Why Does Mother's Schooling

Raise Child Health in Developing

Countries?
Evidence from Morocco

Paul Glewwe

ABSTRACT
Mother's education is often found to be positively correlated with child
health and nutrition in developing countries, yet the causal mechanisms
are poorly understood. Three possible mechanisms are: (I) Fonnal educa-
tion directly teaches health knowledge to future mothers: (2) Literacy and
numeracy skills acquired in school assist future mothers in diagnosing
and treating child health prohlems: and (3) Expo.sure to modern society
from forma! schooling makes women more receptive to modern medical
treatments. This paper uses data from Morocco to assess the role played
by these different mechanisms. Mother's health knowledge alone appears
to be the crucial skill for raising child health. In Morocco, such knowl-
edge is primarily obtained outside the classroom, although it is obtained
using literacy and numeracy skills learned in school: there is no evidence
that health knowledge is directly taught in schools. This suggests that
teaching of health knowledge skills in Moroccan schools could substan-
tially raise child health and nutrition in Morocco.

I. Introduction
Child health is a key indicator of the quality of life iti developing
coutitries. Mother's years of education is often positively associated with improved
child health and nutritional status (see Behmian. 1990). There are a variety of mecha-

Paul Glewwe is a Senior Economist in tbe Development Research Group at the The World Bank. He
would like to ihank Hanan Jacohy. Martin Ravallion, and two anonymous reviewers for helpful com-
ments on previous drafts, and Nauman llias for excellent computational assistance. This research was
supported by a grant from the World Batik Research Committee (RPO 679-84). The finding.^, interpreta-
tions and conclusions expressed in this paper are entirely those of the author. They do not necessarily
represent the views of the World Bank, its Executive Directors, or the countries they represent. The
data used in this article can he obtained beginning May, 1999, through April. 2002, from Paul
Glewwe. The World Bank, 1818 H Street NW. Washington. DC 204.11
(Submitted Ociober 1996; accepted February 1998]

THE JOURNAL OE HUMAN RESOURCES • XXXIV • /

 Glewwe 125

nisms through which tnother's education could raise child health: (1) Direct acqui-
sition oi basic health knowledge in school tnay provide future mothers with itifor-
mation jseful for diagnosing and treating child health problems; (2) Literacy and
numeracy skills learned in school may enhance mothers" abilities to treat child ill-
nesses, conditional on health knowledge, and also should help mothers increase their
stock of health knowledge after leaving school: and (3) Exposure to modem society
in general via schooling may change women's attitudes toward traditional methods
of raising children and treating their health problems.
This paper attempts to assess the relative importance of these three mechanisms,
using the 1990-91 Moroccan Enquete Nationale des Niveaux de Vie des Menages
(ENNVM). Knowledge of the relative importance of these mechanisms can have
important policy implications. For example, if the main impact of education comes
from directly raising mothers' basic health knowledge, such knowledge should be
taught m schools as early as possible (that is, before girls drop out) and perhaps
should aiso be taught in special education courses for women of child-bearing age
who have already left school
The paper is organized as follows. Section II reviews, in broad terms, the impact
of mother's education on child health and briefly reviews the recent literature. Sec-
tion III discusses the data and the estimation strategy. Section IV presents the empiri-
cal results. Section V decomposes the total impact of mother's schooling on child
health. Section VI summarizes the results.

II. Mother's Education and Child Health

A. General Discussion
Figure 1 provides a schematic framework for thinking about the determinants of
child health and nutritional status. As seen at the bottom of that figure, child health
is ultimately determined by three distinct sets of factors: 1. Health and nutritional
inputs provided by the household (arrow i); 2. The local health environment (arrow
/ ) : and 3. The child's health endowment (arrow h). Health and nutritional inputs
provided by the household include prenatal care, breastmilk, breastmilk substitutes
such as infant formula, calories from adult foods (for weaned children), medicines,
and medical care. The quality of household drinking water sources, toilet facilities,
and other hygienic conditions can also be considered as health and nutritional inputs
provided by the household. The local health environment consists of all community
characteristics that directly affect child health and are generally beyond the control
of the parents, such as prevalence of parasites and the incidence of contagious disease
among the general population. Finally, the child health endowment consists of all
components of the child's genetic inheritance that have implications for his or her
health.
Household health and nutritional inputs are determined by household decisions
that reflect the characteristics of the household, the local community, and the child,
such as (initial) household assets, parental schooling, community economic and
health-related characteristics (such as the availability and prices of medical services),
and each child's health endowment. This paper focuses on parental schooling, partic-
126 The Journal of Human Resources
 Glewwe 127

ularly mother's schooling; father's schooling, apart from its income effect, is less
likely to be important for maintaining children's health,
Schullz (1984) argues that mother's education can influence child health in five
ways:' (I) Education may lead to a more efficient mix of health goods used to pro-
duce child health; (2) Better educated mothers may be more effective at producing
child health for a given amount and mix of health goods; (3) Schooling can affect
parents' preferences in systematic ways—for example, educated mothers tend to opt
for fewer but healthier children; (4) More schooling should raise family incomes,
either through higher wages or increased productivity in self-employment, which
should improve child health status; and (5) Education raises the opportunity costs
of time, which tends to increase the time mothers spend working outside tbe home
and thus reduce time for child care—this effect of .schooling could reduce child
health by reducing both maternal time devoted to child care and duration of
breastfeeding. In Figure 1, the third and fourtb pathways are represented by the
arrows ii-a' (and also by a-a") and acd (via a-a'" and c-c'"), respectively.' The first
two pathways, which reflect the direct effect of the health knowledge and cognitive
skills that education impans. have received little attention in the literature. What is
it about schooling that makes mothers more efficient in producing child health?
Figure I presents two mechanisms through whicb schooling could influence tbe
cboice of health and nutritional inputs via tbe knowledge and skills it provides.^
First, schools may directly teacb effective bealth care practices to students. This
pathway is denoted by h-abc. For example, the impact of diarrhea on child health
can be reduced by oral rehydration therapy (ORT), which can be taught even in
primary schools (see Cash 1983). Second., schooling can influence child health inputs
through the cognitive skills imparted, such as literacy and numeracy. Literate moth-
ers are better able to read written instructions for treating of childhood diseases, and
numeracy enables mothers to better monitor illnesses and apply treatments. Tbis
direct effect is shown by c-c" in Figure 1, Literacy and numeracy also enable mothers
to increase tbeir health knowledge by enabling them to gather information from
written sources. Tbis indirect effect is path c-c'-abc in Figure 1.
Figure 1 also depicts how factors other than schooling influence child bealth.
Household physical assets raise household incomes (arrow d).. wbich should bave
a positive effect on both nutritional inpuls {such as calories) and environmental con-
ditions around the home. The choice of health and nutritional inputs will also be
affected by factors associated with the .supply of these inputs in the community
(arrow e). For example, the availability and quality of bealth and non-health commu-
nity facilities affects tbe decisions households make regarding health and nutritional

1, SchuHz's framework is primarily concerned vvith child mortality. Yet broadening it lo include olher.
less sevi;re. aspects of child health docs not require significani modification,
2, The -ift!i pathway, via mother's time, could be added to Figure 1 but i.s omitted to reduce clulter.
Similarly, the impact of the ihird pathway via reduced family size could also be made more explicit (as
a box labeled "family size." another endogenous variable), but this is aiso omitted to reduce clutier,
3, The distinction between pathways {\\ and (2) in the previous paragraph is tbat tbe first concerns an
efiicieni mix of pbysical healtb inputs (for example, medicines) wbile the second adds efficient use of
non-physical inputs (such as care given lo the sick cbild). In this paper. Ihe distinction between physical
and non physical inputs is noi of primary interest, riither the emphasis is on ibedifferent types nf knowledge
and skills learned in school, and bow tbey affect efficient use of both types of inputs.
128 The Joumal of Human Resources

inputs. Finally, the child's health endowment will also affect household bealth and
nutritional inputs (via arrow g), since more sickly cbildren usually receive larger
amounts of health and nutritional inputs.

B. Recent Empirical Evidence

Many recent studies have examined tbe impact of mother's (and father's) education
on cbild health. For comprehensive reviews of the literature see Behrman and Deo-
lalikar (1988). Behrman (1990), and Strauss and Thomas (1995). The discussion
here will be limited to an overview of a few recent studies, focusing on tbe impact
of motber's education on beight-for-age and weight-for-heigbt.
Studies of tbe determinants of child height and weight in many countries have
found positive effects of motber's education. Most of these studies presented reduced
form estimates, but a few went further, examining the pathways by which mother'.s
education improves cbild bealth. In the Philippines, Barrera (1990) found that better
educated mothers tended to wean tbeir children sooner, but they compensated for this
shortened breastfeeding time with better care; overall, tbeir children were healthier as
measured by higher height-for-age ^-scores.'* The only published study that focuses
on tbe "information processing" attributes of schooling is by Thomas. Strauss, and
Henriques (1991), which used Brazilian data tbat included variables for wbetber a
woman reads a newspaper, listens to the radio, or watches television. Motber's
schooling was not significant when dummy variables were included for these "infor-
mation processing" activities; the newspaper and radio variables were significant in
rural areas but only the television variable was significant in urban areas.
Among the most interesting studies are those based on tbe Cebu Longitudinal
Health and Nutrition Survey. Several studies have used these data to model the patb-
way.s by which exogenous variables influence child nutritional status and morbidity.
Tbe Cebu Study Team (1991, 1992) found tbat motber's education leads to improved
waste disposal and bigher non-breastmilk calorie intake, botb of which reduce the
incidence of diarrhea. Maternal education also leads to earlier weaning, wbicb can
increase episodes of diarrhea, but the net effect of maternal education is to reduce
the incidence of diarrhea.
An important critique of findings that mother's education improves cbild health
is tbe hypothesis that education simply reflects unobserved maternal cbaracteristies.
Wolfe and Behrman (1987) used Nicaraguan data on mothers' siblings to control
for unobserved family fixed effects. They found that applying these controls leaves
no significant effect of mother's education on child anthropometric status. However,
Strauss (1990) found in Cote d'lvoire that mother's education raises child heigbt-
for-age and weight-for-heigbt. even after using family fixed effects estimators.
In summary, there is con.siderable evidence that mother's education improves child
bealth. and some evidence on how this occurs. Still, there are no studies that distin-

