characteristics, we are likely to end up overstating the effect of adequate prenatal care
use in reducing the probability of giving birth to a low birth weight infant.
The results from the multilevel model also show that older mothers are less likely to
give birth to low birth weight infants, holding other factors constant. This is consistent
with the finding in the literature of a positive association between mother’s age at birth
of the child and the infant’s birth weight (see, for example, Fraser, Brockert and Ward,
1995; Roth, Hendrickson, Schilling and Stowell, 1998; Borja and Adair, 2003; Mwabu,
2009). Consistent with the findings in the literature (see, for example, Reichman and
Pagnini, 1997; Delpisheh, Brabin, Attia and Brabin, 2008; Swamy et al., 2012), there
appears to be a U−shaped relationship between maternal age at birth of the child and
the probability of delivering a low birth weight infant, holding other factors constant.
Our results also indicate that maternal education reduces the probability of low birth
weight, holding other factors constant. This result is consistent with some of the findings
in the literature (see, for example, Chevalier and O’Sullivan, 2007; Siza, 2008; Muula,
Siziya and Rudatsikira, 2011). There are, however, studies in the literature that fail to
find a significant relationship between maternal education and the incidence of low birth
weight (see, for example, Lawoyin and Oyediran, 1992).
The results further indicate that urban residence is associated with an increased
probability of delivering a low birth weight infant, holding other factors constant. This
result is not surprising since there is evidence in the literature that living in urban areas
does not necessarily guarantee better birth and infant health outcomes (see, for example,
Hillemeier, Weisman, Chase and Dyer, 2007; Simonet et al, 2010).
Prenatal Care and Infant Health: A Multilevel Analysis 60
According to the results, male infants have a lower probability of having low birth
weight as compared to their female counterparts, holding other factors constant. This is
consistent with the findings in the literature (see, for example, Mondal, 1998; Mwabu,
2009).
Infants from twin or multiple births have a higher probability of having low birth
weight as compared to singletons, holding other factors constant. This finding is consistent
with the findings in the literature (see, for example, Adamson, 2007).
Consistent with the findings in the literature (see, for example, Magadi, Madise and
Diamond, 2001; Mwabu, 2009), the results also indicate that first born infants have a
higher probability of having low birth weight than non–firstborns.
The results of the Guevara test show that our instruments are valid.
2.7 Summary, Conclusions and Policy Implications
Summary
In this chapter we investigate the effect of prenatal care on infant health in Kenya using
a multilevel modelling framework. Although there are several indicators of infant health,
we use birth weight because it has been demonstrated to be a good summary measure of
infant health at birth (Mwabu, 2009). There is still controversy in the literature regarding
the effectiveness of prenatal care in preventing adverse birth outcomes and there are also
few studies done in Sub–Saharan Africa that investigate the effect of prenatal care on
birth weight. Although there have been previous attempts to study the effectiveness of
particular types of prenatal care on improving birth outcomes in Kenya (see, for example,
Mwabu, 2009), or antenatal attendance in Kenya (see, for example, Magadi, Diamond,
Madise and Smith, 2004; Brown et al., 2008), no systematic attempt has so far been
made to link adequacy of prenatal care use to low birth weight in Kenya. Previous
studies also mainly used the single–level modelling framework. This chapter adds to the
literature by studying the effectiveness of adequacy of prenatal care use in lowering the
probability of giving birth to a low birth weight infant. It also extends the literature by
using a multilevel modelling framework (see, for example, Steenbergen and Jones, 2002).
Based on WHO definition (see, for example, Zegers–Hochschild et al., 2009; WHO,
2011), we classify an infant’s birth weight as either low or not low. Following WHO
Prenatal Care and Infant Health: A Multilevel Analysis 61
recommendations (Berg, 1995), we construct an index for the adequacy of prenatal care
that classifies prenatal care use as either adequate or inadequate. We then estimate a
model of the effect of the adequacy of prenatal care use on the probability of having a
low birth weight infant, controlling for other factors. The estimation strategy adopted
controls for potential sample selection bias by using an approach suggested by Vella (1992),
potential endogeneity of prenatal care by using the Two–Stage–Residual–Inclusion method
(Terza, Basu and Rathouz, 2008), and potential unobserved heterogeneity by using the
control function approach (Florens, Heckman, Meghir and Vytlacil, 2008; Mwabu, 2009;
Wooldridge, 2010). In the estimation of our models, provincial level average distances to
the nearest health facility and health facilities per 100,000 of population in each province
are used as instruments. The main data set used to estimate our models are the Kenya
Demographic and Household Survey (KDHS) data sets collected in 1993, 1998, 2003
and 2008. The average distances to the nearest health facilities are obtained from the
community data set of the 2005 Kenya Integrated Household and Budget Survey. Health
facilities per 100,000 of population are obtained from the Ministry of Health.
Our results indicate that prenatal care is an endogenous determinant of birth outcomes
but sample selection bias is not an issue in our models. When we do not control for
the endogeneity of prenatal care, adequate use of prenatal care appears to be ineffective
in lowering the probability of low birth weight. Our main finding, after controlling
for the endogeneity of prenatal care, is that adequate use of prenatal care reduces the
probability of delivering a low birth weight infant, holding other factors constant. This
result is consistent with the findings of some of the studies in the literature (see, for
example, Conway and Deb, 2005; Jewell and Triunfo, 2006). The results further indicate
that models that do not control for unobserved mother–specific characteristics overstate
the effect of adequate prenatal care use by the mother in reducing the probability of
delivering an infant with low birth weight. Regarding the determinants of adequate use of
prenatal care, we find that it is significantly influenced by accessibility of health facilities,
availability of health facilities, mother’s age at birth of child, mother’s education, parity,
number of living children born to mother, whether or not mother wanted pregnancy, and
whether or not mother resides in urban areas.
Prenatal Care and Infant Health: A Multilevel Analysis 62
Conclusions
Based on the findings from this chapter, we can draw four main conclusions. First,
adequate use of prenatal care improves birth outcomes, in particular birth weight.
Prenatal care is, however, considered to be adequate if it is obtained from a skilled
provider (either a doctor or a nurse), the mother makes at least four visits to the provider,
and the first visit is made within the first 16 weeks of pregnancy (see, Berg, 1995). The
implication here is that before we judge prenatal care to be ineffective in influencing birth
outcomes, we have to confirm whether it was adequate or not. Second, prenatal care is
an endogenous determinant of birth weight. Prenatal care appears to be ineffective in
influencing birth outcomes in models that do not control for this endogeneity. Attempts
to investigate the effectiveness of prenatal care in influencing birth outcomes must, control
for its endogeneity if they are to arrive at credible conclusions. One way of controlling for
this endogeneity is by using the Two–Stage–Residual–Inclusion method (for details about
this method, see Terza, Basu and Rathouz, 2008). Third, the effect of prenatal care on
birth weight is overstated in models that do not control for unobserved mother–specific
effects. To obtain the accurate effect of prenatal care on birth outcomes, therefore, one
has to control for such effects. One way of doing that is by estimating a multilevel rather
than a single–level model (for details on multilevel modelling, see Steenbergen and Jones,
2002). Fourth, adequate use of prenatal care is significantly influenced by community
characteristics (such as accessibility and availability of health facilities), socio–economic
characteristics of the mother (such as her age at birth of the child, her level of education,
her parity, whether or not she resides in urban areas), household characteristics (such as
wealth status), and maternal psychological factors (such as whether the pregnancy was
wanted or not).
Policy Implications
Since our study shows that adequate use of prenatal care improves infant health, the
implication is that policies for promoting adequate use of prenatal care should be pursued.
Some of these policies include:
i. Encouraging expectant mothers to only seek prenatal care from skilled providers
(that is, doctors and nurses).
ii. Encouraging the expectant mothers to make at least four visits to their prenatal
care provider.
Prenatal Care and Infant Health: A Multilevel Analysis 63
iii. Educating the expectant mothers on the importance of making their first prenatal
care visit within the first 16 weeks of pregnancy.
iv. Improving accessibility and availability of prenatal care services by, for example,
ensuring health facilities are not very far from where expectant women live and
there are enough nurses and doctors to provide the care when it is sought.
v. Discouraging early pregnancies among women.
vi. Investing in girl/maternal education.
vii. Discouraging unwanted pregnancies among women by, for instance, encouraging use
of contraceptives.
viii. Encouraging family planning to reduce the number of children that mothers have
to take care of.
REFERENCES 64
References
[1] Adamson, H., 2007. Low birth weight in relation to maternal age and multiplepregnancies at Muhimbili national hospital. Dar Es Salaam Medical Students’Journal 14(2), pp.55–58.
[2] Agenor, P. and Canuto, O., 2012. Access to infrastructure and women’s timeallocation: evidence and a framework for policy analysis. [pdf] Working Paper/P45. Available at:http://www.ferdi.fr/uploads/sfCmsContent/html/112/P45.pdf [Accessed5 August 2012].
[3] Alderman H. and Behrman, J. R., 2006. Reducing the incidence of low birthweight in low-income countries has substantial economic benefits. World BankResearch Observer 21 (1), pp.25-48.
[4] Alexander, G. R. and Korenbrot, C. C., 1995. The role of prenatal care inpreventing low birth weight. Future of Children 5 (1), pp.103-120.
[5] Arcury T. A., Gesler, W. M., Preissner, J. S., Sherman, J., Spencer, J. and Perin,J., 2005. The effects of geography and spatial behavior on health care utilizationamong the residents of a rural region. Health Services Research 40 (1),pp.135–156.
[6] Baldwin, L., Grossman, D. C., Cases, S., Hollow, W., Sugarman, J. R., Freeman,W. L. and Hart, G., 2002. Perinatal and infant health among rural and urbanAmerican Indians/Alaska natives. American Journal of Public Health 92 (9),pp.1491–1497.
[7] Barker, D. J. and Osmond, C., 1986. Infant mortality, childhood nutrition, andischaemic heart disease in England and Wales. Lancet 1 (8489), pp.1077-1081.
[8] Barros, H., Tavares, M. and Rodrigues, T., 1996. Role of prenatal care in pretermbirth and low birthweight in Portugal. Journal of Public Health 18 (3),pp.321–328.
[9] Bartus, T., 2005. Estimation of marginal effects using margeff. Stata Journal 5(3), pp.309–329.
[10] Becker, G. S., 1965. A theory of the allocation of time. Economic Journal 75(299), pp.493–517.
[11] Berg, C. J., 1995. Prenatal care in developing countries: the World HealthOrganization technical working group on antenatal care. Journal of the AmericanMedical Women’s Association (1972) 50 (5), pp.182–186.
[12] Bollen, K. A., Guilkey, D. K. and Mroz, T. A., 1995. Binary outcomes andendogenous explanatory variables: tests and solutions with an application to thedemand for contraceptive use in Tunisia. Demography 32 (1), pp.111-131.
REFERENCES 65
[13] Borja, J. B. and Adair, L. S., 2003. Assessing the net effect of young maternalage on birthweight. American Journal of Human Biology 15 (6), pp.733–740.
[14] Brown, C. A., Sohani, S. B., Khan, K., Lilford, R. and Mukhwana, W., 2008.Antenatal care and perinatal outcomes in Kwale district, Kenya. BMC Pregnancyand Childbirth 8(2).
[15] Brookhart, M. A., Rassen, J. A. and Schneeweiss, S., 2010. Instrumental variablemethods in comparative safety and effectiveness research. Pharmacoepidemiologyand Drug Safety 19 (6), pp.537–554.
[16] Cameron, A. C. and Trivedi, P. K., 2010. Microeconometrics using stata. RevisedEdition. Stata Press.
[17] Celik, Y. and Hotchkiss, D. R., 2000. The socio-economic determinants ofmaternal health care utilization in Turkey. Social Science & Medicine 50 (12),pp.1797-1806.
[18] Chaikind, S. and Corman, H., 1991. The impact of low birth weight on specialeducation costs. Journal of Health Economics 10 (3), pp.291-311.
[19] Chandhiok, N., Dhillon, B. S., Kambo, I. and Saxena, N. C., 2006. Determinantsof antenatal care utilization in rural areas of India: a cross–sectional study from28 districts (An ICMR task force study). Journal of Obstetrics and Gynecology ofIndia 56 (1), pp.47–52.
[20] Chevalier, A. and O’Sullivan, V., 2007. Mother’s education and birth weight. IZADiscussion Paper No. 2640, [pdf] Institute for the Study of Labour (IZA).Available at: http://ftp.iza.org/dp2640.pdf [Accessed 3 January 2012].
[21] Conway, K. S. and Deb, P., 2005. Is prenatal care really ineffective? Or, is the‘devil’ in the distribution? Journal of Health Economics 24 (3), pp.489–513.
[22] Conway, K. S. and Kutinova, A., 2006. Maternal health: does prenatal care makea difference? Health Economics 15 (5), pp.461–488.
[23] Corman, H. and Chaikind, S., 1998. The effect of low birth weight on the schoolperformance and behavior of school-aged children. Economics of EducationReview 17 (3), pp.307-316.
[24] Cornelissen, T. and Sonderhof, K., 2009. Partial effects in probit and logit modelswith a triple dummy–variable interaction term. Stata Journal 9 (4), pp.571–583.
[25] Currie, J. and Grogger, J., 2002. Medicaid expansions and welfare contractions:offsetting effects on prenatal care and infant health? Journal of HealthEconomics 21 (2), pp.313–335.
[26] Darling, R. D. and Atav, A. S., 2012. Risk factors for low birth weight in NewYork state counties. Policy, Politics & Nursing Practice 13 (1), pp.17–26.
REFERENCES 66
[27] Delgado–Rodriguez, M., Gomez–Olmedo, M., Bueno–Cavanillas, A. andGalvez–Vargas, R., 1997. Unplanned pregnancy as a major determinant ininadequate use of prenatal care. Preventive Medicine 26 (6), pp.834–838.
[28] Delpisheh, A., Brabin, L., Attia, E. and Brabin, B. J., 2008. Pregnancy late inlife: a hospital–based study of birth outcomes. Journal of Women’s Health 17 (6),pp.965–970.
[29] Eggleston, E., 2000. Unintended pregnancy and women’s use of prenatal care inEcuador. Social Science & Medicine 51 (7), pp.1011 – 1018.
[30] Evans, W. N. and Lien, D. S., 2005. The benefits of prenatal care: evidence fromthe PAT bus strike. Journal of Econometrics 125 (1 - 2), pp.207-239.
[31] Florens, J. P., Heckman, J. J., Meghir, C. and Vytlacil, E., 2008. Identification oftreatment effects using control functions in models with continuous, endogenoustreatment and heterogeneous effects. Econometrica 76 (5), pp.1191-1206.
[32] Fortney, J. C., Steffick, D. E., Burgess Jr. J. F., Maciejewski, M. L. and Petersen,L., 2005. Are primary care services a substitute or complement for specialty andinpatient services? Health Services Research 40 (5), pp.1422–1442.
[33] Frank, R. G., Strobino, D. M., Salkever, D. S. and Jackson, C. A., 1992. Updatedestimates of the impact of prenatal care on birthweight outcomes by race.Journal of Human Resources 27 (4), pp.629-642.
[34] Fraser, A. M., Brockert, J. E. and Ward, R. H., 1995. Association of youngmaternal age with adverse reproductive outcomes. New England Journal ofMedicine 332 (17), pp.1113–1117.
[35] Garen, J., 1984. The returns to schooling: a selectivity bias approach with acontinuous choice variable. Econometrica 52 (5), pp.1199–1218.
[36] Garrido, G. G., 2009. The impact of adequate prenatal care in a developingcountry: testing the WHO recommendations. On–line working paper series, [pdf]California Centre for Population Research. Available at:http://escholarship.org/uc/item/1c86q94h [Accessed 30 March 2012].
[37] Goldani, M. Z., Barbieri, M. A., Silva, A. A. M and Bettiol, H., 2004. Trends inprenatal care use and low birthweight in Southeast Brazil. American Journal ofPublic Health 94 (8), pp.1366-1371.
[38] Goldenberg, R. L., Patterson, E. T. and Freese, M. P., 1992. Maternaldemographic, situational and psychosocial factors and their relationship toenrolment in prenatal care: a review of the literature. Women & Health 19 (2),pp.133-151.
[39] Gourieroux, C., Monfort, A., Renault, E. and Trognon, A., 1987. Generalizedresiduals. Journal of Econometrics 34 (1 – 2), pp.5–32.
REFERENCES 67
[40] Greene, W. H., 2008. Econometric analysis. Sixth edition. Upper Saddle River,NJ: Prentice Hall.
[41] Greene, W. H., 2012. Econometric analysis. Seventh edition. Upper Saddle River,NJ: Prentice Hall.
[42] Grossman, M. and Joyce, T., 1990. Unobservables, pregnancy resolution andbirth weight production functions in New York city. Journal of Political Economy98 (5), pp.983–1007.
[43] Guevara, C. A. and Ben–Akiva, M., 2008. A lagrange multiplier test for thevalidity of instruments in MNL models: an application to residential choice. [pdf]European Transport Conference 2008, Leeuwenhorst, The Netherlands. Availableat: http://www.etcproceedings.org/paper/a-lagrange-multiplier-test-for-the-validity-of-instruments-in-mnl-models-an-ap [Accessed 3January 2012].
[44] Guevara–Cue, C. A., 2010. Endogeneity and sampling of alternatives in spatialchoice models. [pdf] Unpublished PhD Thesis, Massachusetts Institute ofTechnology. Available at: http://dspace.mit.edu/bitstream/handle/1721.1/62098/707634193.pdf?sequence=1 [Accessed 5 March 2012].
[45] Guyatt, H. L. and Snow, R. W., 2004. Impact of malaria during pregnancy onlow birth weight in Sub–Saharan Africa. Clinical Microbiology Reviews 17 (4),pp.760–769.
[46] Hack, M., Klein, N. K. and Taylor, G., 1995. Long–term developmental outcomesof low birth weight infants. Future of Children 5 (1), pp.176–196.
[47] Heckman, J., 1979. Sample selection as a specification error. Econometrica 47 (1),pp.153–161.
[48] Hillemeier, M. M., Weisman, C. S., Chase, G. A. and Dyer, A., 2007. Individualand community predictors of preterm birth and low birthweight along therural–urban continuum in central Pennsylvania. Journal of Rural Health 23 (1),pp.42–48.
[49] Institute of Medicine (IOM), 1985. Preventing low birth weight: summary. [pdf]National Academy Press. Available at:http://www.nap.edu/catalog/512.html [Accessed 6 August 2008].
[50] Jewell, R. T., 2009. Demand for prenatal health care in South America. AppliedEconomics 41 (4), pp.469-479.
[51] Jewell, R. T. and Triunfo, P., 2006. The impact of prenatal care on birthweight:the case of Uruguay. Health Economics 15 (11), pp.1245–1250.
[52] Joyce, T., 1994. Self-selection, prenatal care, and birthweight among blacks,whites, and hispanics in New York city. Journal of Human Resources 29 (3),
REFERENCES 68
pp.762-794.
[53] Joyce, T., 1999. Impact of augmented prenatal care on birth outcomes ofmedicaid recipients in New York city. Journal of Health Economics 18 (3),pp.31-67.
[54] Kallan, J. E., 1993. Race, intervening variables, and two components of low birthweight. Demography 30 (3), pp.489-506.
[55] Koroukian, S. M. and Rimm, A. A., 2002. The ‘adequacy of prenatal careutilization’ (APNCU) index to study low birth weight: is the index biased?Journal of Clinical Epidemiology 55 (3), pp.296-305.
[56] Kosimbei, G., 2005. Child healthcare seeking behaviour in Kenya. KIPPRADiscussion Paper No. 50. Nairobi: Kenya Institute for Public Policy Researchand Analysis (KIPPRA).
[57] Kotelchuck, M., 1994a. An evaluation of the Kessner adequacy of prenatal careindex and a proposed adequacy of prenatal care utilization index. AmericanJournal of Public Health 84 (9), pp.1414-1420.
[58] Kotelchuck, M., 1994b. The adequacy of prenatal care utilization index: its USdistribution and association with low birthweight. American Journal of PublicHealth 84 (9), pp.1486-1489.
[59] Kramer, M. S., 1987. Determinants of low birth weight: methodologicalassessment and meta-analysis. Buletin of the World Health Organization, [pdf] 65(5), pp.663–737. Available at: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2491072/pdf/bullwho00076-0086.pdf [Accessed 6 August 2008].
[60] LaVeist, T. A., Keith, V. M. and Gutierrez, M. L., 1995. Black/White differencesin prenatal care utilization: an assessment of predisposing and enabling factors.Health Services Research 30 (1), pp.43-58.
[61] Lawoyin, T. O. and Oyediran, A. B., 1992. A prospective study on some factorswhich influence the delivery of low birth weight babies in a developing country.African Journal of Medicine and Medical Sciences 21 (1), pp.33–39.
[62] Lee, L., 1982. Some approaches to the correction of selectivity bias. Review ofEconomic Studies 49 (3), pp.355–372.
[63] Lewit, E. M., Baker, L. S., Corman, H. and Shiono, P. H., 1995. The direct costof low birth weight. Future of Children 5 (1), pp.35–56.
[64] Lin, S., 2004. Assessing the effect of prenatal care on birth outcome in Taiwan.Social Science Journal 41 (3), pp.493–501.
[65] Long, J. S., 1997. Regression models for categorical and limited dependentvariables. Sage Publications.
REFERENCES 69
[66] Long, J. S. and Freese, J., 2006. Regression models for categorical dependentvariables using Stata. Second edition. College Station, TX: Stata Press.
[67] Lopez, A. D., Mathers, C. D., Ezzati, M., Jamison, D. T. and Murray, C. J. L.eds., 2006. Global burden of disease and risk factors. New York: OxfordUniversity Press and World Bank.
[68] Magadi, M., Madise, N. and Diamond, I., 2001. Factors associated withunfavourable birth outcomes in Kenya. Journal of Biosocial Science 33 (2),pp.199–225.
[69] Magadi, M. A., Madise, N. J. and Rodrigues, R. N., 2000. Frequency and timingof antenatal care in Kenya: explaining the variations between women of differentcommunities. Social Science & Medicine 51 (4), pp.551–561.
[70] Magadi, M., Diamond, I., Madise, N. and Smith, P., 2004. Pathways of thedeterminants of unfavourable birth outcomes in Kenya. Journal of BiosocialScience 36 (2), pp.153–176.
[71] McDonald, T. P. and Coburn, A. F., 1988. Predictors of prenatal care utilization.Social Science & Medicine 27 (2), pp.167–172.
[72] McCormick, M. C., 1985. The contribution of low birth weight to infant mortalityand childhood morbidity. New England Journal of Medicine 312 (2), pp.82–90.
[73] Mitgitti, R., Seanchaisuriya, P., Schelp, F. P., Marui, E. and Yanai, H., 2008.Low birth weight infants born to HIV–seropositive mothers andHIV–seronegative mothers in Chiang Rai, Thailand. Southeast Asian Journal ofTropical Medicine and Public Health 39 (2), pp.273–278.
[74] Mondal, B., 1998. Low birth weight in relation to sex of baby, maternal age andparity: a hospital based study on Tangsa tribe from Arunachal Pradesh. Journalof the Indian Medical Association 96 (12), pp.362–364.
[75] Murray, M. P., 2006. Avoiding invalid instruments and coping with weakinstruments. Journal of Economic Perspectives 20 (4), pp.111-132.
[76] Muula, A. S., Siziya, S. and Rudatsikira, E., 2011. Parity and maternal educationare associated with low birth weight in Malawi. African Health Sciences 11 (1),pp.65–71.
[77] Mwabu, G., 2009. The production of child health in Kenya: a structural model ofbirth weight. Journal of African Economies 18 (2), pp.212-260.
