Intra- and inter-household externalities in children's schooling: evidence from rural residential neighbourhoods in Bangladesh

Abstract

This article tests for neighbourhood effects on children's schooling, using unique data on rural residential neighbourhoods from Bangladesh. We find that school completion of children is positively and significantly affected by the mean grade completion of other children in the neighbourhood. We then present three pieces of evidence that suggest that the social effect offers a valid explanation. Firstly, the evidence we find of inter-household externalities is not driven out by control for a host of neighbourood and household attributes. Secondly, the result remains robust to neighbourhood composition effects: it is unchanged as we purge our main sample of the households within the neighbourhood that are potentially linked in terms of their recent history of partition. Thirdly, a similar peer effect is found for adults who completed schooling before the introduction of existing educational reforms in rural areas suggesting that the observed effect of growing up in educated neighbourhood does not merely capture the influence of common exposure to various government educational interventions. As a by-product, the article also provides evidence of intra-household externality in children's schooling, net of neighbourood externalities. We conclude by discussing the implication of these findings for education policy design.

Acknowledgements

I am thankful to Stefan Dercon, Geeta Kingdon and participants at the Centre for the Study of African Economies (CSAE) seminar (Oxford), 2003 European Union Development Network workshop (Bonn) and 2005 Royal Economic Society Annual Conference (Nottingham) for their helpful comments and suggestions. I would also like to thank ICDDR,B for granting me access to the census data used in this study. However, the usual disclaimers apply.

Notes

¹ Sociologists have discussed a variety of ways through which educational and socio-economic background of constituent members in the social space benefits children's schooling outcomes. For a review, see Mayer and Jencks (1989).

² An exception is Weir (2007), which addresses some aspects of the issue using household survey data from rural Ethiopia.

³ However, children enrolled in the same grade in a school may also compete and decide not to cooperate.

⁴ The framework is extendable beyond household, to neighbourhood and can be disaggregated by characteristics of the literate household member such as gender and age (Basu and Foster, 1998).

⁵ In the calculation of bari averages we take out the index child and all his/her other family members. Hence, despite being calculated at the bari level, for each child E_b [S_bi ] and E_b [x_bi ] vary across households within the bari.

⁶ Support for such a specification also comes from Ginther et al. (2000) who find that the coefficients on neighbourhood variables tend to fall in value and lose statistical significance as the specification of family variables become more complete.

⁷ Apart from the issue of resource pooling among ‘linked’ households, separation of linked households is important for another reason. Some education may have been provided during the period when the linked household members were still co-resident. Even if it is the case that an index child is enrolled in school only in post-partition period, correlation with schooling of children in other households today may simply capture the common home learning environment shared in their pre-school years. Since most of the children (aged 6–17 years) were not born in 1982, this, however, is less of an issue to the extent most of the schooling occurred soon after 1982.

⁸ First, one household was drawn from each of the baris that consisted of a single household. A total of two households were sampled from the remaining baris, each of which had more than one household. This led to data on a total of 24 266 individuals in 4364 households. Detailed information on the MHSS 1996 is available in Rahman et al. (2001).

⁹ We focus on this age group to ensure that all individuals in our 1996 sample were of school age. However, focusing on this age group implies that our dependent variable, grade completion, is left censored. The resulting estimates of the education spillover effects (using censored data) are likely to be biased downward; without accounting for censoring in the data would yield a lower bound of the true spillover effects.

¹⁰ Given this age range, censoring of the dependent variable remains an issue, either due to nonenrollment or current enrollment (so that last grade completed is yet to be observed). To correct for potential bias due to nonenrollment in school, one may estimate a sample selection model. However, in the absence of convincing exclusion restriction, we have not pursued this route. Hence, our results need to be interpreted with a degree of caution.

¹¹ Controlling for mother/head's spouse age is important because older mothers are likely to have greater experience as a child carer or home tutor.

¹² The separation of the full sample data by gender of the head is unavoidable because in almost 95% of cases, female-heads do not have their spouse present in data (either because they are divorced or the male partner is located outside the study area). Splitting sample observations by gender is important because there may be gender differences in the ability to benefit from social interactions.

¹³ Education of bari ‘nonmother female’ adults still have no effect.

¹⁴ Although the effect of mother's education is not necessarily exogenous to the extent that male heads with higher taste for educated children marry educated female so that mother education merely captures unobserved taste for education of the head.

¹⁵ This approach discards all households set up outside the study area i.e. Matlab Thana. These are treated as events of migration.

¹⁶ In this scenario, Foster (1993) further distinguishes between ‘new household’ and ‘inherited household’. If the original head was still in the study area and lived in the household of one of the new heads, then that new head is assumed to have inherited a household. Such a distinction between inherited and new households is not made in our analysis for we do not focus on the impact of partition.

¹⁷ In our analysis, children in linked (nonlinked) households benefit from interactions with children from linked (nonlinked) households only. This imposes a restriction that nonlinked (linked) household children are outside the social space of children from linked (nonlinked) households.

¹⁸ As excluded instruments, we use total market value of assets and total number of cows owned by the household (see column (5)). These instruments are highly significant in the first stage and comfortably pass the validity test. Despite the fact that our excluded instruments pass the validity test, we are cautious in being conclusive about these IV results. The use of cows as instrument, for example, could be contested. For farm households, however, this variable potentially reflects demand for child labour and hence a priori, could be disputed as a choice for valid exclusion restriction.

¹⁹ The test involves specifying that the exogeneity of schooling variable is under suspicion. Under the null hypothesis, the probit model is appropriately specified with all explanatory variables as exogenous. Under the alternative hypothesis, the suspected endogenous schooling variable is expressed as linear projections of a set of instruments (including bari mean schooling variables), and the residuals from those first-stage regressions are added to the model. Under the null hypothesis, these residuals should have no explanatory power.