Intergenerational Consequences of Early Age Marriages of Girls: Effect on Children’s Human Capital * Sheetal Sekhri † and Sisir Debnath ‡ This Draft: December, 2010 Abstract We use a nationally representative data from India on test scores in an instrumental variable framework to isolate the causal effects of early age marriages of girls on human capital of their children. Early age marriages reduce mother’s educational attainment which can adversely impact the education outcomes of their children. On the other hand, better marriage prospects of young brides may compensate and improve chil- dren’s educational outcomes by way of resource provision. Consequently, the effect of early age marriages of girls on their children is theoretically ambiguous and warrants an empirical examination. In our empirical analysis, we use plausibly exogenous variation in age at menarche to instrument for marriage age. Our estimates show that a delay of one year in the marriage age of the mother increases the probability of being able to do the most challenging arithmetic and reading tasks on the administered test by 3 percentage points. Key words: Early-Age Marriages, Child Development, Human Capital * We would like to thank Leora Friedberg, Sarah Turner and the participants of the the Labor Economics Research group at University of Virginia for their insightful suggestions. We also wish to thank Virendra Rao for sharing the Gender, Marriages, and Kinship Survey data with us. † University of Virginia, Email: [email protected]‡ University of Virginia, Email: [email protected]1
49
Embed
Intergenerational Consequences of Early Age Marriages of Girls ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Intergenerational Consequences of Early Age Marriages of Girls:
Effect on Children’s Human Capital ∗
Sheetal Sekhri† and Sisir Debnath‡
This Draft: December, 2010
Abstract
We use a nationally representative data from India on test scores in an instrumental
variable framework to isolate the causal effects of early age marriages of girls on human
capital of their children. Early age marriages reduce mother’s educational attainment
which can adversely impact the education outcomes of their children. On the other
hand, better marriage prospects of young brides may compensate and improve chil-
dren’s educational outcomes by way of resource provision. Consequently, the effect of
early age marriages of girls on their children is theoretically ambiguous and warrants an
empirical examination. In our empirical analysis, we use plausibly exogenous variation
in age at menarche to instrument for marriage age. Our estimates show that a delay
of one year in the marriage age of the mother increases the probability of being able
to do the most challenging arithmetic and reading tasks on the administered test by 3
percentage points.
Key words: Early-Age Marriages, Child Development, Human Capital
∗We would like to thank Leora Friedberg, Sarah Turner and the participants of the the Labor Economics
Research group at University of Virginia for their insightful suggestions. We also wish to thank Virendra
Rao for sharing the Gender, Marriages, and Kinship Survey data with us.
Marriage practices in developing countries, particularly in rural areas, often involve early
marriages of adolescent girls. In many countries around the world, this practice remains
widespread. In India, about 5 percent of girls between the ages of 10 and 14, and over 35
percent girls between the ages of 15 and 19, are married (Census of India, 1991). Similarly,
51 percent of women in Bangladesh and as high as 74 percent of women in Niger are married
before the age of 18 (UNFPA, 2007). Poverty, social norms, lack of security for young
adolescent girls, and parental attitudes toward girls have been identified as potential reasons
for early marriage of females in developing countries. Early age marriages can also have
implications for children. In this paper, we empirically investigate how mother’s age at
marriage influences their children’s welfare.
Early marriages and subsequent early motherhood constraints human capital formation of
women (Field and Ambrus, 2008). Young brides also tend to have less control over resources
in their husband’s families and experience more domestic violence (Jenson and Thornton,
2003). On the other hand, there is a premium on age in the marriage markets. All else equal,
younger brides may be able to marry into relatively richer households. Given this trade-
off, the effect on child welfare is not clear. A number of early studies show that mother’s
education improves the child’s human capital. There is also a growing literature which shows
that income or assets controlled by women are associated with improvements in child health
and greater household spending on nutrients, health and housing (Thomas 1994, Duflo 2003).
Therefore, more schooling and greater control over household resources for women could
translate into greater human capital for the next generation. But, better endowed households
may be able to compensate for the lack of mother’s education. Therefore, theoretically the
effect of mother’s age at marriage on children’s outcomes is ambiguous. Our paper focuses on
the empirical examination of the intergenerational consequences of early marriages of girls.
2
We use a nationally representative data from India, the India Human Development Survey
of 2005, which includes data on test scores of children to isolate the causal effects of early
age marriages of girls on human capital of their children. The main empirical challenge in
identifying the causal effect is that marriage age may be endogenous. In order to address this,
we use age at which girls experience their first menstrual cycle as an instrument for their age
at marriage. Variation in the age at menarche generates a quasi-random difference in the age
at which a girl enters the marriage market.1 Mothers coming from an economically strong
background may receive higher nutrition, and hence would be more likely to menstruate early,
and their offspring might have better health and education resulting from the economic status
of their grandparents. However, conditioning on the nutritional status of the mother, age at
menarche would provide plausibly exogenous source of variation in age at marriage. We use
mothers’ height as a proxy for the nutrition she received in her childhood. It is worthwhile to
point out that since the economic status of the natal family is negatively correlated with age
at menarche, any bias resulting from omission of the economic status will tend to attenuate
the results and the estimates will underreport the effect.
Our estimates show that a delay of one year in the age of marriage of the mother increases
the probability of her child being able to do the most challenging arithmetic and reading
tasks by 3 percentage points, and increases the likelihood of being enrolled in school by 3.5
percentage points.