[78] Nazim, N. H. and Fan, L., 2011. Does prenatal healthcare improve child birthweight outcomes in Azerbaijan? Results of the national demographic and healthsurvey. Economics and Human Biology 9 (1), pp.56-65.
[79] Obermeryer, C. M. and Potter, J. E., 1991. Maternal health care utilization inJordan: a study of patterns and determinants. Studies in Family Planning 22 (3),
REFERENCES 70
pp.177–187.
[80] Ohlsson, A. and Shah, P., 2008. Determinants and prevention of low birth weight:a synopsis of the evidence. Institute of Health Economics (IHE) Report. [pdf]Institute of Health Economics. Available at: http://www.ihe.ca/documents/IHE%20Report%20LowBirthWeight%20final.pdf
[Accessed 3 January 2012].
[81] Olsen, R. J., 1980. A least squares correction for selectivity bias. Econometrica 48(7), pp.1815–1820.
[82] Paneth, N. S., 1995. The problem of low birth weight. Future of Children 5 (1),pp.19–34.
[83] Pollack, R. N. and Divon, M. Y., 1992. Intrauterine growth retardation:definition, classification, and etiology. Clinical Obstetrics & Gynecology 35 (1),pp.99-107.
[84] Qian, D., Pong, R. W., Yin, A., Nagarajan, K. V. and Meng, Q., 2009.Determinants of health care demand in poor rural China: the case of Gansuprovince. Health Policy and Planning 24 (5), pp.324–334.
[85] Raatikainen, K., Heiskanen, N. and Heinonen, S., 2007. Under-attending freeantenatal care is associated with adverse pregnancy outcomes. BMC PublicHealth 7 (268).
[86] Rabe–Hesketh, S. and Skrondal, A., 2008. Multilevel and longitudinal modelingusing Stata. Second edition. College Station, TX: Stata Press.
[87] Rabe–Hesketh, S., Skrondal, A. and Pickles, A., 2002. Reliable estimation ofgeneralized linear mixed models using adaptive quadrature. Stata Journal 2 (1),pp.1-21.
[88] Rabe–Hesketh, S., Skrondal, A. and Pickles, A., 2004. GLLAMM Manual. [pdf]U. C. Berkeley Division of Biostatistics Working Paper Series. Working Paper160. Available at: http://biostats.bepress.com/cgi/viewcontent.cgi?article=1160&context=ucbbiostat [Accessed 3 January 2012].
[89] Rabe–Hesketh, S., Skrondal, A. and Pickles, A., 2005. Maximum likelihoodestimation of limited and discrete dependent variable models with nested randomeffects. Journal of Econometrics 128 (2), pp.301-323.
[90] Raghupathy, S., 1996. Education and use of maternal health care in Thailand.Social Science & Medicine 43 (4), pp.459–471.
[91] Reichman, N. E. and Pagnini, D. L., 1997. Maternal age and birth outcomes:data from New Jersey. Family Planning Perspectives 29 (6), pp.268–272 andp.295.
[92] Reyes, L. and Mannalich, R., 2005. Long-term consequences of low birth weight.
REFERENCES 71
Kidney International 68 (Supplement 97), pp.S107-S111.
[93] Ribeiro, E. R., Guimaraes, A. M. D. N., Bettiol, H., Lima, D. D. F., Almeida, M.L. D., De Souza, L., Silva, A. A. M. and Gurgel, R. Q., 2009. Risk factors forinadequate prenatal care use in the metropolitan area of Aracaju, NortheastBrazil. BMC Pregnancy and Childbirth 9 (31).
[94] Rosenzweig, M. R. and Schultz, T. P., 1982. The behaviour of mothers as inputsto child health: the determinants of birth weight, gestation, and rate of fetalgrowth. In: V. R. Fuchs, ed. Economic aspects of health. [pdf] Chicago:University of Chicago Press, pp.53–92. Available at:http://www.nber.org/chapters/c6545.pdf [Accessed 6 August 2008].
[95] Rosenzweig, M. R. and Schultz, T. P., 1983. Estimating a household productionfunction: heterogeneity, the demand for health inputs, and their effects on birthweight. Journal of Political Economy 91 (5), pp.723-746.
[96] Roth, J., Hendrickson, J., Schilling, M. and Stowell, D. W., 1998. The risk of teenmothers having low birth weight babies: implications of recent medical researchfor school health personnel. Journal of School Health 68 (7), pp.271–275.
[97] Rous, J. F., Jewell, R. T. and Brown, R. W., 2004. The effect of prenatal care onbirth weight: a full-information maximum likelihood approach. Health Economics13 (3), pp.251-264.
[98] Rutstein, S. O. and Rojas, G., 2006. Guide to DHS statistics. [pdf] ORC Macro.Available at: http://www.ophi.org.uk/wp-content/uploads/OPHI-HDCA-SS2012-Guide_DHS_Statistics.pdf [Accessed 3 January 2009].
[99] Schultz, T. P., 1984. Studying the impact of household economic and communityvariables on child mortality. Population and Development Review 10(Supplement), pp.215-235.
[100] Simiyu, D. E., 2004. Morbidity and mortality of low birth weight infants in thenew born unit of Kenyatta national hospital, Nairobi. East African MedicalJournal 81 (7), pp.367–374.
[101] Simonet, F., Wilkins, R., Heaman, M., Smylie, J., Martens, P., Mchugh, N. G. L.,Labranche, E., Wassimi, S., Fraser, W. D. and Luo, Z., 2010. Urban living is notassociated with better birth and infant outcomes among Inuit and First Nationsin Quebec. Open Women’s Health Journal 4, pp.25–31.
[102] Siza, J. E., 2008. Risk factors associated with low birth weight of neonates amongpregnant women attending referral hospital in northern Tanzania. TanzaniaJournal of Health Research 10 (1), pp.1–8.
[103] Snijders, T. A. B. and Bosker, R. J., 1999. Multilevel analysis: an introduction tobasic and advanced multilevel modelling. Sage Publications.
REFERENCES 72
[104] StataCorp., 2009. Stata Statistical Software, Version 11, Special Edition. CollegeStation, TX: Stata Corporation.
[105] Steenbergen, M. R. and Jones, B. S., 2002. Modeling multilevel data structures.American Journal of Political Science 46 (1), pp.218-237.
[106] Swamy, G. K., Edwards, S., Gelfand, A., James, S. A. and Miranda, M. L., 2012.Maternal age, birth order, and race: differential effects on birthweightt. Journalof Epidemiology & Community Health 66 (2), pp.136–142.
[107] Terza, J. V., Basu, A. and Rathouz, P. J., 2008. Two-stage residual inclusionestimation: addressing endogeneity in health econometric modelling. Journal ofHealth Economics 27 (3), pp.531-543.
[108] United Nations Children’s Fund (UNICEF) and World Health Organization(WHO), 2004. Low birthweight: country, regional and global estimates. [pdf] NewYork: UNICEF. Available at:http://www.childinfo.org/files/low_birthweight_from_EY.pdf [Accessed7 August 2008].
[109] Vella, F., 1992. Simple tests for sample selection bias in censored and discretechoice models. Journal of Applied Econometrics 7 (4), pp.413–421.
[110] Vella, F., 1993. A simple estimator for models with censored endogenousregressors. International Economic Review 34 (2), pp.441–457.
[111] Vella, F., 1998. Estimating models with sample selection bias: a survey. Journalof Human Resources 33 (1), pp.127–169.
[112] Warner, G. L., 1995. Prenatal care demand and birthweight production of blackmothers. American Economic Review 85 (2), pp.132-137.
[113] Wehby, G. L., Murray, J. C., Castilla, E. E., Lopez–Camelo, J. S. and Ohsfeldt,R. L., 2009a. Quantile effects of prenatal care utilization on birth weight inArgentina. Health Economics 18 (11), pp.1307-1321.
[114] Wehby, G. L., Murray, J. C., Castilla, E. E., Lopez–Camelo, J. S. and Ohsfeldt,R. L., 2009b. Prenatal care demand and its effects on birth outcomes by birthdefect status in Argentina. Economics and Human Biology 7 (1), pp.84-95.
[115] Wehby, G. L., Murray, J. C., Castilla, E. E., Lopez–Camelo, J. S. and Ohsfeldt,R. L., 2009c. Prenatal care effectiveness and utilization in Brazil. Health Policyand Planning 24 (3), pp.175–188.
[116] Were F. N. and Bwibo, N. O., 2009. The contribution of very low birth weightdeaths to infant mortality. East African Medical Journal 86 (8), pp.374–377.
[117] Wooldridge, J. M., 2002. Econometric analysis of cross section and panel data.MIT Press.
REFERENCES 73
[118] Wooldridge, J. M., 2010. Econometric analysis of cross section and panel data.Second edition. MIT Press.
[119] World Health Organization (WHO), 2006. World health statistics 2006. [pdf]Geneva: World Health Organization. Available at:http://www.who.int/whosis/whostat2006.pdf [Accessed 6 June 2012].
[120] World Health Organization (WHO), 2011. International statistical classificationof diseases and related health problems, 10th Revision (ICD–10): volume 2instruction manual. 2010 edition. [pdf] World Health Organization. Available at:http://www.who.int/classifications/icd/ICD10Volume2_en_2010.pdf
[Accessed 12 October 2012].
[121] World Health Organization (WHO), 2012. World health statistics 2012. [pdf]Geneva: World Health Organization. Available at: http://www.who.int/gho/publications/world_health_statistics/EN_WHS2012_Full.pdf [Accessed 6June 2012].
[122] Zegers–Hochschild, F., Adamson, G. D., de Mouzon, J., Ishihara, O., Mansour,R., Nygren, K., Sullivan, E. and Vanderpoel, S., 2009. International committeefor monitoring assisted reproductive technology (ICMART) and the WorldHealth Organization revised glossary of ART terminology. Fertility and Sterility92 (5), pp.1520–1524.
[123] Zohoori, N. and Savitz, D. A., 1997. Econometric approaches to epidemiologicdata: relating endogeneity and unobserved heterogeneity to confounding. Annalsof Epidemiology 7 (4), pp.251–257.
Chapter 3
Preceding Birth Interval Length and
Maternal Health
3.1 Introduction
As explained in Chapter 1, maternal health includes the health of women when they
are pregnant, during delivery and the period immediately following delivery (Lule et
al., 2005). As also pointed out in Chapter 1, among the maternal health indicators are
pregnancy outcomes (see, for example, Kramer, 2003; Villamor and Cnattingius, 2006).
Adverse pregnancy outcomes (such as stillbirths1, miscarriages2, and preterm births3)
may be indicators of underlying maternal health problems (see, for example, Villamor
and Cnattingius, 2006). We can, therefore, investigate maternal health indirectly by
investigating whether or not a woman experiences an adverse pregnancy outcome.
In this chapter, we measure maternal health using information on whether or not
a woman reports to have experienced an adverse pregnancy outcome, in particular a
miscarriage or a stillbirth.4
1A stillbirth, also referred to as a fetal death, is death of a foetus before birth at or after a gestationalage of 20 completed weeks (Zegers–Hochschild et al., 2009).
2A miscarriage, also known as a spontaneous abortion, is the loss of a foetus of less than 400g or theloss of a clinical pregnancy before a gestational age of 20 completed weeks (Zegers–Hochschild et al.,2009).
3A preterm birth refers to a birth after a gestational period of more than 20 but less than 37 completedweeks (Zegers–Hochschild et al., 2009).
4The survey question, however, asked the women whether they had ever had a pregnancy that miscarried,was aborted, or ended in a stillbirth. There could, therefore, be women in our dataset reporting anabortion which, if not done due to maternal health conditions, is strictly speaking not an adversepregnancy outcome. More information on Demographic and Health Survey (DHS) questionnaires canbe found at http://www.measuredhs.com/What-We-Do/Survey-Types/DHS-Questionnaires.cfm
74
Preceding Birth Interval Length and Maternal Health 75
Table 3.1: Stillbirth rate for Selected Countriesand WHO Regions, 2009
Country/Region Stillbirth rate (per 1,000 total births)
Select Countries
Kenya 22
Namibia 15
Mauritius 9
United Kingdom 4
United States of America 3
WHO Regions
Africa 104
Americas 34
South – East Asia 78
Europe 27
Eastern Mediterranean 74
Western Pacific 37
Source: WHO (2012).
Table 3.1 shows the stillbirth rate for selected countries and WHO regions as of 2009.
From the table, we can observe that the stillbirth rate for Kenya for 2009 was 22 per
1,000 total births. This is higher than that of the other countries shown in the table.
Kenya’s rate is, however, better than the averages for the various WHO regions.
There are several risk factors for adverse pregnancy outcomes. These include short
interpregnancy intervals (see, for example, Lule et al., 2005; Conde–Agudelo, Rosas–
Bermudez and Kafury–Goeta, 2006), long interpregnancy intervals (see, for example,
Conde–Agudelo, Rosas–Bermudez and Kafury–Goeta, 2006), maternal obesity (see, for
example, Cnattingius, Bergstrom, Lipworth and Kramer, 1998; Rosenberg, Garbers,
Lipkind and Chiasson, 2005), maternal morbidity (see, for example, Scannapieco, Bush
and Paju, 2003; Rosenberg, Garbers, Lipkind and Chiasson, 2005) and smoking (see, for
example, Walsh, 1994). In this chapter, we mainly focus on the interpregnancy interval
or generally on birth intervals.
Although there are several different measures of birth spacing (see, for example,
Davanzo et al., 2004 and WHO, 2006 for a detailed discussion), following DaVanzo et
Preceding Birth Interval Length and Maternal Health 76
Table 3.2: Distribution of Preceding Birth Intervals in Kenya
Months Since Preceding Birth % of Mothers
2003 2008–2009
7 – 17 9.4 9.1
18 – 23 13.5 13.5
24 – 35 36.5 34.2
36 – 47 16.9 17.9
48 – 59 9.8
48 and above 23.7
60 and above 15.4
Total 100.0 100.0
Note: The figures may not sum to 100 due to rounding.
Source: Opiyo (2004, Table 4.7); Munguti and Buluma (2010, Table 4.7).
al. (2004) and Rutstein (2005), we use preceding inter–birth interval as a measure of
birth spacing. Birth interval length in our case, therefore, refers to the time, in months,
between the birth of the child in whom we are interested and the immediate previous birth
to the mother (Rutstein, 2005). Rutstein (2005, Table 1) summarizes the distribution of
preceding birth intervals in months for a select group of developing countries and shows
that majority of the birth intervals are of length 24 – 29 months.
Table 3.2 shows the distribution of preceding birth intervals for Kenya based on the
data gathered from the Kenya Demographic and Health Surveys in 2003 and in 2008 –
2009. The table shows that for both periods, the majority of the mothers had a preceding
birth interval length of 24 – 35 months.
The rest of the chapter is organized as follows: Section 3.2 reviews the literature while
Section 3.3 gives the purpose and objectives of the study. The methodology is discussed
in Section 3.4, the descriptive statistics are discussed in Section 3.5 , Section 3.6 presents
the results and Section 3.7 gives the summary, conclusions and policy implications.
Preceding Birth Interval Length and Maternal Health 77
Table 3.3: Determinants of Maternal Health
Determinant Example of Studies
Intermediate or Proximate Determinants
Maternal health status prior to and during a pregnancy McCarthy and Maine, 1992.
Maternal reproductive characteristics McCarthy and Maine, 1992;
(such as age, pregnancy order, birth spacing, etc) Gertler, Rahman, Feifer and Ashley, 1993;
Thaddeus and Maine, 1994; Lule et al., 2005.
Access to health services McCarthy and Maine, 1992.
Health care behaviour/use of health services McCarthy and Maine, 1992;
Thaddeus and Maine, 1994.
Unknown or unpredicted factors McCarthy and Maine, 1992.
Distant determinants
Women’s status in family and community McCarthy and Maine, 1992; Raghupathy, 1996;
Mistry, Galal and Lu, 2009.
Family’s status in community McCarthy and Maine, 1992.
Community’s status McCarthy and Maine, 1992.
Source: Author’s Own Construction Based on McCarthy and Maine (1992, Figure 2).
3.2 Literature Review
In general, a number of factors influence maternal health. These factors are called
intermediate determinants if their influence is direct and distant determinants if their
influence is indirect (McCarthy and Maine, 1992). Some of these factors are summarized
in Table 3.3.
Birth intervals can influence the health of both the mother and her unborn child (Rosen-
zweig and Schultz, 1987; Miller, 1991; Ikamari, 1998; Whitworth and Stephenson, 2002;
Bhargava, 2003; Norton, 2005; Rutstein, 2005; Razzaque et al., 2005; Conde–Agudelo,
Rosas–Bermudez and Kafury–Goeta, 2006; WHO, 2006). For example, compared to
women with inter–pregnancy intervals of 27–50 months, the risk of suffering from several
different maternal morbidities such as pre–eclampsia and high blood pressure is higher
among women having inter–pregnancy intervals of less than 6 months or 75 or more
months (Razzaque et al., 2005). Further, longer birth intervals are associated with a
lower risk for neonatal, infant and child mortality (Rutstein, 2005).
Demographic and medical literature suggests that the best or optimal, in terms of
not compromising both infant and maternal health, birth interval length is between
Preceding Birth Interval Length and Maternal Health 78
36 to 59 months (Conde–Agudelo and Belizan, 2000; Setty–Venugopal and Upadhyay,
2002; Razzaque, et al., 2005; Rutstein, 2005; Conde–Agudelo, Rosas–Bermudez and
Kafury–Goeta, 2007). Interval lengths outside this range are associated with an increased
risk of adverse maternal health outcomes such as increased risk of pre–eclampsia and
eclempsia, increased risk for labour dystocia, increased risks of uterine rupture and
uteroplacental bleeding disorders, increased risk for maternal death, and increased risk
for anaemia (Conde–Agudelo and Belizan, 2000; Conde–Agudelo, Rosas–Bermudez and
Kafury–Goeta, 2007).
A variety of factors influence a mother’s spacing of births such as the health status of
the previous child, maternal demographic and socio–economic characteristics, practices
such as breastfeeding and postpartum abstinence, and cultural norms (Setty–Venugopal
and Upadhyay, 2002). Table 3.4 on page 79 summarizes the risk factors for short birth
intervals. Some studies (see, for example, Rafalimanana and Westoff, 2001), however,
show that in some countries women prefer longer birth intervals than they actually have.
3.3 Purpose and Objectives of the Study
Previous studies concerning birth intervals in Kenya have either examined the link
between birth intervals and child health (see, for example, Ikamari, 1998; Kosimbei, 2005;
Mustafa and Odimegwu, 2008) or the gap between preferred and actual birth intervals
(see, for example, Rafalimanana and Westoff, 2001). To the best of our knowledge, no
study has so far attempted to link preceding birth interval to maternal health in Kenya
or to analyze the determinants of preceding birth interval lengths in Kenya. This study
fills this gap.
The study contributes to the literature by examining the effect of preceding birth
interval length on maternal health in Kenya. Specifically, the objectives of the study
include:
1. Determining the effect of preceding birth interval length on maternal health, where
maternal health is measured by whether or not a woman has ever had a pregnancy
that was terminated, miscarried, or ended in a still birth.
2. Determining the factors that influence the preceding birth interval length chosen by
the mother.
Preceding Birth Interval Length and Maternal Health 79
Table 3.4: Risk Factors for Short Birth Intervals
Factor Example of Studies
Infant death Zenger, 1993; Bohler, 1994; Mturi, 1997;
Udjo, 1997;
Grummer–Strawn, Stupp and Mei, 1998;
Gyimah, 2002.
Poor infant health Bereczkei, Hofer and Ivan, 2000;
Setty–Venugopal and Upadhyay, 2002.
Young maternal age Suwal, 2001;
Setty–Venugopal and Upadhyay, 2002;
Rasheed and Al–Dabal, 2007.
No formal education for mother Setty–Venugopal and Upadhyay, 2002.
Having first birth later in life Upadhyay and Hindin, 2005.
Having fewer children Setty–Venugopal and Upadhyay, 2002.
Low maternal socio–economic status Isvan, 1991;
Setty–Venugopal and Upadhyay, 2002.
Non–employment of mother Setty–Venugopal and Upadhyay, 2002.
Little maternal decision–making autonomy Upadhyay and Hindin, 2005.
Rural residence Mturi, 1997;
Setty–Venugopal and Upadhyay, 2002.
Social pressure for women to prove fertility Setty–Venugopal and Upadhyay, 2002.
Short breastfeeding duration Millman and Cooksey, 1987; Mturi, 1997;
Setty–Venugopal and Upadhyay, 2002;
Hajian–Tilaki, Asnafi and Aliakbarnia–Omrani, 2009.
Not practising postpartum abstinence Setty–Venugopal and Upadhyay, 2002.
Son preference Mace and Sear, 1997;
Graham, Larsen and Xu, 1998;
Setty–Venugopal and Upadhyay, 2002.
Short previous birth intervals Trussell et al., 1985.
Non–contraceptive use Yeakey et al., 2009.
3. Drawing appropriate policy implications from the study findings.
Preceding Birth Interval Length and Maternal Health 80
Figure 3.1: A Conceptual Framework for Analyzing the Effect of Preceding Birth IntervalLength on Maternal Health
Unobserved Variables
Maternal /Household Demographic , Reproductive and Socio – economic characteristics
Preceding Birth Interval Length
Maternal Health Outcome (Miscarriage /Stillbith/ Abortion)
Characteristics of the Previous Child
Maternal/ household preferences
Community Characteristics /Environmental Factors
Maternal Health Endowment
Observed Variables
Source: Own Construction Based on Figure 1.1. and Schultz (1984).
3.4 Methodology
Conceptual Model
Based on the general framework shown in Figure 1.1 of Chapter 1, we can develop a
conceptual framework for analyzing the effect of birth interval length on maternal health.
Such a conceptual framework is shown in Figure 3.1.
According to the figure, the maternal health outcome measure (in our case, whether
or not a woman has ever experienced a pregnancy termination, miscarriage, or stillbirth),
Preceding Birth Interval Length and Maternal Health 81
is influenced by the preceding birth interval length through a maternal health production
function. It is also influenced by maternal health endowment (the component of maternal
health due to either genetic or environmental conditions uninfluenced by maternal
behaviour but partly known to the mother or her family, but not to the researcher).
The preceding birth interval length, in turn, is influenced by unobservable mater-
nal/household preferences, and unobservable maternal health endowment (for example,
unreported pre-existing maternal health conditions). It is further influenced by observ-
able maternal demographic and socio-economic characteristics (such as age, number of
children, education, place of residence, etc), characteristics of the previous child (i.e.,
whether the child is alive or dead, and whether the child is sickly or not sickly), household
characteristics such as household wealth, and community characteristics (such as cultural
norms/practices and health infrastructure).
Theoretical Model
Our theoretical model follows closely that by Rosenzweig and Schultz (1982, 1983) and
Mwabu (2009). We assume that a multiparous mother derives utility, U , from the
consumption of various goods and sevices, X, and from her own health, H. We can
therefore represent the utility function of a typical mother as follows:
U = U (X,H) . (3.1)
The mother’s health, however, is influenced by the preceding interval between one
child and the next, Z, other inputs, Y , and unobservable biological endowments, µ. We
can, therefore, represent the mother’s health production function as follows:
H = H (Y, Z, µ) . (3.2)
The mother faces the following budget constraint:
I = PxX + PyY + PzZ (3.3)
where I is exogenous household income, Px is price per unit of X, Py is price per unit of
Y , and Pz is price per unit of Z.
Preceding Birth Interval Length and Maternal Health 82
Even though Z is not directly purchased from the market, it depends on other goods
that are purchased from the market.
