Mellington 1999

Bulletin of Indonesian Economic Studies Vol 35 No 3, December 1999, pp.
115–44
FEMALE EDUCATION AND CHILD

MORTALITY IN INDONESIA
Nicole Mellington and Lisa Cameron*
University of Melbourne
This paper uses a sample of 6,620 women from the 1994 Indonesian
Demographic and Health Survey to examine the relationship between
female education and child mortality in Indonesia. Female education is
measured in terms of both years of education and literacy. Both primary
education and secondary schooling significantly decrease the probability
of child death, while literacy plays an insignificant role. When the sample
is divided into urban and rural locations, primary and secondary education
are significant in both areas in reducing the likelihood of a mother
experiencing child mortality. The benefits of public and private
infrastructure appear to differ in rural and urban areas. The results confirm
that investment in female human capital lowers the probability of child
mortality.
INTRODUCTION
The Indonesian child mortality rate has fallen significantly over the past
20 years. Infant mortality fell from 125 to 51 deaths per thousand births
between 1960 and 1995 (World Bank 1980, 1997). Information on child
mortality is harder to come by, but there is evidence of a similar decline.1
Apart from the obvious micro-level benefits to mothers and families,
lower child mortality is also desirable at the macro level. High child
mortality is often cited as a significant determinant of high fertility rates
and high population growth. In an environment where there is very little
public financial support for the elderly, it is important for individuals to
ensure that they have offspring who will survive to look after them in
old age. If child mortality rates are high then there is an incentive to have
more children to guarantee that some survive. Risk-averse parents tend
to overcompensate for the risk of child death, and high population growth
results. Thus, lower child mortality rates reduce the incentive to have
116 Nicole Mellington and Lisa Cameron
large families, and over time result in lower fertility rates and a consequent
lowering of the population growth rate.2
A lower rate of population growth has long been a goal of the
Indonesian government. Under its highly successful family planning
program, the fertility rate dropped from five to three children per woman
in the 20 years between 1971 and 1991 (Jensen 1996). Given the research
interest in these fertility changes (for example, Jensen 1996; Gertler and
Molyneaux 1994; and Hull 1993), and the continuing government view
of fertility and population control as an important policy objective, it is
surprising that so little research has been conducted on child mortality.3
We are specifically interested in examining the relationship between child
mortality and the mother’s level of education.
The only previous studies of Indonesian child mortality of which we
are aware are Cho, Suharto, McNicoll and Mamas (1980), McDonald
(1980) and Martin, Trussel, Salvails and Shah (1983). Cho et al. examined
differences in infant mortality rates across urban and rural areas in the
1960s, but did not analyse the determinants of these differentials.
McDonald’s paper is a short note that examines the proportion of children
dying before the age of five in 1976. He presents cross-tabulations of
deaths by the education of the mother and of the father. Child mortality
was highest among households in which the father had no education,
and next highest among those in which the mother had no education. He
found the effect of primary school education to be quite small, however.
The largest decrease in child mortality occurred with secondary education.
Without the ability to control for household expenditure or income,
however, it is not clear whether he was identifying a pure education effect
or the result of higher incomes.
Martin et al. (1983) conducted a more sophisticated statistical analysis
using a hazard model to estimate the probability of child death. Their
study used data from the 1976 Indonesian Fertility Survey, which covered
only Java and Bali. They examined the effect of the mother’s education,
but the data did not allow them to control for household income or
expenditure. Husband’s education level was included as a proxy for
household income. They were also unable to control for household and
community infrastructure. The study found that one to six years of
maternal education did not affect the probability of child death, but more
than six years of education lowered it.
This paper contributes to the literature in a number of ways. First, it
provides only the second multivariate study of child mortality in
Indonesia. Second, it uses data from the 1994 Demographic and Health
Survey (DHS), and so significantly updates this field of inquiry and is
able to study the entire nation, rather than just Java and Bali. Third, we
Female Education and Child Mortality in Indonesia 117
are able to control for household expenditure, and so can separate out
the income-enhancing effect of education from its other roles. Fourth,
we use data on literacy as well as years of formal schooling to assess
more accurately the type of education or knowledge that impacts on child
mortality. Finally, the DHS provides information on household
infrastructure such as toilet facilities and piped water, and we merge this
with provincial level data on the provision of health services. As a result,
we are able to examine the previously unstudied effect of such facilities
on child mortality in Indonesia.
EDUCATION AND CHILD MORTALITY

The role of female education in lowering child mortality has been widely
recognised in the literature. In an early survey of studies in this area,
Cochrane (1979) showed that most empirical research had found maternal
education to reduce the probability of child mortality.
Mother’s education is hypothesised to lower child mortality in a
number of ways. Education reduces the cost of information; thus more
educated women have been found empirically to have a greater
understanding of the value of public health infrastructure and to be better
able to locate health services (Singh 1994: 209–20; Caldwell 1979: 405–9;
Aly and Grabowski 1990: 735). Access to health care facilities and toilet
connections has been found to benefit children of educated mothers more
than those of less educated mothers. Education may also serve to reduce
fatalistic attitudes towards illness, promoting the use of modern
techniques of child care and disease prevention, and thus reducing
reliance upon traditional methods of child care (Caldwell 1979). In both
of these ways, female education is a complement to health services. It
has also been found to compensate for a lack of health facilities: Barrera
(1990) and Benefo and Schultz (1994) found that a higher level of education
amongst mothers counteracted the effects of an unclean community
environment and a lack of safe water connections.
There is evidence that maternal education may also affect the
traditional balance of family relationships, and that this works to the
advantage of children. Several studies have found a relationship between
child outcomes and women’s control of assets and income.4 Given that
education and income are positively related, it is likely that higher levels
of maternal education may also decrease the probability of child mortality
via greater maternal bargaining power within the household. Even
household preferences involving food types consumed, weaning and
methods of child care may alter because of education (Caldwell 1979:
409–10). The positive relationship between education and earnings may

also have a direct effect in terms of increased household expenditure.
Better educated women may be more likely to find partners with greater
earning potential, which would augment this income effect (Schultz 1984).
Barrera (1990) found, however, that an increase in family economic status
is not a significant channel through which education affects child health.
Finally, a woman’s educational attainment reflects the willingness of
her parents to invest in female children, and may be correlated with the
mother having received relatively high nutrition and good training as a
child. This background may improve her capacity to produce healthy
children (Barrera 1990). It is difficult, however, to test this empirically.5
This research focuses on the socio-economic determinants of child
death. We recognise, however, that child death is directly physiological.
As is clear from the discussion above, socio-economic variables such as
the education of the mother affect child death via intermediate factors
that influence the physiological wellbeing of the child. Mosley and Chen
(1984) develop an analytical framework that links the socio-economic
determinants of child death to these intermediate determinants. This set
of variables includes maternal factors, nutrient deficiency, personal illness
control and environmental contamination. In this study, we do not
explicitly model these linkages. We believe this is appropriate, given that
our main aim is to quantify the effect of maternal education on the
probability of child death. If we were to control completely for all
intermediate factors, then the socio-economic variables would be seen to
have no impact on child death. We know that this is not the case, and
that the socio-economic variables simply lie further back along the chain
of causality. Furthermore, as we discuss below, the inclusion of these
intermediate factors is often problematic because of their likely
endogeneity, and in some cases their inclusion is impossible owing to
lack of data. We have included some simple controls for environmental
contamination and immunisation, and examine explicitly whether the
impact of these variables differs with the educational attainment of the
mother.
DATA
The 1994 Indonesian Demographic and Health Survey contains
individual, household and community level information pertaining to
28,168 Indonesian ‘ever-married’ women between the ages of 15 and 49.6
It provides detailed data on health and demographic variables and on
household structure and expenditure patterns over the 12 months prior