4, Height for age ;-scores, whicb will be used in Ibe empirical work below, are based on fitting a standard
normal distribution to tbe growth curves of a bealthy populalion of cbildren, A cbild with a ;-score of
zero is exactly at the median in terms of beigbt for age, wbile children with positive (negative) z-scores
are taller (shorter) than average. Low height for age ;-scores indicate ,stunting due lo repeated episodes
of malnutrilion over the life of tbe child, while low weight for height ;-scores indicate wasting (weight
loss) due to a current episode of tnalnutrition (see Gibson 1990),
 Glewwe 129

guish between the literacy and numeracy impacts of schooling and other, more gen-
eral, impacts. Also, tbere are no studies that attempt to assess directly the impact
of mother's healtb knowledge on child healtb.

III. Analytical Framework, Data and Estimation

Strategy

A. Analytical Framework
Estimation of tbe pathways by which mothers' scbooling affects child healtb is not
necessarily straigbtforward. This subsection provides a framework for thinking about
how to estimate these relationships. Recall Figure 1. The bottom of that figure sbows
how healtb and nutritional inputs, tbe environment and a cbild's bealth endowment
jointly determine cbild healtb. This can be expressed in terms of a production func-
tion for child health:
(1) H,^f{HI,,E,,t,)
where Hi is tbe health of child /, ///, is a vector of health inputs chosen by cbild /'s
household. E, is a vector summarizing the environmental conditions surrounding
child /. and e, is tbe child's genetic healtb endowment. Parents take tbis technological
relationship into account as best they can wben making decisions that affect tbeir
children's healtb. Altbough E, and e, are outside tbe household's control,^ health and
nutritional input choices are chosen by the household.
Estimation of Equation 1 would require detailed information on a large number
of healtb inputs, whicb is not feasible witb the 1990-91 ENNVM data. However,
as seen in Figure I, one can substitute out tbese bealth inputs and obtain a reduced
form relationship tbat shows bow exogenous variables (those sbown at the top of
Figure 1) determine child health:^

(2) ;•/, = g{ES;, MS,, HAi, Ei, E,)

wbere F5, and A/5, are fatber's and mother's scbooling, respectively, and HA, is tbe
initial assets of cbild i's household.
Although Equation 2 is mucb easier to estimate, and often bas heen estimated, it
does not indicate what aspects of motber's schooling lead to improved child health.
Refernng again to Figure 1, one can obtain a better understanding of the impact of
mother's scbooling by replacing it in Equation 2 with tbe educational outcomes it
directly affects, namely cognitive skills, parental values and bealth knowledge:

5, The local health environment is not chosen by parenis if; a) migration for purposes of fmding a belter
bealtb (Environment is rare: and b) bousebolds cannoi pressure local autborities to improve the local bealtb
environment, Tbe former assumption is supported by migration data from tbe 1990-91 ENNVM; only
0,5 percent of respondents report tbat "bealtb reasons" were tbe main reason for their most recent move.
The lalter assumption, while harder to check, is plausible for Morocco because healtb care provision is
highly cenlralized, witb tew funds under tbe control of local govemmenis (see World Bank 1994),
6, The assumpiicin that parental education is exogenous seems reasonable for Morocco, where average
scbooltng for men and woman between the ages of 18 and 65 is only 4,7 and 2,3 years, respectively. Even
so, thi;. assumption will be cbecked in Section IV,
130 The Joumal of Human Resources

(3) H, = h{ES,. L,, N,, V,, HK,. HA,. £ „ e,)

where L,, A',, V, and HK, stand for mother's literacy, numeracy, values and health
knowledge, respectively. This equation is a conditional demand relationship because.
as explained below, health knowledge, and perhaps literacy and numeracy, may be
endogenous.
Estimation of Equation 3 would clarify the pathways by which mother's schooling
affects child health, but such estimation is complicated by .several problems. First,
It is difficult to observe mother's values (V^), and indeed there are no such data in the
1990-91 ENNVM survey. Second, a child's health endowment (e,) is also virtually
impossible to observe. Third, because parents" treatment of their children's health
problems often causes them to acquire additional health knowledge, health knowl-
edge is likely to be an endogenous variable. In particular, heaith knowledge is likely
to be negatively correlated with a child's (unobserved) genetic health endowment
because parents with "healthy" children need not acquire as much health knowledge
as parents with "sickly" children, ceteris paribus.' Fourth, it is also possible that
literacy and numeracy are endogenous because actions to acquire additional health
knowledge may lead to greater use of those skills, though the impact of a child's
genetic health endowment on these variables is likely to be considerably smaller
than its impact on health knowledge. The approaches taken to deal with these prob-
lems will be discussed in detail in Subsection IIIC below.
Another relationship of interest is a variation of Equation 3; when attempting to
assess the pathways by which mother's education affects child health one may wish
to isolate its impact on household income.** In this case one can add household income
iVi) to Equation 3 and remove household assets (since their impact on child health
would operate only through their impact on household income). This yields the fol-
lowing conditional demand relationship for estimation:

(4) H, - h\ES,. L,. N,, V,. HK,, K^, £ „ e,).

The relationships of primary interest in this paper are Equations 3 and 4.
A final issue to consider is tbe possibility that mother's education improves child
health by reducing the number of children women bear—with fewer children, the
mother should be able to allocate more time and health inputs per child. As men-
tioned in Section il. the desire for fewer children can be depicted as the impact of
schooling on parental values. Thus, one could modify Equations 3 and 4 by replacing
Vi with the number of children bom or, more generally, by adding the latter variable
while retaining V, (since values may affect child health in other ways). Of course,
since the number of children bom is clearly an endogenous variable, estimation re-

7. This could he shown in Figure I by an arrow leading from the child's health endowment lo parental
health knowledge.
8. One could go even further. Increased education can raise household income noi only hy increasing
wage rates but also by increasing Ihe amount of time the molher works outside the home. Moreover,
increased lime of the mother away from home may have a direcl, negative impact on child health. Thu.s
one could add hoth household income and mother's time speni working to Equation 4. This was tried in
an earlier version of this paper (see Glewwe 1997). but it proved impossible to find instrumental variables
that could plausibly be excluded from Equation 4 and were also good predictors of mother's lime spent
working. In this paper mother's time spent working has been substituted oui of Equation 4.
 Glewwe 131

quires plausible instrumental variables; this will be discussed further in Subsection

IIIC.

B. The Data
This paper uses data from the 1990-91 Enquete Nationale sur le Niveau de Vie
des Manages (ENNVM), which was implemented by Morocco's Direction de la
Statistique. The survey, which is based on the World Bank's LSMS surveys, covered
3,323 households from all areas of Morocco. The survey collected a variety of infor-
mation from each household, including household expenditures and income, employ-
ment, education, assets, agricultural activities, and much more. A key aspect of these
data for this paper is that they contain the height and weight of all household mem-
bers. Even more important is that a battery of tests was given to household members
in two thirds (2,171) of the sampled households. The tests included: I. Five questions
on health knowledge; 2. Twelve questions on general knowledge (how to mail a
letter, how to read an electricity bill, and so on); 3. An oral mathematics test of ten
questions; 4. A set of written mathematics tests of varying degrees of difficulty;
5. A set of Arabic reading and writing tests; and 6. A set of French reading and
writing tests. The tests are described in detail in Glewwe (1997). The health knowl-
edge test is of particular interest, since it is rarely a part of any household survey.
It consists of five questions on vaccinations, treating infections, polio, diarrhea and
safe drinking water. The test is fully described in Appendix I.
All persons in the 2,171 selected households between the ages of 9 and 69 were
to be lested except: 1. Individuals with a baccalaureate degree** or higher level of
education took only the health knowledge test since it was assumed that they could
obtain nearly perfect scores on all other tests; and 2. The health knowledge test was
taken only by individuals between the ages of 20 and 50. The 2,171 households who
participated contained 1,612 children age 5 or younger, of which 81 had mothers
who did not participate in the tests for one reason or another, leaving a sample of
1,531 children. It is assumed that the 39 mothers with a baccalaureate degree would
have received perfect scores on all the tests (except the health test, which they did
lake), which boosts the sample size to 1,570. Dropping observations with missing
values leaves a sample of 1.495 children.
Table 1 provides descriptive statistics on alt variables used in the analysis. Of
particular interest are the test score variables, which are defined as the number of
questions correctly answered by the respondent. They show substantial variation,
which is necessary to assess the underlying pathways by which mothers' schooling
raises child health. In addition, these scores should not be highly correlated with
years of schooling, or with each other; if they are, regression analysis is less likely
to identify the underlying mechanisms. Table 2 shows correlation of years in school
with the test scores (the table also includes a test on reading a medicine box—this
will be discussed in Section IV). Mathematics, French and Arabic scores are all
highl) correlated with each other and with years in school (correlation coefficients