Our paper makes important contributions to the literature in two ways. First, we identify
the causal estimates of the effect of mother’s age at marriage on children’s educational
outcomes. To the best of our knowledge, our paper is the first to examine the causality. In
addition, the data we use provides comprehensive set of variables including performance in
tests that measure basic human cpaital . We present evidence that mother’s age at marriage
1This instrument was used by Field and Ambrus (2008) to measure the effect of early marriage of girlson their school attainment in Bangladesh. She found that delay in marriage by an additional year increaseseducation by 0.22 years and increases the probability of literacy by 5.66 percent
3
affects school choice, time spent on homework, and household outlays on education related
items. Secondly, we bring together two important strands of literature with important
policy implications. Previous research has shown that women’s age at marriage affects their
educational attainment (Field and Ambrus, 2008). Also, a number of other papers have
shown that mother’s education influences the education and health outcomes of children
(Rosenzweig and Wolpin 1994; Currie and Moretti, 2003). We show that controlling for child,
mother’s and household’s characteristics, the reduced form effect of mother’s age at marriage
on children’s educational outcomes is positive. Our results indicate that women’s age at
marriage affects the likelihood of being enrolled in school, and improves the human capital
of their children over and above the effect of the family resources. These intergenerational
externalities warrant that minimum age laws that prevent under-age marriages of women
be passed and enforced. In India, there is a mandated legal minimum age of marriage.
Prohibition of Child Marriage Act in India of 2006 bans child marriage. It also empowers
civil courts to annul such marriages and to make penal provisions for people who solemnize
these marriages.2 The stated objective of a minimum age at marriage is to prevent maternal
mortality and to increase human capital of women. If effectively enforced, such policies can
improve the educational outcomes of the children as well.
The rest of the paper is organized as follows. Section 2 provides background information
on marriage practises in India. In Section 3, we provide the details of the data used in our
empirical analysis. We specify our identification strategy in Section 4, and the results are
reported in Section 5. We discuss robustness of our estimation in Section 6 and Section 7
discusses suggestive channels through which age at marriage may affect children’s human
capital. We conclude in Section 8.
2Apart from legal provisions, many social welfare programs are designed to provide incentives for parentsto delay marriages of their daughters. A prominent micro-finance program in India excludes borrowers whosedaughters marry before 17, and national education vouchers in Bangladesh exclude married girls.
4
2 Motivation and Background
India, like many other developing countries, is a hot-spot for early age marriages of girls.
Historically, the age at marriage for women in India has been very low. The median age
at marriage was 14.5 in 1951, one of the lowest in the world (Agarwala, 1957). According
to National Family Health survey, in 2005-06, the average age at first marriage for Indian
women stood at 17.96 years and 38% of them were married below the age of 18. A number of
explanations have been provided to account for such widespread practice of early marriage
for girls. Traditional customs that evolved to protect kinship networks are often cited as an
important factor (Dyson and Moore, 1983). Parents enter their daughters in the marriage
markets early, when they are young so they can control spousal choices in order to protect
the kinship network. Another proposed explanation is that parents marry the daughters
young due to economic motives. Marriage outcomes are often determined by the size of the
gifts or dowry that parents of the bride can offer to the groom and his family. These dowry
payments can constitute a substantial burden for poor households.3 Younger brides often get
better matches with lower dowry price (Dyson and Moore 1983, Coale 1992). In addition,
due to a patriarchal social structure, women leave their parents house after their marriage,
and reside with the husband’s family. Marrying young daughters can reduce the number of
children to be looked after and fed. Poorer natal families may exercise this option to reduce
their economic burden.
2.1 Menarche and Marriage Outcomes
Most of the marriages are solemnized soon after the girl child reaches menarche. While 9%
of the women in our sample report pre-menarche marriage, around 52% of the marriages
take place within 3 years of puberty. Parents feel concerned about the safety of their daugh-
3In certain parts of India, dowry prices might be over 50 percent of household assets (Rao, 1993).
5
ters as virginity is highly priced in marriage markets (Caldwell, 2005; Desai and Andrist,
2010; Sheela and Audinarayana, 2003). As the girl starts her menstrual cycle, the kinship
network is informed immediately so that the search for the groom can commence (Sheela
and Audinarayana, 2003).
As shown in Figure 1, the average age of first marriage by age at menarche, age of
marriage is closely followed by age at menarche. In Figure 2, we plot the distributions of
age at marriage and menarche, showing that compared to age at menarche the distribution
of age at marriage has a higher variance but it is shifted to the right. Onset of menarche
predicts age at which girls are married.
2.2 Menarche and Changes in Life Style
Reaching menarche becomes a life changing event for the girls. Typically, religious and
social sanctions are imposed on the girl child. For example, hindu girls are forbidden to
enter temples when menstruating. They are often asked to change the way they dress, and
forbidden to play with male children. Parents are also reported to withdraw their daughters
from schools. Girls are refrained from going out of the house alone. Some studies also report
that girls feel traumatized by these changes and tend to remember the timing of menarche
very well (Nahar et al., 1999).
3 Data
The principal data-set we use in our empirical analysis is the India Human Development
Survey (IHDS) of 2005, a nationally representative data-set spanning 41,554 households
over 25 states and union territories of India (with the exception of Andaman/Nicobar and
Lakshyadweep islands). The survey covered 1503 villages and 971 urban neighborhoods.4
4A stratified random sampling technique was used to construct the sampling frame. See Desai, S., Dubey,A., Vanneman, R., and Banerji, R. (2009) for details about the survey.