The mother’s standard economic problem relates to maximizing her utility function
subject to her health production function and budget constraint (Rosenzweig and Schultz,
1982). The solution to the mother’s utility maximization model leads to the following
reduced input demand equations.
X = X (Px, Py, Pz, I, µ) (3.4)
Y = Y (Px, Py, Pz, I, µ) (3.5)
Z = Z (Px, Py, Pz, I, µ) . (3.6)
Following Mwabu (2009), it can be shown, by the technique of total differentiation,
that price effects on maternal health depend on the effects of changes in prices on the
demand for health production inputs as well as on the marginal products of these inputs in
the production of maternal health. The implication of this is that we must simultaneously
estimate the maternal health production function and input demand parameters, so as
to be able to predict the effect of changes in the prices of the various health inputs on
maternal health (Mwabu, 2009).
Following Rosenzweig and Schultz (1983) and Mwabu (2008), we can formulate a
hybrid maternal health production function that combines equations (3.2) and (3.5) in
an effort to causally link Z to changes in maternal health status. Such a hybrid function
is given by:
H = h (Z, Px, Py, I, µ) . (3.7)
In this equation, Z is potentially endogenous since it may be affected by the initial
health status of the mother (Mwabu, 2008). It can also be shown that the fact that we
cannot observe µ creates complications in interpreting the coefficient of Z as its marginal
product, if not controlled for (see Rosenzweig and Schultz, 1982). This means that the
estimation procedure we adopt for Equation (3.7) should take into account the fact that
Z is potentially endogenous and the fact that µ is unobservable (Mwabu, 2008).
Preceding Birth Interval Length and Maternal Health 83
Estimation Issues
As explained in Section 1.6 of Chapter 1, when estimating our models, we need to worry
about potential sample selection bias, potential endogeneity, and potential unobserved
heterogeneity.
Sample Selection Bias
Our discussion on sample selection bias follows closely that in Vella (1998). In this
Chapter, we measure the health status of a mother by asking her whether or not she
experienced an adverse pregnancy outcome such as a miscarriage, a stillbirth or an
abortion. We are, therefore, only able to observe the health status of a mother if she
reports it in the dataset. For mothers who do not report their health status, we are
unable to analyze the effect of the preceding birth interval length on their health status.
We are, however, interested in determining the effect of the preceding birth interval
length on maternal health for all mothers, not just those who report their health status.
If the unobservable factors that influence a mother’s reporting of her health status
are correlated with the unobservable factors that influence her reported health status,
then our estimates will be biased if we do not include in our maternal health model
an estimate of these unobservable factors (Vella, 1998). An issue of sample selection
bias will, therefore, arise. We must therefore control for potential sample selection bias
in our estimation by including an estimate of these unobservable factors. We use the
method proposed by Olsen (1980) to control for this selection bias. An advantage of the
Olsen approach compared to the Heckman (1979) approach is that, unlike the Heckman
approach which requires an iterative probit in the first step, the Olsen approach only
requires Ordinary Least Squares (OLS) regression techniques in the first step (Olsen,
1980).
The Olsen approach (Olsen, 1980) is implemented as follows:
i. Estimate a linear probability model of the selection equation. There should, however,
be at least one regressor in the linear probability model that is not in the main
model of interest.
ii. Obtain the predicted probability of selection into the sample. Call this predicted
probability, say, P .
Preceding Birth Interval Length and Maternal Health 84
iii. Construct the variable(P − 1
). This variable, referred to as the selection term,
is the additional variable that should be included in the substantive model as a
correction for sample selection bias. If the coefficient of this variable is significantly
different from zero, then we reject the null hypothesis of absence of selection bias.
Endogeneity
In our model, the covariate measuring the preceding birth interval length is potentially
endogenous. To some extent, the preceding birth interval length is a choice variable
(chosen by the mother either singly or jointly with other household members). If
the unobservable factors that influence maternal health status are correlated with the
preceding birth interval length chosen by the mother, then the estimate of preceding
birth interval length in the maternal health production function is inconsistent (Guevara
and Ben–Akiva, 2008) and cannot be given a causal interpretation (Cameron and Trivedi,
2010). We control for this potential endogeneity by using the Two–Stage–Residual–
Inclusion (2SRI) method (Terza, Basu and Rathouz, 2008). This approach involves
including as an additional regressor in our maternal health equation, the generalized
residuals from the preceding birth interval length model (for a discussion on how the
generalized residuals are computed, see Gourieroux, Monfort, Renault and Trognon,
1987). The preceding birth interval length is said to be endogenous if the coefficient of
the preceding birth interval length residuals is statistically significant in the maternal
health status equation (Bollen, Guilkey and Mroz, 1995).
Unobserved Heterogeneity
Unobserved heterogeneity will exist in our maternal health status model if there are
some unobservable factors that interact non-linearly with the preceding birth interval
length causing the effect of the preceding birth interval length on maternal health to
differ amongst mothers in the population (Zohoori and Savitz, 1997). To control for
unobserved heterogeneity, we include in our maternal health equation, the interaction
of the preceding birth interval length with its residuals, as an additional regressor (see,
for example, Mwabu, 2009). Unobserved heterogeneity will exist in our model if the
coefficient of this additional regressor is statistically significant.
Preceding Birth Interval Length and Maternal Health 85
Model Identification
We follow Mwabu (2009) in our discussion of model identification. For us to properly
interpret the estimated coefficients of our maternal health model, the maternal health
effects of the preceding birth interval length (the potentially endogenous input) and
of the sample selection rule, should be identified (Mwabu, 2009). We assume that the
preceding birth interval length is the only potentially endogenous input. Consequently,
identification requires a minimum of two (not one) exclusion restrictions; one for the
endogenous covariate and another one for determining the selection of mothers into the
estimation sample (Mwabu, 2009).
The chosen instruments should be exogenous (that is, they should be uncorrelated
with the error term in the maternal health equation), relevant (that is, they should
significantly affect the preceding birth interval length), and should be excluded from
the maternal health equation (Murray, 2006; Mwabu, 2009; Brookhart, Rassen and
Schneeweiss, 2010).
Some of the instruments that have been used in the literature include miscarriages (see,
for example, Buckles and Munnich, 2012), muslim religion (see, for example, Makepeace,
2006), duration of breastfeeding (see, for example, Maitra and Pal, 2005), sex composition
of previous children (see, for example, Maitra and Pal, 2008), and contraception use (see,
for example, Maitra and Pal, 2008).
We assume exact identification for our models and use two instruments in our
estimation strategy (Murray, 2006). The instruments we use are the average breast
feeding duration, computed from our dataset at the provincial level; and the proportion
of women in each province reporting to have ever used contraception, referred to here as
contraceptive prevalence, also computed from our dataset at the provincial level.
The average duration of breastfeeding is chosen because of the association between
duration of breastfeeding and birth spacing (see, for example, Mturi, 1997; Hajian–Tilaki,
Asnafi and Aliakbarnia–Omrani, 2009; Singh, Singh and Narendra, 2010). It has also
been demonstrated in the literature (see, for example, Millman and Cooksey, 1987) that
lack of breastfeeding increases the risk of conceiving.
The proportion of women using contraceptives is chosen because there is evidence in
the literature that contraceptive use is associated with longer birth intervals (see, for
example, Rafalimanana and Westoff, 2001; Yeakey et al., 2009). Although contraceptive
use at the level of the individual woman has beneficial effects on maternal health (see,
Preceding Birth Interval Length and Maternal Health 86
for example, Ahmed, Qingfend, Liu and Tsui, 2012), there is no reason to expect the
prevalence of contraceptive use in a particular region to have effects on the health of an
individual woman.
We expect both our instruments to be positively related to preceding birth interval
length. We expect contraceptive prevalence to be positively related to preceding birth
interval length because the higher the prevalence of contraceptive use in a particular area,
the more accessible or available contraceptives are in that area and consequently the
higher the chances that any individual woman who wants to use them will be accessible
to them in that area. We expect higher average breastfeeding in an area to impact on
the duration of breastfeeding of an individual woman through peer effects.
We use provincial averages to remove the likelihood of the instruments being choice
variables by the mother. We argue that although an individual woman can choose the
duration to breastfeed, she is unlikely to influence the average breastfeeding duration in
the entire province where she lives. Similarly, although an individual woman can choose
whether or not to use contraceptives, she is also unlikely to influence the proportion of
women in the entire province using any type of contraception at any given point in time.
Empirical Model
We measure maternal health using a binary variable. For a typical mother i, we define
the following health status measure:
Hi =
1 if mother i experienced a miscarriage, stillbirth, or abortion,
0 otherwise.(3.8)
where Hi is the observed health status of mother i.
Being a binary variable, the appropriate model for maternal health status is the
binary regression model (Long, 1997; Long and Freese, 2006).
We assume that underlying this observed maternal health status is a continuous latent
variable, H∗i that is related to the observed health status for mother i via the following
Preceding Birth Interval Length and Maternal Health 87
equation:
Hi =
1 if H∗i > 0,
0 otherwise.(3.9)
The latent variable is in turn linked to the covariates via the following equation:
H∗i = β0 + β1Y + β2Zi + ε1i (3.10)
where Y is a vector of controls, Z is preceding birth interval length and ε1 is the stochastic
error term.
When we have values of Y and Z, it can be demonstrated that (see, for example,
Long and Freese, 2006; Cameron and Trivedi, 2010):
Pr (Hi = 1 | Y, Z) = Pr (H∗i > 0 | Y, Z) (3.11)
where Pr stands for “probability”.
Substituting Equation (3.10) into Equation (3.11) and performing further manipula-
tions gives:
Pr (Hi = 1 | Y, Z) = Pr (−ε1i ≤ β0 + β1Y + β2Z | Y, Z) = F (β0 + β1Y + β2Z) (3.12)
where F (·) is the cumulative density function of −ε1i, which is the same as that of ε1i if
ε1i is symmetric.
Assuming a standard normal distribution for ε1i leads to a probit model given by:
Pr (Hi = 1 | Y, Z) = F (β0 + β1Y + β2Z) = Φ (β0 + β1Y + β2Z) (3.13)
where Φ(·) is the standard normal cumulative distribution function.
Since Z is potentially endogenous, the 2SRI method requires that we obtain its
generalized residuals and include them as an additional regressor in our model. We must,
therefore, also formulate a model for Z that can be estimated.
Recall from Section 3.2 that the best birth interval, in terms of enhancing maternal
and infant health, is 36 to 59 months. Based on this observation, we define our birth
Preceding Birth Interval Length and Maternal Health 88
interval measure as follows:
Zi =
1 if the preceding birth interval is 36 to 59 months,
0 otherwise.(3.14)
where Zi is the observed preceding birth interval length for reference child i.
We assume that this observed preceding birth interval length is linked to an unob-
servable continuous variable Z∗i via the following equation.
Zi =
1 if Z∗i > 0,
0 otherwise.(3.15)
We can link Z∗i to the various covariates using the equation
Z∗i = α1 + α2Y + α3Q+ ε2i (3.16)
where Y is a vector of controls, Q is a vector of instruments, and ε2 is the stochastic
error term.
Assuming a standard normal distribution for ε2 leads to the following probit model:
Pr (Zi = 1) = Φ (α1 + α2Y + α3Q) . (3.17)
We estimate this model, obtain the generalized residuals for Z, and then include the
generalized residuals as an additional regressor in our structural model of interest to
control for endogeneity. Unobserved heterogeneity is controlled for by including in our
structural model of interest, an interaction between preceding birth interval length and
its generalized residual as an additional regressor.
The variable for inclusion of mothers in the sample is defined as follows:
Ii =
1 if mother i reports her health status,
0 otherwise.(3.18)
Preceding Birth Interval Length and Maternal Health 89
Following Olsen (1980), we formulate the following sample selection equation:
Ii = γ1 + γ2Y + γ3Q+ ε3i (3.19)
where Y is a vector of controls, Q is a vector of instruments, and ε3 is the stochastic
error term.
The Olsen approach requires that we estimate this model, obtain the predicted
probabilities for inclusion into the sample, P , construct the selection term, P − 1, and
include this selection term as an additional regressor in our structural equation of interest.
To control for potential endogeneity of birh interval length, potential sample selection
bias, and potential unobserved heteorgeneity, Equation (3.10) is extended as follows:
H∗i = β0 + β1Y + β2Zi + β3ε2i + β4
(Pi − 1
)+ β5Ziε2i + ε1i (3.20)
where Y is a vector of controls, Z is the preceding birth interval length,(P − 1
)is the
sample selection term, ε2i are the generalized residuals from the preceding birth interval
length model, and ε1 is the stochastic error term.
Assuming a standard normal distribution for ε1 leads to the following probit model,
which is our structural equation of interest:
Pr (Hi = 1) = Φ(β0 + β1Y + β2Zi + β3ε2i + β4
(Pi − 1
)+ β5Ziε2i
). (3.21)
Our equations are estimated using Stata 11 software (StataCorp, 2009).
Data
Our main dataset comes from the Kenya Demographic and Health Surveys (KDHS)
conducted in Kenya in 1998, 2003, and 2008.5 These are nationally representative
household surveys that collect a wide range of household level data on child and maternal
health. In Kenya, they are carried out by the Kenya National Bureau of Statistics
(KNBS) in collaboration with various international organizations.
5More information on the Demographic and Health Surveys can be obtained at http://www.
measuredhs.com/What-We-Do/Survey-Types/DHS.cfm. The 1993 dataset for Kenya does not con-tain information on adverse pregnancy outcomes.
Preceding Birth Interval Length and Maternal Health 90
The analytic sample consists of women aged between 15 and 49 years who reported
whether or not they had experienced an adverse pregnancy outcome (specifically either a
miscarriage, stillbirth, or an abortion). The unit of observation is a multiparous mother,
aged 15 - 49 years, who reports whether or not she had experienced an adverse pregnancy
outcome.
Table 3.5 on page 91 shows the definition of the variables used in our models.
3.5 Descriptive Statistics
Descriptive statistics are shown in Table 3.6 on page 92. There are variations in the
number of observations for the various variables because some of the variables are missing
for some of the observations. The statistics show, for instance, that the youngest mother
gave birth at the age of 12 years while the oldest mother gave birth at the age of 48
years. We can also see from the table that contraceptive prevalence in our dataset varies
from 0.22% to 59.07%. Average breastfeeding duration, on the other hand, varies from
about 13 months to about 18 months.
Table 3.7 on page 92 shows the distribution of the maternal health status indicator
by year of survey. According to the table, about 10% of the women in our overall sample
report ever having experienced a miscarriage, a stillbirth or an abortion.
Table 3.8 on page 93 shows the distribution of maternal health status by preceding
birth interval length in our overall sample. The table shows that about 27.3% of the
women who had preceding birth interval lengths of 36–59 months did not experience
adverse pregnancy outcomes. The table also shows that about 71.2% of the women who
experienced adverse pregnancy outcomes had birth interval lengths outside the 36–59
months bracket.
Preceding Birth Interval Length and Maternal Health 91Table
3.5:
Var
iab
leD
efin
itio
ns
for
Mat
ern
alH
ealt
hM
od
els
Var
iable
Defi
nit
ion
Mat
ernal
hea
lth
stat
us
1if
mot
her
has
ever
exp
erie
nce
da
mis
carr
iage
,a
stillb
irth
,
oran
abor
tion
;0
other
wis
e.
Pre
cedin
gbir
thin
terv
alle
ngt
h1
ifth
epre
cedin
gbir
thin
terv
alis
36to
59m
onth
s,0
other
wis
e.
Rep
ort
hea
lth
stat
us
1if
mot
her
rep
orts
her
hea
lth
stat
us;
0ot
her
wis
e.
Pri
mar
yed
uca
tion
1if
mot
her
’shig
hes
tle
vel
ofed
uca
tion
ispri
mar
y;
0ot
her
wis
e.
Sec
ondar
yed
uca
tion
1if
mot
her
’shig
hes
tle
vel
ofed
uca
tion
isse
condar
y;
0ot
her
wis
e.
Hig
her
educa
tion
1if
mot
her
’shig
hes
tle
vel
ofed
uca
tion
ishig
her
;0
other
wis
e.
Curr
entl
ym
arri
ed1
ifm
other
iscu
rren
tly
mar
ried
;0
other
wis
e.
Mal
ere
fere
nce
child
1if
refe
rence
child
ism
ale;
0ot
her
wis
e.
Mot
her
’sag
eat
bir
thM
other
’sag
eat
bir
thof
refe
rence
child
inye
ars.
Num
ber
oflivin
gch
ildre
nN
um
ber
oflivin
gch
ildre
nb
orn
tom
other
.
Urb
anre
siden
ce1
ifurb
an;
0ot
her
wis
e.
Ass
etin
dex
Ass
etin
dex
,ra
nge
sin
incr
easi
ng
order
from
0to
10.
1998
1if
surv
eyye
aris
1998
;0
other
wis
e.
2003
1if
surv
eyye
aris
2003
;0
other
wis
e.
2008
1if
surv
eyye
aris
2008
;0
other
wis
e.
Con
trac
epti
vepre
vale
nce
Per
centa
geof
wom
enre
por
ting
touse
any
contr
acep
tive
met
hod.
Com
pute
dat
the
pro
vin
cial
leve
l.
Ave
rage
dura
tion
ofbre
astf
eedin
gA
vera
gedura
tion
ofbre
astf
eedin
gin
mon
ths.
Com
pute
dat
the
pro
vin
cial
leve
l.
Squar
eof
dura
tion
ofbre
astf
eedin
gT
he
squar
eof
the
aver
age
dura
tion
ofbre
astf
eedin
g.
Sel
ecti
onte
rmT
erm
tom
easu
rese
lect
ivit
ybia
s.C
onst
ruct
edbas
edon
Ols
en(1
980)
.
Pre
cedin
gbir
thin
terv
alle
ngt
hre
sidual
Gen
eral
ized
resi
dual
from
the
bir
thin
terv
alm
odel
.
Pre
cedin
gbir
thin
terv
alin
tera
cted
wit
hre
sidual
Inte
ract
ion
ofbir
thin
terv
alle
ngt
hw
ith
its
resi
dual
.
Preceding Birth Interval Length and Maternal Health 92
Table 3.6: Descriptive Statistics for Maternal Health Models
Variable Number of Mean Standard Minimum maximum
Observations Deviation
Maternal health status 15,554 0.103 0.304 0 1
Preceding birth interval length 11,773 0.274 0.446 0 1
Report health status 15,559 0.9997 0.0179 0 1
Primary education 15,554 0.588 0.492 0 1
Secondary education 15,554 0.183 0.387 0 1
Higher education 15,554 0.041 0.197 0 1
Currently married 15,554 0.797 0.403 0 1
Male reference child 15,554 0.509 0.500 0 1
Mother’s age at birth 14,424 26.235 6.466 12 48
Number of living children 15,554 3.406 2.163 0 13
Urban residence 15,554 0.227 0.419 0 1
Asset index 14,472 3.576 1.966 0 10
1998 15,554 0.227 0.419 0 1
2003 15,554 0.382 0.486 0 1
2008 15,554 0.391 0.488 0 1
Contraceptive prevalence 15,554 32.502 13.125 0.22 59.07
Average duration of breastfeeding 15,554 16.169 1.571 13.04 18.58
Square of duration of breastfeeding 15,554 263.897 50.151 170.042 345.216
Table 3.7: Distribution of Maternal Health Status by Survey Year, Percentages in Parentheses
Year of Survey Ever Experienced Miscarriage, Stillbirth or Abortion Total
No Yes
1998 3,249 279 3,528
(92.1%) (7.9%) (100%)
2003 5,287 661 5,948
(88.9%) (11.1%) (100%)
2008 5,416 662 6,078
(89.1%) (10.9%) (100%)
Total 13,952 1,602 15,554
(89.7%) (10.3%) (100%)
Preceding Birth Interval Length and Maternal Health 93
Table 3.8: Distribution of Maternal Health Status by Preceding Birth Interval Length, Per-centages in Parentheses
Ever Experienced a Miscarriage, Preceding Birth Interval Length 36 – 59 months Total
Stillbirth or Abortion No Yes
No 7,561 2,833 10,394
(72.7%) (27.3%) (100%)
Yes 982 397 1,379
(71.2%) (28.8%) (100%)
Total 8,543 3,230 11,773
(72.6%) (27.4%) (100%)
3.6 Results
First–Stage Models
We estimate our models in two stages. In the first stage we estimate the sample selection
equation together with the preceding birth interval equation. In the second stage we
estimate the maternal health production functions.
As indicated in Chapter 2 Section 2.6, we report the average marginal effects for our
models. Table 3.9 on page 94 shows the results of the sample selection model and the
preceding birth interval model. The results from the sample selection model are shown
in the column labelled (1) while those from the preceding birth interval length model are
shown in the column labelled (2).
Focusing on the preceding birth interval model in column (2), we observe that the
coefficients on the instrumental variables have the expected signs. In particular, the
higher the contraceptive prevalence rate in the province in which the mother lives, the
higher the probability of having a preceding birth interval length of 36 to 59 months,
holding other factors constant. Similarly, the higher the average duration of breastfeeding
in the province in which the mother lives, the higher the probability of having a preceding
birth interval of length 36 to 59 months, holding other factors constant.
A careful look at the results for the preceding birth interval model reveals that
significant determinants of the 36 to 59 months preceding birth interval include secondary
Preceding Birth Interval Length and Maternal Health 94
Table 3.9: Average Marginal Effects from Sample Selection and Preceding Birth IntervalLength Models, Robust Z Statistics in Parentheses
Variable Sample Selection Model Preceding Birth Interval Model
(Report health status =1) (36 – 59 months =1)
(1) (2)
Primary education 0.00025 -0.012
(0.67) (-1.12)
Secondary education 0.00065 -0.035
(1.87) (-2.61)
Higher education 0.00043 -0.0067
(1.81) (-0.23)
Currently married 0.00018 -0.044
(0.45) (-4.14)
Male reference child -0.00030 -0.014
(-0.99) (-1.97)
Mother’s age at birth 0.000069 0.009
(1.55) (10.80)
Number of living children -0.0002 -0.015
(-0.88) (-5.67)
Urban residence -0.00010 0.028
(-0.14) (2.40)
Asset index -0.00008 -0.003
(-1.16) (-1.25)
2003 0.00089 -0.024
(1.93) (-2.57)
2008 0.0011 -0.0017
(1.86) (-0.18)
Contraceptive prevalence 0.00001 0.0009
(1.43) (2.29)
Duration of breastfeeding 0.0015 0.046
(1.49) (1.71)
Square of duration of breastfeeding -0.00005 -0.0014
(-1.69) (-1.70)
Number of observations 13,417 14,692
Preceding Birth Interval Length and Maternal Health 95
education, marital status, sex of the reference child, mother’s age at birth, number of
living children, and type of residence (that is, whether rural or urban).
According to the results, mothers whose highest level of education is secondary have a
lower probability of having preceding birth intervals of 36 – 59 months compared to those
without formal education, holding other factors constant. The results also show that
currently married mothers have a lower probability of having preceding birth intervals of
length 36 – 59 months compared to their not currently married counterparts, holding
other factors constant.
The results further indicate that the probability of the preceding birth interval being
36 – 59 months is lower if the reference child is male as opposed to female, holding other
factors constant. This finding is similar to the finding in the literature (see, for example,
Mace and Sear, 1997). This would tend to suggest that mothers looking out for male
children have shorter preceding birth intervals. There are, however, other studies in the
literature (see, for example, Chakraborty, Sharmin and Islam, 1996) that do not find the
sex of the reference child to be a significant determinant of birth interval length.