9. Roughly speaking, a baccalaureate degree lies somewhere between a U.S. high school degree and a
college degree. It is only awarded alter passing a rigorous set of examinalions.
132 The Joumal of Human Resources

Table 1
Descriptive Suui.sTics of Variables Used

Standard
Variable Mean Deviation

Height for age Z-score -0.94 1.86

Per capita expenditure 5,398.82 5,081.68
Sex of child (female) 0.53 0.50
Age of child (in months) 35.86 20.00
Mother's height 157.02 6.20
Father's height 168.49 5.95
Father's height missing 0.31 0.46
Mother's years schooling 1.98 2.97
Father's years schooling 3.14 4.10
Health knowledge 2.89 1.57
General knowledge 1.53 3.18
Oral mathematics 1.71 1.90
Reading and writing mathematics 1.18 2.62
Arabic reading 2.33 4.98
Arabic writing 0.57 1.54
French reading 1.48 4.44
French writing 0.26 1.06
Rental income 1,460.79 13,920.17
Children overseas 0.01 0.09
Irrigated crop land (hectares) 2.36 12.62
Unirrigated crop land (hectares) 30.88 94.59
Tree crop land (hectares) 0.32 1.42
Mother's married sisters 1.88 1.56
Father's married sisters 1.51 1.56
Father's married sisters mi.ssing 0.22 0.41
Father bom here 0.64 0.48
Number of televisions 0.55 0.58
Number of radios 0.89 0.55
Availability of newspapers 0.21 0.41
Mother's father's schooling 0.03 0.16
Mother's mother's schooling 0.01 O.IO

Sample Size: 1,493.

 Glewwe 133

Table 2
Correlation Among Schooling and Test Score Variables of Mothers

Reading
Years Arabic French Health Box of
Schooling Literacy Literacy Numeracy Knowledge Medicine

Years schcoling 1.0000

Arabiu literacy 0.8938 1.0000
French literacy 0.8869 0.8681 1.0000
Numeracy 0.8665 0.8695 0.8383 1.0000
Health knowledge 0.3343 0.3867 0.3138 0.4356 1.0000
Reading bi)x of medicine 0.8152 0.8553 0.7775 0.8198 0.4387 l.(K)O0

No[e: All variables are in logarithms.

from 0.84 to 0,89).'" Health knowledge is less highly correlated with these other
variable;; {correlation coefficients from 0.31 to 0.44). Whether regression analysis
can distinguish between the impacts of the most highly correlated skills is uncertain
and will become clear only by examining estimation results.

C Estimation Strategy: General

Consider estimation of equation (4) using the 1990-91 ENNVM." Child health
can be measured by the Z-score of child height for age (see endnote 4), which indi-
cates chronic malnutrition over a child's lifetime (stunting). In principle, one
could also use weight for height, but the 1990-91 ENNVM weight data suffer
from serious measurement error because weight was recorded only to the nearest
kilogram.
Several estimation problems arise conceming the explanatory variables in Equa-
tion 4. Begin with the child's local health environment. £,. Although the 1990-
91 ENNVM data contain information about local health clinics and other pertinent
variables, some of the data are missing and comparability is a problem. In addition,
the sampling procedure used for choosing the health facilities covered by the survey
is unclear. Because the main interest of this paper focuses on household level vari-
ables (the skills and education levels of mothers), a simple community fixed effects
procedure is used to avoid bias caused by omitted community level health environ-
ment variables- This is possible because ihe sampled households were drawn from
140 primary sampling areas.
Of the remaining explanatory variables in Equation 4, two are unobserved: mater-
nal values acquired in school (V,) and the child's health endowment (e,). If the effect

10. For simplicity, iti the remainder of this paper (the logarithms of) the (wo mathematics scores are
summed t J create a single mathematics variable, and the reading and writing scores are summed for French
and Arabic. For an analysis of tbe mere disaggregated scores, see Glewwe (1997). the lindings of wbich
are basicilly tbe same as tbose in tbis paper.
11. The tollowing paragraphs also apply to Equation 3, except tbe discussion on cboosing instrumental
variables for bousebold income i.s irrelevant (housebold Income has been substituted out of Equation 3).
134 The Joumal ol" Human Resources

of schooling on maternal values is important., one could detect this by adding years
of schooiing to Equation 4. A positive effect of years of schooling on height for age
would indicate that values (or perhaps some other aspect of schooling other than
literacy, numeracy, and health knowledge) is an important determinant of child
health. If the years of schooling variable has no perceptible effect, it is unlikely
that values acquired by mothers from schooling is ati important pathway by which
schooling affects child health.
The inability to directly observe a child'.s health endowment (e,) could lead to
biased parameter estimates due to its correlation with observed variables. One way
to reduce such bias is to enter the heights of both parents as explanatory variables,
since taller parents are likely to have better health endowments, which in tum are
inherited by their children. In addition, parents (and their children) display variation
in height that is not related to health status—healthy people can vary in height.
Entering parental height in the regre.ssions controls for this as well, purging the
dependent variable of variation in height that is not indicative of health status.
Note that father's height is missing for about one third of the children, either be-
cause the father did not live with the children or was unavailable for measurement
at the time of the interview. To avoid losing this portion of the sample, which
could lead to sample selectivity biases, a dummy variable is created indicating
that father's height was missing; in such cases father's height variable is set equal
to the mean.
Even after adding parental height to reduce bias caused by unobserved child health
endowments, it is still possible that the inability to observe children's health endow-
ments could bias estimated impacts of observed variables. Health knowledge, house-
hold income, and perhaps literacy and numeracy, may well be endogenous. Particu-
larly worrisome is the fact that health knowledge may be negatively correlated with
the child's unobserved health endowment. In principle, using instrumental variable
methods can remove bias, but this requires plausible in.stmments. Household assets
can be used to instrument curTent income.'^ The 1990-91 ENNVM contains data
that can be used to construct several household asset variables The following are
used in this paper: 1. Three variables on agricultural land (in hectares) owned by
the household; 2. Household rental income (from land, buildings and durable goods);
and 3. The number of adult children of household members living overseas (who
may send sizable remittances).
Finding instmmental variables for mother's health knowledge is more difficult.
Three different types seem plausible: indicators of the existence of close relatives
who could be sources of health knowledge; exposure to mass media; and mother's
education (which can be excluded from Equation 4 if one finds that it is nol needed
as a proxy of the impact of values on child health). One way that mothers can acquire
health knowledge is from close relatives, especially those who have had children.
The idea here is that mothers consult with other relatives conceming their children's
health, and by doing so they add to their stock of health knowledge. Most of these
relatives do not directly care for the mother's young children, except perhaps the
patemal grandmother, so their impact on the child's health comes about only by

\2, In the estimates given in later sections, household expenditure is used insread of household income
t)ecausc it is likely to be more accurate and more ctosely relaled to households' permanent incomes.
 Glewwe 135

raising the mother's health knowledge (Ihus they can be excluded from Equations
3 and 4l.'- The 1990-91 ENNVM contains data on the number of married sisters
of the mother and of her husband. It can also be used to add a dummy variable
indicating whether the husband was born m the current place of residence; if he was,
there should be several members of his family in the area from whom his wife can
obtain health knowledge. Finally, the education of the mother's own parents could
also affect her health knowledge (and her cognitive skills as well), but can be ex-
cluded Irom Equations 3 and 4 because the mother no longer lives with her parents.
Mass media is also a useful source of health knowledge information (see Thomas,
Strauss and Henriques, 1991), It is unlikely that these variables have any effect on
child health apart from their impact on mother's health knowledge; in particular, it
would be rare for parents to purchase televisions or radios in respon.se to having a
sick ch-ld. The ENNVM data collect data on the number of radios and televisions
in the household and on the availability of local newspapers. These mass media
variables could also be used as instmments for numeracy and literacy.
Finally, this paper will also investigate whether part of the impact of mother's
education on child health works by reducing family size, and since family size is also
endogenous one needs instmmental variables for children ever bom. It is particularly
difficult to find plausible instruments for this variable. Some possibilities are the
number of married sisters of both the woman and her husband, which could reflect
preferences for children on both sides of the family, the education levels of the
womar's parents, which again may reflect family preferences for children, and finally
the age of the woman, since older women will have had more time to bear children.
Of course, one can imagine plausible reasons for why these variables may directly
affect child health, but there are no better instmmenta! variables available from this
data set.
The instmmental variables described in the previous paragraphs generally appear
reasonable, but one cannot prove that they do not belong in Equation 4. This problem
plagues most, if not nearly all. applications of instrumental variables. Thus one
.should apply a specification test to check the plausibility of the underlying exclusion
restrictions. This is done using standard overidentification tests (see Davidson and
MacKinnon 1993).