6
The survey provides comprehensive data that we need for our analysis at both individual
and household level. Ever married women between ages of 15 and 49 were asked to provide
complete information about their marital and reproductive history.
Among other modules, the survey also covered topics concerning health, education, and
employment for all members of the household. Data were collected on children’s school
enrollment, type of school, medium of instruction and hours spent in school, homework,
private tuition, and number of days absent from school in the last month. Expenditure
incurred on school fees, private tuition and on other school accessories were collected as well.
In addition to these self reported measures, children aged 8 to 11 were administered short
reading, writing and arithmetic tests. Children were classified according to their ability to
read, in one of the following five categories: (a) Cannot read at all; (b) Can recognize letters
but cannot read words; (c) Can read words but cannot read entire sentence; (d) Can read a
short paragraph of two to three sentences but cannot read a short story; (e) Can read a one
page short story. The mathematical skill of the children were classified into four categories:
(a) Cannot read numbers above 10; (b) Can read two digit numbers but unable to do more
complex number manipulations; (c) Can subtract a two digit number from another; (d) Can
divide a three digit number by a single digit number. In addition to the maths and reading
tests children were also administered writing tests. The writing scores were classified into
two categories: (a) Unable to write; (b) Can write with two or less mistakes.5
3.1 Sample Construction and Main Outcomes of Interest
The main outcome variables that we analyze are the test scores of the children. Since the
tests were administered only to 8 to 11 year children, we restrict our sample to them while
analyzing test score outcomes. We restrict our sample to children without missing values
5Since these categories are not exhaustive, the scores are not meaningfully reported for all children.Hence, we do not analyze effects on writing scores.
7
for the test scores. The IHDS data consists of 29,263 children. Among them, 9126 children
are between ages 8 to 11. However, the tests were administered to only 75.4% of them.
Children could not be tested without parental consent which shrinks the sample size for this
analysis. We examine whether the parents and households where these variables are missing
are systematically different from those which report the test scores and find no evidence of
observable differences. Table 1.1. compares the summary statistics across all children, those
who took the test and those who did not. From a comparison of Columns (iii) and (v), we
find no systematic differences between the children whose parents consented versus those
whose parents did not consent to administer the tests.
3.2 Summary Statistics
Summary statistics for the children between 8 to 11 years with test scores are presented in
the last two columns in Table 1.1. There are 6884 children in this age group, belonging to
5787 mothers. The average age of marriage for the mothers is 17, which is one year below
the legal age. The average years of schooling for mothers is 3.93. Fathers are 5.2 year older
than mothers on average. 80 % of the households are Hindu and 26 % of them are below the
official poverty line. The average number of sibling for a child is 3.5. 53 % of the children
are boys.
Table 1.2. provides the summary statistics of children’s outcome by the age at first
marriage of their mother. The table shows that the the mothers who got married at the
age 18 or later are more likely to enroll their children in private schools and their offsprings
spend more time at school, homework and private tuition. On an average their children
score 0.37 and 0.43 points more in math and reading tests respectively, compared to those
who got married at 17 or earlier.
We further explore the relationship between age at marriage of mothers on test score
outcomes of their children in Figures 3 and 4. These figures plot the distribution of children’s
8
math and reading skills by age of marriage of their mother. In Figure 3, the average number
of children who Cannot do math are unfamiliar with basic mathematical concepts and those
who know only Numbers decrease as the age at marriage of the mother increases. However,
the average number of children who can do Subtraction and Division, which require greater
skills increases steadily with the age at marriage of the mother. The same patterns are
reflected in Figure 4 for reading scores. We formally test these observations in the following
sections.
4 Identification Strategy
Our goal is to evaluate the intergenerational effects of early age marriages of women. We
want to isolate the causal effect on the human capital of their children. The empirical model
where Yij is the outcome of child i born to mother j, Aj is the age at marriage of the mother,
Fij are the characteristics of the father, Mj are the characteristics of the mother, Hj are
household characteristics, Xij are the characteristics of the child i and εij is a random error
term. The coefficient βa on age at marriage Aj, is the parameter of interest. We rewrite
equation (1) as follows:
Yij = β0 + βaAj + βwWij + εij (2)
where βw = (βf βm βh βx) and Wij = [Fij Mj Hj Xij]′. Wij includes all regressors except
age at marriage of the mother and βw is the coefficient on Wij.
The are two main empirical challenges. First, age of marriage may be endogenous.
Omitted variables may affect both the age at marriage of the mother and the child outcomes.
9
For example, a father who prefers to invest in his children may also have stronger preferences
not to marry a very young woman. In principal, there might be other potential omitted
variables which are not orthogonal to age of marriage of the mother and might be correlated
with the children’s outcomes. The second issue relates to the accuracy of the report of age
of marriage. During the survey age at marriage was self reported. Inaccurate reports would
generate measurement error in the explanatory variable and could attenuate the estimates
of the coefficient of interest. To address these concerns, we follow an instrument variable
(IV) approach. We use age of menarche as an instrument for marriage age of the mother.