to the survey.
The sample used in this paper is restricted to women who had given
birth in the five years before the DHS survey and for whom there are
expenditure data. Owing to cost constraints the expenditure data were
collected only for a random 50% of the sample. As our study relies heavily
on this information, the working sample is restricted to this proportion
of the population. The final sample includes 6,620 women.
We limited the sample to women who had given birth relatively
recently because we wished to examine more recent child deaths. This
was necessary because the DHS provides information on the value of the
explanatory variables only in the year of the survey. For instance, we
know what per capita household expenditure was in 1994, but not in
previous years. We are thus examining the relationship between past
deaths and the current circumstances of the mother. For variables that
we expect to be largely non-time-varying, such as maternal education
and province of residence, this is not problematic. Even the time-varying
variables used, such as per capita expenditure and control for socio-
economic status, are likely to be highly correlated across time. Thus, for
example, per capita household expenditure today is likely to be highly
correlated with past expenditure, and hence with past child deaths.
Restricting the sample to mothers who have young children in order to
focus the analysis on relatively recent deaths ensures the relevance of the
current values of the variables.7 In the restricted sample, 75% of the
mothers who have experienced child death have done so in the last 10
years, and the reported results are robust to restricting child deaths to
those that occurred only in the last 10 years. We also present results
estimated using only non-time-varying variables.8
It is not possible to restrict the sample to only very recent births (for
example, those in the past three years) because this exacerbates the
censoring of observations in the sample; that is, any child aged less than
five at the time of the survey has not had the full opportunity to die
before his or her fifth birthday. Hence, the choice of sample reflects the
need to balance the biases that arise from censoring (so including children
born more than five years ago) against the need to examine relatively
recent births because of the lack of information on past values of the
explanatory variables.9
The DHS data are supplemented with regional information from the
Statistical Yearbook of Indonesia (BPS 1995) on local market prices, the
number of public health centres and weather patterns in 1994.
METHODOLOGY
Child mortality is here defined as the death of a child before the age of
five. Of the women in the sample, 22% have had a child die before this
age.10 There are a number of ways to model child death. We model the
probability of a mother experiencing the death of a child using probit
estimation. Another alternative would have been to estimate tobits, with
the proportion of the woman’s children who died as the dependent
variable, and upper censoring at 1.0 and lower censoring at zero. A tobit
specification theoretically uses the data more efficiently because it utilises
more information than just the 0/1 nature of accounting for a child death.
We estimated tobits but found the results did not differ significantly from
the probit results. This suggests that little differentiates a woman who
has had one child die from women who have experienced multiple child
deaths. In the interests of clarity we have chosen to present the probit
results.11
A further alternative would have been to structure the data with one
observation per ever-born child, and then to have estimated the
probability of each child having died before the age of five. The main
advantage of this would have been to allow us to control for some of the
intermediate determinants of child death. However, as discussed above,
we wish to capture the full effect of socio-economic determinants on child
death and so do not wish to apportion them over the various intermediate
factors. In addition, data limitations make it impossible to control for
variables such as the duration of breast-feeding (reported only for children
born in the last three years) and immunisation (reported only for children
who were alive at the time of the survey). The DHS does provide data on
preceding birth intervals for all children and the birth age of the mother,
but these variables are likely to be endogenous. It is well established that
fertility and mortality are jointly determined, so any variable that is a
function of fertility is endogenous to child mortality. For instance, the
birth age of the mother is a function of her fertility behaviour and so is
endogenous, as is any variable that is a function of previous or current
mortality experience. Barrera (1991) argues that breast-feeding is
endogenous because whether to breast-feed or not is a function of child
health. Since the length of the preceding birth interval is likely to be a
function of whether the previous child survived or not, it is similarly
endogenous.
Given that little is gained from using the child as the unit of
observation, and that the focus of the paper is on the effect of maternal
education on child mortality, we present results from a sample with one
observation per woman. Results obtained from samples with one
observation per child are qualitatively identical, and are available from
the authors on request.
The probit model assumes that there is a latent variable yi* which can
be written as a linear function of variables that affect the probability of a
child dying. Hence, we can write:
yi∗ = βX i + ε i (1)
where Xi is a vector of explanatory variables, β is the vector of coefficients

that will be estimated and ε i is a random error term. The latent variable
is unobservable and instead the dummy variable yi = 1 if a child has
died, and zero otherwise, is observed. It is defined as:
yi = 1 if yi* > 0
= 0 otherwise
The probit model assumes that the error term, ε i , is distributed
according to the cumulative normal distribution function. If this is the
case, then the probability of at least one child dying can be written as:
β X i −0.5t 2
1 e dt
2π ∫
P= (2)
−∞
where t is a standardised normal variable. Maximum likelihood

estimation produces estimates of the β s.
Explanatory Variables
Household Level Variables. The probability of a child dying will reflect
characteristics of both the mother and father and the household’s socio-
economic status. The geographic region in which the household lives
and the extent of social and household infrastructure in the region may
also be important determinants of child mortality.
The characteristic of the mother that has received the most attention
as a determinant of child mortality is her educational attainment (Aly
and Grabowski 1990; Singh 1994), which can affect child mortality through
the channels discussed above. We are interested in determining the level
of education that is effective in reducing child mortality. Some studies
have suggested that it is literacy rather than years of formal education
that is of importance. United Nations (1991), for example, states that
‘literacy is a better measure of education than enrollment since it usually
reflects a minimal level of completing schooling’. Obviously, literacy and
years of education are closely related variables. In the DHS data they are,
however, independent enough to allow us to include both as explanatory
variables. A dummy variable (literate), which equals 1 if the mother is

able to read and zero otherwise, is included in the regressions. We also
define three variables (primary, secondary, tertiary), which indicate the
number of years of schooling received by the mother at each of the
respective stages of the education process. By defining three variables,
we allow the effect of an additional year of schooling to differ depending
on whether it is at primary, secondary or tertiary level.
If it is literacy per se that has a negative effect on the incidence of child
mortality, then we would expect literate to be statistically significant in
the regressions. This might be the case, for example, if by being literate
one has access to information on health that is not available to the illiterate.
However, it may be that through additional formal education women
can further reduce the likelihood of a child dying. For example,
information on health procedures may be disseminated through schools,
or it could be that through participation in the formal education process
one establishes the contacts and confidence to use public health facilities
more effectively.
The woman’s education is also likely to affect her earning capacity,
and hence the income and expenditure of the household. The DHS does
not provide income data. We are able to control for household
expenditure, however, and include the logarithm of per capita annual
household expenditures as an explanatory variable. Hence, the
coefficients on the education terms will pick up the pure effect of
education on child mortality, not the effect via income or expenditure.
Expenditure is likely to be endogenous because it will vary with
household size, which is a function of child mortality. Even per capita
expenditure is likely to vary with household size, because of economies
of scale within the household.12 The estimation procedure deals with the
likely endogeneity of the logarithm of per capita expenditure. An
expenditure equation is estimated, and the predicted value, rather than
the actual value, of the logarithm of per capita expenditure is used in the
child mortality probits. To ensure identification we need to include
instruments in the prediction regression that do not belong directly in
the child mortality probit. The instruments used are: whether the family
owns its own home; whether it has a mortgage or pays rent; and whether
it has electricity, a television set, a kerosene stove, a motor cycle or a
motor boat.13
We further control for the socio-economic status of the household by
including variables reflecting the husband’s occupation. The sample
includes only ever-married women, and we control for the occupation of
the husband, regardless of whether he still resides in the household. This
is because the husband’s occupation is likely to be correlated with the
socio-economic status of the woman whether he is resident or not. We