D. Estimation Strategy: Decomposing the Separate Impacts of

Mother's Education
Finally, consider how estimates of Equations 3 and 4 can be used to assess the
mechanisms by which mother's schooling leads to improved child health. As seen
in Equation 4, and Figure I. the overall impact of mother's schooling on child health
can be decompo.sed as follows:

13. In Moroccan culture, when wonen marry they join their husband's household. Thus the mother moves
away from her parents and sihlmgs, and her husband's married sisters have moved away from his family.
Only tlie husband's mother, his unmarried sisters, and ihe wives of any married brothers he may have
belong to his {and thus to his wife s) household.
136 The Joumal of Human Resources

dHK. ^ 3 W A ; 3 L ^ ^ dHK,^\ ^ a//, BY,

dHK,\dMS, dL^ BMS, dN, d

The first, second, third, fourth, and fifth terms to the right of the equal sign show
the impact of mother's schooiing via its impact on literacy, numeracy, values, health
knowledge, and income, respectively. Note that the impacts via literacy and numer-
acy (the first and second terms) are direct effects only (arrow c" in Figure I); the
indirect effects via the impact on health knowledge (arrow c' in Figure I) are repre-
sented by the second and third terms inside the parentheses of the fourth term. A
similar decomposition based on Equation 3 is identical except that the income term
is dropped; the remaining partial derivatives may differ from those in Equation 5
because the income effect is "divided u p " among the remaining terms (for example,
the impact of literacy now incorporates both a direct effect and the indirect effect
of literacy through its effect on household income).
Assume simple linear functional forms for Equations 2, 3, and 4:

(2') //, = |3,, + ^,FS, + P3A/5, + p,W/\, + H,

(3') H, = Y,) + liFS, + Y:^, + l^^i + YJK + y',HK, + y,HA, + u,
(4') H, = 5n + 5,F5, + 5:L, + b.N, + b,V, + b^HK, + b^Y^ + u,.

In each specification, an error term is added to account for random measurement

error in child height and the child's unobserved health endowment (recall that com-
munity fixed effects estimation allows one to drop E^ from each equation). In Equa-
tion 2', pi estimates the overall impact of mother's schooling on child health,
dHildMSi, which is the left hand side of Equation 5. Estimates of ail the partial
derivatives on the right hand side of Equation 5 come from: I. Equation 4. or Equa-
tion 3 if the income term is dropped from Equation 5; 2. Reduced form estimates
of the impact of mother's schooling on literacy (L,), numeracy (A',), values (V,),
health knowledge (HK,), and household income (Y,). which can be denoted as a^,,
a^-, Ov, a^^, and ay., respectively; and 3. Estimates of the impact of literacy
and numeracy on health knowledge, which can be denoted as Ti^ and r\n. respec-
tively.'"
Thus Equation 5 becomes:

(5') p, - S.a^ + 8,a,v + 840^ + b^ian^ + n i a ^ + ^^oi„) + b^ay.

This decomposition is based on Equation 4' and thus explicitly accounts for the
impact of mother's education on income, if income is substituted out, as in Equation
3', the last term is dropped from Equation 5 and the decomposition becomes:

{5") P: - Y 2 a t + YJOCA' + Y^oCv + y^iCLHK +

14, These impacts (Tit and il,,) are shown as arrow c' in Figure 1, If literacy and numeracy can be consid-
ered exogenous with respect to health knowledge, namely, literacy and numeracy change very little after
one leaves school but health knowledge can change, then this relationship is a reduced form. On the other
hand, if literacy and numeracy are endogenous then this is a conditional demand function, Thi.s will be
further di,scus.sed when this relationship is estimated in Section IV.
 Glewwe 137

Estimation of the different parameters of Equations 5' and 5" allows one to examine
not on ,y which pathways are important, but also to assess their relative contributions.

IV. Estimation Results

A. Reduced Form Estimates of the Determinants of Child Height
Table 3 presents ordinary least squares (OLS) and community fixed effects (FE)
estimates of Equation 2, that is reduced form estimates of the determinants of child
height for age.'^ The OLS atid FE estimates in Table 3 are similar, but specification
tests favor the fixed effects specification."' Because of this, all remaining estimates
in this paper will incorporate community fixed effects.
Table 3 shows that mother's years of schooling has a significantly positive effect
on child height, which is consistent with evidence from other countries discussed
in Set:tion 11.'^ This is the main finding of interest in Table 3; the rest of this paragraph
briefly discusses the other explanatory variables. A dummy variable for female chil-
dren • s negative but insignificant, yielding no strong evidence of gender discrimina-
tion in child health. Child height for age varies substantially with child age (in
months); this refiects common pattems whereby malnutrition rises with age in the
first two years of life (until weaning ends) but then levels off (and may even decline).
Both mother's and father's height are positively correlated with height for age. which
picks up both variation in the child's unobserved health endowment (e,) and natural
variation in height among healthy children. Father'.s years of schooling is positive
but insignificant, which suggests little direct effect on child height. Perhaps a weak
indir;ct effect exists via household income; this will be checked below. The house-
hold asset variables are all insignificant, though three of the iive have the expected
signs,.
Before moving to conditional demand estimates, it is useful to check whether
mother's years of schooling can really be specified as exogenous, and whether part
of the effect of mother's schooling on child health works through reduced family
size. To investigate the former, mother's years of education was instrumented using
the educational levels of both her parents, as well as the number of married sisters
she has. These instruments were good predictors of mother's schooling (joint F-
statistics over 20 for both OLS and fixed effects estimates, with p-values of 0.000).

15, All years of schooling and lest score variables are ,specirted in logarithmic form because: 1, It seems
reasonable to assume that attainment of the most basic skills would have larger impact than would attain-
ment of additional skills among {arsons who already have basic skills; and 2, tn general, taking the logs
of these variables almost always fit the data better (as measured by R" statistics). If years of schooling or
any test .score is zero, the log of it is set to zero, and the same applies to rental income,
16, A Hausman te.stof fixed effects versus the null hypothesis ofrandom effects yields a X'(d,f, 13) statistic
of 20.29, which is statistically significant at the 10 percent level. Since Hausman lests often have low
power to reject the null, il is prudent to reject ihe random effects specification. In tum, that specification
(not shown in Table 3, but similar to the OLS results) is favored over OLS: the Breusch-Pagan Lagrange
multiplier tesE has a X'^fd,f, 1) statistic of 46,70, clearly rejecting the null hypothesis of homoscedaslicity.
17, In regressions not shown here, a squared term of (the logarithm of) mother's years of schooling was
added to both specifications in Table 3 to allow for a more flexible functional form, Tbe linear term
remained significant (at the 10 percent level), but the squared term was completely insignificant. In the
rest of this paper only the linear term is used.
138 The Joumal of Human Resources

Table 3
Reduced Form Estimates of Determinants of Height for Age Z-Scores

Community
OLS . Fixed Effects

Variable Coefficient f-Statistic Coefficient ir-Statistic

Constant -16.931 -10.83

Sex (female) -0.149 -1.69 -0.116 -1.31
Age (months) -0.081 -9.14 -0.076 -8.55
Age- (months) 0.001 7.70 0.001 7.11
Mother's height 0.066 8.95 0.052 6.39
Father's height 0.041 5.39 0.041 4.96
Father's height missing 0.155 1.61 0.167 1.60
Log mother's schooling 0.226 3.17 0.165 1,97
Log father's schooling 0.076 1.43 0.050 0.82
Log rental income 0,020 1.03 -0.016 -0.74
Children overseas 0.517 0.99 0.659 1.21
Irrigated crop land -0.001 -0.28 0.001 0,14
Unirrigated crop land 0.001 1.21 0.000 0.61
Tree crop land -0.038 -1.23 -0.014 -0.39
R^ 0, 170 0,299
Lagrange multiplier test (OLS vs. random effects), x ' ( l ) = 46.70. /)-value --- 0 . 0 0 0
Hausman test (random effects vs. fixed effects), x'C-^) = 20.29, /7-value = 0.088

Sample size: 1.495.

The impact of mother's .schooling on child health was substantially lower, and not
significantly different from zero for both OLS and fixed effects estimates. However,
the standard errors were much larger (0.256 for OLS and 0,374 for fixed effects),
so that standard Hausman tests could not reject the hypothesis that the instmmental
variable estimates were equal to those given in Table 3.
Regarding the impact of education on family size, simple reduced form OLS esti-
mates (not shown here) confirm that mother's years of schooling has a strong and
significantly negative impact on family size. Unfortunately, the only instrumental
variable that had a significant impact on fertility was mother's age. This was true
for two distinct fertility variables tried, number of children and age of mother at
first birth. The fixed effects specification in Table 3 was reestimated adding each
variable separately, and both instrumented and uninstrumcnted versions were tried.
In each case, the fertility variable was not statistically significant and the parameter
estimate for mother's years of school showed little change. Thus it appears that
although mother's education does reduce family size, there is no strong impact of
family size on child health after controlling for mother's education.
 Glewwe 139

B. Conditional Demand Estimates of Child Height (not Conditioning

on Income)
Table 4 presents estimates of Equation 3, the conditional demand determinants of
child height where income is not one of the conditioning variables. Three kinds of
estima;es are shown: community fixed effects (FE); two stage least squares with
community fixed effects (2SLSFE) with only health knowledge specified as endoge-
nous; and 2SLSFB with all four skills test (mathematics. Arabic. French, and health
knowledge) specified as endogenous. The last two kinds were e.stimated twice, once
with mother's education and once without it. for reasons explained below.
The FE estimates shown in Table 4 are identical to the FE estimates in Table 3,
except that the four test score variables have been added as (exogenous) explanatory
variables. Mother's schooling is now completely insignificant, and even slightly neg-
ative, which suggests that the four test score variables fully capture the impact of
mother's education on child health. However, although all four test score variables
show positive effects none is even close to being statistically significant, and neither
are they jointly significant. These results are quite puzzling. One problem with this
specification is that health knowledge, and perhaps the other test score variables, may
be endogenous. The remaining specifications in Table 4 use 2SLSFE specifications to
address this problem.
Tht: second and third columns of Table 4 are identical to the first except that health
knowledge is specified as endogenous, using the instrumental variables described in
subsection IIIB, The one difference between the second and third columns is that
the former includes mother's schooling (to account for the impact of values) while
the latter excludes it (to provide an additional instrumental variable for health knowl-
edge). The main finding when health knowledge is specified as endogenous is that
such knowledge has a strong and statistically significant impact on child health.
Indeed, in the specification that includes mother's education the point estimate in-
creases 50-fold and is significant at the 1 percent level."* When mother's education
is excluded from tbe specification the impact of health knowledge is somewhat
smaller, but still highly significant (/-value of 2,48), None of the other test score
variables (mathematics, Arabic, and French) have significantly positive effects, and
in fact both Arabic and mathematics have negative effects.
Ore issue that arises at tbis point is what is meant by health knowledge. The five
questions that comprise the healtb knowledge test are shown in Appendix I. Could
it be ihat what really matters is only a subset of these questions? Of tbe five questions,
the one conceming vaccinations provided very little information because 95 percent
of the women answered it correctly. The percentage of women answering the other
four questions correctly varied between 44 percent and 54 percent. These four ques-
tions were also fairly highly correlated with each other, witb correlation coefficients
between 0.40 and 0.50. To check whether some questions mattered more than others,
the regression in tbe second column of Table 4 was reestimated five times, each
time with the response to a single question replacing the health knowledge variable.

