Multiplying Siblings: Exploring the Trade-off Between Family Size and Child Education in Rural Bangladesh

Abstract

The question of whether large families have a subsequent negative impact on child health, education, and welfare is of pressing concern for development and public health policy. We tackle this question by empirically exploring whether parents face a trade-off between increasing the size of their family and investing in their children. Using data from a rural sub-district in Bangladesh (the 1996 Matlab Health and Socioeconomic Survey), we estimate the effect of siblings on school attendance, literacy, and numeracy. We use an instrumental variables approach, instrumenting first for children’s sibship size their mothers’ menarche age, and also instrumenting for children’s sibship size with the household’s treatment status (in the Matlab family planning experiment). Although we find no effect of siblings on school attendance, we find that additional siblings increase the likelihood that children are literate and numerate. These effects are greater for girls and younger siblings, consistent with positive numeracy and literacy spillovers from older siblings to younger ones. The results provide counter-evidence to the quality-quantity trade-off theory and demonstrate that sibling education spillovers may dominate any reduced investment on the part of parents.

Keywords:

• Education
• family size
• fertility

JEL CLASSIFICATION CODES:

• D1
• O1
• I2

Disclosure statement

The authors declare that they have no conflict of interest.

Notes

1 We use sibship size and family size interchangeable, both referring to the number of biological children in a family.

2 Black et al. (2005) observe a negative correlation between family size and education that disappears when birth order controls are included. They conclude that family size itself has little direct impact on child quality, and instead negative marginal effects on quality are experienced through birth order effects. The literature on siblings, human capital and birth order confirms differential outcomes by birth order (Emerson & Souza, 2008).

3 We also perform supplementary analysis using the nationally-representative Bangladesh Integrated Household Survey data (BIHS). Similarly, in this data we test whether the number of siblings has an effect on literacy, for all children, for boys and for girls. The resulting tables from this analysis are in Table 11 in the Supplementary Materials. The results are discussed in the results section.

4 The Demographic Surveillance Survey data provided exact dates for events such as births, deaths, and marriages, which were used to check the accuracy of respondents in the MHSS.

5 Although it may seem possible that parents may influence which children are selected to be interviewed, the MHSS documentation specifically clarifies that enumerators list all household members, each with a line number as the very first step of enumeration, then begin the more detailed interviewing process. Adults are interviewed first, and then the two children are selected randomly using their line numbers from the household roster in conjunction with a random number table to select participating children. Although we cannot entirely rule out the possibility that randomization could be undermined by persistent parents, the MHSS documentation specifically indicates that the problem of select samples of children in larger households is avoided.

6 From the survey questionnaire, can read means that the child can read a Bangla newspaper and can write means that the child can write a letter in Bangla.

7 Although LPM may produce predicted values outside of [0,1], we use this estimation strategy because it does not impose a functional form on the error term. Moreover, we do not make forecasts on the outcome variables.

8 We follow Peters et al. (2014) in using this definition of sibship size to ensure that any two children in the same family have the same reported number of siblings regardless of whether any siblings have died.

9 Mother’s education attainment variables include indicators for literacy and numeracy, and a set of indicators for completed years of education.

10 We use an wealth index to determine quintiles, using the methods outlined in Davila, McCarthy, Gondwe, Kirdruang, & Sharma (2022).

11 Birth order is top-coded at birth 10.

12 A similar downward bias would appear in the case of marital age and intra-household bargaining power. Women with older menarche ages enter the marriage market later, and this may make them closer in age to their husband, potentially increasing their intra-household bargaining power. These women may have fewer children, and pursue more education for their children, creating the same downward bias as omitted socioeconomic status. However, if for example, women who marry later are more likely to work outside the household, it is possible that they may be occupied with work duties, potentially leaving less time to invest in the educational skills and attainment of their children. This second situation would imply an introduction of bias into our estimates.

13 We experimented with adding controls for father’s education attainment in the descriptive statistics and regression estimates as well. Unfortunately, these variables are only available for a subset of the sample (those whose father was living in the household when the MHSS survey was administered) thus their inclusion would decrease the sample size substantially.