4.1 Instrumental Variable Approach
The IV approach involves estimating a two stage model which is specified as follows:
Aj = α0 + αzZj + αwWij + ηj (3)
Yij = β0 + βaAj + βwWij + εij (4)
The first stage is given by the equation (3), and equation (4) is the structural equa-
tion. The mother’s age at marriage Aj is instrumented by Zj, her age at menarche, and
Yij are the children’s outcome of interest. As above, Wij is a set of control variables that
include child’s, mother’s and father’s characteristics, household background and socioeco-
nomic status. Child’s characteristics include number of siblings, gender, grade and birth
order. Mother’s characteristics include age and height. Father’s characteristics include age
and education, and in the background of the family we include number of household mem-
bers, place of residence (urban/rural), land ownership and dummies for below poverty line
status and religion.
We use a standard two stage estimation procedure when the outcome variable is con-
tinuous, ordered probit when the outcome variable is an ordered categorical variable, and a
10
probit model when the outcomes are binary. We cluster standard errors at the village level.
We perform a number of robustness checks to test for the validity of the instruments.
4.2 Validity of the Instrument
First, we examine whether age at menarche predicts age at marriage which is the endogenous
regressor. Consistent with Field and Ambrus’s findings for Bangladesh, we find that age at
menarche significantly predicts age at marriage in India. The results from the regression of
women’s age at first marriage are presented in Table 2. Column (i) reports the coefficient
on age at menarche without additional controls. The coefficient of 0.32 is highly statistically
significant, and the F Statistic is 133.7, eliminating concerns about ‘weak instruments’.
Next, we examine the threats to the validity of this instrument. Acute malnutrition in
early childhood can result in delayed onset of menarche. Exposure to acute malnutrition
could potentially affect long term health of the mother and consequently her child. This
could undermine our instrument. Medical evidence suggests that severe loss in food intake
can result in stunting, and in some cases delayed onset of puberty. The changes in nutrition
that could result in delayed onset of menarche are also likely to result in stunting (Stathopolu
et al, 2003). We explore this correlation in our sample. Figure A2 in the Appendix shows
adult heights by age at menarche among the mothers in our sample. We do not observe any
evidence of correlation between adult height and age at menarche.
Volatility in exposure to malnutrition also affects maturation (Field and Ambrus, 2008).
Agriculture and allied activities, that employ majority of the Indians, are overly weather
dependent. Extreme weather conditions in the mother’s birth year like drought and flooding
might lead to crop failure resulting in transitory but severe malnutrition. Therefore, females
born during these unprecedented weather events may experience delayed age at menarche
as they are more likely to be malnourished. We control for this possibility in our first stage
regression. In column (ii), we add birth year fixed effects for the mother to account for
11
extreme weather events at the time of birth.6 The point estimates and standard errors are
remarkably similar across columns (i) and (ii). Next, we include adult height in the regression
in column (iii) as a proxy for acute malnutrition in childhood. Neither the point estimates,
nor the standard errors change. We condition all subsequent results on adult height and
mother’s birth year fixed effects.
It is conjectured that hard physical labor in early childhood can influence menarcheal
age (Pellerin-Massicotte et al., 1997). However, the children who work in India do not do
strenuous physical work like construction. Detailed data on child labor collected by Das
from northern India show that 99.8 % of working girls of age 6 to 14 are engaged in domestic
work while 0.001 % of them work for wage (Basu, Das and Dutta, 2010). Economic status
of the woman’s natal family might affect the age at which she reacheed puberty as it might
affect whether the women worked strenuously as a child or not. We do not directly observe
the economic status of the parents of the mother. However, married women were asked if
the economic status of their natal family was the same as their husband’s family. After
restricting the sample to those women who were married within same economic status, we
control for an asset index of the husband’s family to account for the socioeconomic status of
the natal family.7 For this sample, we report the OLS results for age at menarche conditioned
on adult height with mother’s birth year fixed effects in column (iv). The results from the
regression in which we additionally control for the family’s asset index are reported in column
(v). The coefficient of age of menarche on age at marriage is still highly significant.
To further substantiate the observation that conditional on mother’s height and birth year
fixed effects, the age of menarche is not influenced by the characteristics of the natal family,
we examine the relation between age of menarche and characteristics of the natal family using
an additional survey data from India.8 We first show that the distributions of age of menarche
6Birth year fixed effects can also account for any exposures to environmental determinants of age atmenarche.
7We construct an asset index using Principal Component Analysis from the asset data.8We use data from the Gender, Marriage,and Kinship Survey conducted by NCAER in 1995. This data
12
and age of marriage across the two datasets - IHDS and Gender, Marriage, and Kinship
Survey are similar. We plot the kernel densities of age at menarche and age at marriage in
the top panel in the appendix Figure A3. This indicates a similar relation between the two
variables as indicated by Figure 2, based on IHDS-2005 (the data we use in our empirical
analysis). The bottom panel shows the relationship between the age at menarche by literacy
of father and whether father owned irrigated land before the marriage of the girl in the
Gender, Marriage, and Kinship survey data. None of figures in the bottom panel show any
systematic relation between socioeconomic characteristics of the natal family and the age of
menarche of the girl. This is suggestive that conditional on height and birth year fixed effects,
age at menarche may not be influenced by the socioeconomic characteristics of the natal
family. Most importantly, since economic status and age at menarche would be negatively
correlated, exclusion of socioeconomic status of mother’s natal family, would attenuate our
estimates of age at marriage on child outcomes. If anything, we would underreport the effect
of age of marriage on child outcomes.
Geographical features like temperature and altitude also may influence age at puberty.9
The data do not report the location of the mother’s natal family. However, we rely on
estimated size of the marriage markets to predict natal family’s location. We re-estimate
the first stage and regress age at marriage on age at menarche controlling for the average
temperature and elevation of the proxied natal locations. A detailed discussion of the method
followed and the results is presented in section 6.2.