Mother’s age at birth is shown to positively affect the probability of having a preceding
birth interval of length 36 – 59 months, holding other factors constant. This result is
consistent with the findings reported in the literature (see, for example, Chakraborty,
Sharmin and Islam, 1996; Hajian–Tilaki, Asnafi and Aliakbarnia–Omrani, 2009). The
results also indicate that the higher the number of living children born to a mother,
the lower the probability of having preceding birth intervals of length 36 – 59 months,
holding other factors constant. Further, we observe from column (2) of the table that
mothers residing in urban areas have a higher probability, as compared to those living in
the rural areas, of having a preceding birth interval of length 36 to 59 months, holding
other factors constant.
Maternal Health Models
The estimated versions of the maternal health production functions are shown in Table
3.10 on page 96. The table shows the estimation results of four maternal health status
models. In column (1) of the table we present the results of the maternal health status
model that does not correct for sample selection bias, endogeneity of preceding birth
interval length or unobserved heterogeneity. The results in column (2) are for a model
that corrects for sample selection bias. In column (3) we correct for both sample selection
Preceding Birth Interval Length and Maternal Health 96
Table 3.10: Average Marginal Effects for Maternal Health Status Models, Robust Z Statisticsin Parentheses
Variable Miscarriage, stillbirth, or abortion =1
(1) (2) (3) (4)
Preceding birth interval length 0.008 0.008 -0.325 -0.325
(1.16) (1.15) (-104.38) (-104.44)
Primary education 0.012 0.001 -0.022 -0.031
(1.27) (0.08) (-1.57) (-1.92)
Secondary education -0.008 -0.029 -0.082 -0.10
(-0.08) (-1.63) (-4.55) (-4.56)
Higher education 0.014 -0.002 -0.028 -0.037
(0.60) (-0.08) (-1.30) (-1.78)
Currently married 0.029 0.026 -0.029 -0.057
(3.55) (2.87) (-1.49) (-1.73)
Male reference child 0.0015 0.010 0.011 0.011
(0.24) (1.15) (1.24) (1.28)
Mother’s age at birth 0.006 0.004 0.009 0.012
(7.84) (2.22) (4.14) (3.68)
Number of living children 0.003 0.008 0.002 -0.001
(1.49) (1.98) (0.40) (-0.24)
Urban residence -0.013 -0.011 0.022 0.037
(-1.51) (-1.23) (1.58) (1.88)
Asset index -0.0013 0.002 0.007 0.009
(-0.65) (0.67) (2.03) (2.33)
2003 0.047 0.021 -0.045 -0.071
(5.04) (1.09) (-1.83) (-2.18)
2008 0.045 0.013 -0.040 -0.061
(4.94) (0.54) (-1.54) (-1.95)
Selection term 29.212 82.869 105.216
(1.42) (3.33) (3.36)
Preceding birth interval length residual 0.599 0.911
(3.74) (2.88)
Preceding birth interval interacted with residual -0.142
(-1.13)
Number of observations 10,233 10,233 10,233 10,233
bias and potential endogeneity of preceding birth interval length. Column (4) shows
results where sample selection bias, potential endogeneity of preceding birth interval
length and potential unobserved heterogeneity are controlled for.
Preceding Birth Interval Length and Maternal Health 97
In column (2) we notice that the coefficient of the selection term is not statistically
significant. This indicates that there is no sample selection bias in our model. Preceding
birth interval length is indeed endogenous in our model as indicated by the statistical
significance of the coefficient of the preceding birth interval residuals in column (3).
Notice from column (3) that, after controlling for the endogeneity of preceding birth
interval length, we now notice that there is actually sample selection bias in our model
and failure to control for it would bias the results. From the results in column (4), we can
conclude that their is no unobserved heterogeneity in our model since the coefficient of
the interaction between preceding birth interval length and the preceding birth interval
length residuals is not statistically significant. Our preferred model is, therefore, the one
shown in column (3). We, therefore, focus our attention on the results in column (3).
We can observe from column (3) that preceding birth interval length is a significant
determinant of maternal health status. Specifically, the results show that having a
preceding birth interval length of 36 – 59 months reduces the probability of reporting
an adverse pregnancy outcome (a miscarriage, stillbirth, or an abortion) by 0.325. This
indicates that optimal birth spacing is beneficial to maternal health. This finding is
consistent with the findings reported in the literature where birth intervals outside 36
– 59 months are found to be associated with an increased risk of stillbirths (see, for
example, Williams et al., 2008), where short or long birth intervals are found to be
detrimental to maternal health (see, for example, Conde–Agudelo and Belizan, 2000;
Conde–Agudelo, Rosas–Bermudez and Kafury–Goeta, 2007; Stamilio et al., 2007), and
where long interpregnancy intervals are found to be associated with an increased risk of
stillbirths (see, for example, Stephansson, Dickman and Cnattingius, 2003).
Other significant determinants of maternal health status include secondary education,
mother’s age at birth, and household wealth (as measured by the asset index). Mothers
whose highest level of education is secondary have a lower probability of experiencing
a miscarriage, stillbirth, or an abortion, compared to those without formal schooling,
holding other factors constant. This finding is supported by the findings in the literature
where stillbirth rates are found to be higher for mothers with lower levels of education
(see, for example, Luque–Fernadez, Gutierrez–Garitano and Bueno–Cavanillas, 2012;
Auger, Delezire, Harper and Platt, 2012), and where education is shown to have beneficial
effects on maternal health (see, for example, McAlister and Baskett, 2006; Karlsen et al.,
2011).
Although marital status is not statistically significant in our model, the sign of the
coefficient implies that currently married mothers are less likely to experience adverse
Preceding Birth Interval Length and Maternal Health 98
pregnancy outcomes compared to those who are not currently married, holding other
factors constant. This suggests that being married has beneficial effects on maternal
health. This finding is consistent with the finding in the literature (see, Gunilla, Haglund
and Rosen, 2000) that lone mothers are at a higher risk of premature death than mothers
with partners. Other studies (see, for example, Osborn, Cattaruzza and Spinelli, 2000)
show that unmarried women have an increased risk for spontaneous abortion.
The older the mother at the time of birth, the higher the probability of experiencing
an adverse pregnancy outcome, holding other factors constant. This result is consistent
with the finding in the literature that women who give birth at an advanced age have a
higher risk of a stillbirth (see, for example, Miller, 2005; Huang et al., 2008; Reddy, Ko
and Willinger, 2006), and the finding that women aged 35 and above are at a higher risk
of experiencing a miscarriage (De la Rochebrochard and Thonneau, 2002).
According to the results, mothers from wealthy households are less likely to experience
adverse pregnancy outcomes, holding other factors constant. This result is consistent
with the finding in the literature that poor women are at a higher risk of having stillbirths
(Ha et al., 2012).
3.7 Summary, Conclusions and Policy Implications
Summary
This chapter examines the effect of preceding birth interval length on maternal health
in Kenya. We measure maternal health using an indicator of whether or not a woman
has ever had a pregnancy that was terminated, miscarried, or ended in a still birth
since adverse pregnancy outcomes may point towards some underlying maternal health
problems (see, for example, Villamor and Cnattingius, 2006).
The literature identifies key risk factors for adverse pregnancy outcomes as short
interpregnancy intervals, long interpregnancy intervals, maternal obesity, maternal mo-
bidity, and smoking (see, for example, Walsh, 1994; Cnattingius, Bergstrom, Lipworth
and Kramer, 1998; Rosenberg, Garbers, Lipkind and Chiasson, 2005; Conde–Agudelo,
Rosas–Bermudez and Kafury–Goeta, 2006).
According to the literature, birth interval lengths of between 36 to 59 months do
not compromise both infant and maternal health (see, for example, Setty–Venugopal
Preceding Birth Interval Length and Maternal Health 99
and Upadhyay, 2002; Conde–Agudelo, Rosas–Bermudez and Kafury–Goeta, 2007). Key
factors that influence birth intervals include the health status of the previous child,
mother’s demographic and socioeconomic characteristics, contraceptive use and cul-
tural/traditional practices such as breastfeeding and postpartum abstinence (see, for
example, Setty–Venugopal and Upadhyay, 2002; Yeakey, et al., 2009).
Earlier studies concerning birth intervals in Kenya (see, for example, Ikamari, 1998;
Rafalimanana and Westoff, 2001; Kosimbei, 2005; Mustafa and Odimegwu, 2008) do not
link preceding birth interval to maternal health. Our study fills this gap.
We adopt an estimation strategy that controls for potential endogeneity of preceding
birth interval length, potential sample selection bias and potential unobserved hetero-
geneity. In the estimation, the proportion of women who have ever used any type of
contraceptive in each province and the average duration of breastfeeding in each province,
are used as instruments. We estimate the model using the Demographic and Health
Survey (DHS) datasets for Kenya for 1998, 2003 and 2008.
Our main finding is that preceding birth interval lengths of 36 to 59 months reduce
the probability of a mother experiencing a miscarriage, stillbirth, or an abortion, holding
other factors constant. This result is consistent with the findings in the literature (see,
for example, Williams et al., 2008). We further find that preceding birth interval length
is an endogenous determinant of maternal health. Maternal health is also shown to be
significantly influenced by secondary education, mother’s age at birth, and household
wealth (as measured by the asset index). We also find that preceding birth interval
lengths are influenced by secondary education, sex of reference child, mother’s age at
birth of reference child, number of living children and mother’s region of residence (either
urban or rural).
Conclusions
We can draw a number of conclusions from the findings in this Chapter. First, consistent
with the findings in the literature, preceding birth intervals of length 36 to 59 months
improve maternal health. Second, preceding birth interval is, however, an endogenous
determinant of maternal health. This endogeneity must be controlled for in models that
attempt to link preceding birth interval length to maternal health. Failure to control
for the endogeneity may lead to results showing that preceding birth intervals of length
36 to 59 months worsen maternal health. Finally, preceding birth interval length is
Preceding Birth Interval Length and Maternal Health 100
significantly influenced by maternal characteristics (such as her level of education, her
age at birth of the child, her marital status, whether she resides in urban or rural areas,
and number of living children), child characteristics (such as the sex of the child), and
community/cultural characteristics such as contraceptive prevalence and duration of
breastfeeding.
Policy Implications
Our findings imply that policies that encourage mothers to maintain a preceding birth
interval of 36 – 59 months should be pursued. This can be done by, for instance, making
family planning services easily accessible to mothers, and sensitizing mothers on the
health–benefits of adequate spacing of births. Mothers should also be encouraged to
breastfeed their children for longer periods of time. Investment in maternal education is
also a plausible policy in this case.
REFERENCES 101
References
[1] Ahmed, S., Qingfend, L., Liu, L. and Tsui, A. O., 2012. Maternal deaths avertedby contraceptive use: an anlysis of 172 countries. Lancet 380 (9837), pp.111–125.
[2] Auger, N., Delezire, P., Harper, S. and Platt, R. W., 2012. Maternal educationand stillbirth: estimating gestational–age–specific and cause–specific associations.Epidemiology 23 (2), pp.247–254.
[3] Bereczkei, T., Hofer, A. and Ivan, Z., 2000. Low birth weight, maternalbirth–spacing decisions, and future reproduction: a cost benefit analysis. HumanNature 11 (2), pp.183–205.
[4] Bhargava, A., 2003. Family planning, gender differences and infant mortality:evidence from Uttar Pradesh, India. Journal of Econometrics 112 (1), pp.225-240.
[5] Bohler, E., 1994. Has primary health care reduced infant mortality in EastBhutan? The effects of primary health care and birth spacing on infant and childmortality patterns in East Bhutan. Journal of Tropical Pediatrics 40 (5),pp.256-260.
[6] Bollen, K. A., Guilkey, D. K. and Mroz, T. A., 1995. Binary outcomes andendogenous explanatory variables: tests and solutions with an application to thedemand for contraceptive use in Tunisia. Demography 32 (1), pp.111-131.
[7] Brookhart, M. A., Rassen, J. A. and Schneeweiss, S., 2010. Instrumental variablemethods in comparative safety and effectiveness research. Pharmacoepidemiologyand Drug Safety 19 (6), pp.537–554.
[8] Buckles, K. S. and Munnich, E. L., 2012. Birth spacing and sibling outcomes.Journal of Human Resources 47 (3), pp.613–642.
[9] Cameron, A. C. and Trivedi, P. K., 2010. Microeconometrics using stata. RevisedEdition. Stata Press.
[10] Chakraborty, N., Sharmin, S. and Islam, M. A., 1996. Differential pattern ofbirth intervals in Bangladesh. Asia–Pacific Population Journal 11 (4), pp.73–86.
[11] Cnattingius, S., Bergstrom, R. Lipworth, L. and Kramer, M. S., 1998. Pregnancyweight and the risk of adverse pregnancy outcomes. New England Journal ofMedicine 338 (3), pp.147–152.
[12] Conde-Agudelo, A. and Belizan, J. M., 2000. Maternal morbidity and mortalityassociated with interpregnancy interval: cross sectional study. British MedicalJournal (BMJ) 321 (7271), pp.1255-1259.
[13] Conde–Agudelo, A., Rosas–Bermudez, A. and Kafury–Goeta, A. C., 2006. Birthspacing and risk of adverse perinatal outcomes: a meta–analysis. Journal ofAmerican Medical Association 295 (15), pp.1809–1823.
REFERENCES 102
[14] Conde-Agudelo, A., Rosas–Bermudez, A. and Kafury-Goeta, A. C., 2007. Effectsof birth spacing on maternal health: a systematic review. American Journal ofObstetrics & Gynecology 196 (4), pp.297-308.
[15] DaVanzo, J., Razzaque, A., Rahman, M., Hale, L., Ahmed, K., Khan, M. A.,Mustafa, A. H. M. G. and Gausia, K., 2004. The effects of birth spacing on infantand child mortality, pregnancy outcomes and maternal morbidity in Matlab,Bangladesh. RAND Labour and Population Working Paper WR–198, [pdf]RAND. Available at: http://pdf.usaid.gov/pdf_docs/pnadt346.pdf[Accessed 3 January 2010].
[16] De la Rochebrochard, E. and Thonneau, P., 2002. Paternal age and maternal ageare risk factors for miscarriage: results of a multicentre European study. HumanReproduction 17 (6), pp.1649–1656.
[17] Gertler, P., Rahman, O., Feifer, C. and Ashley, D., 1993. Determinants ofpregnancy outcomes and targeting of maternal health services in Jamaica. SocialScience & Medicine 37 (2), pp.199-211.
[18] Gourieroux, C., Monfort, A., Renault, E. and Trognon, A., 1987. Generalizedresiduals. Journal of Econometrics 34 (1 – 2), pp.5–32.
[19] Graham, M. J., Larsen, U. and Xu, X., 1998. Son preference in Anhui province,China. International Family Planning Perspectives 24 (2), pp.72–77.
[20] Grummer–Strawn, L. M., Stupp, P. W. and Mei, Z., 1998. Effect of a child’s deathon birth spacing: a cross–national analysis. In: M. R. Montgomery and B. Cohen,eds. From Death to Birth. Washington, D.C.: National Academy Press, pp.39–73.
[21] Guevara, C. A. and Ben–Akiva, M., 2008. A lagrange multiplier test for thevalidity of instruments in MNL models: an application to residual choice. [pdf]European Transport Conference 2008, Leeuwenhorst, The Netherlands. Availableat: http://www.etcproceedings.org/paper/a-lagrange-multiplier-test-for-the-validity-of-instruments-in-mnl-models-an-ap [Accessed 3January 2012].
[22] Gunilla, R. W., Haglund, B. and Rosen, M., 2000. Mortality among lone mothersin Sweden: a population study. Lancet 355 (9211), pp.1215–1219.
[23] Gyimah, S. O., 2002. Lagged effect of childhood mortality on reproductivebehaviour in Ghana and Kenya. PSC Discussion Papers Series, [pdf] 16 (3),Article 1. Available at: http://ir.lib.uwo.ca/pscpapers/vol16/iss3/1[Accessed 3 March 2010].
[24] Ha, Y. P., Hurt, L. S., Tawiah–Agyemang, C., Kirkwood, B. R. and Edmond, K.M., 2012. Effect of socioeconomic deprivation and health service utilization onantepartum and intrapartum stillbirth: population cohort study from ruralGhana. PloS ONE, [pdf] 7 (7), 1–8. Available at: http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0039050 [Accessed 12
REFERENCES 103
October 2012 ].
[25] Hajian–Tilaki, K. O., Asnafi, N. and Aliakbarnia–Omrani, F., 2009. The patternsand determinants of birth intervals in multiparous women in Babol, NorthernIran. Southeast Asian Journal of Tropical Medicine and Public Health 40 (4),pp.852–860.
[26] Heckman, J., 1979. Sample selection as a specification error. Econometrica 47 (1),pp.153-161.
[27] Huang L., Sauve, R., Birkett, N., Fergusson, D. and van Walraven, C., 2008.Maternal age and risk of stillbirth: a systematic review. Canadian MedicalAssociation Journal 178 (2), pp.165–172.
[28] Ikamari, L., 1998. Birth intervals and child survival in Kenya. African Journal ofHealth Sciences 5 (1), pp.15–24.
[29] Isvan, N. A., 1991. Productive and reproductive decisions in Turkey: the role ofdomestic bargaining. Journal of Marriage and Family 53 (4), pp.1057-1070.
[30] Karlsen, S., Say, L., Souza, J., Hogue, C. J., Calles, D. L., Gulmezoglu, A. M.and Raine, R., 2011. The relationship between maternal education and mortalityamong women giving birth in health care institutions: Analysis of thecross–sectional WHO global survey on maternal and perinatal health. BMCPublic Health 11 (606).
[31] Kosimbei, G., 2005. Child healthcare seeking behaviour in Kenya. KIPPRADiscussion Paper No. 50. Nairobi: Kenya Institute for Public Policy Researchand Analysis (KIPPRA).
[32] Kramer, M. S., 2003. The epidemiology of adverse pregnancy outcomes: anoverview. Journal of Nutrition 133 (5), pp.1592S–1596S.
[33] Long, J. S., 1997. Regression models for categorical and limited dependentvariables. Sage Publications.
[34] Long, J. S. and Freese, J., 2006. Regression models for categorical dependentvariables using Stata. Second edition. College Station, TX: Stata Press.
[35] Lule, E., Ramana, G. N. V., Oomman, N., Epp, J., Huntington, D. and Rosen, J.E., 2005. Achieving the millennium development goal of improving maternalhealth: determinants, interventions and challenges. Health, Nutrition andPopulation (HNP) Discussion Paper. World Bank.
[36] Luque–Fernadez, M. A., Gutierrez–Garitano, I. and Bueno–Cavanillas, A., 2012.Stillbirth risk by maternal socio–economic status and country of origin: apopulation–based observational study in Spain, 2007 – 2008. European Journal ofPublic Health 22 (4), pp.524–529.
[37] Mace, R. and Sear, R., 1997. Birth interval and the sex of children in a
REFERENCES 104
traditional african population: an evolutionary analysis. Journal of BiosocialScience 29 (4), pp.499–507.
[38] Maitra, P. and Pal, S., 2005. Birth spacing and child survival: comparativeevidence from India and Pakistan. [pdf] Available at:http://www.researchgate.net/publication/23746845_Birth_Spacing_
and_Child_Survival_Comparative_Evidence_from_India_and_Pakistan*
[Accessed 7 November 2012].
[39] Maitra, P. and Pal, S., 2008. Birth spacing, fertility selection and child survival:analysis using a correlated hazard model. Journal of Health Economics 27(3),pp.690–705.
[40] Makepeace, G., 2006. Effects of birth interval on child mortality: evidence from asequential analysis. World Health & Population 8 (2), pp.69–82.
[41] McAlister, C. and Baskett, T. F., 2006. Female education and maternal mortality:a worldwide survey. Journal of Obstetrics and Gynaecology Canada 28 (11),pp.983–990.
[42] McCarthy, J. and Maine, D., 1992. A framework for analyzing the determinantsof maternal mortality. Studies in Family Planning 23 (1), pp.23-33.
[43] Miller, D. A., 2005. Is advanced maternal age an independent risk factor foruteroplacental insufficiency? American Journal of Obstetrics & Gynecology 192(6), pp.1974–1980.
[44] Miller, J. E., 1991. Birth intervals and perinatal health: an investigation of threehypotheses. Family Planning Perspectives 23 (2), pp.62–70.
[45] Millman, R. and Cooksey, E. C., 1987. Birth weight and the effects of birthspacing and breastfeeding on infant mortality. Studies in Family Planning 18 (4),pp.202-212.
[46] Mistry, R., Galal, O. and Lu, M., 2009. Women’s autonomy and pregnancy care inrural India: a contextual analysis. Social Science & Medicine 69 (6), pp.926-933.
[47] Mturi, A. J., 1997. The determinants of birth intervals among non–contraceptingTanzanian women. African Population Studies 12 (2).
[48] Munguti, J. N. and Buluma, R., 2010. Fertility levels, trends, and differentials. In:Kenya National Bureau of Statistics (KNBS) and ICF Macro. KenyaDemographic and Health Survey 2008–09. Calverton, Maryland: KNBS and ICFMacro, pp.47–56.
[49] Murray, M. P., 2006. Avoiding invalid instruments and coping with weakinstruments. Journal of Economic Perspectives 20 (4), pp.111-132.
[50] Mustafa, H. E. and Odimegwu, C., 2008. Socioeconomic determinants of infantmortality in Kenya: analysis of Kenya DHS 2003. Journal of Humanities &
REFERENCES 105
Social Sciences 2 (2), pp.1–16.
[51] Mwabu, G., 2008. Health economics for low-income countries. In: T. P. Schultzand J. Strauss, eds. Handbook of development economics, Volume 4. Amsterdam:Elsevier/North–Holland, pp.3305–3374.
[52] Mwabu, G., 2009. The production of child health in Kenya: a structural model ofbirth weight. Journal of African Economies 18 (2), pp.212-260.
[53] Norton, M., 2005. New evidence on birth spacing: promising findings forimproving newborn, infant, child, and maternal health. International Journal ofGynecology & Obstetrics 89 (Supplement 1), pp.S1-S6.
[54] Olsen, R. J., 1980. A least squares correction for selectivity bias. Econometrica 48(7), pp.1815-1820.
[55] Opiyo, C., 2004. Fertility levels, trends, and differentials. In: Central Bureau ofStatistics (CBS)[Kenya], Ministry of Health (MOH)[Kenya] and ORC Macro.Kenya Demographic and Health Survey 2003. Calverton, Maryland: CBS, MOHand ORC Macro, pp.51–62.
[56] Osborn, J. F., Cattaruzza, M. S. and Spinelli, A., 2000. Risk of spontaneousabortion in Italy, 1978 – 1995, and the effect of maternal age, gravidity, maritalstatus, and education. American Journal of Epidemiology 151 (1), pp.98–105.
[57] Rafalimanana, H. and Westoff, C. F., 2001. Gap between preferred and actualbirth intervals in Sub–Saharan Africa: implications for fertility and child health.DHS Analytical Studies No. 2. [pdf] Calverton, Maryland: ORC Macro. Availableat: http://www.measuredhs.com/pubs/pdf/AS2/AS2.pdf [Accessed 6November 2012].
[58] Raghupathy, S., 1996. Education and the use of maternal health care in Thailand.Social Science & Medicine 43 (4), pp.459-471.
[59] Rasheed, P. and Al–Dabal, B. K., 2007. Birth interval: perceptions and practicesamong urban–based Saudi Arabian women. Eastern Mediterranean HealthJournal 13 (4), pp.881–892.
[60] Razzaque, A., Da Vanzo, J., Rahman, M., Gausia, K., Hale, L., Khan, M. A. andMustafa, A. H. M. G., 2005. Pregnancy spacing and maternal morbidity inMatlab, Bangladesh. International Journal of Gynecology & Obstetrics 89(Supplement 1), pp.S41-S49.
[61] Reddy, U. M., Ko, C. and Willinger, M., 2006. Maternal age and the risk ofstillbirth throughout pregnancy in the United States. American Journal ofObstetrics & Gynecology 195(3), pp.764–770.