14 Tables 1–6 in the Supplementary Materials demonstrate the results of this analysis for a subsample of women who have completed fertility. In this case, we examine the effect of sibship size on key education outcomes for children of mothers older than 35 years. Although the magnitude and direction of the effects are very similar to the main results of this study, the effects are mostly not statistically significant (smaller sample). Importantly, however, the first stage statistical significance remains: there is a negative and significant effect of menarche age on total number of children at the end of a mothers’ fertile period.

15 Under certain characteristics, it is possible that age of menarche is an indicator for a mother’s overall health as a child (young women with a higher BMI are likely to have lower menarche ages). For example, if age of menarche is correlated with an individual-level socioeconomic status, which is also a strong predictor of numeracy and literacy, then our two-stage least squares estimates would be contain bias. It is worthwhile to note the direction of this bias. If socioeconomic status remains an omitted variable, this would actually downward bias the IV estimate of the causal effect of family size on educational abilities (due to a downward bias in the first-stage estimate) (Field & Ambrus, 2008).

16 Attendance results disaggregated by gender are omitted, but also showed no effect of sibship. We further tested the heterogeneous effect of gender by interacting gender with the effect of sibship size. We find qualitatively similar results to the split sample analysis.

17 Because Qian (2009) studies the effect of the relaxation of the One-Child Policy, she measures the effect of gaining a singular sibling (from a sibship size of 1 to 2), whereas my effects are more generalizable in that they show a positive average effect on reported skills of gaining (any) sibling.

18 We replicated this analysis using the nationally-representative BIHS data, and these results are shown in the Supplementary Materials. Column 1 of Table 11 shows the first stage of the analysis—the effect of mother’s menarche age on the number of children. The effect is negative, but it is not statistically significant. The second stage regressions show mixed results. For the mixed gender sample, the effect of family size on literacy is small, positive, and not statistically significant. In the analysis for just girls, the second stage regression shows a small negative effect on literacy. The effect is actually positive for boys, however the standard error is large and thus the effect is also statistically insignificant. In sum, the effect of sibship size on literacy and years of schooling using BIHS data is zero. Although the heterogeneous findings between BIHS vs. Matlab does limit the external validity of our results, there are some important distinctions between the BIHS and the Matlab data. First, birth order is only available in the 2015 BIHS data for mothers and children under 5. Birth order has been shown to be an important covariate to reduce bias in measuring the true effect of family size (Black et al., 2005). Because children under five are generally illiterate, we are unable to include this as a control variable. Numeracy is also unavailable in the 2015 BIHS, so we only examined the effect on literacy. Secondly, the nationally representative data that had menarche, sibship size and literacy and was available to use was from 2015, significantly further along the demographic transition in Bangladesh. However, this replication exercise increases the robustness checks in the analysis.

19 In the analysis for oldest siblings, we removed the controls for birth order. However, because the effect of sibling spillover may have differential effects for the second child relative to the eighth child, we do include birth order controls in the subset of the younger children sample.

20 We completed this analysis using the interaction of oldest/younger with sibship size as well; this yielded qualitatively similar results.

21 We explored labor force participation as another possible avenue of numeracy and literacy spillovers from older to younger siblings. However, in this sample, children with older teenage siblings in the labor force were on average 10 percentage points less likely to know how to read, write, add and multiply (these differences are statistically significant).

22 We do test for the quality-quantity trade-off by re-running the main analysis using enumerator-tested cognitive skills as the outcome variables (the sample size drops to ∼1500) and we find no significant effect of sibship size on education skill outcomes. The sample of cognitive test respondents is much smaller than the sample of mother-reported and self-reported literacy and numeracy. Thus, we focus our results on the larger sample of reported abilities.

23 In Table 14 we test for differences in characteristics of those who over-report and those who under-report education skills. A systematic pattern of over-reporting or under-reporting may also bias our results. We find that over-reporting mothers and teens are more literate and educated, and more likely to be mistaken about their daughters’ (rather than sons’) skills. Importantly for the validity of our results, the mothers and teens who overreport do not have systematically larger family sizes or earlier associated menarche ages than those who underreport.