18. Al the suggestion of one reviewer, separate regressions were run by sex and age (the iwo age categories
being 0-35 months and 36-71 months) to see whether the impact of health knowledge varied by sex or
age. Althotjgh there were some differences (the impact was larger for girls than for boys, and larger for
yotiiii;er children), they were not statistically significant.
140 The Journal of Human Resources

o —' o o o o O O O
I I I I I I I I I

g ^ c^
^ t^ o^
!3
C — O O O O O O O O — O O O O O
II ^11 — I I I I I

CJ ^ -^ '-^ m '-, ro — O
m 00 00 tn r^ r-,
Tf rf o 00 o "*
L O O O O O o — CJ o o o <D -rt d
I I I I I ^ 1 1 1
C O

TJ" .^-^ f^\ ,..—.., ^ \ ^-K f^ ^^^ T ^ '•^^ ^ ^

'—' r*J CJ ^f o*^ V"^ [—' f"*, •/*! c ^ CJ o ^ o -^
c p r-^ p 00 p
™— m —^ ^ 5 '—' m rn o*^ rn ^f r^ 'JOM;—'
p r^ p p
O o — o — o — o — o — ci o o o vo o O rn O "Tf
I I ! I " ' 1 1 ^ I I I I

CN 00 ^
'•— t^ ^ m ^ "/^ c-l — oo m
O r N O ^ O ' ' f O ( N 3
o o o o o c o o o o o o o
I I I I

bo

s
kno
ing

00
n "o •5
o DO
ths

OJ
J= 'S
(A X! e
x:
(moi
UOUl

in VI
'Uc 'Uc *i-
W u U 'u
x: ^^ -^^ u
-s OJ u
00
o
x:
rt U- <: <
 Glewwe 141

\o o
n 1 3O o >n ^ CC
O ' ^ " " — " O w ^ O O O l —
^ 1,1,1 - w

O —. CD ,^~, m ^-N
c i r- — O O c^ m Cl ^ o r- o
CJ w-i o O \C oc o t^ o
O o o o o o o
I I I 1 I

O Cl S^o o
O t~~ C l IT) —
p p —
Os p
o —o o wS d

ciO^O—'mv^O'^iO
o —d — d d d d d
^11 ""II

w «j

C
o u
- ^0^-10^
p ^ j
C3

^1

c
JZ u
0
C o c
JZ -a o
ati

u
T3 J2
— o T3
•S rt
x: D. t; _rt
OJJ
o _ ap O <U
^ -S
=3 2 00 <
•S c
U X
142 The Joumal of Human Resources

As one would expect given its lack of variation, the answer to the vaccination ques-
tion had no significant effect on child health. In contrast, each of the four other
questions had a significantly positive effect on child health, with r-statistics ranging
from 1.71 to 1.99 and parameter estimates ranging from 1.24 to 1.79. This suggests
that the health knowledge embedded in each of these questions is important, and
because the different types of health knowledge are fairly highly correlated with
each other they may well reflect the impact of other types of basic health knowledge
not measured in this test. In addition, these results also suggest that specifying the
overall impact of health knowledge as the sum of the scores on this test is a reason-
able way to aggregate health knowledge into a single variable.
Given how the results change when health knowledge is specified as endogenous
in the second and third columns of Table 4, it is worthwhile to apply some specifica-
tion tests to these regressions. In general, instrumental variables must not be corre-
lated with the error tetm of the equation of interest (H; in equation 3'), and they
must provide strong explanatory power for the endogenous variable(s). The first
requirement can be checked by an overidentification test. In both columns the exclu-
sion restrictions are not rejected (the /j-values being 0.254 and 0.120 in the second
and third columns, respectively). The second requirement is verified by F-tests on
the explanatory power of the identifying instruments, the null hypothesis of no ex-
planatory power is resoundingly rejected {see Appendix II for the first stage regres-
sions and partial R- statistics, the latter of which are recommended by Bound, Jaeger,
and Baker 1993). Finally, a Hausman test is used to examine whether the 2SLSFE
specification is preferred to the FE specification. This test rejects the FE specifica-
tion.''' Overall, the specification tests indicate that healtb knowledge should be treated
as endogenous and that the instrumental variables used satisfy both requirements.
There is one more set of regressions to check before concluding that health knowl-
edge is tbe most important pathway by which mother's education leads to improved
child health. Perhaps the mathematics. Arabic and French test score variables in
Equation 3 are also endogenous, so that wben they are specified as .sucb during
estimation they will also yield significant impacts on cbild health. This is examined
in the last two columns of Table 4. using the same instrumental variables used for
health knowledge. The basic results are unchanged—health knowledge has a large,
positive and statistically significant impact on child health, and none of tbe other
variables does. This is true regardless of whether mother's education is included as
an explanatory variable.
More specifically, when mother's education is included as a regressor (Column
4) Arabic has a small (relative to the impact of health knowledge) positive effect.
but it is completely insignificant {/-statistic of 0.16); wben mother's education is
excluded (Column 5) the impact is slightly negative and even less significant
((-statistic of -0.00). French language skills have implausible negative effects in
both Columns 4 and 5, and are completely insignificant. Finally, the impact of mathe-
matics is stronger {in terms of the size of the coefficient) than Arabic or French, but

19, This Hausman test examined only the coefliclent on health iaiowledge in order to increase ihe power
of the test to reject the null hypothesis. Ali Hausman tests in the remainder of this paper are applied only
to the parameters associated with potentially endogenous variables, for the same reason, Hausman tests
were also mn on the etitire set of explanalory variables; in every case they failed to reject the null hypothe-
sis, which probably reflects their low power when jointly testing a large number of parameters.
 Glewwe 143

it is still statistically insignificant {f-statistics of 0.92 and 0.57 in Columns 4 and 5.

respectively). To see whether the statistical significance of mathematics skills might
improve: if only it and health knowledge were instrumented, the regression in Column
4 was reestimated (not shown here) with only these two test score variables as endog-
enous. Tbe coefficient on mathematics skills drops to half its value {0.402) and the
r-statistic is correspondingly much lower (0.55), Tbus specifying mathematics. Ara-
bic, anc, French as endogenous does not change the basic results found when only
health tjiowledge is treated as endogenous.
Tbe specification tests done for Columns 2 and 3 were also done for Columns 4 and
5. Both specifications passed the overidentification test, indicating that the exclusion
restrictions are reasonable and the explanatory power of the identifying instruments
is quite good {especially when mother's education is not used as a regressor). Unlike
the spet:ifications when only health knowledge is endogenous. Hausman tests of the
joint endogeneity of the four test score variables only weakly reject the FE specifica-
tion (p-values of 0.065 and 0.073 for Columns 4 and 5, respectively). In fact, a
Hausman test of tbe joint endogeneity of the mathematics. Arabic, and French test
scores does not reject the null hypothesis of exogeneity (the X' statistic with 3 degrees
of freedom is 3.83, which has ap-value of 0.280). Thus it is not clear that the specifi-
cation with all test scores endogenous is preferable to that with only healtb knowl-
edge endogenous.
A final issue to consider regarding Table 4 is wbether part of the effect of mother's
education works through reductions in the number of children born. Using age of
the mother, and its square, as additional identifying variables, tbe regressions in the
last four columns of Table 4 were reestimated twice, once adding the number of
children and again trying age of mother at first birth. Neither variable was significant
at the 5 percent level, and the impact of mother's healtb knowledge did not change
appreciably (thougb in one of the eight cases it lost statistical significance). More-
over, the same results hold when tbese two variables are treated as exogenous. Tbus
there is no evidence that most, or even some, of the effect of mother's education
on child health operates through reductions in family size.
To summarize the results of Table 4. it appears tbat healtb knowledge is tbe main
pathway by which mother's education leads to healthier children. This was seen
only when mother's health knowledge was specified as endogenous; treating it as
exogenous greatly underestimated its impact on child health. Numeracy and literacy
skills Tiever showed any significantly positive impact, whether they were specified
as endogenous or exogenous. In tbe tbree specifications that included mother's
schooling as a regressor, its coefficient was never significant, whicb casts doubt on
the hypothesis that an important pathway by which mother's schooling affects child
health is by changing the values of the mother.

C. Conditional Demand Estimates of Child Height (Conditioning on Income)

Tum now to tbe final pathway to investigate, tbat via household income. Table 5
shows three regressions that estimate Equation 4, the conditional demand for child
healtb that includes income as a conditioning variable. Tbe FE estimates in the first
column are in many ways similar to those in Table 4; in particular, although all four
test score variables show positive effects none is close to being statistically signifi-
144 The Joumal of Human Resources

3 «=

o o o o o o o o o —
3 .2i I

<3 V.
a- X
00 U.