[62] Rosenberg, T. J., Garbers, S., Lipkind, H. and Chiasson, M. A., 2005. Maternalobseity and diabetes as risk factors for adverse pregnancy outcomes: differences
REFERENCES 106
among 4 racial/ethnic groups. American Journal of Public Health 95 (9),pp.1545–1551.
[63] Rosenzweig, M. R. and Schultz, T. P., 1982. The behaviour of mothers as inputsto child health: the determinants of birth weight, gestation, and rate of fetalgrowth. In: V. R. Fuchs, ed. Economic aspects of health. [pdf] Chicago:University of Chicago Press, pp.53–92. Available at:http://www.nber.org/chapters/c6545.pdf [Accessed 6 August 2008].
[64] Rosenzweig, M. R. and Schultz, T. P., 1983. Estimating a household productionfunction: heterogeneity, the demand for health inputs, and their effects on birthweight. Journal of Political Economy 91 (5), pp.723-746.
[65] Rosenzweig, M. R. and Schultz, T. P., 1987. Fertility and investments in humancapital : estimates of the consequences of imperfect fertility control in Malaysia.Journal of Econometrics 36 (1 – 2), pp.163–184.
[66] Rutstein, S. O., 2005. Effects of preceding birth intervals on neonatal, infant andunder–five years mortality and nutritional status in developing countries:evidence from the demographic and health surveys. International Journal ofGynecology & Obstetrics 89 (Supplement 1), pp.S7–S24.
[67] Scannapieco, F. A., Bush, R. B. and Paju, S., 2003. Periodontal disease as a riskfactor for adverse pregnancy outcomes: a systematic review. Annals ofPeriodontology 8 (1), pp.70–78.
[68] Schultz, T. P., 1984. Studying the impact of household economic and communityvariables on child mortality. Population and Development Review 10(Supplement), pp.215-235.
[69] Setty–Venugopal, V. and Upadhyay, U. D., 2002. Birth spacing: three to fivesaves lives. Population Reports, [pdf] 30 (3), pp.1–23. Available at:http://www.k4health.org/sites/default/files/l13.pdf [Accessed 3 March2010].
[70] Singh, S. N., Singh, S. N. and Narendra, R. K., 2010. Demographic andsocio–economic determinants of birth interval dynamics in Manipur: a survivalanalysis. Online Journal of Health and Allied Sciences, [pdf] 9 (4), pp.1–5.Available at: http://cogprints.org/7245/1/2010-4-3.pdf [Accessed 4March 2012].
[71] Stamilio, D. M., DeFranco, E., Pare, E., Odibo, A. O., Peipert, J. F., Allsworth,J. E., Stevens, E. and Macones, G. A., 2007. Short interpregnancy interval: riskof uterine rupture and complications of vaginal birth after cesarean delivery.Obstetrics and Gynecology 110 (5), pp.1075–1082.
[72] StataCorp., 2009. Stata Statistical Software, Version 11, Special Edition. CollegeStation, TX: Stata Corporation.
REFERENCES 107
[73] Stephansson, O., Dickman, P. W. and Cnattingius, S., 2003. The influence ofinterpregnancy interval on the subsequent risk of stillbirth and early neonataldeath. Obstetrics and Gynecology 102 (1), pp.101–108.
[74] Suwal, J. V., 2001. Socio–cultural dynamics of birth intervals in Nepal.Contributions to Nepalese Studies 28 (1), pp.11-33.
[75] Thaddeus, S. and Maine, D., 1994. Too far to walk: maternal mortality incontext. Social Science & Medicine 38 (8), pp.1091–1110.
[76] Terza, J. V., Basu, A. and Rathouz, P. J., 2008. Two-stage residual inclusionestimation: addressing endogeneity in health econometric modelling. Journal ofHealth Economics 27 (3), pp.531-543.
[77] Trussell, J., Martin, L. G., Feldman, R., Palmore, J. A., Concepcion, M., Laily, D.and Bakar, D. A., 1985. Determinants of birth-interval length in the Philippines,Malaysia, and Indonesia: a hazard-model analysis. Demography 22 (2),pp.145-168.
[78] Udjo, E. O., 1997. The effect of child survival on fertility in Zimbabwe: amicro–macro level analysis. Journal of Tropical Pediatrics 43 (5), pp.255–266.
[79] Upadhyay, U. D. and Hindin, M. J., 2005. Do higher status and moreautonomous women have longer birth intervals? Results from Cebu, Philippines.Social Science & Medicine 60 (11), pp.2641-2655.
[80] Vella, F., 1998. Estimating models with sample selection bias: a survey. Journalof Human Resources 33 (1), pp.127–169.
[81] Villamor, E. and Cnattingius, S., 2006. Interpregnancy weight change and risk ofadverse pregnancy outcomes: a population–based study. Lancet 368 (9542),pp.1164–1170.
[82] Walsh, R., 1994. Effects of maternal smoking on adverse pregnancy outcomes:examination of the criteria of causation. Human Biology 66 (6), pp.1059–1092.
[83] Whitworth, A. and Stephenson, R., 2002. Birth spacing, sibling rivalry and childmortality in India. Social Science & Medicine 55 (12), pp.2107-2119.
[84] Williams, E. K., Hossain, M. B., Sharma, R. K., Kumar, V., Pandey, C. M. andBaqui, A. H., 2008. Birth interval and risk of stillbirth or neonatal death:findings from rural India. Journal of Tropical Pediatrics 54 (5), pp.321–327.
[85] World Health Organization (WHO)., 2006. Report of a WHO technicalconsultation on birth spacing. [pdf] Geneva: World Health Organization.Available at: http://cdrwww.who.int/maternal_child_adolescent/documents/birth_spacing.pdf [Accessed 3 January 2010].
[86] World Health Organization (WHO), 2012. World health statistics 2012. [pdf]Geneva: World Health Organization. Available at: http://www.who.int/gho/
REFERENCES 108
publications/world_health_statistics/EN_WHS2012_Full.pdf [Accessed 6June 2012].
[87] Yeakey, M. P., Muntifering, C. J., Ramachandran, D. V., Myint, Y., Creanga, A.A. and Tsui, A. O., 2009. How contraceptive use affects birth intervals: results ofa literature review. Studies in Family Planning 40 (3), pp.205–214.
[88] Zegers–Hochschild, F., Adamson, G. D., de Mouzon, J., Ishihara, O., Mansour,R., Nygren, K., Sullivan, E. and Vanderpoel, S., 2009. International committeefor monitoring assisted reproductive technology (ICMART) and the WorldHealth Organization revised glossary of ART terminology. Fertility and Sterility92 (5), pp.1520–1524.
[89] Zenger, E., 1993. Siblings’ neonatal mortality risks and birth spacing inBangladesh. Demography 30 (3), pp.477–488.
[90] Zohoori, N. and Savitz, D. A., 1997. Econometric approaches to epidemiologicdata: relating endogeneity and unobserved heterogeneity to confounding. Annalsof Epidemiology 7 (4), pp.251–257.
Chapter 4
Smoking and General Health
4.1 Introduction
As discussed in Chapter 1, one of the indicators of general health status at the individual
level is self–reported health status. In studies where the health status of individuals
is measured by their own self–assessments, the individuals are typically asked to rate
their own health, relative to some reference group (e.g., peers), in one of several ordinal
categories (Krause and Jay, 1994).1 Several empirical studies of individual health have
used self-assessed health status as the indicator for health (see for example, Damian,
Ruigomez, Pastor and Martin–Moreno, 1999; Zimmer, Natividad, Lin and Chayovan,
2000; Breidablik, Meland and Lydersen, 2008; Kyobutungi, Ezeh, Zulu and Falkingham,
2009; Yen, Shaw and Yuan, 2010; Kodzi, Gyimah, Emina and Ezeh, 2011; Giron, 2012;
Hanibuchi, Nakaya and Murata, 2012).
The main weakness of self–rated health measures is that even though they may reflect
the individual’s actual health status, they could be dependent on the specific individual
perceptions concerning health (e.g., whether health concerns physical functioning, suf-
fering from particular health problems or practising certain health behaviours) and the
information available to the individual concerning health (Krause and Jay, 1994; Strauss
and Thomas 1998, 2008; Bailis, Segall and Chipperfield, 2003).
Despite this weakness, self-rated health status has been shown to be a predictor
of mortality, morbidity, and disability (see, for example, Idler and Benyamini, 1997;
Burstrom and Fredlund, 2001; Mansson and Rastam, 2001; Osler et al., 2001; Benyamini,
Blumstein, Lusky and Modan, 2003; Ahmad, Jafa and Chaturvedi, 2005; Ford, Spallek
1The frequently used categories are: poor, fair, good, and excellent (Krause and Jay, 1994).
109
Smoking and General Health 110
and Dobson, 2008; Jylha, 2009; McFadden et al., 2009). Other studies (see, for example,
Delpierre et al., 2009) show that the self–reported health status is closer to the true
health status amongst individuals who are highly educated. In this chapter, we use
self–rated health status as an indicator of individual general health.
Active smoking is harmful to health because it is associated with several diseases such
as cancers, cardiovascular diseases, and respiratory diseases (USDHHS, 2004; USDHHS,
2010). Research (see, for example, Strachan and Cook, 1997; Chen, Dales, Krewski and
Breithaupt, 1999; Adab, McGhee, Hedley and Lam, 2005) shows that active smoking is
generally associated with poor health status.
Involuntary or passive smoking2 also has several health implications such as its
association with premature deaths and disease among children and non–smoking adults
(USDHHS, 2006). Other effects of passive smoking include reduced health–related quality
of life (see, for example, Bridevaux et al., 2007) and increased risk of cognitive impairment
in non–smoking women (see, for example, Chen et al., 2012).
Table 4.1 on page 111 summarizes some key facts about tobacco use and its conse-
quences in Kenya. The table shows that about 3% of the male deaths in 2004 can be
attributed to tobacco. The table further shows that about 23% of adult males smoke
while under 1% of adult females smoke. Smoking in Kenya is, therefore, mainly a male
phenomenon. The table further shows that about 11% of the boys aged between 13 and
15 years smoke and about 5% of the girls aged between 13 and 15 years smoke.
The rest of the chapter is organized as follows: Section 4.2 presents a review of
the literature while Section 4.3 discusses the purpose and objectives of the study. The
methodology is discussed in Section 4.4 while Section 4.5 presents some descriptive
statistics. The results are presented and discussed in Section 4.6 while Section 4.7 gives
the summary, conclusions and policy implications.
4.2 Literature Review
Factors influencing an individual’s rating of his or her own health include socio–demographic
factors (such as age, gender, and marital status), socio–economic factors (such as educa-
tion and income), psychosocial factors, social capital (measured by interpersonal trust
2Involuntary or passive smoking refers to the exposure of non–smokers to tobacco combustion productsin the indoor environment (USDHHS, 2006).
Smoking and General Health 111
Table 4.1: Some Facts about Tobacco Use and its Consequences in Kenya
Indicator Value
Tobacco use consequences
Estimated male deaths due to tobacco (2004) 2.9%
Tobacco use
Male cigarette use (age 18 years and above, 2009) 22.5%
Female cigarette use (age 18 years and above, 2009) < 1%
Male smoking any tobacco product (age 18 years and above, 2009) 25.5%
Female smoking any tobacco product (age 18 years and above, 2009) 1.5%
Boys’ current cigarette use (ages 13 – 15) 11.2%
Girls’ current cigarette use (ages 13 – 15) 5.2%
Cigarettes per person 144
Youth exposed to secondhand smoke in homes (ages 13 – 15) 24.7%
Adult smokeless tobacco use 1.7%
Smoking prevalence of health professional students 9.8%
Source: Eriksen, Mackay and Ross (2012).
and social participation), health care access, and lifestyle factors (such as, smoking) (Cott,
Gignac and Bradley, 1999; Kawachi, Kennedy and Glass, 1999; Lima–Costa, Barreto,
Firmo and Uchoa, 2003; Abu-Omar, Rutten and Robine, 2004; Subramanian, Kim and
Kawachi, 2005; Murata et al., 2006; Molarius et al., 2006; Poortinga, 2006; Giordano and
Lindstrom, 2010; Omariba, 2010; Kodzi, Gyimah, Emina and Ezeh, 2011; Prus, 2011;
Delpierre et al., 2012; Giron, 2012; Wang et al., 2012).
Individuals typically choose to actively smoke (Audrian-McGovern et al., 2004; Heikki-
nen, Patja and Jallinoja, 2010). Table 4.2 on page 112 summarizes some of the deter-
minants of smoking. According to the table, smoking is mainly influenced by cigarette
prices, demographic characteristics of the individual, socio–economic characteristics,
psychological factors, addiction, anti–smoking sentiments, and anti–smoking restrictions.
Research shows that those individuals who actively smoke are more likley to rate
their own health as poor compared to non–smokers (see, for example, Cott, Gignac
and Bradley, 1999; Kawachi, Kennedy and Glass, 1999; Birch, Jerret and Eyles, 2000;
Johnson and Richter, 2002; Ho, Lam, Fielding and Janus, 2003; Langhammer et al.,
2003; Kirkland, Greaves and Devichand, 2004; Birch et al., 2005; Nakata et al., 2009;
Smoking and General Health 112
Table
4.2:
Det
erm
inan
tsof
Sm
okin
g
Det
erm
inan
tE
xam
ple
sL
iter
atu
reS
ou
rce
Cig
aret
tep
rice
sC
halo
up
ka,
1999;
Ch
alo
up
kaan
dW
arn
er,
2000;
Ross
an
dC
halo
up
ka,
2003;
Caw
ley,
Mark
owit
zan
dT
au
ras,
2004;
Ch
enan
dX
ing,
2011;
Non
nem
ake
rand
Farr
elly
,2011.
Dem
ogra
ph
icch
arac
teri
stic
sag
e,ge
nd
er,
mar
ital
stat
us,
and
are
aof
resi
den
ceR
ap
tou
,M
att
as
an
dK
atr
akil
idis
,2009.
Fato
ye,
2003.
Soci
oec
onom
icch
arac
teri
stic
socc
up
atio
n,
edu
cati
on,
reli
gion
,and
inco
me
Jon
esan
dK
irig
ia,
1999;
Cow
ell,
2006;
Vir
tan
enet
al,
2007;
Ken
kel,
Lil
lard
an
dL
iu,
2009;
Rap
tou
,M
att
as
an
dK
atr
akil
idis
,2009;
Fra
nce
scon
i,Jen
kin
san
dS
iedle
r,2010;
Kje
llss
on
,G
erd
tham
an
dL
ytt
ken
s,2011.
Psy
chos
oci
alfa
ctor
sfa
mil
ym
emb
ers
wh
osm
oke,
Pow
ell,
Tau
ras
an
dR
oss
,2005;
Van
Loon
,T
ijhu
is,
Su
rtee
san
dO
rmel
,2005.
frie
nd
sor
pee
rsw
ho
smok
e,fe
elin
gst
ress
ed,
Kem
pp
ain
enet
al.
,2006;
ciga
rett
ead
ver
tisi
ng,
infl
uen
ceby
cigare
tte
off
erin
gs,
Kro
snic
ket
al.
,2006;
Naka
jim
a,
2007;
wea
kfa
mil
yb
ond
s,H
arr
isan
dLop
ez-V
alcarc
el,
2008;
bel
iefs
abou
tth
eh
ealt
hri
sks
of
smokin
gR
ap
tou
,M
att
as
an
dK
atr
akil
idis
,2009;
Goh
lman
n,
Sch
mid
tan
dT
au
chm
an
n,
2010;
Lou
reir
o,
San
z–d
e–G
ald
ean
o,
2010;
Fle
tch
er,
2010.
Ad
dic
tion
Yen
an
dJon
es,
1996;
US
DH
HS
,2010.
Anti
–sm
okin
gse
nti
men
tsD
eCic
caet
al.
,2008;
Rap
tou
,M
att
as
an
dK
atr
akil
idis
,2009.
Sm
okin
gre
stri
ctio
ns
DeC
icca
etal.
,2008;
Rap
tou
,M
att
as
an
dK
atr
akil
idis
,2009.
Smoking and General Health 113
Darviri, Artemiadis, Tigani and Alexopoulos, 2011; Giron, 2012; Wang et al., 2012).
Other studies, however, find that heavy smokers are more likely to rate their own health
as excellent (see for example, Yen, Shaw and Yuan, 2010). Further, individuals exposed
to second–hand smoke have poorer self–rated health status compared to those individuals
not exposed to second–hand smoke (see, for example, Nakata et al., 2009).
4.3 Purpose and Objectives of the Study
Previous studies on self–rated health in Kenya (see, for example, Kyobutungi, Ezeh,
Zulu and Falkingham, 2009; Kodzi, Gyimah, Emina and Ezeh, 2011) have not explicitly
examined the effect of smoking on health. The studies mainly examine the relationship
between HIV/AIDS and self–rated health (Kyobutungi, Ezeh, Zulu and Falkingham,
2009), or the relationship between religious involvement and social engagement and self–
rated health (Kodzi, Gyimah, Emina and Ezeh, 2011). Our study, therefore, contributes
to the literature by documenting the effect of smoking on self–rated health in Kenya.
Specifically, the objectives of the study include:
1. Determining the effect of smoking on self-rated health status in Kenya.
2. Determining the factors that influence an individual’s decision on whether or not to
smoke.
3. Based on the study findings, drawing some policy implications.
4.4 Methodology
Conceptual Model
Figure 4.1 on page 114 shows a conceptual framework for analyzing the effect of smoking
on general health. The framework is based on the general framework presented in Figure
1.1, that is in turn based on Schultz (1984).
According to the framework, the general health of an individual (as measured by the
individual’s self–rated health status) is influenced by whether the individual smokes or
not and unobservable biological endowments of the individual. However, an individual’s
smoking status is, in turn, influenced by unobservable biological endowments of the
Smoking and General Health 114
Figure 4.1: A Conceptual Framework for Analyzing the Effect of Smoking on General Health
Unobserved Variables
Individual /Household Demographic, Psychological and Socio – economic characteristics
Smoking
Health Outcome
(Self – Rated Health Status)
Individual/ household preferences
Community Characteristics /Environmental Factors
Individual Health Endowment
Observed Variables
Source: Author’s Own Construction Based on Figure 1.1 and Schultz (1984).
individual; individual’s demographic, socio–economic, and psychosocial characteristics;
and the characteristics of the community or the environment in which the individual
lives.
Theoretical Framework
Following the general modelling framework presented in Chapter 1 that was based on
Rosenzweig and Schultz (1982, 1983) and Mwabu (2009), we assume that a typical
individual derives utility from the consumption of various goods and services that have
no direct effect on his/her health, Y , his/her smoking status (which also affects health
directly), S, and his/her health status, H. The individual’s utility function can, therefore,
Smoking and General Health 115
be written as follows:
U = U (Y, S,H) . (4.1)
The health status of the individual is, in turn, influenced by the individual’s smoking
status, S, other inputs, X, and unobservable biological endowments, µ. The individual’s
health production function can, therefore, be expressed as follows:
H = H (S,X, µ) . (4.2)
The individual faces a budget constraint given by:
I = PyY + PsS + PxX (4.3)
where I is exogenous household income, Py is unit price of Y , Ps is unit price of S, and
Px is unit price of X.
The individual maximizes his or her utility function subject to the health production
function and budget constraint. As shown by Mwabu (2009), this process leads to the
following reduced–form input demand equations.
X = X (Px, Py, Ps, I, µ) (4.4)
Y = Y (Px, Py, Ps, I, µ) (4.5)
S = S (Px, Py, Ps, I, µ) . (4.6)
Following Rosenzweig and Schultz (1983) and Mwabu (2008), we can formulate a
hybrid general health production function by combining Equations (4.2) and (4.4). The
resulting hybrid health production function is given by:
H = H (S, Px, Py, I, µ) . (4.7)
In this equation, S is potentially endogeneous and the inability to observe µ creates
further estimation problems (Rosenzweig and Schultz, 1982; Mwabu, 2008).
Smoking and General Health 116
Estimation Issues
As discussed in Section 1.6 of Chapter 1, we need to worry about potential sample selection
bias, potential endogeneity, and potential unobserved heterogeneity when estimating our
models.
Sample Selection Bias
We are only able to obtain information on self–rated health status of a particular
individual if the individual reports it. However, information on self–rated health status
is missing for some observations in our dataset. If the unobservable factors that influence
the reporting of self–rated health status are correlated with the unobservable factors that
influence the reported self–rated health status, then a selection issue arises and failure
to control for it may bias our results (Vella, 1998). We use the Heckman procedure
(Heckman, 1979) to control for the sample selection bias.
The Heckman procedure typically involves the estimation of two equations: an
equation of primary interest, and the selection equation. An example of an equation of
primary interest is the self–rated health status equation in this chapter. Mwabu (2009)
summarizes the steps involved in the Heckman procedure. We extend this summary and
provide the following steps for implementing the Heckman procedure.
i. Estimate the selection equation using the probit formulation.
ii. Obtain the predicted probit estimates from the estimated equation.
iii. Construct a probability density function for each observation.
iv. Construct a cumulative density function for each observation.
v. Construct the inverse Mill’s ratio by taking the ratio of the probability density
function to the cumulative density function, for each observation.
vi. Include the inverse Mill’s ratio as an additional regressor in the equation of primary
interest.
Smoking and General Health 117
Endogeneity
As observed in Section 4.2 above, people typically choose to start smoking (see, for
example, Audrian-McGovern et al., 2004; Heikkinen, Patja and Jallinoja, 2010). We
therefore expect the variable measuring smoking in our model to be potentially endogenous.
We use the Two–Stage Residual Inclusion (2SRI) method (Terza, Basu and Rathouz,
2008) to control for this potential endogeneity. In line with this approach, we include in
our self–rated health status model the generalized residuals3 from the smoking status
model, as an additional regressor. If the coefficient of these generalized residuals is
statistically significant, then smoking would indeed be endogenous in our model (Bollen,
Guilkey and Mroz, 1995).
Unobserved Heterogeneity
Unobserved heterogeneity will exist in our self–rated health status model if some un-
observable factors interact non–linearly with the smoking status variable causing the
effect of smoking on self–rated health to differ amongst the various individuals in the
population (Mullahy, 1997). We use the control function approach, as applied by Mwabu
(2009), to control for unobserved heterogeneity in our model.
Model Identification
Since we have one endogenous variable in our model and a sample selection rule, we
need two instrumental variables4 for estimating our first–stage models (Mwabu, 2009).
Common variables used as instruments for smoking in the literature include an encour-
agement to stop smoking versus no encouragement (see, for example, Permutt and Hebel,
1989), observed cigarette price per region (see, for example, Leigh and Schembri, 2004),
random assignment to the intervention or control group (see, for example, Eisenberg and
Quinn, 2006), and per–pack excise tax on cigarettes (see, for example, Mullahy, 1997).
We do not have information on the kinds of instruments that have been used in the
literature in our dataset. This calls for innovation on our part for the type of variables
to serve as instruments. We, therefore, use two instruments. The first is the smoking
3Gourieroux, Monfort, Renault and Trognon (1987) provide a good discussion on how to computegeneralized residuals for various non–linear models.
4More information on the application of instrumental variables in health research can be found inMartens et al. (2006).
Smoking and General Health 118
prevalence in each district. Using the information on whether an individual in our data
set smokes or not, we construct a smoking prevalence indicator for each district, which
is simply the proportion of individuals in our dataset from the district who reported
that they smoke. We do not expect any single individual in our dataset to influence the
smoking prevalence of the district in which he or she lives. We expect smoking prevalence
in the district where the individual lives to be correlated with the smoking behaviour of
the individual, but not directly related to the self–rated health status of the individual
(since the individual may not be smoking after all or may not be exposed to second–hand
smoke). We further expect that the higher the smoking prevalence in the district where
the individual lives, the higher the probability that the individual smokes, holding other
factors constant. This is because, a high smoking prevalence rate signals the ease with
which individuals can access cigarettes in the particular district.