control for the presence of the husband using a dummy variable (husband)
that equals 1 if the husband lives in the household and 0 otherwise.
The woman’s age is also likely to be an important determinant of the
probability of her having experienced the death of a child. We define five
age category variables, which range from the lowest age of 15 to the
highest of 49. The woman’s age is a proxy for her physical health and her
reproductive opportunities. Because the likelihood of genetic disorders
in the child increases with the age of the mother, we would expect a
positive relationship between the mother’s age and the probability of
child death. In addition, the woman’s age identifies her birth cohort, and
so may also capture the effects on child mortality of Indonesia’s
development stage. For example, a woman who is 45 is more likely to
have had children 20 years ago, when there was much less in the way of
public health education and health centres.14
The overriding factor with respect to age, however, is likely to be that
older women will have had more children on average than younger
women, and so will have had greater exposure to the possibility of child
death. Ideally, we would control for the number of children ever born to
each woman. However, the number of children ever born is likely to be
endogenous, because women whose children have died will be more
likely, other things being equal, to have had more children than those
who have not experienced the death of a child. Modelling the number of
children ever born is a complex undertaking and beyond the scope of
this paper.15 The estimation procedure used here does, however, make
use of data on the number of children ever born to each woman. The
inability to control directly for the number of children gives rise to a
heteroscedasticity problem. The variance of the residuals is likely to be
smaller for women who have had more children, because more
information is contained in these observations.16 The variance of the error
term is thus inversely related to the number of children; that is:
σ2
Var(ε i ) = (3)
childreni
To correct for this, we estimate a weighted probit where the weights

are the square root of the number of children ever born.17 All of the
possible relationships between the mother’s age and child mortality
discussed above give rise to a positive association between the two
variables. We will be unable, however, to differentiate empirically between
these hypotheses.
The sanitation facilities of the household are also likely to affect the
probability of child death. We control for access to piped water and for
whether the household has a toilet.
We also control for the woman’s religion. Child mortality may differ
across religious groups because of differences in traditions (such as the
practice of circumcision) and in the cultural importance of children.
Provincial Level Variables. In addition to household level variables,
we supplement the DHS data with some provincial level indicators. Rice
is the dominant subsistence food in Indonesia, so we include provincial
retail rice prices to control for the effect on child health of differences in
purchasing power across regions. We also control for different inflation
rates across provinces by including a clothing price index, a general price
index and a housing price index.18 The average monthly rainfall in each
province in the wet and dry seasons may also be relevant. Rainfall may
impact upon child mortality, as it is linked to the prevalence of certain
infectious diseases and parasites, and is correlated with agricultural
conditions (Benefo and Schultz 1994: 9–10).19
We also control for the number of public health centres per province.20
A potential problem with the inclusion of infrastructure variables is the
possible endogeneity of government policy. Interregional variation in
policies and programs may not be independent of household resources
and preferences. For example, health programs may be set up in regions
with particularly serious health problems. Immunisation and disease
control may be targeted to poorer areas where the populace is less
educated. Hence some public health programs could be associated with
higher child mortality. Additionally, there is the problem of endogenous
migration. People may relocate toward healthier environments and
specially targeted public health programs. If these migrants also invest
relatively more in the health of their family for unobserved reasons, then
this type of migration may lead to bias in regionally based policy
evaluation studies. We do not attempt to endogenise these policy
variables, but such issues need to be kept in mind when interpreting
their coefficients.
We also wished to control for child immunisation. The proportion of
each woman’s children who were immunised is likely to be endogenous,
however, because women who have had a child die may be more inclined
to have their remaining children immunised. Also, it is an inappropriate
measure in practice, because the DHS reports it only for living children.
To overcome this problem, the provincial average immunisation rate was
constructed from the DHS data. The provincial average reflects the
prevalence of immunisation in the province, and so will control for
community attitudes to immunisation and the existence and effectiveness
of immunisation programs, while avoiding the problem of endogeneity.
TABLE 1 Summary Statistics
Variable Mean Standard Min. Max.
Dependent variable
At least one of the mother’s offspring died 0.221 0.415 0 1
before the age of five
Explanatory variables at the individual and household levels
Women’s completed years of schooling
Primary 4.398 2.253 0 6
Secondary 1.437 2.306 0 6
Tertiary 0.151 0.906 0 6
Literacy dummy 0.810 0.393 0 1
Logarithm of annual household 13.02 0.587 10.65 15.81
expenditure per capita
Predicted logarithm of annual household 13.00 0.394 12.29 14.45
Woman’s age: 15–24 0.252 0.434 0 1
25–29 0.288 0.453 0 1
30–34 0.236 0.425 0 1
35–39 0.149 0.356 0 1
40–49 0.075 0.263 0 1
Husband’s occupation
Administration 0.010 0.100 0 1
Agriculture 0.452 0.498 0 1
Clerical 0.065 0.247 0 1
Industry 0.222 0.416 0 1
Professional 0.074 0.261 0 1
Sales 0.104 0.306 0 1
Service 0.048 0.215 0 1
Other 0.004 0.059 0 1
Current rural resident 0.734 0.442 0 1
The family owns its own land 0.309 0.462 0 1
Religion: Catholic 0.083 0.276 0 1
Christian 0.089 0.285 0 1
Buddhist 0.009 0.092 0 1
Hindu 0.034 0.180 0 1
Other 0.002 0.049 0 1
Toilet facility in residence 0.426 0.495 0 1
Protected water source, including piped water 0.119 0.324 0 1
Provincial level variables
Average monthly rainfall: Dry season 2.181 1.929 0 7.02
Wet season 9.469 2.778 3 16.3
Average immunisation rate 0.444 0.102 0.27 0.67
Price of: Rice 743.0 85.97 589.2 969.5
Housing 209.5 219.1 128.0 1167.2
Clothing 135.6 11.57 119.7 166.9
General items 153.4 6.405 139.9 169.2
Number of public health centres 287.6 270.5 78 951
N = 6,620
Two estimation strategies were adopted. First, the regressions were

estimated with the full set of explanatory variables detailed above.
Second, provincial dummy variables were included. These control for
all observable and unobservable differences across provinces. Their
inclusion meant, however, that we could not include the provincial level
variables that may be of policy interest, such as the immunisation rate
and the number of public health centres.21 Table 1 presents the descriptive
statistics of the dependent and explanatory variables.
Urban versus Rural Areas. It may be argued that the effect of female
education is likely to differ between rural and urban areas. Rural people
are more likely to rely on traditional methods for dealing with health
problems. Health facilities are not so readily accessible, and perhaps not
of the same quality as in urban areas. The information communicated
through education may thus be more important in rural than in urban
areas, where information is more readily available and clinics are more
accessible (Aly and Grabowski 1990: 73). In the first specifications we
used a rural/urban dummy to control for whether the household was
situated in a rural area. To allow for the possible differential impact of
education in rural and urban areas, we then estimated separate rural
and urban equations.
Interacting Education with Infrastructure. In the same way that the value
of education may differ across rural and urban settings, the value of social
infrastructure may vary depending on the educational attainment of the
population. To examine this, we ran additional regressions that interacted
the years of education with the variables that reflect the existence of toilet
facilities, the availability of piped water, the number of public health
centres and the average provincial immunisation rate.
EMPIRICAL RESULTS
Preliminary Results
Figures 1a and 1b present the raw relationship between female education
and child mortality in the DHS data. The negative relationship is clearly
depicted. The results suggest that the benefits of education are
concentrated in the first six years of education (the completion of primary
school). This is at odds with the conclusions drawn in Martin et al. (1983)
and McDonald (1980). Women with no education experience child
mortality rates 35% higher than those who have entered, but not
completed, primary school, and 97% higher than those who have
completed primary school. The mean child mortality rate drops even
further from 5% for women who have entered secondary education to
FIGURE 1 Mean Child Mortality by Educational Attainment
(a) Total Sample