These robustness tests lend further credibility to our hypothesis that conditional on adult
height and birth year fixed effects, the residual variation in age at menarche is plausibly
exogenous.
is collected from rural areas of two large states of India (Uttar Pradesh and Karnatka).9See Field and Ambrus, 2008 for a detailed discussion
13
5 Empirical Results
In this section, we provide empirical evidence that a mother’s age at marriage affects the
human capital, school outcomes of her children and household investments in their human
capital. Human capital is measured by a child’s performance in arithmetic, and reading tests
that were administered during the interview. School outcomes are measured by enrollment
status and the type of the school. Investment in human capital is measured by outlays on
education related items and time spent studying.
5.1 Children’s Human Capital
To measure the impact of early marriage on a child’s human capital, we estimate the effect
of a mother’s age at marriage on a child’s test scores. The tests were administered during
the survey to 8 to 11 year old children which measured arithmetic and reading skills. These
scores are ordered categorical variables.10
The results from OLS and Ordered Probit regressions for test scores on mother’s age at
marriage are reported in Table 3. These estimates are not causal, as age at marriage of
mother is potentially endogenous. The OLS coefficient for math scores, reported in column
(i), is 0.013 and it is highly significant at 1% level.11 Since the scores are ordered we present
the estimates from an Ordered Probit model in column (ii). The Ordered Probit coefficient
for math score is 0.018 and it is significant at 1% level. The coefficient does not measure
the direct effect of mother’s age at marriage on child’s math score but it provides crucial
information about the sign of the effect for the lowest and highest categories of the score.
10Children were awarded scores based on their skills, these scores were integers between 0 and 4. Forexample children who could recognize only single digit number were given a score of zero, the lowest score,and those who could divide a three digit number with a two digit number were awarded a score of 3, thehighest math score.
11Arithmetic score, is an ordered categorical variable. The four categories of the score are: (a) Score 0:cannot read numbers above ten; (b) Score 1: can recognize two digit numbers but not able to do morecomplex number manipulation; (c) Score 2:can subtract a two digit number from another; (d) Score 3: candivide a three digit number between by a single digit number.
14
The positive coefficient suggests that increase in mother’s age at marriage increases the
probability that her child will score the highest possible score and decreases the probability
that the child will score the lowest possible score.
The results from the OLS and Ordered Probit regressions on reading scores are reported in
columns (iii) and (iv). The coefficients from both the regressions are positive and significant
at 1% level.12 The coefficient estimates for mother’s age at marriage are the same from both
the models at 0.019. This suggests that the marginal effect of mother’s age at marriage is
negative for the lowest math and reading score categories, but it is positive for the highest
category.
As the estimates presented in Table 3 potentially suffer from endogeneity, we use age
at menarche as an instrument for the endogenous regressor and since the scores are ordinal
we estimate the effect of marriage timing with a IV-Ordered Probit model. We estimate
a Seemingly Unrelated Regression model, given by equation (3) and (4). Only the final
stage, equation (4) is structural, and the estimators are obtained by maximizing a Limited
Information Likelihood function (LIML). The results from the IV-Ordered Probit regression
on math score are reported in Table 4.1. Panel A of the table reports the first stage of the
regression. The coefficient on age at menarche is highly significant and positive. One year
delay in onset of menarche increases age at marriage by 0.34 years. Panel B in the same table
reports the marginal effect of mother’s age at marriage for several categories of math score.
All the estimates are highly statistically significant. The estimates show that an increase in
age at marriage by one additional year decrease the probability that the child will receive the
lowest score (a child will be able to recognize a two digit number but will not be able to do
more complex number manipulation) by 1.8 percentage points and increases the probability
12Scores in reading are classified under five categories; (a)Score 0: cannot read at all; (b)Score 1: canrecognize letters but cannot read words; (c) Score 2:can read words but cannot read entire sentence; (d)Score 3: can read a short paragraph of two to three sentences but cannot read a short story; (e) Score 4:can read a one page short story.
15
of receiving the highest score (a child will be able to divide a three digit number by a single
digit number) by 3 percentage points.
Table 4.2. reports the estimates for reading scores. The first stage of the estimation is
presented in Panel A. The first stage estimates differ from those in the Table 4.1. as we
include test language fixed effects in our model. The test language for math test and reading
test could be different. age at menarche is again significant and positively influences age at
marriage. The F statistics is high at 95.9. Panel B reports marginal effects of mother’s age
at marriage for several categories of reading score. Similar to the previous result, delay in
marriage increases the chance that the child will fare better in terms of reading skills. In
particular, delay in marriage of a mother by one year increases the probability that her child
will have the highest reading score (a child will be able to read a one page story) by 3%
points.
The results presented above indicate that after controlling for parents, household, and
child characteristics, mothers’ age at marriage influences children’s human capital.13 If we
assume these effects to be linear with age at marriage, a 2 year delay in marriage could
translate into 6 percentage point increase in probability that a child will have the division
skills and the skill to read a one page story.14
5.2 School Enrollment and School Choices
In this section, we provide estimates of the effects of mother’s age at marriage on school
enrollment and school choices for her children. We explore if the choice of schools in terms
of public versus private, and in terms of English medium versus local language, are affected
by mother’s age at marriage.