Another instrument that we use is the average number of cigarettes smoked in the
province where the individual lives. Using information from the male dataset of the
Kenya Demographic and Health Survey of 2003, we are able to construct a variable of the
average number of cigarettes smoked by individuals in the last 24 hours in each province.
We do not expect any particular individual to influence the average number of cigarettes
smoked in the province in 24 hours. We expect the average number of cigarettes smoked
to be correlated with the individual’s decision to smoke but we do not expect this average
number of cigarettes smoked in the province to influence the individual’s rating of his
or her own health. Consequently, we expect that, holding other variables constant, the
higher the average number of cigarettes smoked in the province where the individual
lives, the higher the probability of the individual smoking. A higher number of cigarettes
smoked in any province also indicates the affordability and accessibility of the cigarettes
in the province.
Empirical Model
We measure general health status using the individual’s self–rated health status. In our
dataset, individuals were asked to rate their own health as compared to age–mates in
one of four categories: poor, satisfactory, good, and very good. Using these ratings, we
Smoking and General Health 119
construct the following observed health status indicator variable:
Hi =
1 if inividual i rates own health as “Poor”,
2 if inividual i rates own health as “Satisfactory”,
3 if inividual i rates own health as “‘Good”,
4 if inividual i rates own health as “Very Good”.
(4.8)
Since this variable is ordinal, the appropriate model for it is the ordered choice
regression model (Long, 1997; Long and Freese, 2006).
Following Long (1997) and Long and Freese (2006), we assume that there is a
continuous latent variable, H∗i , that is related to the observed health status indicator via
the following equation:
Hi =
1 if τ0 = −∞ ≤ H∗i < τ1,
2 if τ1 ≤ H∗i < τ2,
3 if τ2 ≤ H∗i < τ3,
4 if τ3 ≤ H∗i < τ4 =∞,
(4.9)
where τ1, τ2, and τ3 are cutpoints (or thresholds) to be estimated.
The latent variable is in turn related to the various covariates via the equation:
H∗i = β1 + β2S + β3Q+ ε1i (4.10)
where S is the smoking status of individual i, Q is a vector of controls, and ε1 is a
stochastic disturbance term.
The probability of observing a given outcome for given values of the independent
variables is given by:
Pr (Hi = m | S,Q) = Pr (τm−1 ≤ H∗i < τm) (4.11)
where m = 1, 2, 3, 4.
Smoking and General Health 120
Substituting Equation (4.10) into Equation (4.11) and simplifying leads to the following
formula for predicted probabilities of the observed outcomes:
Pr (Hi = m | S,Q) = F (τm − β1 − β2S − β3Q)− F (τm−1 − β1 − β2S − β3Q) , (4.12)
where m = 1, 2, 3, 4, and F is the cumulative distribution function for ε1. If we assume
a standard normal distribution for ε1 we obtain the ordered probit model, which we
estimate in our case.
Since the smoking status of the individual, S, is potentially endogenous, we need to
obtain its generalized residuals to include them as an additional regressor in our health
status model so as to correct for this potential endogeneity. We must, therefore, estimate
a smoking status model. In our dataset, individuals indicate whether or not they smoke
cigarettes or pipe. The observed value of S for a typical individual i is, therefore, given
by:
Si =
1 if individual i smokes,
0 otherwise.(4.13)
We formulate the smoking status model as a binary regression model because smoking
status is a binary variable (Long, 1997; Long and Freese, 2006).
We assume that the observed smoking status indicator for individual i, Si, is linked
to a latent continuous variable, S∗i , via the equation
Si =
1 if S∗i > 0,
0 otherwise.(4.14)
S∗i is, in turn, linked to the covariates by the following equation:
S∗i = α1 + α2Q+ α3R + ε2i (4.15)
where Q is a vector of controls, R a vector of instruments, and ε2 is a stochastic disturbance
term. We assume a standard normal distribution for ε2 and consequently estimate the
probit model. The probit model, in this case, is given by:
Pr (Si = 1) = Φ (α1 + α2Q+ α3R) . (4.16)
Smoking and General Health 121
The generalized residuals from our smoking status model are included as an additional
regressor in our structural model of interest to control for the potential endogeneity of
smoking status. We control for potential unobserved heterogeneity by including the
interaction of smoking participation with its generalized residuals, as an additional
regressor in our model.
Following Heckman (1979), the selection equation for individuals in our sample is
given by:
Ii =
1 if individual i reports self–rated health status,
0 otherwise.(4.17)
Based on this equation we use the procedure described in the subsection on Model
Identification above to construct the inverse Mill’s ratio (IMR) that is then included in
our model of interest as an additional regressor.
To control for potential endogeneity of smoking status, potential unobserved hetero-
geneity, and potential sample selection bias, Equation (4.10) is therefore extended as
follows:
H∗i = β1 + β2Si + β3Q+ β3ε2i + β4IMR + β5Siε2i + ε1i (4.18)
where Si is the smoking status of individual i, Q is a vector of controls, ε2 are generalized
residuals from the smoking status model, IMR is the inverse Mill’s ratio, and ε1 is a
stochastic disturbance term.
Substituting Equation (4.18) in Equation (4.11) and assuming a standard normal
distribution for ε1, we obtain the model that we estimate.5
We estimate all our equations using Stata 11 software (StataCorp, 2009).
Data
The dataset we use comes from a nationally representative household Health Expenditure
and Utilization survey carried out by the then Ministry of Health in 2003. The survey
5The model is given by:
Pr (Hi = m | S,Q) = F (τm − β1 − β2S − β3Q− β3 ε2i − β4IMR− β5Si ε2i)− F(τm−1 − β1 − β2S − β3Q− β3 ε2i − β4IMR− β5Si ε2i
),
where m = 1, 2, 3, 4, and F is the cumulative distribution function for ε1.
Smoking and General Health 122
Table 4.3: Variable Definitions for Smoking and General Health Models
Variable Definition
Self–rated health status Rating of own health compared to age–mates.
1 = Poor, 2 = Satisfactory, 3 = Good,
4 = Very Good.
Report self–rated health status 1 if individual reported self–rated health status,
0 otherwise.
Currently smokes 1 if individual currently smokes, 0 otherwise.
Male 1 if individual is male, 0 otherwise.
Age Individual’s age in completed years.
Primary education 1 if individual’s highest level of education is
primary, 0 otherwise.
Secondary education 1 if individual’s highest level of education is
secondary, 0 otherwise.
University education 1 if individual’s highest level of education is
university, 0 otherwise.
Currently married 1 if individual is currently married, 0 otherwise.
Wealth index 1 = Poorest, 2 = Second, 3 = Middle,
4 = Fourth, 5 = Richest.
Smoking prevalence Percentage of individuals in the sample who smoke.
Measured at the district level.
Average number of cigarettes smoked Average number of cigarettes smoked in
the last 24 hours. Computed at the provincial level.
Inverse Mill’s ratio Inverse Mill’s ratio.
Smoking residuals Generalized residuals from the smoking status model.
Smoking interacted with residuals Interaction of smoking status with smoking residuals.
collected a wide range of information including health related expenditures, self–rated
health status, smoking status, among other variables. In addition to this dataset, we
obtained information on the number of cigarettes smoked in the last 24 hours from the
Kenya Demographic and Health Survey male dataset of 2003. This information was then
used to calculate the average number of cigarettes smoked in the past 24 hours for each
province.
Table 4.3 shows the variable definitions.
Smoking and General Health 123
Table 4.4: Descriptive Statistics for Smoking and General Health Models
Variable Number of Mean Standard Minimum Maximum
Observations Deviation
Self–rated health status 37,765 3.040 0.706 1 4
Report self–rated health status 38,121 0.991 0.096 0 1
Currently smokes 37,400 0.056 0.230 0 1
Male 38,020 0.494 0.5 0 1
Age 38,042 22.742 18.072 0 121
Primary education 37,901 0.471 0.499 0 1
Secondary education 37,901 0.153 0.360 0 1
University education 37,901 0.014 0.118 0 1
Currently married 37,458 0.328 0.469 0 1
Wealth index 38,121 2.734 1.398 1 5
Smoking prevalence 37,647 5.650 3.146 0.91 18.43
Average number of cigarettes smoked 38,121 8.083 2.085 5.02 12.46
4.5 Descriptive Statistics
The descriptive statistics are shown in Table 4.4. This table shows for each variable,
the number of observations, its mean, its standard deviation, its maximum value, and
its minimum value. The number of observations differ across variables since the values
of certain variables are missing for some of the individuals. The table shows that, for
instance, the average smoking prevalence is 5.65%, the minimum smoking prevalence is
0.91%, while the maximum smoking prevalence is 18.43%. Similarly, the table shows
that the mean age in our sample is about 23 years, the youngest individual in the sample
is aged less than a year, while the oldest individual is aged 121 years.
Table 4.5 on page 124 shows the distribution of self–rated health status in our dataset,
while Table 4.6 also on page 124 shows the distribution of self–rated health status by
smoking status in our dataset. From Table 4.5, we can observe that 61.48% of the
individuals in our sample rate their own health as good compared to age–mates. Only
3.85% of the individual rate their own health as poor compared to age–mates. From
Table 4.6, we can observe that majority of the individuals (about 94%) in our data set
do not smoke. We can also further observe that a large majority of the individuals rate
their health as good compared to age–mates, even if the individuals are smokers.
Smoking and General Health 124
Table 4.5: Distribution of Self–Rated Health Status in the Dataset
Self–rated health status Number of observations Percent (%)
Poor 1,453 3.85
Satisfactory 4,343 11.50
Good 23,217 61.48
Very good 8,752 23.17
Table 4.6: Distribution of Self–Rated Health Status by Smoking Status in the Dataset
Self–rated health status Currently Smokes Total
No Yes
Poor 1,298 129 1,427
Satisfactory 3,954 306 4,260
Good 21,694 1,245 22,939
Very good 8,195 411 37,232
Total 35,141 (94.4%) 2,091 (5.6%) 37,232 (100%)
4.6 Results
First–Stage Models
We employ a two–stage estimation strategy. In the first stage, we estimate the sample
selection model and the smoking status model. The results of these estimations are
shown in Table 4.7 on page 125. The second stage involves estimation of the self–rated
health status model.
We note from Table 4.7 that with respect to the smoking status model, our instruments
have the expected signs and are statistically significantly different from zero. We can
also observe that, according to the table, smoking status is influenced by the sex of the
individual, age, education level, marital status and wealth.
Specifically, the table indicates that male individuals have a higher probability of
smoking as compared to female individuals, holding other variables constant. This
finding is supported by the observation in the literature that the smoking behaviour
of men is different from that of women (see, for example, Gilmore et al., 2004; Bauer,
Smoking and General Health 125
Table 4.7: Average Marginal Effects for Sample Selection and Smoking Status Models, RobustZ Statistics in Parentheses
Variable Sample Selection Model Smoking Status Model
(Report self–rated health status =1) (Currently smokes =1)
(1) (2)
Male 0.00078 0.089
(0.82) (32.30)
Age 0.00014 0.002
(3.00) (22.34)
Primary education 0.0083 0.028
(7.08) (9.57)
Secondary education 0.0035 0.030
(2.22) (8.26)
University education 0.0054 0.005
(1.16) (0.60)
Currently married 0.000082 0.036
(0.05) (11.86)
Wealth index -0.00063 -0.004
(-1.69) (-4.05)
Smoking prevalence -0.00022 0.007
(-1.38) (23.48)
Average number of cigarettes smoked 30.002 0.002
(7.51) (3.13)
Number of observations 36,705 36029
Gohlmann and Sinning, 2007) and that males are more likely to smoke than females (see,
for example, Rachiotis et al., 2008; El Ansari and Stock, 2012).
The results also indicate that older individuals are more likely to be smokers than
younger ones, holding other factors constant. This finding contradicts the findings
reported in some of the studies in the literature that show that elderly individuals have
a lower probability of smoking (see, for example, Pomerleau et al., 2004) and also the
findings from studies that show that young women are more likely to be smokers than
older women (see, for example, Lu, Tong and Oldenburg, 2001).
According to the results in the table, individuals whose highest level of education is
primary and those whose highest level of education is secondary have a higher probability
of smoking as compared to those individuals without formal education. This result
contradicts the findings from some studies in the literature that show that individuals
Smoking and General Health 126
with low levels of education have higher probabilities of being smokers (see, for example,
Hosseinpoor, Parker, d’Espaignet and Chatterji, 2011).
The results also indicate that currently married individuals are more likely to smoke
than their counterparts who are not currently in marital relationships, holding other
factors constant. This result is contradicted by findings from some studies in the literature
that show that the probability of smoking is higher among divorced, separated or widowed
individuals (see, for example, Mwenifumbo, Sellers and Tyndale, 2008).
Self–Rated Health Status Models
We estimate the self–rated health status model using ordered probit. We report the aver-
age marginal effects of the probabilities of rating own health as either poor, satisfactory,
good, or very good. The results are presented in Table 4.8 on page 127, Table 4.9 on
page 129, Table 4.10 on page 130, and Table 4.11 on page 131. In each table we show
the baseline model; the model controlling for selection bias; the model controlling for
selection bias and endogeneity of smoking; and the model controlling for selection bias,
endogeneity of smoking, and unobserved heterogeneity. We mainly interpret the results
in Table 4.8 and in Table 4.11, the two extremes.
Table 4.8 on page 127 presents the results of four models. Column (1) of the table
shows the results for the model that does not control for potential sample selection bias,
potential endogeneity of smoking, and potential unobserved heterogeneity. Column (2)
shows the results of the model that controls for potential sample selection bias. Column
(3) presents the results of the model in which both potential sample selection bias and
potential endogeneity of smoking are controlled for. Finally, column (4) shows the results
of the model that controls for potential sample selection bias, potential endogeneity of
smoking, and potential unobserved heterogeneity.
The results in column (2) suggest the presence of selection bias that biases the results
presented in column (1). The results in column (3), on the other hand, indicate that
smoking status is endogenous in the model for self–rated health status. The results
presented in column (4) also indicate the presence of unobserved heterogeneity which
biases the results in models that do not control for it. Consequently, our preferred model
is the model that controls for all the three potential problems, which in our case is
Smoking and General Health 127
Table 4.8: Average Marginal Effects for Probability of Rating Own Health as “Poor”, RobustZ Statistics in Parentheses
Variable Self–rated health = Poor
(1) (2) (3) (4)
Currently smokes 0.009 0.009 0.022 0.018
(3.94) (4.05) (3.04) (2.35)
Male -0.004 -0.003 -0.004 -0.005
(-4.02) (-3.40) (-3.95) (-4.33)
Age 0.001 0.001 0.001 0.001
(22.50) (22.72) (21.25) (20.00)
Primary education -0.010 -0.004 -0.004 -0.004
(-8.80) (-2.82) (-2.98) (-3.10)
Secondary education -0.021 -0.018 -0.018 -0.019
(-12.62) (-10.78) (-10.88) (-10.97)
University education -0.039 -0.035 -0.034 -0.034
(-9.29) (-8.25) (-8.14) (-8.13)
Currently married -0.013 -0.013 -0.013 -0.014
(-9.29) (-9.09) (-9.29) (-9.44)
Wealth index -0.004 -0.004 -0.004 -0.004
(-10.46) (-11.18) (-11.17) (-11.19)
Inverse Mill’s ratio 0.349 0.347 0.347
(7.55) (7.50) (7.50)
Smoking residuals -0.007 -0.015
(-1.89) (-2.53)
Smoking interacted with residuals 0.012
(1.69)
Number of observations 36,338 35,883 35,883 35,883
the model shown in column (4). We, therefore, concentrate on the results presented in
column (4).
According to the results in column (4) of Table 4.8, smoking increases the probability
of rating own health as “Poor” compared to age–mates by 0.018. This means that,
holding other factors constant, the probability of individuals who smoke rating their
own health as “Poor” is higher than that of the individuals who do not smoke rating
their own health as “Poor” by 0.018. Consequently, smoking increases the probability
of rating own health as “Poor” compared to age–mates, holding other factors constant.
This result is consistent with the finding in the literature that individuals who smoke
are more likely to rate their own health as “Poor” (see, for example, Meurer, Layde and
Guse, 2001; Osler et al., 2001; Ahmad, Jafa and Chaturvedi, 2005; Asfar et al., 2007;
Smoking and General Health 128
Lim et al., 2007; Demirchyan and Thompson, 2008). Comparing the effect of smoking on
self–rated health shown in column (1) with that shown in column (4), we notice that the
result appearing in column (4) is twice that appearing in column (1). We can, therefore,
conclude that failure to control for sample selection bias, endogeneity of smoking and
unobserved heterogeneity leads to an understatement of the effect of smoking on the
probability of rating own health as “Poor” compared to age–mates.
The results in column (4) of the table also indicate that males have a lower probability
of rating their own health as “Poor” compared to females. In particular, holding other
factors constant, the probability of males rating their own health as “Poor” is lower than
that of females by about 0.005. This result is consistent with the findings from studies in
the literature that show that women have a higher probability than men of rating their
own health as “Poor” (see, for example, Ahmad, Jafa and Chaturvedi, 2005; Asfar et
al., 2007; Lim et al., 2007; Montazeri, Goshtasebi and Vahdaninia, 2008; Shin, Shin and
Rhee, 2012). The results further indicate that the probability of rating own health as
“Poor” increases with age. This result is consistent with the findings of some studies in
the literature that indicate that older individuals are more likely to rate own health as
“Poor” (see, for example, Asfar et al., 2007; Lim et al., 2007; Shin et al., 2012).
The results also indicate that individuals whose highest level of education is either
primary, secondary, or university are less likely to rate their own health as “Poor”
as compared to individuals without formal education, holding other factors constant.
Although some studies in the literature (see, for example, Lim et al., 2007) find no
relationship between educational attainment and self–rated health status, other studies
(see, for example, Demirchyan and Thompson, 2008; Montazeri et al., 2008) do find such
a relationship and report that highly educated individuals are more likely to perceive
their own health as being worse. We can also observe that the probability of rating own
health as “Poor” is lower for currently married individuals as compared to those who are
not currently married, holding other factors constant. This result contradicts the finding
in the literature that married individuals are more likely to rate own health as “Poor”
(see, for example, Asfar et al., 2007).
As mentioned earlier, Table 4.11 on page 131 shows the results of the models for the
probability of rating own health as “Very Good” compared to age–mates. We show the
average marginal effects. The results in column (1) do not control for potential sample
selection bias, potential endogeneity of smoking, and potential unobserved heterogeneity.
Smoking and General Health 129
Table 4.9: Average Marginal Effects for Probability of Rating Own Health as “Satisfactory”,Robust Z Statistics in Parentheses
Variable Self–rated health = Satisfactory
(1) (2) (3) (4)
Currently smokes 0.017 0.017 0.043 0.035
(3.96) (4.06) (3.04) (2.35)
Male -0.007 -0.006 -0.009 -0.010
(-4.03) (-3.41) (-3.95) (-4.34)
Age 0.002 0.002 0.002 0.002
(25.92) (26.41) (24.33) (22.47)
Primary education -0.018 -0.007 -0.008 -0.008
(-9.03) (-2.83) (-2.99) (-3.11)
Secondary education -0.040 -0.035 -0.035 -0.036
(-13.27) (-11.19) (-11.30) (-11.40)
University education -0.075 -0.067 -0.066 -0.066
(-9.45) (-8.36) (-8.25) (-8.23)
Currently married -0.024 -0.024 -0.025 -0.026
(-9.45) (-9.24) (-9.44) (-9.59)
Wealth index -0.007 -0.008 -0.008 -0.008
(-10.72) (-11.52) (-11.50) (-11.53)
Inverse Mill’s ratio 0.671 0.667 0.667
(7.68) (7.63) (7.62)
Smoking residuals -0.014 -0.029
(-1.89) (-2.53)
Smoking interacted with residuals 0.022
(1.69)
Number of observations 36,338 35,883 35,883 35,883
Those in column (2) only control for potential sample selection bias. The ones in column
(3) control for both potential sample selection bias and potential endogeneity of smoking,
while those in column (4) control for potential sample selection bias, potential endogeneity
of smoking, and potential unobserved heterogeneity.
Looking closely at the results in column (2) of Table 4.11 on page 131 shows that
sample selection bias is present and failure to control for it biases the results. A closer
look at the results in column (3) of the same table also shows that smoking status is
endogenous and this endogeneity should be controlled for. The results in column (4) of
the same table also imply that there is unobserved heterogeneity. Thus, the best model
for our purposes is the one shown in column (4) of the table. We will, therefore, only
interpret the results of this model.
Smoking and General Health 130
Table 4.10: Average Marginal Effects for Probability of Rating Own Health as “Good”, RobustZ Statistics in Parentheses
Variable Self–rated health = Good
(1) (2) (3) (4)
Currently smokes 0.008 0.008 0.020 0.016
(3.91) (4.01) (3.01) (2.33)
Male -0.003 -0.003 -0.004 -0.005
(-3.99) (-3.38) (-3.89) (-4.26)
Age 0.001 0.001 0.001 0.001
(18.86) (19.28) (18.55) (17.74)
Primary education -0.008 -0.003 -0.004 -0.004
(-8.53) (-2.80) (-2.96) (-3.07)
Secondary education -0.018 -0.016 -0.016 -0.017
(-11.81) (-10.23) (-10.31) (-10.38)
University education -0.034 -0.031 -0.031 -0.030
(-8.98) (-8.03) (-7.94) (-7.92)
Currently married -0.011 -0.011 -0.012 -0.012
(-8.86) (-8.69) (-8.83) (-8.95)
Wealth index -0.003 -0.004 -0.004 -0.004
(-10.23) (-10.92) (-10.91) (-10.93)
Inverse Mill’s ratio 0.310 0.308 0.308
(7.49) (7.44) (7.44)
Smoking residuals -0.007 -0.014
(-1.88) (-2.51)
Smoking interacted with residuals 0.010
(1.69)
Number of observations 36,338 35,883 35,883 35,883
The results indicate that individuals who smoke are less likely to rate their own health
as “Very Good” as compared to age–mates, holding other factors constant. Specifically,
the results imply that the probability of individuals who smoke rating their own health
as “Very Good” is lower by 0.069 compared to that of individuals who do not smoke,
holding other factors constant. Consequently, smoking reduces the probability of rating
own health as “Very Good” compared to age–mates, holding other factors constant. This
result is consistent with the finding in the literature that the probability of individuals
who smoke rating their own health as “Very Good” or “Excellent” is very low (see,
for example, Kirkland, Greaves and Devichand, 2004; El Ansari and Stock, 2012). If
we compare the results in column (4) to those in column (1), we notice that even
though in column (1) smoking reduces the probability of rating own health as “Very
Smoking and General Health 131
Table 4.11: Average Marginal Effects for Probability of Rating Own Health as “Very Good”,Robust Z Statistics in Parentheses
Variable Self–rated health = Very good
(1) (2) (3) (4)
Currently smokes -0.033 -0.034 -0.085 -0.069
(-3.96) (-4.07) (-3.04) (-2.35)
Male 0.015 0.013 0.017 0.021
(4.04) (3.41) (3.95) (4.34)
Age -0.004 -0.004 -0.004 -0.004
(-26.64) (-27.25) (-25.01) (-23.04)
Primary education 0.036 0.014 0.015 0.016
(9.03) (2.82) (2.99) (3.10)
Secondary education 0.079 0.0698 0.070 0.071
(13.28) (11.18) (11.29) (11.38)
University education 0.149 0.133 0.131 0.131
(9.51) (8.40) (8.29) (8.27)
Currently married 0.048 0.048 0.050 0.052
(9.47) (9.26) (9.45) (9.61)
Wealth index 0.015 0.016 0.016 0.016
(10.85) (11.67) (11.65) (11.68)
Inverse Mill’s ratio -1.331 -1.322 -1.321
(-7.72) (-7.66) (-7.66)
Smoking residuals 0.028 0.058
(1.89) (2.53)
Smoking interacted with residuals -0.044
(-1.69)
Number of observations 36,338 35,883 35,883 35,883
Good” compared to age–mates by 0.033, that in column (4) shows the corresponding
reduction in probability to be 0.069. This means that in models where sample selection
bias, endogeneity of smoking and unobserved heterogeneity are not controlled for, the
effect of smoking on rating own health as “Very Good” as compared to age–mate is
underestimated.