0.16
0.12
0.08
0.04
0
None Incomplete Complete Incomplete Complete Higher
primary primary secondary secondary
(b) Urban and Rural Samples

0.16
Urban
0.12
Rural
0.08
0.04
0
None Incomplete Complete Incomplete Complete Higher
primary primary secondary secondary
3% for women who have completed secondary education or entered

higher education.
The figures for urban and rural areas suggest that the relationship
between the mother’s education and child mortality differs with the
region of residence. The rural child mortality rate is higher at each level
of education, and the benefits of education do not seem to drop away so
sharply after the completion of primary school in rural areas.
Estimation Results
Table 2 presents the probit results. The marginal effects and z-statistics
are reported; the probit coefficients are not reported, but are available
from the authors on request.
The Endogeneity of Expenditure in the Child Mortality Regression.
Column 1 presents the results without instrumenting for the logarithm
of per capita expenditure and without the inclusion of provincial
dummies. A comparison of the results in column 1 with those in column
2, which instrument for expenditure, confirms the expectation that
expenditure is endogenous. When expenditure is not treated as
endogenous, the coefficient on expenditure is positive and statistically
significant. This suggests that the greater the level of expenditure per
capita in a household, the higher child mortality is likely to be. This is
clearly counter-intuitive.22 Once one addresses the possible endogeneity
of the variable, the coefficient remains statistically significant but becomes
negative. The marginal effect indicates that a household with 10% higher
current per capita household expenditure than another household has
on average a 0.97 percentage point lower probability of having had a
child die.23 The first stage regression is reported in appendix table A1.
The instruments were jointly significant (F-stat = 36.92; p-value = 0.000).
Education Variables. The results in column 2 of table 2 show that
additional years of either primary or secondary school education
significantly reduce the probability of child death. (Tertiary education is
negatively correlated with child mortality, but the relationship is not
significant.) In contrast, the literacy dummy is insignificant. If literacy is
the only control for education in the regression, it comes in as being
strongly negatively related to child mortality, but when we also control
for years of education, literacy is no longer statistically significant. This
suggests that literacy is an important determinant of child death, but
that it is more than just literacy skills that reduces the probability of child
death. Involvement in formal education, even beyond the years in which
literacy skills are acquired, is an important factor.24
TABLE 2 Weighted Probit Resultsa

(dependent variable = 1 if the woman had a child who died aged ≤5 years, 0 otherwise)
N = 6,620 (1) (2) (3)

Instrumenting Incl. Provincial
for Log Exp Dummies &
Instrumenting
Variable dF/dx z-stat. dF/dx z-stat. dF/dx z-stat.
Household level variables

Female education in years
Primary –0.019 –4.71 –0.018 –4.28 –0.017 –3.98
Secondary –0.026 –6.69 –0.021 –5.12 –0.020 –4.85
Higher –0.018 –1.65 –0.011 –1.02 –0.009 –0.82
Literacy dummy 0.006 0.29 0.009 0.42 0.008 0.38
Logarithm of annual 0.032 2.61 –0.097 –2.39 –0.126 –3.03
Family owned land –0.041 –2.48 –0.047 –2.88 –0.046 –2.75
Woman’s age: 25–29 0.124 5.98 0.118 5.72 0.116 5.58
30–34 0.229 11.03 0.223 10.72 0.225 10.75
35–39 0.284 12.66 0.273 12.16 0.271 11.95
40–49 0.361 14.02 0.350 13.56 0.350 13.40
Husband present –0.061 –1.77 –0.050 –1.46 –0.044 –1.30
Professional –0.034 –0.74 –0.023 –0.50 –0.023 –0.48
Administration 0.050 0.67 0.085 1.09 0.085 1.09
Clerical 0.037 0.74 0.054 1.05 0.064 1.23
Sales 0.016 0.36 0.023 0.52 0.026 0.57
Service 0.080 1.56 0.089 1.73 0.091 1.76
Agriculture 0.034 0.80 0.021 0.50 0.022 0.51
Industry –0.010 –0.23 –0.007 –0.17 –0.007 –0.17
Other –0.198 –1.99 –0.205 –2.10 –0.209 –2.23
Rural residence 0.040 2.44 0.005 0.24 0.011 0.51
Toilet facility in residence –0.066 –5.22 –0.052 –3.88 –0.044 –3.16
Piped water –0.079 –3.62 –0.067 –2.99 –0.065 –2.91
Religion: Buddhist –0.167 –2.72 –0.148 –2.25 –0.162 –2.52
Catholic –0.049 –2.25 –0.052 –2.37 –0.013 –0.44
Christian –0.065 –3.18 –0.070 –3.43 –0.074 –3.35
Hindu –0.130 –3.53 –0.134 –3.70 –0.186 –3.62
Other –0.148 –1.55 –0.151 –1.58 –0.165 –1.80
Provincial level variables
Dry season rainfall 0.006 1.45 0.008 1.94
Wet season rainfall –0.011 –3.91 –0.010 –3.79
Public health centres/100 –0.006 –2.48 –0.005 –2.12
Average immunisation rate –0.251 –5.22 –0.327 –4.36
Price of housing –0.000 –0.23 –0.000 –0.94
Price of clothing –0.003 –5.42 –0.003 –4.50
Price of general items 0.004 2.74 0.005 3.71
Price of rice 0.000 0.45 0.000 1.76
Pseudo-R2 0.12 0.12 0.13
a
The excluded categories are: aged 15–24; husband not working; Muslim. Note that the mar-
ginal effects correspond to a one-unit increase in continuous variables and a discrete change
from 0 to 1 for dummy variables.
An extra year of primary school education decreases the probability

of a mother experiencing child death by 1.9 percentage points. The
marginal effect for secondary schooling is slightly higher (2.6 percentage
points), but we cannot reject the hypothesis that an additional year of
secondary schooling has the same effect as an additional year of primary
schooling (p-value = 0.57).
These results are consistent with the findings of McDonald (1980)
and Martin et al. (1983), in that we find secondary education to be an
important determinant of child mortality. Like these studies, the point
estimates suggest that an additional year of secondary school might have
a slightly larger effect than a year of primary schooling, but, as stated
above, we cannot reject the hypothesis that the coefficients on primary
and secondary schooling are equal. The finding that primary schooling
plays a significant role is contrary to the results of the earlier studies.
Other Variables. As expected, there is a positive relationship between
maternal age and child death, and it is strongly statistically significant.
The estimated marginal effect on the probability of child death increases
monotonically through the age categories. A woman aged 25–29 has a
12.4 percentage point higher probability of having had a child die than a
woman aged 15–24. A woman aged 40–49 has a 36.1 percentage point
higher probability of one of her children having died. This is consistent
with the idea, discussed above, that older women have a higher risk of
child mortality for biological and reproductive cycle reasons. Additionally,
the effect of Indonesia’s development path is captured in these variables,
with younger families having greater access to health services, and the
quality of these services being higher.
Owning land decreases the probability of child death by 4.1 percent-
age points, and this is statistically significant. Land ownership confers
greater power over the family’s livelihood, and may be correlated with
more stable incomes. Husband’s occupation, included as another meas-
ure of socio-economic status, does not seem to be an important determi-
nant of child mortality. Only in the ‘other’ category is the relationship
statistically significant at the 5% level. Women with husbands in this cat-
egory have a 20.5 percentage point lower probability of having had a
child die. We experimented with also including husband’s educational
attainment, and found that the paternal education variables were statis-
tically insignificant.25 The presence of the husband in the household has
a negative but statistically insignificant effect on child mortality.
The results for the provincial price variables are mixed. These variables
were included to reflect differences in the cost of living across provinces.
They were thus expected to be positively correlated with child mortality.
The general price index has a positive effect and is statistically significant.
Higher prices, if not immediately compensated for by higher incomes,