-J 3

S5

1
1
•T3 O
« C ,^ o
GO (N
d r- O t^ o
O O O O O — l o -^ o o o r-
IS I '"' 1 1 1

O — - O O ^ O ^ N O l — ^ O f ^ O f N f N
o — r~- n ^
— CN p q
d o o c J d d o d d d o d d o — d od
II ' 'II' '
I 1 I I

s;
capi

"o
o

T3
C

a.
Sex

oo
0- X <
 Glewwe 145

(6)
00 m fM ns
o O 1. -) O OS
O n-J O m o — ON o

O
d

3S
u
P2 bO
•n
M

c „
a
o
ssi

-Si
'S c 5
o C u
ttO u
•o
'S 'C '^ 'l-l
c 'o
u
TD
u "C </i
o
TO
:S
rt
u
>•
B £ < < a.
[T
o
146 The Joumal of Human Resources

cant, and mother's schooling has no explanatory power. Household expenditures per
capita has a significantly positive effect, as expected. Of course, household income,
health knowledge, and possibly mathematics. Arabic and French, may be endoge-
nous. The second and third specifications in Table 5 allow for this.'''
The second column of Table 5 treats both health knowledge and household per
capita expenditures as endogenous, using the instrumental variables described in
Subsection HIB. As in Table 4, health knowledge has a significantly positive effect
on child health, but the other test scores are insignificant and/or negative. Household
income has a positive effect that is significant at the 10 percent level, and more than
twice as large as the effect shown in the FE estimates of Column 1.-' The same
specification tests shown in Table 4 are also shown here. The overidentification test
does not reject the exclusion restrictions implied by the choice of instruments, and
those instruments have strong explanatory power. Finally, the Hausman te.st shows
that the 2SLSFE results are significantly different from the FE results (p-value of
0.026). Thus, assuming that mother's education affects household income (which is
verified below), there is a pathway other than health knowledge by which mother's
education affects child heaith.
The third column of Table 5 treats all test score variables, as well as per capita
expenditures, as endogenous. As in Table 4, this does not change the general finding
that health knowledge is the key skill that educated mothers possess that raises their
children's health. The specification tests show that the instrumental variables appear
reasonable, and the Hausman test does not support the hypothesis that all five endog-
enous variables are indeed endogenous. Indeed, when only the mathematics, Arabic,
and French test scores are specified as endogenous, the Hausman test cannot reject
the null hypothesis of exogeneity (the %' statistic with 3 degrees of freedom is 2.57.
which has a p-value of 0.463). Note finally that the coefficient on per capita expendi-
tures is almost the same in Column 3 as in Column 2. though no longer significant
at the 10 percent level (the /statistic is 1.45). Overall, the results in Table 5 show
that household income is another pathway by which mother's education can affect
child health (and the only pathway by which lather's schooling affects child health,
since the first stage regressions in Table Al of Appendix II show a significant impact
of father's schooling on household expenditures). The relative magnitudes of the
different impacts of mother's schooling will be examined in Section V.

D. Health Knowledge or the Ability to Read a Medicine Box?:

A Brief Digression
As mentioned in Section III. another test given to respondents in the 1990-91
ENNVM was one on "general knowledge." This test consisted of 12 questions

20. In addition to the three specification.s shown, analogous regressions were run in which mother's educa-
tion was iitnitced as an explanatory variable. They were very sitnilar to the results shown here atid thus
are tiot presented k» reduce clutter in this table.
21. To check for nontlnearities. this regression was repeated with a squared expenditures term added
(not shown here). The squared term was completely insignificant. Because adding a squared term has the
disadvantage of adding another endogenous variable lo the regression, ihc squared term was not used in
the estimation results reported here.
 Glewwe 147

conceming everyday life in which a common object is handed to the respondent and
two or three questions are asked about it. The objects were a national identity card,
a letter, a box of medicine, a newspaper and an electricity bill. Of particular interest
for this study are the questions conceming a box of medicine. The three questions
asked were: 1. Where does it show how many pills are in the box?; 2. Where are the
instructions for using the medicine?; and 3. Where is the date of expiration indicated?
Presumably, mothers who are more able to answer these questions correctly can
provide better health care for their children.
For purposes of this paper, Iwo questions arise. Do the medicine box questions
measure something that the other tests do not? And if they do, what are the implica-
tions regarding the relationship between mothers' health knowledge and child health?
A mother's ability to read a medicine box may simply reflect her ability to read
Arabic (all writing on the medicine box was in Arabic), the impact of which on
child health has already been examined. Altematively, it may be that the ability to
read a medicine box is a fomi of health knowledge. In this case, it would be interest-
ing to examine whether reading a medicine box is the "most important" kind of
health knowledge (in which case that ability would "displace" most of the explana-
tory power of health knowledge) or whether it plays only a small role (and thus
would not "displace" the explanatory power of health knowledge). The correlations
shown il Table 2 suggest that the ability to read a medicine box may be little more
than another version of the test for Arabic literacy, since the correlation coefficient
between these two variables is 0.85. However, it is also fairly highly correlated with
health knowledge (correlation coefficient of 0.44). Further investigation requires re-
gression analysis.
The role of the ability to read a medicine box is examined in Table 6. In order to
focus on the relationship between that ability and health knowledge, the mathematics,
French and Arabic test score variables have been omitted (recall that they had no
explanatory power in Tables 4 and 5). The first two columns show FE estimates,
with (Column I) and without (Column 2) mother's schooling. As usual with FE
estimation, mother's health knowledge is completely insignificant. However, moth-
er's ability to read a medicine box is significant at the 10 percent level when mother's
schooling is included as a regressor and nearly at the 1 percent level when mother's
schooling is excluded. Thus this variable appears to be capturing more than just the
ability to read Arabic, which was never significant in any of the previous regressions.
Because the previous regressions presented fairly convincing evidence that moth-
er's health knowledge is endogenous, that variable should be instrumented. In addi-
tion, it is prudetit, and intuitively plausible (parents with sickly children have more
experience reading medicine boxes), to specify the ability to read a medicine box
as endogenous. These regressions are shown in Columns 3 and 4 of Table 6. When
mother s education is included in the regression, both mother's health knowledge
and the ability to read a medicine box have much larger effects." While neither is
significant at the 5 percent level, health knowledge is significant at the 10 percent

22. Note that mother's ability lo read a medicine bottle has not been transformed into logarithms because
the original variable ranges only frotn 0 to 3. Moreover, iransforming it to logs wotild have eqtiated re-
sponses i)f 0 and I. which logelher accounted for S5 percent of the responses.
148 The Joumal of Human Resources

Table 6
Conditional Demand Estimates of Height for Age Z-Scores, Health Knowledge
and Ability to Read a Medicine Box
Two Stage Least Squares
(Including Community
Fixed Effects)

Health Knowledge
and Ability to
Community Read Medicine
Fixed Effects Box Endogenous
Father's schooling 0.028 0.027 -0.205 -0.105
(0,48) (0.47) (-1.82) (-1.19)
Mother s health knowledge 0.034 0.036 1.563 2.223
(0.39) (0.41) (1.69) (2.82)
Mother's ability to read a medicine box 0.164 0.149 1.670 -0.120
(1.87) (2.53) (1.40) (-0.83)
Mother's schooling -0.027 — -1.743 —
(-0.22) (-1.51)
Sex (female) -0.114 -0.114 -0.088 -0.084
(-1.33) (-1.33) (-0.79) (-0.78)
Age (months) -0.076 -0.076 -0.071 -0.074
(-8.25) (-8.26) (-6.30) (-6.96)
Age^ (months) 0.001 0.001 0.001 0.001
(7.13) (7.14) (5.15) (5-90)
Mother's height 0.051 0.051 0.036 0.037
(6.67) (6.68) (3.15) (3.33)
Father's height 0.041 0.041 0.038 0.040
(5.55) (5.55) (3.65) (3.97)
Father's height missing 0.155 0.156 0.202 0-278
(1.55) (1.57) (1.37) (2.07)
Rental income -0.015 -0.014 -0.017 -0.022
(-0.58) (-0.58) (-0.59) (-0.83)
Children overseas 0.692 0.688 0.968 0.414
(1.12) (1.11} (1.25) (0.63)
Irrigated crop land (hectares) -0.000 -0.000 -0.001 -0.002
(-0.00) -0.00) (-0.18) (-0.36)
Unirrigated crop land (hectares) 0.000 0.000 0.001 0.001
(0.77) (0.77) (1.48) (1.39)
Tree crop land (hectares) -0.014 -0.014 0.013 0.059
(-0.41) (-0.41) (0.22) (1.18)
Overidentification tests (d.f.) — — 5.89(7) 8.99(8)
[0.553] [0-343]
F-tests of identifying instruments
Mother's health knowledge — — 4.41 7.66
[O.OOOJ [0-000]
Ability to read a medicine bottle — — 2.37 100.08
[0.012J [0.000]
Hausman tests (d.f.) — — 9.59(2) 7.83(2)
(Endogenous parameters only) [0.008] [0.020]
Notes:
1. Sample size is 1,473.
2. Asymptotic f-statistics shown in parentheses.
3. P-values of specification tests shown in brackets.
 Glewwe 149

level, and the eslimaled coefficient is about the same magnitude as it was in the
regressions in Tables 4 and 5. Note, however, that mother's education has a very
large negative coefficient, with a ?-statistic of - 1 . 5 1 . This suggests possible col-
inearit) problems. When mother's education is dropped from the regression the
impact of Ihe ability to read a medicine box becomes completely insignificant,
and even becomes slightly negative, while health knowledge remains strongly
significant.
Overall, the results shown in Table 6 do not alter the conclusion reached above
that health knowledge is the key aspect of mothers education that leads to improved
child health. Although the ability of mothers to read a medicine box initially appeared
to contain information not picked up in health knowledge, 2SLSFE estimation casts
serious doubt on this proposition. In contrast, as long as health knowledge is specified
as endogenous it yields statistically significant results (at least at the 10 percent level)
and the parameter estimates are fairly stable.
To summarize Section IV, Ihe fundamental result is that mother's health knowl-
edge is the key mechanism by which mother's education leads to improved child
health. A second finding is of a more methodological nature: ignoring the endogene-
ity of mother's health knowledge may seriously underestimate its role in promoting
child health.

