Reconsidering Gender Bias in Intrahousehold Allocation in India

Abstract

Finding evidence of gender discrimination among children in the intrahousehold allocation of goods has often proven to be difficult. This article uses data on education expenditures in India to test whether data aggregation, data reliability and the statistical method used help explain this pattern. Results suggest that discrimination against girls is increasing in age and robust to the statistical method and the expenditure measure at the all-India level, although state-level results are more sensitive. I find that data aggregation and statistical method are important factors in detecting gender bias, while data reliability does not seem to play a major role.

Acknowledgements

I thank Ashwini Deshpande, Suresh Tendulkar and participants at the Midwest Economics Association Annual Conference in Chicago for valuable comments and feedback.

Notes

1. For this and other evidence on gender bias in developing countries see for example Klasen (1996), Sauerborn et al. (1996), Messer (1997), Miller (1997), Asfaw et al. (2010), and Knight et al. (2010).

2. See for example Deaton (1989), Ahmad and Murdoch (1993), Haddad and Reardon (1993), Subramaniam (1996), Gibson (1997), Murdoch and Stern (1997), Gibson and Rozelle (2004), Liu and Hsu (2004), Fuwa et al. (2006), Lee (2008), Himaz (2010).

3. These results are consistent with evidence on gender differences in the Indian education system as summarised by Kingdon (2007).

4. Bhalotra and Attfield (1998), for instance, estimate semi-parametric Engel curves and test whether potential effects are washed out by too large age categories, but still find hardly any difference between the treatments of children. Gong et al. (2005) also use semi-parametric analysis for different goods and conclude that gender bias seems to occur more through higher mortality rates and lower school enrolment for girls than boys, rather than through the consumption of common goods.

5. See, for example, Rosenzweig and Wolpin (2000) for a more detailed discussion of the quantity–quality trade-off and its empirical importance.

6. In my dataset, education expenditures do indeed have a lognormal distribution.

7. See Kingdon (2005) for a more extensive explanation and justification of these specifications.

8. The raw results are available from the author on request. In the conditional OLS specification, the result is the outcome of bootstrapping the difference in the calculated marginal effects with 1000 replications. F tests are used to test the equality of marginal effects in the unconditional OLS and probit specifications.