reduce a household’s purchasing power and so may reduce expenditure
on inputs to child health. The price of rice, although also on average
being positively correlated with child mortality, is insignificant at the 5%
level (p-value = 0.08). This is somewhat surprising, given the role of rice
as a staple food. The clothing price index is also statistically significant,
but has a negative sign. Both of the price index effects are quantitatively
small.26
Infrastructure and Environment. We included the dry season and wet
season rains as a proxy for the prevalence of disease in each province.
Child mortality is higher in regions where there are high dry season rains
(although this variable has a p-value of 0.052) and lower in regions with
high wet season rains (p-value = 0.00). This suggests that a more
monsoonal climate may be beneficial for child health.
In terms of household infrastructure, we found that both having a
toilet and having piped water in the residence reduced the probability of
child death. The coefficients for presence of these facilities were strongly
statistically significant and quantitatively quite important. The presence
of a toilet decreased the probability of child death by 5.2 percentage points.
Piped water was even more important, reducing the probability of one
or more of a woman’s children dying by 6.7 percentage points.
The number of public health centres in the province also has a
statistically significant negative effect on the probability of a mother
experiencing child death.27 The coefficient on the variable is small,
however, with the addition of 100 public health centres reducing the
probability of child mortality by a factor of 0.5 percentage points. As
discussed above, this quantitatively small effect is difficult to interpret,
given the possible endogeneity of health centre placement.
The provincial immunisation rate has a large and strongly statistically
significant negative effect on the probability of child death. A woman
living in a province where there is no immunisation has a 32.7 percentage
points higher probability of having a child die than a woman who lives
in a province where all of the children are immunised.
The coefficients on the religion dummies indicate that there are
substantial religious differences in the incidence of child mortality.
Muslims are significantly more likely to experience child death than non-
Muslims. Catholics are the next most likely, then Protestants, Buddhists
and Hindus. The magnitude of the effect is quite large. For example, the
probability of having a child die is 7.0 percentage points lower for a
Protestant than for a Muslim.
Inclusion of Provincial Dummies

Column 3 of table 2 presents the results of the probit regressions in which
provincial dummy variables were used to absorb any across-province
differences. The coefficients on the other variables are largely invariant
to whether we control for all provincial differences using the dummy
variables or include provincial level explanatory variables. This suggests
either that the provincial level variables in column 2 were picking up a
lot of the across-province variation, or that provincial level differences
are largely uncorrelated with the other explanatory variables. Only the
coefficients on dummy variables that reflect some of the minority religions
change appreciably. This is not surprising, given that the religions differ
across geographic zones. There are still statistically significant differences
across religious groups, although there is no longer any statistically
significant difference in the prevalence of child death for Muslims and
Catholics. The coefficients and standard errors of the provincial dummies
are reported in appendix table A2. Geographic differences in child
mortality are found. The provinces of Lampung, North Sulawesi, Central
Java, Central Kalimantan and Yogyakarta are all associated with a
significantly lower likelihood of child mortality than the omitted province
of Jakarta. Among mothers in the sample, the probability of having had
at least one child die is actually highest in Bali, although this is not
statistically significantly different from that for Jakarta at the 5% level (p-
value = 0.081). Child mortality is significantly higher in West Kalimantan
and West Nusa Tenggara than in Jakarta. The differences across provinces
are large. For example, the probability of a mother in West Kalimantan
having one or more children die is 27.0 percentage points higher than
that for a mother in Yogyakarta.
Non-Time-Varying Variables Only

Our sample was restricted to women who had given birth in the last five
years in order to focus attention on relatively recent deaths, because the
data set measured the explanatory variables only in the survey year. To
assess the sensitivity of our results to the use of the current values of
time-varying variables, we estimated the regressions using only non-time-
varying variables, including provincial dummies. We also included the
variables that reflect the age of the woman, because these to some extent
capture the time path of mortality changes. Appendix table A3 shows
that the results of this experiment are very similar to those produced
with the full set of explanatory variables. Of the educational variables,
only primary and secondary education are statistically significant. The
estimated marginal effects of extra years of schooling at these levels are,
not surprisingly, slightly larger than when we include the full set of
explanatory variables. An extra year of primary school is now estimated
to reduce the probability of a woman experiencing child death by 2.2
percentage points (compared to the earlier 1.7 percentage points in column
3 of table 2) and a year at secondary school is estimated to lower the
probability by 3.3 percentage points (compared to the previous 2.0 points).
The education categories now pick up the effects of all of the omitted
time-varying variables with which they are correlated. So, for instance,
as education generally results in higher incomes and hence higher
expenditure, the education variables are now also picking up the negative
effect of expenditure on child mortality. To avoid this omitted variable
bias, because we expect the time-varying variables to be highly correlated
across time, and furthermore because, as mentioned earlier, the results
were found to be robust to restricting attention to just a subset of more
recent deaths, we proceed using both time-varying and non-time-varying
variables.28
Rural and Urban Comparisons

Table 3 presents the results for the estimation over the rural and urban
areas separately. Although the rural dummy is not significant in the
estimation over the full sample (once one endogenises expenditure), it is
likely that an additive dummy cannot effectively capture the difference
in the probability of child mortality between rural and urban regions.
The estimation results in table 2 estrict
r the coefficients on the variables,
other than the constant, to be invariant across these two regions. It may
be, for example, that female education plays a different role in rural and
urban areas. In fact a different relationship is suggested in figure 1, where
child mortality falls more slowly with female education in rural than in
urban areas. Estimating the equations over the separate samples allows
us to test this.
The pattern of benefits of female education is very similar in urban
and rural areas. Literacy per se is insignificant in both regressions, as is
tertiary education. Primary and secondary education significantly reduce
the probability of child death in rural and urban areas, and in both areas
we cannot reject the hypothesis that the benefit of an extra year of
secondary education is the same as that of an extra year of primary
education (as was concluded for the whole sample). However, the point
estimates suggest that the benefits of additional maternal education are
greater in urban than in rural areas. An extra year of primary (secondary)
education reduces the probability of child death by 1.7 (1.9) percentage
points in rural areas, and by 2.3 (2.0) percentage points in urban areas. A
TABLE 3 Rural–Urban Comparisonsa
Variable Rural Urban
dF/dx z-stat. dF/dx z-stat.
Female education (years)