V. Decomposing the Impact of Mother's Schooling

on Child Health

The finding in the previous section that mother's health knowledge

plays by far the most important role in detennining child health does not necessarily
imply that literacy and numeracy do not matter. As explained in Section II, mother's
heaith knowledge may develop after leaving school, and in a way that will depend
on mother's literacy, and perhaps on numeracy as well. This section attempts to
decompose the overall impact of mother's schooling on child health by decomposing
the parameter 3 ; given in Equation 2' in Section III.
Recall from Section Hi that the overall effect of mother's education on child health
in (2'), p2' t:an ^^ decomposed in two ways. If one explicitly includes the pathway
that operates via household income, the decomposition is that given in Equation 5'
of Section III. If income effects are substituted out, the decomposition is that in
Equation 5". The results in Section IV indicate that the direct effect of literacy,
numeracy and values on child health (that is the coefficients §2, 5, and 84) on child
health were not statistically significant from zero. The 2SLSFE estimates in Table
5 indicate that the impact of household income on child health (measured by Sg)
was approximately equal to 0,59.'' Finally, the best point estimates for the impact

23. Tables 4, 5. and Al present several specifications. Based on the results of the previous section, the
specifications used in this section are those where: !. Heath knowledge is endogenous but Arabic. French,
and mathematics skills are exogmous; and 2. The direct effect of years of schooling on child health is
consirained to equal zero. An altemative is to he more agnostic, taking averages across different specifica-
tions, yet doing so produced results very similar lo those given here.
150 The Joumal of Human Resources

of health knowledge on child health (65) are 1.44 when income effects are explicit
and 1.45 when income effects are substituted out.
To complete the decomposition of p; one needs estimates of the impact of school-
ing on health knowledge {either directly via a^A or indirectly via r\iai and Ti^^a^)
and the impact of mother's schooling on hou.sehold income (ay). Table Al in the
Appendix provides a point estimate of 0.153 for a^. The remaining parameter esti-
mates are shown in Table 7. The first three columns present reduced form estimates
of the determinants of health knowledge under the assumption that Arabic, French,
and mathematics scores can be considered as exogenous.-'' The regression in column
1 shows that Arabic and mathematics skills have significantly positive impacts on
health knowledge. French skills have an unexpected negative effect, but this is sig-
nificantly different from zero only at the 10 percent level. Finally, after controlling
for these effects years in school has a significantly negative effect. Taking these
estimates at face value implies that a^*; = -0.185, T;^ = 0.184 and the two parts
of T\i are 0.088 and -0.052 for Arabic and French, respectively. Estimates of the
impact of years of schooling on literacy (a^) and numeracy (a^-) skills are shown
in columns 4 - 6 of Table 7. Briefly, a,v ^ 1.268 and the two parts of a^ are 1.734
and 1.266 for Arabic and French, respectively.
Using the estimates of a^K. r\, and a^ given in the first column of Table 7 yields
an estimate of |3; of 0.196, as shown in the first row of the last column of Table 8.
Given the simple functional forms used and the imprecision of the estimates, this
is surprisingly similar to the estimated figure of 0.165 from the FE reduced form
estimate in Table 3. The other columns in Table 8 show how this is decomposed
according to Equation 5". Perhaps the most unusual finding is that years in school
does not raise health knowledge; indeed, it has a strongly significant negative effect.
Is this plausible? It may be. In Moroccan schools, basic health knowledge is not
part of the standard curriculum,'- so one should not be surprised that the impact
of schooling is not positive. Even a negative impact may occur—because school
attendance reduces the time girls .spend at home with their mothers, it may reduce
opportunities for them to acquire health knowledge at home. That is, time girls .spend
at home is an omitted variable that is negatively correlated with girls' schooling;
schooling itself does not reduce health knowledge, but it implies an allocation of
time that results in lower health on knowledge.
Turning to the rest of the decomposition, the main avenue by which schooling
raises health knowledge is by raising Arabic and mathematics skills, particularly the
latter, which can in tum be u.sed to acquire health knowledge. In contrast, French
skills have a small negative impact, based on a parameter that was significantly
different from zero only at the 10 percent level. While the positive impact of Arabic

24. Intuitively, while it is plausihie thai a sickly child will increase parents' health knowledge, there is
less reason to think that parents obtain greater literacy and numeracy skills in response to bouts of sickness
in their children. This is consistent with the findings in Section IVB; the Hausman tesis clearly rejected
the exogeneity of health knowledge but could not reject the joint exogeneity of mathematics. Arabic and
French skills. Finally, as a practical matter it is very difficult (o find instmmental variahles for these three
skills in the 1990-91 ENNVM thai do not also affect the acquisition of health knowledge.
25. This statement is based on discussions with Worid Bank staff who have worked on health and education
issues in Morocco.
 Glewwe 151

literacy is plausible (literate women can acquire health knowledge be reading various
lnpwritten materials), the positive impact of numeracy is less intuitive. It is probably
the case that mathematical skills help mothers monitor their children's illnesses and
more accurately apply medicines and treatments. In addition, it may be that mathe-
matical skills develop mothers" abstract reasoning abilities, which in turn helps them
to orgarize and refine the health knowledge they acquire.
Given the high colineadty between schooling, literacy and numeracy, as shown
in Table 2, more precise estimates might be obtained by dropping insignificant vari-
ables. Thus in the second regression reported in Table 7 the French literacy variable
was dropped. The resulting decomposition is shown in the second line of Table 8.
Overall, the results do not change very much. In particular, the significantly negative
impact of years in school does not "go away." Dropping years of schooling as well,
which is hard to justify econometrically, was done in the third column of Table 7,
and the associated decompositions are shown in the third row of Table 8. The overall
result is not very satisfying because now Arabic skills have no significant effect on
health knowledge, and the estimate of p; shown in the last column of Table 8 (0.273)
is much larger than the estimate given in Table 3 (0.165).
The bottom half of Table 8 examines decompositions based on Equation 5', which
includes the impact of income. The first row (that is, the fourth row of Table 8) is
based on the first regression in Table 7. The total impact of education is estimated
to be 0.285. This is also much higher than the estimate of 0.165 in Table 3, so the
decompositions underlying it may not be very precise. That being said, the decompo-
sition indicates that the effect of education via its effect on income is 0.089, which
is about one third of the total effect. The rest of the decomposition (that is, the
different impacts via health knowledge) is very similar to the case where income is
substituted out. Thus the previous discussion applies here.
The findings of this section can be summarized as follows. The evidence suggests
that education improves child health primarily by increasing health knowledge. It
also has an impact by raising household income, but rough estimates indicate that
this income effect is only about one third of the total effect. This is similar to the
findings of Thomas, Strauss, and Henriques (1991), who found little impact through
improved household income. There is no direct effect of either Arabic or French
literacy skills, nor of numeracy skills, on child health. Neither is there evidence that
other aspects of schooling, particularly changes in mothers" values, have any direct
effect. The question then arises as to how mothers obtain health knowledge. School-
ing alone has no contribution, and may even have a negative effect (due to reduced
time spent at home by girls in school). The lack of a positive effect is consistent
with the fact that Moroccan schools do not teach health knowledge to students. In-
stead, children acquire health knowledge by acquiring literacy and numeracy skills
in school, which they then use to attain health knowledge outside of school. Only
Arabic literacy appears to matter; French literacy has no significant effect.
Overall, the.se findings are quite interesting and have some immediate policy im-
plicaiions. However, further research to confirm, or possibly refute, these findings
is in order. The decompositions here are based on simple functional forms and the
point estimates are not particularly precise. They should be treated as suggestive but
not definitive.
152 The Journal of Human Resources

-—- o —^
E~- ^ <n
" n CN o^
o o o
I I I

II d -^
I I

00 ^-s fN ^ Tf .—
^ — O r- in r-
Q p p rN Tf o ^ O '^
d d -^ "^ d d d d
rn ^-' •—'

r- r- — t-^
p — p >3;
u d d d d

— m r- 00
c> <Zi Ci r-
d d d »N

O O O
o o

O
7! *3

o
[i.
u
 Glewwe 153

00 fN fN 00 r~
(N o »n — 00
O O O C N O O O — O O O O — O fN

ts,-.(N,-v^'-^"^--vO^r-
ocooooxmrnin<noooo—l
O f N p u - f N r i p r o o o p

(N
o^ CN
m —
p
d d d d d d d
I I I I - -

tn '-v o -—
^pp 8 5"'T •-* QQ QQ
p — <n
— dddfNdrJC'—•
d d d d "^
II -- o

d pp
d
II — d d d d

£ H

50
.S 1 52
I
1
schoc

Ht
vspapt

1
telev: ions

Ul
.. to u
radio
other

here

n
o
•s -s
S .$"
o 'o 'o Si S
Ul Ul
. Ul M ex
OJ Ul RJ 0)
5 as
vail

5
o
"5
•s
1 a
2
o •s
UM Z - c-i
1
154 The Journal of Human Resources

Table 8
Decompositions of Impact of Mother's Schooling on Child Health

Through Health Knowledge

Via Via Via Through

Direct Arabic French Numeracy Income Total
(Tli-Cti,) (56a4) Effect

Not conditioning on income -0.268 0.221 -0.095 0.338 0.196

-0.319 0.194 — 0.322 — 0.197
— 0.028 — 0.245 — 0.273
Conditioning on income -0.267 0.220 -0.095 0.337 0.090 0.285
-0.317 0.193 — 0.320 0.090 0.286
— 0.028 — 0.243 0.090 0.361