We can also observe that males are more likely than females to rate their own heath as
“Very Good”, holding other factors constant. This result is consistent with the findings in
the literature (see, for example, McCullough and Laurenceau, 2004). The results also
indicate that rating of own health as “very good” declines with age, holding other factors
constant.
Smoking and General Health 132
We can further observe that those individuals whose highest level of education is
primary, secondary or university are more likely to rate their own health as “Very Good”
as compared to individuals without formal education, holding other factors constant.
These results are consistent with the finding in the literature that individuals with higher
levels of education are more likely to report “Good” or “Very Good” self–rated health
(see, for example, Phillips, Hammock and Blanton, 2005; Von dem Knesebeck and Geyer,
2007).
We can further observe that currently married individuals are more likely than their
non–married counterparts to rate own health as “Very Good” compared to age–mates,
holding other factors constant. The findings from the literature are mixed with some
studies finding no relationship between marital status and self–rated health (see, for
example, Phillips, Hammock and Blanton, 2005) while others finding married individuals,
especially women, to be more likely to rate own health as “Poor” (see, for example, Asfar
et al., 2007).
4.7 Summary, Conclusions and Policy Implications
Summary
This chapter investigates the effect of smoking on general health status in Kenya. General
health status is measured using self–rated health status. Previous studies have shown
that self–rated health status can predict mortality, morbidity, and disability (see, for
example, McFadden et al., 2009). An individual’s rating of own health may, however,
be influenced by the individual’s perceptions about health and the health information
available to the individual (see, for example, Strauss and Thomas 1998, 2008).
The literature shows that self–rated health status is influenced by socio–demographic
factors, socio–economic factors, psychosocial factors, social capital, health care access,
and lifestyle factors (for example, smoking) (see, for example, Delpierre et al., 2012;
Giron, 2012; Wang et al., 2012).
Most studies show that both active and passive smoking are associated with an
increased probability of rating own health as “Poor” (see, for example, Wang et al., 2012)
although there are a few exceptions where those who smoke are reported to rate their
own health as “Excellent” (see, for example, Yen, Shaw and Yuan, 2010).
Smoking and General Health 133
This study fills a gap in the literature created by the failure of previous studies
on self–rated health status in Kenya to explicitly investigate the relationship between
smoking and self–rated health (see, for example, Gyimah, Emina and Ezeh, 2011). Using
household level data collected by the Ministry of Health in 2003, we estimate an ordinal
model linking smoking to rating of own health. The estimation strategy adopted for
the study controls for potential sample selection bias using the approach suggested
by Heckman (1979), potential endogeneity of smoking using the Two–Stage–Residual–
Inclusion (2SRI) method (Terza, Basu and Rathouz, 2008), and potential unobserved
heterogeneity using the control function approach, as applied by Mwabu (2009).
The main finding is that smoking increases the probability of rating own health as
“Poor” but reduces the probability of rating own health as “Very Good”, holding other
factors constant. This result is consistent with the findings reported in the literature
(see, for example, Demirchyan and Thompson, 2008). The study also demonstrates that
when we do not control for sample selection bias, endogeneity of smoking and unobserved
heteorgeneity we end up underestimating the effect of smoking on self–rated health.
Conclusions
We can draw three main conclusions from the findings in this Chapter. First, smoking
is harmful to health, consistent with the findings in the literature. Second, failure to
control for sample selection bias, endogeneity of smoking and unobserved heterogeneity
leads to an understatement of the negative effect of smoking on self–rated health. Third,
the risk factors for smoking include male gender, older ages, having primary education,
having secondary education, low wealth and being currently married. Smoking is also
influenced by availability of cigarettes.
Policy Implications
Our findings point towards policies aimed at discouraging smoking so as to improve
general health. An example of such a policy is restrictions on the availability and
accessibility of cigarettes to high risk groups.
REFERENCES 134
References
[1] Abu-Omar, K., Rutten, A. and Robine, J., 2004. Self-rated health and physicalactivity in the European Union. Social and Preventive Medicine 49 (4),pp.235-242.
[2] Adab, P., McGhee, S. M., Hedley, A. J. and Lam, T. H., 2005. Smoking,respiratory disease and health service utilisation: the paradox. Social Science &Medicine 60 (3), pp.483-490.
[3] Ahmad, K., Jafa, T. H. and Chaturvedi, N., 2005. Self–rated health in Pakistan:results of a national health survey. BMC Public Health 5 (51).
[4] Asfar, T., Ahmad, B., Rastam, S., Mulloli, T. P., Ward, K. D. and Maziak, W.,2007. Self–rated health and its determinants among adults in Syria: a model fromthe Middle East. BMC Public Health 7 (177).
[5] Audrian-McGovern, J., Rodriguez, D., Tercyak, K., Epstein, L. H., Goldman, P.and Wileyto, E. P., 2004. Applying a behavioural economic framework tounderstanding adolescent smoking. Psychology of Addictive Behaviors 18 (1),pp.64-73.
[6] Bailis, D. S., Segall, A. and Chipperfield, J. G., 2003. Two views of self-ratedgeneral health status. Social Science & Medicine 56 (2), pp.203-217.
[7] Bauer T., Gohlmann, S. and Sinning, M., 2007. Gender differences in smokingbehavior. Health Economics 16 (9), pp.895–909.
[8] Benyamini, Y., Blumstein, T., Lusky, A. and Modan, B., 2003. Genderdifferences in the self–rated health – mortality association: is it poor self–ratedhealth that predicts mortality or excellent self–rated health that predictssurvival? Gerontologist 43 (3), pp.396–405.
[9] Birch, S., Jerrett, M. and Eyles, J., 2000. Heterogeneity in the determinants ofhealth and illness: the example of socioeconomic status and smoking. SocialScience & Medicine 51 (2), pp.307-317.
[10] Birch, S., Jerrett, M., Wilson, K., Law, M., Elliot, S. and Eyles, J., 2005.Heterogeneities in the production of health: smoking, health status and place.Health Policy 72 (3), pp.301-310.
[11] Bollen, K. A., Guilkey, D. K. and Mroz, T. A., 1995. Binary outcomes andendogenous explanatory variables: tests and solutions with an application to thedemand for contraceptive use in Tunisia. Demography 32 (1), pp.111-131.
[12] Breidablik, H., Meland, E. and Lydersen, S., 2008. Self–rated health inadolescence: a multifactorial composite. Scandinavian Journal of Public Health36 (1), pp.12–20.
REFERENCES 135
[13] Bridevaux, P., Cornuz, J., Gaspoz, J., Burnard, B., Ackermann–Liebrich, U.,Schindler, C., Leuenberger, P., Rochat, T. and Gerbase, M. W., 2007.Secondhand smoke and health–related quality of life in never smokers: resultsfrom the SAPALDIA cohort study 2. Archives of Internal Medicine 167 (22),pp.2516–2523.
[14] Burstrom, B. and Fredlund, P., 2001. Self rated health: is it as good a predictorof subsequent mortality among adults in lower as well as in higher social classes?Journal of Epidemiology & Community Health 55 (11), pp.836–840.
[15] Cawley, J., Markowitz, S. and Tauras, J., 2004. Lighting up and slimming down:the effects of body weight and cigarette prices on adolescent smoking initiation.Journal of Health Economics 23 (2), pp.293-311.
[16] Chaloupka, F. J., 1999. Macro–social influences: the effects of prices andtobacco–control policies on the demand for tobacco products. Nicotine & TobaccoResearch 1 (Supplement 1), pp.S105–S109.
[17] Chaloupka, F. J. and Warner, K. E., 2000. The economics of smoking. In: A. J.Culyer and J. P. Newhouse, eds. Handbook of health economics, Volume 1B.Amsterdam: Elsevier/North-Holland, pp.1539–1627.
[18] Chen, Y., Dales, R., Krewski, D. and Breithaupt, K., 1999. Increased effects ofsmoking and obesity on asthma among female Canadians: the nationalpopulation survey, 1994-1995. American Journal of Epidemiology 150 (3),pp.255-262.
[19] Chen, Y. and Xing, W., 2011. Quantity, quality, and regional price variation ofcigarettes: demand analysis based on household survey in China. ChinaEconomic Review 22 (2), pp.221-232.
[20] Chen, R., Zhang, D., Chen, Y., Hu, Z. and Wilson, K., 2012. Passive smokingand risk of cognitive impairment in women who never smoke. Archives of InternalMedicine 172 (3), pp.217–273.
[21] Cott, C. A., Gignac, M. A. M. and Badley, E. M., 1999. Determinants of selfrated health for Canadians with chronic disease and disability. Journal ofEpidemiology & Community Health 53 (11), pp.731–736.
[22] Cowell, A. J., 2006. The relationship between education and health behaviour:some empirical evidence. Health Economics 15 (2), pp.125-146.
[23] Damian, J., Ruigomez, A., Pastor, V. and Martin-Moreno, J. M., 1999.Determinants of self assessed health among Spanish older people living at home.Journal of Epidemiology & Community Health 53 (7), pp.412-416.
[24] Darviri, C., Artemiadis, A. K., Tigani, X. and Alexopoulos, E. C., 2011. Lifestyleand self–rated health: a cross–sectional study of 3,601 citizens of Athens, Greece.BMC Public Health 11.
REFERENCES 136
[25] Decicca, P., Kenkel, D., Mathios, A., Shin, Y. and Lim, J., 2008. Youth smoking,cigarette prices, and anti-smoking sentiment. Health Economics 17 (6),pp.733-749.
[26] Delpierre, C., Lauwers–Cances, V., Datta, G. D., Lang, T. and Berkman, L.,2009. Using self–rated health for analysing social inequalities in health: a risk forunderestimating the gap between socioeconomic groups? Journal of Epidemiology& Community Health 63 (6), pp.426–432.
[27] Delpierre, C., Datta, G. D., Kelyy–Irving, M., Lauwers–Cances, V., Berkman, L.and Lang, T., 2012. What role does socio–economic position play in the linkbetween functional limitations and self–rated health: France vs. USA? EuropeanJournal of Public Health 22 (3), pp.317–321.
[28] Demirchyan, A. and Thompson, M. E., 2008. Determinants of self–rated health inwomen: a population–based study in Armavir Marz, Armenia, 2001 & 2004.International Journal for Equity in Health 7 (25).
[29] Eisenberg, D. and Quinn, B. C., 2006. Estimating the effect of smoking cessationon weight gain: an instrumental variable approach. Health Services Research 41(6), pp.2255–2266.
[30] El Ansari, W. and Stock, C., 2012. Factors associated with smoking, quitattempts and attitudes towards total smoking bans at university: a survey ofseven universities in England, Wales, and Northern Ireland. Asian PacificJournal of Cancer Prevention 13 (2), pp.705–714.
[31] Eriksen, M., Mackay, J. and Ross, H., 2012. The tobacco atlas, Fourth Edition.American Cancer Society.
[32] Fatoye, F. O., 2003. Psychological correlates of substance use amongst secondaryschool students in south western Nigeria. East African Medical Journal 80 (3),pp.154–158.
[33] Fletcher, J. M., 2010. Social interactions and smoking: evidence using multiplestudent cohorts, instrumental variables, and school fixed effects. HealthEconomics 19 (4), pp.466–484.
[34] Ford, J., Spallek, M. and Dobson, A., 2008. Self-rated health and a healthylifestyle are the most important predictors of survival in elderly women. Age andAging 37 (2), pp.194-200.
[35] Francesconi, M., Jenkins, S. P. and Siedler, T., 2010. The effect of lonemotherhood on the smoking behaviour of young adults. Health Economics 19(11), pp.1377-1384.
[36] Gilmore, A., Pomerleau, J., Mckee, M., Rose, R., Haerpfer, C. W., Rotman, D.and Tumanov, S., 2004. Prevalence of smoking in 8 countries of the former SovietUnion: results from the living conditions , lifestyles and health study. American
REFERENCES 137
Journal of Public Health 94 (12), pp.2177–2187.
[37] Giordano, G. N. and Lindstrom, M., 2010. The impact of changes in differentaspects of social capital and material conditions on self-rated health over time: alongitudinal cohort study. Social Science & Medicine 70 (5), pp.700-710.
[38] Giron, P., 2012. Determinants of self–rated health in Spain. European Journal ofPublic Health 22 (1), pp.36–40.
[39] Gohlmann, S., Schmidt, C. M. and Tauchmann, H., 2010. Smoking initiation inGermany: the role of intergenerational transmission. Health Economics 19 (2),pp.227-242.
[40] Gourieroux, C., Monfort, A., Renault, E. and Trognon, A., 1987. Generalizedresiduals. Journal of Econometrics 34 (1 – 2), pp.5–32.
[41] Hanibuchi, T., Nakaya, T. and Murata, C., 2012. Socio–economic status andself–rated health in East Asia: a comparison of China, Japan, South Korea andTaiwan. European Journal of Public Health 22 (1), pp.47–52.
[42] Harris, J. E. and Lopez-Valcarcel, B. G., 2008. Asymmetric peer effects in theanalysis of cigarette smoking among young people in the United States,1992-1999. Journal of Health Economics 27 (2), pp.249-264.
[43] Heckman, J., 1979. Sample selection as a specification error. Econometrica 47 (1),pp.153-161.
[44] Heikkinen, H., Patja, K. and Jallinoja, P., 2010. Smokers’ accounts on the healthrisks of smoking: why is smoking not dangerous for me? Social Science &Medicine 71 (5), pp.877-883.
[45] Ho, S. Y., Lam, T. H., Fielding, R. and Janus, E. D., 2003. Smoking andperceived health in Hong Kong Chinese. Social Science & Medicine 57 (9),pp.1761-1770.
[46] Hosseinpoor, A. R., Parker, L. A., Tursan d’Espaignet, E. and Chatterji, S., 2011.Social determinants of smoking in low– and middle–income countries: resultsfrom the world health survey. PLoS ONE, [pdf] 6 (5), pp. 1–7. Available at:http://www.plosone.org/article/info:
doi%2F10.1371%2Fjournal.pone.0020331 [Accessed 10 March 2012].
[47] Idler, E. L. and Benyamini, Y., 1997. Self-rated health and mortality: a review oftwenty-seven community studies. Journal of Health and Social Behaviour 38 (1),pp.21-37.
[48] Johnson, P. B. and Richter, L., 2002. The relationship between smoking, drinking,and adolescents’ self-perceived health and frequency of hospitalization: analysesfrom the 1997 National Household Survey on Drug Abuse. Journal of AdolescentHealth 30 (3), pp.175-183.
REFERENCES 138
[49] Jones, A. M. and Kirigia, J. M., 1999. Health knowledge and smoking amongSouth African women. Health Economics 8 (2), pp.165-169.
[50] Jylha, M., 2009. What is self-rated health and why does it predict mortality?Towards a unified conceptual model. Social Science & Medicine 69 (3),pp.307-316.
[51] Kawachi I., Kennedy, B. P. and Glass, R., 1999. Social capital and self–ratedhealth: a contextual analysis. American Journal of Public Health 89 (8),pp.1187–1193.
[52] Kemppainen, U., Tossavainen, K., Vartiainen, E., Jokela, V., Puska, P.,Pantelejev, V. and Uhanov, M., 2006. Environmental factors as predictors ofsmoking among ninth–grade adolescents in Pitkaranta (Russian Karelia).Research in Nursing & Health 29 (6), pp.543-555.
[53] Kenkel, D., Lillard, D. R. and Liu, F., 2009. An analysis of life–course smokingbehaviour in China. Health Economics 18 (Supplement 2), pp.S147-S156.
[54] Kirkland, S., Greaves, L. and Devichand, P., 2004. Gender differences in smokingand self reported indicators of health. BMC Women’s Health 4 (Suppl 1),pp.S1–S7.
[55] Kjellsson, G., Gerdtham, U. and Lyttkens, C. H., 2011. Breaking bad habits byeducation – smoking dynamics among Swedish women. Health Economics 20 (7),pp.876-881.
[56] Kodzi, I. A., Gyimah, S. O., Emina, J. and Ezeh, A. C., 2011. Religiousinvolvement, social engagement, and subjective health status of older residents ofinformal neighborhoods of Nairobi. Journal of Urban Health 88 (Suppl 2),pp.370–380.
[57] Krause, N. M. and Jay, G. M., 1994. What do global self-rated health itemsmeasure? Medical Care 32 (9), pp.930-942.
[58] Krosnick, J. A., Chang, L., Sherman, S. J., Chassin, L. and Presson, C., 2006.The effects of beliefs about the health consequences of cigarette smoking onsmoking onset. Journal of Communication 56 (Supplement 1), pp.s18-s37.
[59] Kyobutungi, C., Ezeh, A. C., Zulu, E. and Falkingham, J., 2009. HIV/AIDS andthe health of older people in the slums of Nairobi, Kenya: results from a crosssectional survey. BMC Public Health 9 (153).
[60] Langhammer, A., Johnsen, R., Gulsvik, A., Holmen, T. L. and Bjermer, L., 2003.Sex differences in lung vulnerability to tobacco smoking. European RespiratoryJournal 21 (6), pp.1017–1023.
[61] Leigh, J. P. and Schembri, M., 2004. Instrumental variables technique: cigaretteprice provided better estimate of effects of smoking on SF–12. Journal of Clinical
REFERENCES 139
Epidemiology 57 (3), pp.284–293.
[62] Lim, W., Ma, S., Heng, D., Bhalla, V. and Chew, S. K., 2007. Gender, ethnicity,health behaviour & self–rated health in Singapore. BMC Public Health 7 (184).
[63] Lima–Costa, M. F., Barreto, S. M., Firmo, J. O. A. and Uchoa, E., 2003.Socioeconomic position and health in a population of Brazilian elderly: theBambui health and aging study (BHAS). Pan American Journal of Public Health13 (6), pp.387–394.
[64] Long, J. S., 1997. Regression models for categorical and limited dependentvariables. Sage Publications.
[65] Long, J. S. and Freese, J., 2006. Regression models for categorical dependentvariables using Stata. Second edition. College Station, TX: Stata Press.
[66] Loureiro, M. L., Sanz-de-Galdeano, A. and Vuri, D., 2010. Smoking habits: likefather, like son, like mother, like daughter? Oxford Bulletin of Economics andStatistics 72 (6) pp.717-743.
[67] Lu, Y., Tong, S. and Oldenburg, B., 2001. Determinants of smoking and cessationduring and after pregnancy. Health Promotion International 16 (4), pp.355–365.
[68] Mansson, N. O. and Rastam, L., 2001. Self-rated health as a predictor ofdisability pension and death – a prospective study of middle-aged men.Scandinavian Journal of Public Health 29 (2), pp.151-158.
[69] Martens, E. P., Pestman, W. R., de Boer, A., Belitser, S. V. and Klungel, O. H.,2006. Instrumental variables: application and limitations. Epidemiology 17 (3),pp.260–267.
[70] McCullough, M. E. and Laurenceau, J., 2004. Gender and the natural history ofself–rated health: a 59–year longitudinal study. Health Psychology 23 (6),pp.651–655.
[71] McFadden, E., Luben, R., Bingham, S., Warenham, N., Kinmonth, A. L. andShaw, K. T., 2009. Does the association between self-rated health and mortalityvary by social class? Social Science & Medicine 68 (2), pp.275-280.
[72] Meurer, L. N., Layde, P. M. and Guse, C. E., 2001. Self–rated health status: anew vital sign for primary care? Wisconsin Medical Journal 100 (7), pp.35–39.
[73] Molarius, A., Berglund, K., Eriksson, C., Lamber, M., Nordstrom, E., Eriksson,H. G. and Feldman, I., 2006. Socioeconomic conditions, lifestyle factors, andself–rated health among men and women in Sweden. European Journal of PublicHealth 17 (2), pp.125-133.
[74] Montazeri, A., Goshtasebi, A. and Vahdaninia, M., 2008. Educational inequalitiesin self–reported health in a general Iranian population. BMC Research Notes1(50).
REFERENCES 140
[75] Mullahy, J., 1997. Instrumental–variable estimation of count data models:applications to models of cigarette smoking behavior. Review of Economics andStatistics 79 (4), pp.586–593.
[76] Murata, C., Kondo, T., Tamakoshi, K., Yatsuya, H. and Toyoshima, H., 2006.Determinants of self-rated health: could health status explain the associationbetween self-rated health and mortality? Archives of Gerontology and Geriatrics43 (3), pp.369-380.
[77] Mwabu, G., 2008. Health economics for low-income countries. In: T. P. Schultzand J. Strauss, eds. Handbook of development economics, Volume 4. Amsterdam:Elsevier/North–Holland, pp.3305–3374.
[78] Mwabu, G., 2009. The production of child health in Kenya: a structural model ofbirth weight. Journal of African Economies 18 (2), pp.212–260.
[79] Mwenifumbo, J. C., Sellers, E. M. and Tyndale, R. F., 2008. Socioeconomic anddrug use determinants of smoking status in an urban adult population of blackAfrican descent. Nicotine & Tobacco Research 10 (8), pp.1319–1325.
[80] Nakajima, R., 2007. Measuring peer effects on youth smoking behaviour. Reviewof Economic Studies 74 (3), pp.897-935.
[81] Nakata, A., Takahashi, M., Swanson, N. G., Ikeda, T. and Hojou, M., 2009.Active cigarette smoking, secondhand smoke exposure at work and at home, andself–rated health. Public Health 123 (10), pp.650–656.
[82] Nonnemaker, J. M. and Farrelly, M. C., 2011. Smoking initiation among youth:the role of cigarette excise taxes and prices by race/ethnicity and gender. Journalof Health Economics 30 (3), pp.560-567.
[83] Omariba, D. W. R., 2010. Neighbourhood characteristics, individual attributesand self-rated health among older Canadians. Health & Place 16 (5), pp.986-995.
[84] Osler, M., Heitmann, B. L., Høidrup, S., Jørgensen, L. M. and Schroll, M., 2001.Food intake patterns, self rated health and mortality in Danish men and women.a prospective observational study. Journal of Epidemiology & Community Health55 (6), pp.399–403.
[85] Permutt, T. and Hebel, R., 1989. Simultaneous–equation estimation in a clinicaltrial of the effect of smoking on birth weight. Biometrics 45 (2), pp.619–622.
[86] Phillips, L. J., Hammock, R. L. and Blanton, J. M., 2005. Predictors of self–ratedhealth status among Texas residents. Preventing Chronic Disease, [pdf] 2 (4),pp.1–10. Available at:http://www.cdc.gov/pcd/issues/2005/oct/pdf/04_0147.pdf [Accessed 12March 2012].
[87] Pomerleau, J., Gilmore, A., McKee, M., Rose, R. and Haerpfer, C. W., 2004.
REFERENCES 141
Determinants of smoking in eight countries of the former Soviet Union: resultsfrom the living conditions, lifestyles and health study. Addiction 99 (12),pp.1577–1585.
[88] Poortinga, W., 2006. Do health behaviors mediate the association between socialcapital and health? Preventive Medicine 43 (6), pp.488–493.