Primary –0.017 –3.28 –0.023 –3.40
Secondary –0.019 –3.37 –0.020 –3.78
Higher –0.014 –0.68 –0.011 –1.07
Literacy dummy 0.002 0.08 0.022 0.60
Logarithm of annual –0.140 –2.60 0.0003 0.01
per capita expenditure
Family owned land –0.053 –2.81 0.0182 0.32
Number of public –0.003 –0.97 –0.011 –2.46
health centres/100
Average provincial –0.276 –2.92 –0.547 –4.37
immunisation rate
Toilet facility –0.062 –3.79 –0.023 –1.07
in residence
Piped water –0.033 –0.89 –0.070 –3.11
Dry season rainfall 0.012 2.38 –0.006 –0.71
Wet season rainfall –0.013 –4.01 –0.002 –0.39
Pseudo-R2 0.10 0.18
N 4,860 1,699
a
Log per capita expenditure has been instrumented for. These regressions also
controlled for all of the other variables shown in table 2. Only variables of specific
interest and of distinct difference between the models have been reported. The
results for the unreported coefficients are available from the authors on request.
possible explanation for this difference is that rural education is of lower

quality. 29 Excessively high student–teacher ratios and a shortage of
fundamental resources such as libraries in rural areas (relative to urban
areas) may reduce the impact on child mortality of years of education
completed. It may be that the effect of a year of a standardised quality of
education does not differ in urban and rural areas. An alternative
possibility is that students may receive education relevant to appropriate
child care practices in both regions, but that it may be more difficult to
implement these practices in a rural environment.
Coefficients on some of the other explanatory variables also differ

between rural and urban areas. Expenditure is only significant in rural
areas. Health care services and other facilities are more accessible in urban
areas, whereas greater expenditure may be required in rural areas to
compensate for isolation. 30 The family’s ownership of land is also
significant only in rural samples. Land ownership is likely to be of greater
importance in rural areas, owing to its role as a production input and its
effect on the level and stability of income streams. In urban areas, although
reflecting wealth, the ownership of land is less likely to have such a direct
impact on livelihood.
The infrastructure variables reflecting waste disposal and cleanliness
of water supply suggest interesting variations across urban and rural
areas. Both variables have negative signs for both samples, implying that
toilet and protected water connections tend to reduce child mortality.
However, the toilet connections variable is significant only in the rural
sample, whereas the protected water variable is significant only in urban
areas. Piped water may be more important in urban areas, where
congested conditions may lower the cleanliness of naturally available
water.
The rural areas were driving the coefficients on the weather variables
in the estimation over the entire sample. In the wet season higher rainfall
reduces the prevalence of child mortality in rural areas, and in the dry
season it increases it. Rainfall is not a significant determinant of child
mortality in urban areas in either season. This may reflect the effect of
weather conditions on agricultural output. Alternatively, rural diseases
may be more dependent on rain conditions than diseases that are
prevalent in urban areas.
The coefficient for the average immunisation rate bears a negative
sign and is significant in both areas, although quantitatively immunisation
is a much more important determinant in urban areas. Congested
conditions in cities may make immunisation (not just of one’s own but
also of neighbours’ children) an important factor in the spread of disease.
Another possibility is that immunisation programs may be more reliable
and of a higher quality in the city.
The number of public health centres is significant only in urban areas,
and the magnitude of the effect is much larger than when estimated over
the whole sample. The presence of an extra 100 health centres per province
decreases the probability of a mother losing a child by 1.1 percentage
points. The difference between rural and urban areas in the importance
of public health centres probably reflects clustering of health facilities in
urban centres. Rural residents benefit little from greater numbers of health
centres if they remain geographically isolated from such services.
Interaction of Public Health Programs with Female Education

One of the hypothesised benefits of higher levels of education among
mothers has been increased knowledge of the worth of public and private
facilities. For example, greater knowledge of the benefits of health and
sanitation services may cause more educated women to use these facilities
or to use them more effectively, and hence may lower child mortality
rates. Alternatively, education may serve as a substitute for community
health programs. Both relationships suggest that the variables reflecting
access to facilities should be interacted with the mothers’ years of
education.
Table 4 reports the results of regressions that include these interaction
terms. Only the coefficients on the interacted variables are reported,
although the full set of explanatory variables was included in the
regression. None of the interaction terms is statistically significant at the
5% level. The urban/rural results in table 3 suggested that it might be
worth including the interacted variables in separate rural and urban
regressions. This was done, but the interaction terms remained
insignificant in both cases. Hence we find no evidence that the benefits
of health infrastructure and immunisation are greater or less for women
with higher levels of education.
TABLE 4 Interaction of Female Education with Infrastructure Variablesa
Odds Ratio z-statistic
Total years of education interacted with:

Toilet 0.007 1.91
Piped water 0.010 1.72
Number of public health centres/100 0.000 –0.01
Average provincial immunisation rate 0.014 0.90
a
The other explanatory variables used in the previous regressions were also con-
trolled for in this regression. In addition, this regression allowed the effect of years
of education to differ across rural and urban areas.
CONCLUSION
Child mortality and the health of young children are important issues in
all countries, but particularly in developing nations, where child mortality
rates are high by international standards. The empirical results obtained
in this study suggest that increased investment in the human capital of
women is one way of lowering these mortality rates. In Indonesia, the
benefits are obtained via primary and secondary level education. An extra
year of maternal primary schooling is estimated to reduce the probability
of child death by 1.7 percentage points, and an extra year of secondary
schooling by 2.0 percentage points.
The separate regressions for urban and rural women display patterns
similar to those of the full sample. However, education in years completed
appears to have a greater influence over the probability of child mortality
in urban areas. This may be due to lower educational quality in remote
rural regions. Moreover, the relatively low educational attainment of
females in rural areas suggests that much can still be achieved in reducing
child mortality through rural education programs aimed at women.
Public health variables were found to be statistically significant, but
the impact differed by rural/urban residence. While immunisation rates
are significant in both regions, the benefits are greatest in urban areas. In
contrast, the number of public health centres per province is significant
only in the urban sample. Having a toilet in the household reduces child
mortality in rural areas, whereas in urban areas access to piped water is
more important.
The results obtained in this study indicate that encouraging women
to complete secondary education is likely to be effective in reducing child
mortality in Indonesia. The extent to which lowered child mortality results
in a consequent lowering of fertility rates and a slowing of population
growth is an interesting area for further research.
NOTES
* We thank Joe Hirschberg and Chris Worswick for their helpful comments and
assistance with the data, and acknowledge funding from Australian Research
Council. Any errors are our own. The authors may be contacted by e-mail at:
<[email protected]>.
1 Child mortality is defined in this study as the death of a child aged up to five
years. World Bank (1980) provides figures for deaths of children aged between
one and four years. These fell from 31 per thousand in 1960 to 20 per thousand
in 1978.
2 Ray (1998) and Dasgupta (1994) discuss these linkages in detail.
3 None of the empirical studies of which we are aware includes child mortality
as a determinant of fertility.
4 Hodinott and Haddad (1994) and Thomas (1990) find that mother’s income
has a positive effect on child anthropometric measures (weight for height,
height for age). Thomas (1997) and Quisumbing and de la Briere (1998) find
that households in which mothers have higher income spend a larger
proportion of the household budget on education. Doss (1997) finds that the
completed level of schooling of children is positively related to current assets
owned by women. Galasso (1999) finds that child labour is less likely in
Indonesian households where the woman would receive more assets if the
couple were to divorce.
5 Better educated mothers may be better able to process information and
therefore less likely to repeat the fertility behaviour of previous generations.
Lower child mortality rates may be factored into fertility decisions more readily
by women with higher levels of education. They may also place more
confidence in modern medicine and child care practices. In this way, education
can serve to reduce expected child mortality as well as actual child mortality,
thus lowering the ex ante response of fertility to child mortality.
6 The data collection was managed by Macro International Inc. and funded by
USAID.
7 Interestingly, current per capita expenditure has greater explanatory power
over deaths more than 10 years ago than over more recent deaths. This may
be due to the lower coverage of public health programs in years past and the
greater reliance on families’ own financial resources.
8 A lack of data corresponding to the period of death is common in studies of
this kind (Martin et al. 1983).
9 Restricting the sample in this way results in an over-representation of high
fertility women, because they are more likely to have had a child in any given
period. However, this is also the case in studies that use the child as the unit
of observation.
10 We do not attempt to differentiate between the determinants of infant mortality
and those of the death of older children. Benefo and Schultz (1996) and Barrera
(1990) studied child mortality and infant mortality separately, and found their
determinants not to be significantly different.
11 Another possibility was to estimate ordinary least squares with the proportion
of child deaths as the dependent variable. This has the disadvantage of not
restricting the proportion to lie between zero and one. We could, of course,
have estimated logits rather than probits, the only difference being that the
underlying distribution of the errors is assumed to be the cumulative logistic
distribution function. Probits were chosen because their estimation makes it
easier to calculate the marginal effect on the dependent variable of changes in
the explanatory variables.
12 Ideally equivalence scales would be used in order to take into account the fact
that consumption is likely to differ among household members. For example,
one would expect that children would consume less than their parents. The
use of equivalence scales is highly controversial, however, because of the
difficulty of determining appropriate scales.
13 No attempt has been made in this paper to correct any heteroscedasticity that
may arise from the use of the predicted regressor.
14 Even though the sample includes only women who have had a child in the
past five years, the dependent variable equals one if any of her children has
died before the age of five.
15 See Schultz (1984) for a study that recognises the endogeneity of the number
of children, and models child mortality and fertility simultaneously.
16 To see this, recognise that a zero value of the dependent variable for a woman
who has had eight children has a higher information content than a zero value
for a woman who has only ever had one child.
17 In practical terms, the weighted and non-weighted probit results differ only a
little.
18 All price data are for the capital city of each province, as published in BPS
(1995).
19 For instance, greater rainfall acts against water-borne diseases such as cholera,
although malaria is more prevalent during the wet season.
20 The rainfall data and the number of public health centres are the 1994 figures
from Statistik Indonesia (BPS 1995).
21 Because they are perfectly collinear with the provincial dummies.
22 The more likely causality is that if a child dies, there are fewer household
members and so expenditure per capita is higher.
23 Note that the interpretation of the marginal effect on this variable differs from
that for all of the other variables, because expenditure is measured in
logarithms. As a result, the marginal effect corresponds to a 100% increase in
the explanatory variable, as opposed to a one-unit increase, as is the case for
variables measured in levels.
24 In the data it is possible to be literate and not to have attended school, and
vice versa. The correlations between literacy and years spent in primary,
secondary and tertiary education are as follows: 0.80, 0.30, 0.08.
25 Specifically, we included dummy variables reflecting the husband’s highest
level of schooling as primary, secondary or higher, with and without the
inclusion of husband’s professional status. In both formulations, the coefficient
on husband’s education was positive but statistically insignificant.
26 The price data, like the other variables, are for the year 1994, and so may
differ from those that prevailed around the time of the offspring’s early
childhoods. We are implicitly assuming that the woman has not moved
provinces since the birth of her children and that price differences and other
provincial level variables are relatively stable across time and provinces. This
is a common problem in studies of this kind (Martin et al. 1983, for example),
and must be borne in mind when interpreting the results.
27 It could be argued that we should study the effect of the number of health
centres per head of population. We have used the absolute number of centres
because we believe that this better captures access to a health centre: individuals
who live in a sparsely populated province with a high number of centres per
head may not have easy access to a centre, because they are likely to live at a
long distance from one.
28 Although the results are robust to restricting attention to births in just the last
10 years, we prefer to report the results where the dependent variable reflects
all child deaths that a woman has experienced: it is somewhat arbitrary to
treat as having experienced child death a woman whose child died 10 years
ago and not one whose child died 11 years ago. Also, using all deaths results
in the regressions having much higher explanatory power.
29 See Hill (1996: 211–12) for an outline of problems relating to the quality of
education in Indonesia’s more remote areas.
30 Hill (1996: 213) notes that public funding of hospitals (normally based in urban
centres) absorbs a larger portion of the health budget than does funding of
rural health centres and services.
REFERENCES
Aly, H.Y., and R. Grabowski (1990), ‘Education and Child Mortality in Egypt’,
World Development 18 (5): 733–43.
Barrera, A. (1990), ‘The Role of Maternal Schooling and Its Interaction with Public
Health Programs in Child Health Production’, Journal of Development Economics
32: 69–91.
—— (1991), ‘The Interactive Effects of Mother’s Schooling and Unsupplemented
Breastfeeding on Child Health’, Journal of Development Economics 34: 81–98.
Benefo, K., and T.P. Schultz, (1994), ‘The Determinants of Fertility and Child
Mortality in Côte d’Ivoire and Ghana’, LSMS Working Paper no. 103, World
Bank, Washington DC.
—— (1996), ‘Fertility and Child Mortality in Côte d’Ivoire and Ghana’, The World
Bank Economic Review 10 (1): 123–58.
BPS (1995), Statistik Indonesia 1994, Jakarta.
Caldwell, J. (1979), ‘Education as a Factor in Mortality Decline: An Examination
of Nigerian Data’, Population Studies 33, November: 345–413.
Cho, L.J., S. Suharto, G. McNicoll and S.G. Made Mamas (1980), Population Growth
of Indonesia, University Press of Hawaii, Honolulu.
Cochrane, S.H. (1979), Fertility and Education: What Do We Really Know?, Johns
Hopkins University Press, Baltimore MD.
Dasgupta, M. (1994), ‘What Motivates Fertility Decline? A Case Study from the
Punjab, India’, in B. Egro and M. Hammarskjold (eds), Understanding
Reproductive Change, Lund University Press, Lund.
Doss, C. (1997), The Effects of Women’s Bargaining Power on Household Health
and Education Outcomes: Evidence from Ghana, Williams College,
Williamstown MA (mimeo).
Galasso, E. (1999), Intra-Household Allocation and Child Labor in Indonesia,

Boston College, Boston MA (mimeo).
Gertler, P., and J.W. Molyneaux (1994), ‘How Economic Development and Family
Planning Programs Combined to Reduce Indonesian Fertility’, Demography
31 (1): 33–63.
Hill, H. (1996), The Indonesian Economy since 1966, Cambridge University Press,
Cambridge.
Hodinott, J., and L. Haddad (1994), ‘Women’s Income and Boy-Girl Anthropometric
Status in the Côte d’Ivoire’, World Development 22 (4): 543–53.
Hull, T. (1993), 'East Java's Family Planning Revolution, 1961–1987', in H. Dick,
J. Fox and J. Mackie (eds), Balanced Development: East Java in the New Order,
South-East Asian Social Science Monographs, Oxford University Press,
Singapore, Oxford, New York and Toronto: 54–74.
Jensen, E.R. (1996), ‘The Fertility Impact of Alternative Family Planning
Distribution Channels in Indonesia’, Demography 33 (2): 153–65.
McDonald, P. (1980), ‘The Equality of Distribution of Child Mortality: Java–Bali,
1950–1976’, Bulletin of Indonesian Economic Studies 16 (3): 115–19.
Martin, L., J. Trussel, F.R. Salvails and N.M. Shah (1983), ‘Co-variates of Child
Mortality in the Philippines, Indonesia, and Pakistan: An Analysis Based on
Hazard Models’, Population Studies 37 (3): 417–32.
Mosley, W., and L. Chen (1984), ‘An Analytical Framework for the Study of Child
Survival in Developing Countries’, in W. Mosley and L. Chen (eds), Child
Survival: Strategies for Research, Cambridge University Press, Cambridge and
New York.
Quisumbing, A.R., and B. de la Briere (1998), ‘Women’s Assets and Intrahousehold
Allocation in Rural Bangladesh: Testing Measures of Bargaining Power’, Paper
presented at the NEUDC Conference, October 16–17, Yale University, New
Haven CT.
Ramanathan, R. (1995), Introductory Econometrics with Applications, 3rd ed., Harcourt
Brace, Fort Worth FL.
Ray, D. (1998), Development Economics, Princeton University Press, Princeton NJ.
Schultz, T.P. (1984), 'Studying the Impact of Household Economic and Community
Variables on Child Mortality', Population and Development Review 10 (0),
Supplement 1984: 215–35.
Singh, R. (1994), ‘ Fertility–Mortality Variations across LDCs: Women’s Education,
Labour Force Participation and Contraceptive-Use’, Kyklos 47: 209–29.
Thomas, D. (1990), ‘Intra-household Resource Allocation: An Inferential Approach’,
Journal of Human Resources 25 (4): 635–54.
—— (1997), ‘Incomes, Expenditures and Health Outcomes: Evidence on
Intrahousehold Resource Allocation’, in L. Hadda et al., Intrahousehold Resource
Allocation in Developing Countries: Models, Methods and Policy, Johns Hopkins
University Press, Baltimore MD and London.
United Nations (1991), The World’s Women 1970–1990: Trends and Statistics, United
Nations, New York NY.
World Bank (1980), World Development Report, Washington DC.
—— (1997), World Development Report, Washington DC.
APPENDIX TABLE A1 First Stage Regression