VI. Summary and Conclusion

Three major conclusions can be drawn from the empirical work in
this paper. First, health knowledge appears to be the most important skill that mothers
(indirectly) obtain from their schooling that prepares them to provide for their chil-
dren's health. Second, estimating the impact of health knowledge on child health
could suffer from substantial endogeneity bias that can underestimate its true impact
if one does not instrument the health knowledge variable. Third, the analysis suggests
that schooling contributes to mother's health knowledge in Morocco only indi-
rectly—health knowledge is not directly learned in school but instead is learned
using literacy and numeracy skills acquired in school.
The above conclusions have direct policy implications for Morocco. First, they
suggest that health knowledge should be directly taught in Moroccan schools.''^ They
should be taught at an early age because girls who drop out early may never acquire
sufficient numeracy and literacy skills to allow them to acquire health knowledge
on their own. Many girls in Morocco leave school at a very early age; only 28 percent
of women aged 18-20 in the 1990-91 ENNVM continued their schooling beyond
primary school. Even more disturbing is that 51 percent of these women never at-
tended school at ail—this latter fact suggests the need for a major effort to teach
young women basic health knowledge in adult education programs.
These results and their policy implications may well apply to other developing
countries. If the finding that health knowledge is the key skill for improving child
health is confirmed in other countries (an important task for future research), any
country where a large proportion of women do not go beyond primary school should

26. One objectioti to this policy recommendation h that health knowledge taught in school will merely
displace health knowledge obtained elsewhere, so thai eventual health knowledge attained will not change.
Yet if literacy and numeracy skills acquired in school are low, there may be liiile acquisition of heallh
knowledge later in life. Moreover, for women who leave school with good lileracy and numeracy skills,
a higher "initial stock" of health knowledge will allow them to reach a higher level of health knowledge
than they would have reached with a low or nonexistent initial stock.
 Glewwe 155

add basic health education to its primary school curriculum, ln addition, if a large
proportion of women do not even attend primary school, health education programs
for women of child-bearing age should also receive high priority. Finally, the find-
ings here support two general policy recommendations for developing countries:
1. Education of girls should be a high priority; and 2. School quality must not be
neglected, since women will not be able to raise their level of health knowledge
after their schooling is completed if they leave school without basic literacy and
numeracv skills.

Appendix 1
Description of Health Knowledge Test
The health knowledge test used in the 1990-91 Moroccan Enquete Nationale des
Niveaux de Vie des Menages (ENNVM) consisted of the following five questions,
given to the respondent in his or her maternal language:

1. Is it possible to get vaccinations for children without paying money? If yes,

where?
(Answer: Yes, at public hospitals, Red-Cross centers, or visits by nurses to
villages).
2. What should one do to a wound to avoid infection?
(Answer: Wash it well with soap, apply alcohol and cover it).
3. What is the best way to prevent children from getting polio?
(Answer: Vaccination).
4. If a child develops diarrhea, what should one do if no doctor is available?
(Answer: Use boiled water, feed rice or carrots, give salts, avoid milk and fats).
5. In places where the water is not safe to drink, what should one do to it before
drinking?
(Answer: Boil it or add drops of "javeP")-
156 The Joumal of Human Resources

^6 oo o —
O •—' O ^ prjooooqpppr-;
< a^ o d d d o — o —od
I I I l -^ -^ w

00 "g
S o '^ p -^ —;
O w-> O
T- 5 y o—
u CJ ^ £

•a
.1-1
i^d -^ <^ -^ d d>

O rn O OOOOOOOCNOO
I I t I w w w
tu

II I I
OOOO
I l l l

o o
o (J
.r:
o
1/)

i §• ca .b
Ul
U
5
o
IS u
u
c
•s
o
 Glewwe 157

O O C > 0 0 ( N 0 0 0 — O—'COO

— OrMO — O^OO — O — O — O O O t N O O O r J O O O m O o G O O O

OO
r-lu-j — " ^
o
Oo—
o O—
(NOOO\DO —O —O —

S in'o m Q —Q
— o ir, o o O
—'OOOOOOOrJOfNO—iO — O — OOOOOrNOi^, O
I " I ! " I l l l — I

S m(^ O
O fN
—
f ,
—
O ~ O — O—'OOO — 0 ' < N O O O < N O ' ^ 0 ^ 0 — O C O O O ^ O r O O

00
a
(at)
c
oli

0 1/3
OJ) o b-
o
j =
x:
u
u
t!r a
1/1
c o
l/l lA
1/1
ra a. 'uc •X3

u T3
U
•C
u
"C
u o u u
>
a. o .c ra
O u. E E
ated

u o O ;2
Q. I/-. o ^.
£ c O
u.
u u 'u k..
the

the
Chi Idr

ra (3 u
nn E E
.5
Inrij

Uni

c ^ <<
o o 3 3
ID tl. Z Z o
u- 2 - : -N
158 The Journal of Human Resources

Table A2
Partial R' Statistics for Instrumented Variables

Instrument Sets
Nine Health
Nine Health Knowledge
Nine Health Knowledge Instruments
Knowledge Instruments Plus Plus Five Income
Instrumented Variable Instruments Mother's Schooling Instruments

Health knowledge 0.143 0.179 0.158

Mathematics 0.273 0.773 0.283
French 0.218 0.788 0.224
Arabic 0.200 0.808 0,203
Income — — 0.394
Reading a medicine bottle 0.231 0.682 —

Notes:
1. The nine health knowledge instnimenis are: mother's manied sisters, father's married sisters, father's
married sislers missing, father born here, number of televisions, number of radios, availability of newspa-
pers,, mother's father's schooling, and mother's mother's schooling.
2. The live income instmmenis are; rental income, number of children living overseas, irrigated crop land,
unirrigated crop land, and tree crop land.

References
Barrera, Albitio. 1990. "The Interactive Effects of Mother's Schooling and Unsupple-
mented Breastfeeding on Child Health." Journal of Development Economics 34(1):81-
98.
Behrman. Jere. 1990. "The Action of Human Resources and Poverty on One Another:
Whal We Have Yet lo Leam." Living Standards Measurement Study Working Paper
No. 74. The World Bank. Washington. D.C.
Behrman, Jere. and Anil Deolalikar. 1988. "Health and Nutrition." In Handbook of Devel-
opment Economics, Vol. I. ed. Hollis Chenery and T.N. Srinivasan. 631-711. Amster-
dam: North Holland.
Bound. John. David Jaeger and Regina Baker. 1993. "The Cure Can Be Worse than the
Disease: A Cautionary Tale Regarding Instrumental Variables." Technical Paper No.
137. Cambridge, Mass.: National Bureau of Economic Research.
Cash. Richard A. 1983. "Oral Rehydraiion in the Treatment of Diarrhea: Issues in the Im-
plementation of Diarrhea Treatment Programs." In Diarrhea and Malnutrition: Interac-
tions, Mechanisms, and lnter\'entions. ed. Lincoln Chen and Neville Scrimshaw. New
York: Plenum Press.
Cebu Study Team. 1991. "Underlying and Proximate Determinants of Child Health: The
Cebu Longitudinal Health and Nutrition Study." American Journal of Epidemiology
133(2): 185-201.
. 1992. "A Child Health Production Function Estimated from Longitudinal Data."
Journal of Development Economics 38(2):323-51.
 Glewwe 159

Davidson. Russell, and James MacKinnon. 1993. Estimation and Inference in Economet-
rics. Oxford University Press.
Gibson. Rosalind. 1990. Principles of Nutritional Assessment. Oxford University Press.
Glewwe. Paul. 1997. "How Does Schooling of Mothers Improve Child Health? Evidence
from Morocco." Living Standards Measurement Study Working Paper No. 128. Wash-
ington. D.C: The World Bank
Schultz. T. Paul. 1984. "Sttidying the Impact of Household Economic and Community
Variables on Child Mnrtaliiy." Population and Development Review 10(suppi.):2l.'i-35.
Slrauss. John. 1**90. "Households. Communities and Preschool Child Nutrition Outcomes:
Evidence from Cote d'lvoire." Economic Development and Cultural Change 38(2):23I-
61.
Strauss. lohn, and Duncan Thomas. 1995. "Human Resources: Household Decisions and
Markets." In Handbook of Development Economics, ed. Jere Behrman and T.N. Sriniva-
san. Volume 3.
Thomas, Duncan. John Strauss, and Maria Helena Henriques. 1991. "How Does Mother's
Education Affect Child Height?" The Journal of Human Resources 26(2):183-211.
Wolfe, Barbara, and Jere Behrman. 1987. "Women's Schooling and Children's Health:
Are ihe Effects Robust with Adult Sibling Control for the Women's Childhood Back-
ground?" Journal of Health Economics 6:239-.'i4.
World Bank. 1994. "Kingdom of Morocco, Public Expenditure: Issues and Outlook."
Washington. D.C: Middle East and North Africa Region, World Bank.