[89] Powell, L. M., Tauras, J. A. and Ross, H., 2005. The importance of peer effects,cigarette prices and tobacco control policies for youth smoking behaviour.Journal of Health Economics 24 (5), pp.950-968.
[90] Prus, S. G., 2011. Comparing social determinants of self-rated health across theUnited States and Canada. Social Science & Medicine 73 (1), pp.50–59.
[91] Rachiotis, G., Muula, A. S., Rudatsikira, E., Siziya, S., Kyrlesi, A.,Gourgoulianis, K. and Hadjichristodoulou, C., 2008. Factors associated withadolescent cigarette smoking in Greece: results from a cross sectional study(GYTS study). BMC Public Health 8 (313).
[92] Raptou, E., Mattas, K. and Katrakilidis, C., 2009. Investigating smoker’s profile:the role of psychosocial characteristics and the effectiveness of tobacco policytools. American Journal of Economics and Sociology 68 (2), pp.603-638.
[93] Rosenzweig, M. R. and Schultz, T. P., 1982. The behaviour of mothers as inputsto child health: the determinants of birth weight, gestation, and rate of fetalgrowth. In: V. R. Fuchs, ed. Economic aspects of health. [pdf] Chicago:University of Chicago Press, pp.53–92. Available at:http://www.nber.org/chapters/c6545.pdf [Accessed 6 August 2008].
[94] Rosenzweig, M. R. and Schultz, T. P., 1983. Estimating a household productionfunction: heterogeneity, the demand for health inputs, and their effects on birthweight. Journal of Political Economy 91 (5), pp.723-746.
[95] Ross, H. and Chaloupka, F. J., 2003. The effect of cigarette prices on youthsmoking. Health Economics 12 (3), pp.217-230.
[96] Schultz, T. P., 1984. Studying the impact of household economic and communityvariables on child mortality. Population and Development Review 10(Supplement), pp.215-235.
[97] Shin, H., Shin, M. and Rhee, J., 2012. Gender differences in the associationbetween self–rated health and hypertension in a Korean adult population. BMCPublic Health 12 (135).
[98] StataCorp., 2009. Stata Statistical Software, Version 11, Special Edition. CollegeStation, TX: Stata Corporation.
[99] Strachan, D. P. and Cook, D. G., 1997. Parental smoking and lower respiratoryillness in infancy and early childhood. Thorax 52 (10), pp.905-914.
REFERENCES 142
[100] Strauss, J. and Thomas, D., 1998. Health, nutrition, and economic development.Journal of Economic Literature 36 (2), pp.766-817.
[101] Strauss, J. and Thomas, D., 2008. Health over the life course. In: T. P. Schultzand J. Strauss, eds. Handbook of development economics, Volume 4. Amsterdam:Elsevier/North–Holland, pp.3375–3474.
[102] Subramanian, S. V., Kim, D. and Kawachi, I., 2005. Covariation in thesocioeconomic determinants of self rated health and happiness: a multivariatemultilevel analysis of individuals and communities in the USA. Journal ofEpidemiology & Community Health 59 (8), pp.664-669.
[103] Terza, J. V., Basu, A. and Rathouz, P. J., 2008. Two-stage residual inclusionestimation: addressing endogeneity in health econometric modelling. Journal ofHealth Economics 27 (3), pp.531-543.
[104] United States Department of Health and Human Services (USDHHS)., 2004. Thehealth consequences of smoking: a report of the surgeon general. [pdf] Atlanta,GA: U.S. Department of Health and Human Services, Centers for Disease Controland Prevention, National Center for Chronic Disease Prevention and HealthPromotion, Office on Smoking and Health. Available at: http://www.cdc.gov/tobacco/data_statistics/sgr/2004/complete_report/index.htm [Accessed3 March 2012].
[105] United States Department of Health and Human Services (USDHHS)., 2006. Thehealth consequences of involuntary exposure to tobacco smoke: a report of thesurgeon general. [pdf] Atlanta, GA: U.S. Department of Health and HumanServices, Centers for Disease Control and Prevention, Coordinating Center forHealth Promotion, National Center for Chronic Disease Prevention and HealthPromotion, Office on Smoking and Health. Available at:http://www.surgeongeneral.gov/library/reports/secondhandsmoke/
report-index.html [Accessed 3 March 2012].
[106] United States Department of Health and Human Services (USDHHS)., 2010. Howtobacco smoke causes disease: the biology and behavioral basis forsmoking-attributable disease: a report of the surgeon general. [pdf] Atlanta, GA:U.S. Department of Health and Human Services, Centers for Disease Control andPrevention, National Center for Chronic Disease Prevention and HealthPromotion, Office on Smoking and Health. Available at: http://www.surgeongeneral.gov/library/reports/tobaccosmoke/full_report.pdf
[Accessed 3 March 2012].
[107] Van Loon, A. M., Tijhuis, M., Surtees, P. G. and Ormel, J., 2005. Determinantsof smoking status: cross–sectional data on smoking initiation and cessation.European Journal of Public Health 15 (3), pp.256–261.
[108] Vella, F., 1998. Estimating models with sample selection bias: a survey. Journalof Human Resources 33 (1), pp.127–169.
REFERENCES 143
[109] Virtanen, M., Kivimaki, M., Kouvonen, A., Elovaino, M., Linna, A., Oksanen, T.and Vahtera, J., 2007. Average household income, crime, and smoking behaviourin a local area: the Finnish 10 - Town study. Social Science & Medicine 64 (9),pp.1904-1913.
[110] Von dem Knesebeck, O. and Geyer, S., 2007. Emotional support, education andself–rated health in 22 European countries. BMC Public Health 7 (272).
[111] Wang, M. P., Ho, S. Y., Lo, W. S., Lai, M. K. and Lam, T. H., 2012. Smoking isassociated with poor self–rated health among adolescents in Hong Kong. Nicotine& Tobacco Research 14 (6), pp.682–687.
[112] Yen, S. T. and Jones, A. W., 1996. Individual cigarette consumption andaddiction: a flexible limited dependent variable approach. Health Economics 5(2), pp.105-117.
[113] Yen, S. T., Shaw, W. D. and Yuan, Y., 2010. Cigarette smoking and self-reportedhealth in China. China Economic Review 21 (4), pp.532-543.
[114] Zimmer, Z., Natividad, J., Lin, H. and Chayovan, N., 2000. A cross-nationalexamination of the determinants of self-assessed health. Journal of Health andSocial Behaviour 41 (4), pp.465-481.
Chapter 5
Summary, Conclusions, and Policy
Implications
5.1 Summary
Health encompasses physical, mental, and social well–being (see, for example, Strauss
and Thomas, 2008). Good health is of interest to economists mainly because it is
a component of human capital that contributes to wealth creation (see, for example,
Mwabu, 2001; Mwabu, 2008). It may be sometimes informative to study the health of
specific population groups, such as infants and mothers, alongside general health.
Health can be investigated at the individual or population level, which leads to
population and individual measures of health. This thesis is concerned with individual
health.
We cannot observe an individual’s true health status but can approximate it with the
help of one or several observable indicators (Stein, 2004). Examples of the indicators of
individual health include birth weight (for infants), pregnancy outcomes (for mothers),
and self–reported general health status (for general health) (see, for example, Strauss
and Thomas 1998, 2008; Villamor and Cnattingius, 2006; Mwabu 2008, 2009).
Individuals can be thought of as producing health that maximizes the utility they
derive from health and other goods (see, for example, Rosenzweig and Schultz 1982,
1983; Mwabu 2008, 2009). The health must, however, be produced with the help of
various inputs (see, for example, Mosley and Chen, 1984; Schultz, 1984; Fuchs, 2004).
144
Summary, Conclusions, and Policy Implications 145
The individuals do not, however, know a priori which inputs are health enhancing and,
therefore, need this information. This thesis aims at contributing to this knowledge.
The thesis investigates the effects of various health inputs on health outcomes in
Kenya. It does this in three essays. The first essay, discussed in Chapter 2, investigates
the effect of the adequacy of prenatal care use on infant health. In this essay, birth
weight is used as a measure of infant health (for a justification, see , for example, Mwabu,
2009). Following the WHO classification, infants are classified as either having low birth
weight or normal weight (Zegers–Hochschild et al., 2009; WHO, 2011). The birth weight
of infants is then linked to a measure of adequacy of prenatal care use, constructed
based on WHO recommendations (Berg, 1995) and other individual and socio–economic
characteristics. Single–level and multilevel (see, Steenbergen and Jones, 2002) models
are then estimated controlling for potential sample selection bias, potential endogeneity
of prenatal care, and potential unobserved heterogeneity using data from the Kenya
Demographic and Household Surveys (KDHS) collected in 1993, 1998, 2003 and 2008.
We find that adequate use of prenatal care improves infant health. In particular,
adequate use of prenatal care lowers the chances of delivering a low birth weight infant.
The literature (see, for example, Conway and Deb, 2005; Jewell and Triunfo, 2006)
supports this finding. The multilevel model results show that if we do not control for
unobserved mother–specific effects, then we run the risk of overstating the effect of
prenatal care on infant health.
The second essay, contained in Chapter 3, investigates the effect of preceding birth
interval length on maternal health. Maternal health is measured in the essay using an
indicator of whether or not a woman experienced an adverse pregnancy outcome (in
particular a miscarriage, a stillbirth or an abortion) since adverse pregnancy outcomes
may be pointers towards the existence of some underlying maternal health problems (see,
for example, Villamor and Cnattingius, 2006). Previous studies on birth intervals in
Kenya (see, for example, Ikamari, 1998; Rafalimanana and Westoff, 2001; Kosimbei, 2005;
Mustafa and Odimegwu, 2008) have not linked preceding birth intervals to maternal health
and the essay fills this gap. We construct a model that links maternal health to preceding
birth interval and other variables. The estimated model controls for potential sample
selection bias, potential endogeneity of preceding birth interval length, and potential
unobserved heterogeneity. We use data from the Kenya Demographic and Health Survey
(KDHS) collected in 1998, 2003, and 2008 to estimate our model. Consistent with the
literature (see, for example, Conde–Agudelo and Belizan, 2000; Stephansson, Dickman
and Cnattingius, 2003; Conde–Agudelo, Rosas–Bermudez and Kafury–Goeta, 2007;
Summary, Conclusions, and Policy Implications 146
Stamilio et al., 2007; Williams et al., 2008), we find that preceding birth intervals of
length 36 – 59 months improve maternal health. This conclusion is, however, only arrived
at when we control for the endogeneity of preceding birth interval.
The third essay, presented in Chapter 4, looks at smoking and general health. General
health, in this essay, is measured using self–rated health status. Studies show that
self–rated health status is a predictor of actual health status (see, for example, Ahmad,
Jafa and Chaturvedi, 2005; Ford, Spallek and Dobson, 2008; Jylha, 2009; McFadden et
al., 2009). In the literature, both active and passive smoking are shown to be harmful
to health (see, for example, USDHHS, 2004; Adab et al., 2005; USDHHS, 2006; Chen
et al., 2012). Since previous studies on self–rated health in Kenya (see, for example,
Kyobutungi, Ezeh, Zulu and Falkingham, 2009; Kodzi, Gyimah, Emina and Ezeh, 2011)
have not explicitly examined the effect of smoking on self–rated health status, this
essay fills this gap. We estimate a model linking the individual’s smoking status to
the individual’s self–rated health status using data mainly from the National Health
Expenditure and Utilization Survey collected by the Ministry of Health in 2003. We
control for potential sample selection bias, potential endogeneity of smoking, and potential
unobserved heterogeneity in the estimation process. The results confirm the findings
from the literature of a negative effect of smoking on health (see, for example, Montazeri,
Goshtasebi and Vahdaninia, 2008; El Ansari and Stock, 2012; Shin, Shin and Rhee, 2012).
We particularly find that the probability of rating one’s own health as “Poor” compared
to age–mates is higher among smokers than among non–smokers, holding other factors
constant. This probability is, however, underestimated in models that do not control
for sample selection bias, endogeneity of smoking and unobserved heterogeneity. We
also find that the probability of rating one’s own health as “Very Good” is lower among
smokers as compared to non–smokers, holding other factors constant. This probability is
also underestimated in models that do not control for sample selection bias, endogeneity
of smoking and unobserved heterogeneity.
5.2 Conclusions
We can draw several conclusions based on the findings in this thesis. First, prenatal care,
when adequately used, improves infant health. The beneficial effects of prenatal care on
infant health are, however, overstated in models that do not control for unobserved mother–
specific effects. Second, preceding birth intervals of length 36 to 59 months improve
Summary, Conclusions, and Policy Implications 147
maternal health. However, preceding birth interval is an endogenous determinant of
maternal health and its positive effect on maternal health is only observed after controlling
for this endogeneity. Third, smoking is harmful to health. The harmful effects of smoking
on health are, however, underestimated in models that do not control for sample selection
bias, endogeneity of smoking and unobserved heterogeneity. There is need to carry
out further research that seeks to shed more light on the barriers to adequate use of
prenatal care, barriers to optimal birth spacing, and the determinants of smoking among
different population groups. Such research will help in designing policies that promote
health–enhancing behaviours and actions.
5.3 Policy Implications
The findings from the thesis have several implications for policies that could improve health
in Kenya. First, policies that promote adequate use of prenatal care are likely to improve
infant health by, lowering the number of low birth weight infants, and ultimately infant
and child mortality. This will help towards the attainment of Millennium Development
Goal (MDG) No. 4 on reducing child mortality (United Nations, 2010). Second, policies
that make it easier for mothers to optimally space their births should be pursued.
Such policies are likely to contribute to the attainment of MDG No. 5 of improving
maternal health (United Nations, 2010). Third, policies that discourage active and
passive smoking would lead to an improvement in general health. Examples of such
policies are documented in the World Health Organization’s International Framework
Convention on Tobacco Control (WHO, 2003), which Kenya has already ratified.
REFERENCES 148
References
[1] Adab, P., McGhee, S. M., Hedley, A. J. and Lam, T. H., 2005. Smoking,respiratory disease and health service utilisation: the paradox. Social Science &Medicine 60 (3), pp.483-490.
[2] Ahmad, K., Jafa, T. H. and Chaturvedi, N., 2005. Self–rated health in Pakistan:results of a national health survey. BMC Public Health 5 (51).
[3] Berg, C. J., 1995. Prenatal care in developing countries: the World HealthOrganization technical working group on antenatal care. Journal of the AmericanMedical Women’s Association (1972) 50 (5), pp.182–186.
[4] Chen, R., Zhang, D., Chen, Y., Hu, Z. and Wilson, K., 2012. Passive smokingand risk of cognitive impairment in women who never smoke. Archives of InternalMedicine 172 (3), pp.217–273.
[5] Conde-Agudelo, A. and Belizan, J. M., 2000. Maternal morbidity and mortalityassociated with interpregnancy interval: cross sectional study. British MedicalJournal (BMJ) 321 (7271), pp.1255-1259.
[6] Conde-Agudelo, A., Rosas–Bermudez, A. and Kafury-Goeta, A. C., 2007. Effectsof birth spacing on maternal health: a systematic review. American Journal ofObstetrics & Gynecology 196 (4), pp.297-308.
[7] Conway, K. S. and Deb, P., 2005. Is prenatal care really ineffective? Or, is the‘devil’ in the distribution? Journal of Health Economics 24 (3), pp.489–513.
[8] El Ansari, W. and Stock, C., 2012. Factors associated with smoking, quitattempts and attitudes towards total smoking bans at university: a survey ofseven universities in England, Wales, and Northern Ireland. Asian PacificJournal of Cancer Prevention 13 (2), pp.705–714.
[9] Ford, J., Spallek, M. and Dobson, A., 2008. Self-rated health and a healthylifestyle are the most important predictors of survival in elderly women. Age andAging 37 (2), pp.194-200.
[10] Fuchs, V. R., 2004. Reflections on the socio-economic correlates of health.Journal of Health Economics 23 (4), pp.653-661.
[11] Ikamari, L., 1998. Birth intervals and child survival in Kenya. African Journal ofHealth Sciences 5 (1), pp.15–24.
[12] Jewell, R. T. and Triunfo, P., 2006. The impact of prenatal care on birthweight:the case of Uruguay. Health Economics 15 (11), pp.1245–1250.
[13] Jylha, M., 2009. What is self-rated health and why does it predict mortality?Towards a unified conceptual model. Social Science & Medicine 69 (3),pp.307-316.
REFERENCES 149
[14] Kodzi, I. A., Gyimah, S. O., Emina, J. and Ezeh, A. C., 2011. Religiousinvolvement, social engagement, and subjective health status of older residents ofinformal neighborhoods of Nairobi. Journal of Urban Health 88 (Suppl 2),pp.370–380.
[15] Kosimbei, G., 2005. Child healthcare seeking behaviour in Kenya. KIPPRADiscussion Paper No. 50. Nairobi: Kenya Institute for Public Policy Researchand Analysis (KIPPRA).
[16] Kyobutungi, C., Ezeh, A. C., Zulu, E. and Falkingham, J., 2009. HIV/AIDS andthe health of older people in the slums of Nairobi, Kenya: results from a crosssectional survey. BMC Public Health 9 (153).
[17] McFadden, E., Luben, R., Bingham, S., Warenham, N., Kinmonth, A. L. andShaw, K. T., 2009. Does the association between self-rated health and mortalityvary by social class? Social Science & Medicine 68 (2), pp.275-280.
[18] Montazeri, A., Goshtasebi, A. and Vahdaninia, M., 2008. Educational inequalitiesin self–reported health in a general Iranian population. BMC Research Notes1(50).
[19] Mosley, W. H. and Chen, L. C., 1984. An analytical framework for the study ofchild survival in developing countries. Population and Development Review 10(Supplement), pp.25-45.
[20] Mustafa, H. E. and Odimegwu, C., 2008. Socioeconomic determinants of infantmortality in Kenya: analysis of Kenya DHS 2003. Journal of Humanities &Social Sciences 2 (2), pp.1–16.
[21] Mwabu, G., 2001. Health status in Africa: a regional profile. South AfricanJournal of Economics 69 (2), pp.319-335.
[22] Mwabu, G., 2008. Health economics for low-income countries. In: T. P. Schultzand J. Strauss, eds. Handbook of development economics, Volume 4. Amsterdam:Elsevier/North–Holland, pp.3305–3374.
[23] Mwabu, G., 2009. The production of child health in Kenya: a structural model ofbirth weight. Journal of African Economies 18 (2), pp.212-260.
[24] Rafalimanana, H. and Westoff, C. F., 2001. Gap between preferred and actualbirth intervals in Sub–Saharan Africa: implications for fertility and child health.DHS Analytical Studies No. 2. [pdf] Calverton, Maryland: ORC Macro. Availableat: http://www.measuredhs.com/pubs/pdf/AS2/AS2.pdf [Accessed 6November 2012].
[25] Rosenzweig, M. R. and Schultz, T. P., 1982. The behaviour of mothers as inputsto child health: the determinants of birth weight, gestation, and rate of fetalgrowth. In: V. R. Fuchs, ed. Economic aspects of health. [pdf] Chicago:University of Chicago Press, pp.53–92. Available at:
REFERENCES 150
http://www.nber.org/chapters/c6545.pdf [Accessed 6 August 2008].
[26] Rosenzweig, M. R. and Schultz, T. P., 1983. Estimating a household productionfunction: heterogeneity, the demand for health inputs, and their effects on birthweight. Journal of Political Economy 91 (5), pp.723-746.
[27] Schultz, T. P., 1984. Studying the impact of household economic and communityvariables on child mortality. Population and Development Review 10(Supplement), pp.215-235.
[28] Shin, H., Shin, M. and Rhee, J., 2012. Gender differences in the associationbetween self–rated health and hypertension in a Korean adult population. BMCPublic Health 12 (135).
[29] Stamilio, D. M., DeFranco, E., Pare, E., Odibo, A. O., Peipert, J. F., Allsworth,J. E., Stevens, E. and Macones, G. A., 2007. Short interpregnancy interval: riskof uterine rupture and complications of vaginal birth after cesarean delivery.Obstetrics and Gynecology 110 (5), pp.1075–1082.
[30] Steenbergen, M. R. and Jones, B. S., 2002. Modeling multilevel data structures.American Journal of Political Science 46 (1), pp.218-237.
[31] Stein, R. E. K., 2004. Measurement of children’s health. Ambulatory Pediatrics 4(4), pp.365-370.
[32] Stephansson, O., Dickman, P. W. and Cnattingius, S., 2003. The influence ofinterpregnancy interval on the subsequent risk of stillbirth and early neonataldeath. Obstetrics and Gynecology 102 (1), pp.101–108.
[33] Strauss, J. and Thomas, D., 1998. Health, nutrition, and economic development.Journal of Economic Literature 36 (2), pp.766-817.
[34] Strauss, J. and Thomas, D., 2008. Health over the life course. In: T. P. Schultzand J. Strauss, eds. Handbook of development economics, Volume 4. Amsterdam:Elsevier/North–Holland, pp.3375–3474.
[35] United Nations, 2010. Millennium development goals: at a glance. [pdf] UNDepartment of Public Information. Available at:http://www.un-kampagne.de/fileadmin/downloads/presse/MDGs-at-a-
Glance-engl-03-2010_02.pdf [Accessed 3 March 2012].
[36] United States Department of Health and Human Services (USDHHS)., 2004. Thehealth consequences of smoking: a report of the surgeon general. [pdf] Atlanta,GA: U.S. Department of Health and Human Services, Centers for Disease Controland Prevention, National Center for Chronic Disease Prevention and HealthPromotion, Office on Smoking and Health. Available at: http://www.cdc.gov/tobacco/data_statistics/sgr/2004/complete_report/index.htm [Accessed3 March 2012].
REFERENCES 151
[37] United States Department of Health and Human Services (USDHHS)., 2006. Thehealth consequences of involuntary exposure to tobacco smoke: a report of thesurgeon general. [pdf] Atlanta, GA: U.S. Department of Health and HumanServices, Centers for Disease Control and Prevention, Coordinating Center forHealth Promotion, National Center for Chronic Disease Prevention and HealthPromotion, Office on Smoking and Health. Available at:http://www.surgeongeneral.gov/library/reports/secondhandsmoke/
report-index.html [Accessed 3 March 2012].
[38] Villamor, E. and Cnattingius, S., 2006. Interpregnancy weight change and risk ofadverse pregnancy outcomes: a population–based study. Lancet 368 (9542),pp.1164–1170.
[39] Williams, E. K., Hossain, M. B., Sharma, R. K., Kumar, V., Pandey, C. M. andBaqui, A. H., 2008. Birth interval and risk of stillbirth or neonatal death:findings from rural India. Journal of Tropical Pediatrics 54 (5), pp.321–327.
[40] World Health Organization (WHO), 2003. WHO framework convention ontobacco control. [pdf] Geneva: WHO. Available at:http://whqlibdoc.who.int/publications/2003/9241591013.pdf [Accessed3 March 2012].
[41] World Health Organization (WHO), 2011. International statistical classificationof diseases and related health problems, 10th Revision (ICD–10): volume 2instruction manual. 2010 edition. [pdf] World Health Organization. Available at:http://www.who.int/classifications/icd/ICD10Volume2_en_2010.pdf
[Accessed 12 October 2012].
[42] Zegers–Hochschild, F., Adamson, G. D., de Mouzon, J., Ishihara, O., Mansour,R., Nygren, K., Sullivan, E. and Vanderpoel, S., 2009. International committeefor monitoring assisted reproductive technology (ICMART) and the WorldHealth Organization revised glossary of ART terminology. Fertility and Sterility92 (5), pp.1520–1524.
Colophon
This thesis was typset in LATEX 2ε using the “hepthesis” class.
153