(dependent variable = logarithm of annual per capita household expenditure)
Explanatory Variables Coefficient t-statistic
Female education in years completed

Primary 0.0095 2.5392
Secondary 0.0194 3.4646
Higher 0.0401 5.0933
Woman’s age: 25–29 –0.0462 –2.0492
30–34 –0.0682 –3.1956
35–39 –0.0857 –3.9203
40–49 –0.0948 –2.2935
Family owned land –0.0349 –1.8725
Average provincial immunisation rate –0.5265 –5.1658
Number of public health centres/100 0.0132 0.3486
Administration 0.1425 1.7740
Agriculture –0.0537 –1.1560
Clerical 0.0160 0.2759
Industry 0.0087 0.1878
Professional –0.0055 –0.0870
Sales –0.0102 –0.1846
Service 0.0467 0.9469
Other –0.1258 –1.2678
Husband present –0.0031 –0.0803
Price of rice 0.0007 9.2388
Price of clothing 0.0026 2.8724
Price of general items 0.0112 4.6833
Price of housing –0.0002 –5.1725
Rural –0.1595 –6.9427
Toilet connection 0.0577 3.1753
Protected water source –0.0038 –0.1048
Dry season rainfall 0.0142 2.0037
Wet season rainfall 0.0009 0.3552
Religion: Buddhist 0.1767 2.7573
Catholic 0.0266 0.7797
Christian –0.0307 –0.9611
Hindu –0.0927 –1.8349
Other –0.0753 –0.6677
Woman headed householda –0.1004 –1.8236
Owns motor boat or motor cyclea 0.1465 7.0618
Owns television seta 0.1050 5.0358
Owns bicyclea 0.0004 0.0237
Has electricitya 0.1226 5.4071
Owns refrigeratora 0.3284 10.5020
Owns kerosene stovea 0.0868 4.2009
Pays mortgage/renta 0.0900 2.5463
Owns the buildinga –0.1046 –3.4716
Constant 10.6345 35.2900
R2 0.2831
a
Indicates the instruments.
APPENDIX TABLE A2 Coefficients and Standard Errors

on Provincial Dummy Variables
(corresponding to table 2, column 3; Jakarta is the omitted province)
Province dF/dx z-statistic
Bali 0.137 1.74

West Kalimantan 0.110 2.38
West Nusa Tenggara 0.105 2.32
West Java 0.080 1.95
Southeast Sulawesi 0.074 1.43
Bengkulu 0.054 1.15
Sth Kalimantan 0.053 1.07
Nth Sumatra 0.047 1.04
Maluku 0.034 0.70
Irian Jaya 0.033 0.62
Riau 0.030 0.71
East Nusa Tenggara 0.010 0.19
East Kalimantan 0.009 0.19
Central Sulawesi 0.004 0.09
West Sumatra 0.003 0.07
South Sulawesi –0.033 –0.75
Aceh –0.035 –0.84
Jambi –0.040 –0.86
East Timor –0.042 –0.84
South Sumatra –0.064 –1.48
East Java –0.070 –1.64
Lampung –0.106 –2.51
North Sulawesi –0.115 –2.48
Central Java –0.118 –3.18
Central Kalimantan –0.144 –3.40
Yogyakarta –0.164 –3.83
APPENDIX TABLE A3 Non-Time Varying Explanatory Variablesa
Variable dF/dx z-statistic
Female education (years)

Primary –0.022 –5.35
Secondary –0.033 –9.24
Higher –0.020 –1.87
Literacy dummy 0.004 0.20
Woman’s age
25–29 0.110 5.38
30–34 0.214 10.35
35–39 0.259 11.68
40–49 0.341 13.31
Pseudo-R2 0.12
N 6,620
a
Dummy variables reflecting the province of residence and religion of the house-
hold were also included.

Mellington 1999

Uploaded by

Copyright:

Available Formats

Mellington 1999

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mellington 1999

Uploaded by

Copyright:

Available Formats

Bulletin of Indonesian Economic Studies Vol 35 No 3, December 1999, pp.

FEMALE EDUCATION AND CHILD

Nicole Mellington and Lisa Cameron*

EDUCATION AND CHILD MORTALITY

409–10). The positive relationship between education and earnings may

household structure and expenditure patterns over the 12 months prior

where Xi is a vector of explanatory variables, β is the vector of coefficients

where t is a standardised normal variable. Maximum likelihood

variables. A dummy variable (literate), which equals 1 if the mother is

socio-economic status of the woman whether he is resident or not. We

To correct for this, we estimate a weighted probit where the weights

TABLE 1 Summary Statistics

Variable Mean Standard Min. Max.

Two estimation strategies were adopted. First, the regressions were

FIGURE 1 Mean Child Mortality by Educational Attainment

(a) Total Sample

(b) Urban and Rural Samples

3% for women who have completed secondary education or entered

TABLE 2 Weighted Probit Resultsa

N = 6,620 (1) (2) (3)

Household level variables

An extra year of primary school education decreases the probability

Higher prices, if not immediately compensated for by higher incomes,

Inclusion of Provincial Dummies

Non-Time-Varying Variables Only

Rural and Urban Comparisons

TABLE 3 Rural–Urban Comparisonsa

Variable Rural Urban

dF/dx z-stat. dF/dx z-stat.

Female education (years)

possible explanation for this difference is that rural education is of lower

Coefficients on some of the other explanatory variables also differ

Interaction of Public Health Programs with Female Education

TABLE 4 Interaction of Female Education with Infrastructure Variablesa

Odds Ratio z-statistic

Total years of education interacted with:

Galasso, E. (1999), Intra-Household Allocation and Child Labor in Indonesia,

APPENDIX TABLE A1 First Stage Regression

Explanatory Variables Coefficient t-statistic

Female education in years completed

APPENDIX TABLE A2 Coefficients and Standard Errors

Province dF/dx z-statistic

Bali 0.137 1.74

APPENDIX TABLE A3 Non-Time Varying Explanatory Variablesa

Variable dF/dx z-statistic

Female education (years)

You might also like