13We evaluated the effects on nutrition and health outcomes as well. While the nutritional outcomesare statistically significantly influenced by mother’s age of marriage, health outcomes such as incidence ofrespiratory diseases are not. Results are available on request.
14The average age at marriage for our sample is 17 years. We also check for non-linear effects of mother’sage at marriage on test scores of her child. Our estimations reject presence of any non-linear effects.
16
The results of the estimation are reported in Table 5. Panel A of the table reports the first
stages of the estimation and Panel B reports the second stages. In column (i) the dependent
variable is school enrollment status of a child, which is binary. We estimate an iv-probit
model and use mother’s age at menarche as an instrument for her age at marriage. Column
(i) in Panel B reports the marginal effect of mothers age at marriage on the probability that
a child will be enrolled in school. A one year delay in marriage increases the probability
that a child will be enrolled in school by 3.5 percentage points. The estimate is significant
at 10% level. Similarly, in Column (ii) the dependent variable is whether the chosen school
is private. The estimate is highly significant and one year delay in marriage increases the
probability that a child will be enrolled in a private school by 6.3 percentage points. Finally,
Column (iii) reports the effects of age at marriage of mother on medium of instruction in the
school of her children. The coefficient is negative but it is statistically insignificant. These
estimates suggest that delay in marriage age is likely to improve the chance that the child is
in school, and is more likely to be enrolled in a private school.
5.3 Effects on Investments in Human Capital
We also investigate whether age of marriage of a mother influences expenditure on education
for her children. In particular we measure the effect on school fees, books, and private
tuition.15 We also examine the effect of mothers’ age at marriage on time spent by children
at school, private tuition and homework.
The estimates of the effects of mothers’ age at marriage on outlays on human capital
are reported in Table 6.1. As before, we use mothers’ age at menarche as an instrument.
Columns (i), (ii), and (iii) in Panel A report the first stage of the estimation. The coefficient
on age at menarche is positive and highly significant. Columns (i), (ii) and (iii) in Panel
B report the second stage of the instrumental variable estimation on expenditure on school
15Expenditure on books include expenditure on uniform and transportation (eg. bus fare).
17
fees, books, and private tuition respectively. Our estimates suggest that one year delay
in marriage increases outlays on school fees and books by Rs. 88 and 138 respectively.16
The effect of mothers’ age at marriage on expenditure on private tuition is positive but not
significant.
We also examine the effect of age at marriage on allocation of children’s time for study.
Table 6.2 reports the results. Columns (i), (ii), and (iii) in Panel A report the first stage
of the instrumental variable estimation. The coefficient on age at menarche is positive and
highly significant. Columns (i), (ii), and (iii) in Panel B report the second stage of the
instrumental variable estimation on time spent by children at school, private tuition, and
homework per week respectively. We find that conditional on parent, household, and child
characteristics, a child would spend 40 additional minutes at school per week if his mother
were married one year later. The estimate is marginally significant at 10 % level. Time
spent by children at private tuition, and on homework seems to decline with mothers’ age
at marriage. But this is inconclusive since the estimates are insignificant.
6 Robustness
In this section, we address two additional concerns about our empirical analysis. First, two
variables that we use intensively, age at marriage and menarche, are collected retrospectively.
This raises a concern about strategic misreporting and recall bias in them. Secondly, variety
of medical literature suggests that other than socioeconomic factors puberty is also affected
by geographical and climatic conditions. If the unobserved error in the structural equation
is correlated with climatic variables, such as elevation and temperature, then our instrument
could be potentially invalid. We present substantiating evidence to allay these concerns.
16$1 is equivalent to Rs. 47.
18
6.1 Recall Bias & Strategic misreporting
One concern with our estimates is the presence of systematic measurement error in the
variables that were collected retrospectively, for example, the age of a woman at her marriage
and menarche. One concern might be that the measurement error in the reported age at
marriage may increase with current age of a woman. We conduct a test to check if this
might be of concern. Marriages in India are closely followed by motherhood and pregnancies
outside marriages are rare. Therefore, one should expect the age gap between first birth and
marriage to be small and positive. We calculate the age at first birth for each woman from
current age profile of their children. For 8.75 % women in the full sample, the difference
between age at marriage and age at birth is negative indicating that there could be some
reporting error in the variable. Figure A1 (Appendix) plots the difference between age at
first birth and reported age at marriage by the current age of a women. However, this figure
does not reveal any stark relationship between current age and time gap between age at first
birth and age at marriage. The lowess smoother is almost horizontal at zero, implying there
is no systematic recall bias that increases with age of the women.
Another concern about reported age at marriage is that women who were married under
the legal minimum age of marriage might overstate their marriage age. The legal minimum
age at marriage for females in India is 18. According to Indian Penal Code, conjugal rela-
tionship with a minor girl is a punishable offence. Women who were married below 18 might
not reveal their true age at marriage on the suspicion that the information might be revealed
to the authorities. If such strategic misreporting were prevalent in our data,we would see
a break in the distribution of age at marriage at age 18. We formally test if there is any
discontinuity in the distribution of age at marriage at 18 using McCrary’s DC Density test.17
We do not find any significant jump in the reported age at marriage at 18.
The variable age at menarche is also collected retrospectively. If respondents do not
17Results are available upon request.
19
remember their age at menarche, they might approximate it with their age at marriage and
that might confound the correlation between age at marriage and menarche. The concerns
about systematic measurement error in age at menarche is less severe. First, medical lit-
erature suggests that women are generally able to recall their age at menarche accurately
(Field and Ambrus, 2008). Secondly, as we discussed earlier menarche ushers dramatic life
style changes for a Indian girls. Hindu girls are forbidden to enter temples and to participate
in any other religious activity when they are menstruating. Muslim girls are instructed to
pray five times a day, to keep fast and cover most of their body. In some parts of India,
menarche is celebrated with gifts of jewelry and traditional dresses to the girl. Additionally,
anthropological accounts suggest that most of the girls are unaware of menstruation before
it begins and are traumatized by the event. Therefore, women tend to remember their age
at puberty.
An additional concern might be that menarchial and marriage ages might be misreported
in rural areas or not remembered by women in rural areas. We separate the families according
to their area of residence (urban or rural) and plot the distribution of age at marriage and
age at menarche for these two groups of women in Figure 5. The top panel of the figure
shows the distribution of age at marriage by area of residence and the bottom panel shows
the distribution of age at menarche by area of residence. It is clear that women in urban
areas marry later but no such difference is noticed in the distribution of age at menarche.
This also suggests that age of menarche is less prone to measurement error.
Finally, we follow Field and Ambrus’s strategy to check if we discern differences in age of
menarche across two groups distinguished by a pre-existing preference for different marriage
ages but presumably orthogonal to age at puberty. Since menarche is unrelated to the
preference for early marriage in first group, we do not expect to find significant difference in
the distribution of reported age of menarche. If we notice a difference in age at menarche,
it would be suggestive of recall bias or strategic misreporting. To test this we compare the
20
distribution of age at marriage and age at menarche by parent’s literacy (Field and Ambrus,
2008 used literacy of the mother) using another survey data from India and subsequently
compare the distributions of those variables with IHDS data.18 The idea is to show that
there is no recall bias in age at menarche in the Gender, Marriages,and Kinship Survey data
(refered as NCAER data in the figures) and show that the distribution of age at menarche
and age of marriage is similar across IHDS and this data-set.19 The top panel of Appendix
Figure A4 shows the the distribution of age at marriage varies with literacy of the parent, but
the distribution of age at menarche is similar across these groups characterized by literacy
of the parents. The bottom panel compares the distributions of residuals of age at marriage
and age at menarche across (i) the Gender, Marriages,and Kinship Survey data, (ii) IHDS
data restricted to the districts from which the Gender, Marriages,and Kinship Survey data
was collected, and (iii) the entire sample from the IHDS data. The distributions of residuals
of age at marriage are different across the three data-sets, but the distribution of residual of
age at menarche are remarkably similar.
6.2 Climate & Age at Menarche
According to medical literature, exposure to endocrine disrupting chemicals (direction of
in diet in utero or in childhood (delay puberty), altitude and temperature (high altitude and
cold weather delay puberty) are the major determinants of age at puberty.20 The first two
factors are less likely to confound our results. Exposure to chemicals are strongly correlated
with area of residence, since we control for a indicator for residence, exposure to chemicals
18IHDS does not collect information for women’s natal family. We use data from the Gender Marriage,Kinship Survey conducted by NCAER in 1995. This data is collected from rural areas of two large states ofIndia (Uttar Pradesh and Karnatka).
19Since the states in Gender, Marriages,and Kinship Survey vary in terms of geography, climate, andmany other attributes, we plot the distributions of residual age at menarche after controlling for state fixedeffects and height of the women.
20See Field and Ambrus, 2008 for a comprehensive discussion
21
are also unlikely to confound our results. We also control for mothers’ birth year fixed effects
in all our specifications, thus eliminating concerns for extreme weather events affecting age
at menarche of mother. Lastly, our sample spans all over India with significant geographical
and climatic variation. Therefore, if high altitude delays age at puberty, and thereby age at
marriage, and if schools are difficult to access in high altitude areas, then our results could
be driven by a spurious correlation in these variables introduced by omission of altitude.
Unfortunately, we do not have the geographical location of mother’s natal family. But
in order to confirm whether our estimates are robust to these concerns, we predict the natal
family from estimated size of marriage markets in India. Bolch et al (2002) estimates that
the average distance between husband’s home and wife’s natal home is 21.1 miles for India.
Given that estimate and a 2077 square miles average area of a district, we can assume that
a woman in India is most likely to get married within her natal district. Therefore, we
include geographic and climatic control for the district in which a woman was surveyed to
our regressions. In particular we include altitude and temperature averaged at the district
level of the current location of the women in the sample. In addition, we also construct the
averages of these variables for all neighboring districts that border the district of current
residence. Figure A5 in the Appendix shows a map to serve as an example of neighboring
districts used to construct this measure.21
Appendix Table A1, Panel A reports the coefficients from a regression of age at menarche
on probable predictors of menarche including district averages of temperature and altitude.
In Column (i), we control for the average altitude and temperature of the current residing
district and in Column (ii), we control these variables averaged for the residing and the
neighboring districts. Across these specifications, adult height of the women is statistically
2196% of the women in IHDS report that it takes 10 hours or less to reach their natal home. Therefore,by including the average temperature and altitude of the current residence district and the neighboringdistricts in the regressions, we try to maximize the probability that the climate condition of the natal homeis controlled.
22
significantly correlated with age at menarche, though the correlation is small. However, the
temperature and altitude do not seem to be correlated with age at menarche. The coefficients
are very small and only altitude is marginally significant in the case where we include the
averages of the current residing district.
Panel B reports the results from a regression of age at marriage on age at menarche
including the geographical variables in the regression. This shows that the correlation be-
tween age at menarche and age at marriage is highly significant even after controlling for the
climatic variables. In column (i), we report the coefficients from a specification where we
control for average temperature and elevation in the current residing district. The coefficient
and the standard error on age at menarche is similar to the benchmark case reported in col-
umn (iii) of Table 2. Marriage age is independently correlated with average temperature in
the residing district. We see the same patterns in column (ii) where we include the average
for current residing and neighboring districts. Columns (iii) to (v) restrict the sample to
those women who report that the economic status of their husband’s family is similar to
the natal family. In column (iv), we additionally control for assets owned by the husband’s
family. Neither the coefficient nor the standard error on age at menarche changes.
Finally, we include these geographical variables in our second stage regressions. Appendix
Table A2 reports the results of our estimates for mathematics and reading scores of the child
when we include additional geographical controls. Column (i) and (iii) report the estimates
for maths score while column (ii) and (iv) report the same for reading scores. In column
(i) and (ii), we control for average temperature and altitude of the residing district, and in
column (iii) and (iv), we control the same variables averaged for the residing district and
the neighboring districts. Panel A reports the first stage estimates. The coefficient on age
at menarche is positive and highly significant for both the scores. The coefficient on age at
menarche is different across columns as we control for test language fixed effects. Panel B
reports the second stage estimates. Column (i) shows that an increase in age at marriage
23
by one additional year decreases the probability that the child will receive the lowest math
score (a child will be able to recognize a two digit number but will not be able to do more
complex number manipulation) by 1.9 percentage points and increases the probability of
receiving the highest score (a child will be able to divide a three digit number by a single
digit number) by 3 percentage points. Similarly, column (iii) shows that an increase in age
at marriage by one additional year increases the probability of receiving the highest reading
score by 2.7 percentage points. These results are similar to the benchmark results reported
in Panel B of Tables 4.1 and 4.2.
7 Mechanisms
This section provides suggestive evidence on the possible channels through which early mar-
riage of mothers’ can affect human capital of their children. Previous literature identifies
lower human capital and lower autonomy of women as a consequence to early marriage (Field
and Ambrus 2008, Jenson and Thornton 2003). We explore if early marriage translates into
lower human capital for the next generation through these channels.
The approach we follow to understand the channels involves OLS estimation of the coef-
ficients of mother’s age at marriage on test scores of their children with additional controls
measuring mother’s autonomy and her human capital successively. Subsequently, we com-
pare the results across the specifications with the baseline case where none of the additional
controls are used. If inclusion of a set of additional controls leads to a decline in the coef-
ficient on mother’s age at marriage, then it would suggest that the additional control may
be one of the operational intermediate pathways. This method provides only suggestive
insights to identify the channels. Therefore, the results described in this section should be
interpreted as suggestive.
Table 7 reports the OLS estimates of arithmetic and reading scores of children with addi-
24
tional controls measuring autonomy and human capital of mothers. We use four indicators
as measure of mothers’ autonomy. These variables take the value one if a mother decides (a)
what is to be cooked in the household on a daily basis, (b) whether to purchase an expensive
item, (c) the number of children she bears, and (d) what ought to be done if her children fall
sick. Mother’s years of education is used to measure her human capital. Column (i) and (v)
report the coefficient of mother’s age at marriage on arithmetic and reading scores without
any additional controls.22 A one year delay in mother’s marriage increases arithmetic and
reading scores of the child by 1.2 and 2.1 percentage points respectively. None of the esti-
mates change significantly after controlling for mother’s autonomy additionally as reported
in column (ii) and (vi). However, in column (iii) and (vii), the coefficient of mother’s age
at marriage changes both in magnitude and significance when mother’s years of education
is used as an additional control. Column (iv) and (viii) report the estimates when both
mother’s autonomy and years of education are controlled. The value of R-square across all
the specifications remain remarkably unchanged. These results suggest that a substantial
part of inter-generational effects of early marriage are mediated through lower human capi-
tal of mothers’ as inclusion of mother’s years of education in the regressions attenuates the
effects of mother’s age at marriage. We find little evidence that the early marriage effects
on test scores are mediated through mother’s autonomy.
8 Conclusion
This paper provides empirical evidence that early marriage of girls affects educational and
health outcomes of her children. Delay in age at marriage of a woman leads to an improve-
ment in her children’s human capital. A one year delay in woman’s marriage increases the
22All regressions control for mother’s height, birth year FE and age, indicator for poverty, land ownership,residence (urban/rural), religion, # household members, father’s age and education, # sibling of the child,birth order, gender, grade and test language FE.
25
probability that her children will be able to perform higher level cognitive tasks by 3 per-
centage points. We also show that mothers who marry later are more likely to send their
children to private schools and they spend more on education related items. These effects are
over and above the compensation offered by the household to the child in terms of resources.
Our results suggest that mandating a minimum marriage age and strictly enforcing it will
improve the education outcomes of children.
26
References
[1] Agarwala,S. N. (1957), “The age at marriage in India”, Population Index, Vol. 23, pp.
96-107
[2] Aguero, J. and Ramachandran M. (2010), “The Intergenerational Effects of Increasing
Parental Schooling: Evidence from Zimbabwe”, Working Paper
[3] Basu, K., Das, S., and Dutta, B. (2010), “Child labor and household wealth: Theory
and empirical evidence of an inverted-U”, Journal of Development Economics, Vol. 91,