Gender and household education expenditure in Pakistan

Abstract

Pakistan has very large gender gaps in educational outcomes. One explanation could be that girls receive lower educational expenditure allocations than boys within the household, but this has never convincingly been tested. This article investigates whether the intra-household allocation of educational expenditure in Pakistan favours males over females. It also explores two different explanations for the failure of the extant ‘Engel curve’ studies to detect gender-differentiated treatment in education even where gender bias is strongly expected. Using individual level data from the latest household survey from Pakistan, we posit two potential channels of gender bias: bias in the decision whether to enrol/keep sons and daughters in school, and bias in the decision of education expenditure conditional on enrolling both sons and daughters in school. In middle and secondary school ages, evidence points to significant pro-male biases in both the enrolment decision as well as the decision of how much to spend conditional on enrolment. However, in the primary school age-group, only the former channel of bias applies. Results suggest that the observed strong gender difference in education expenditure is a within rather than an across household phenomenon.

Acknowledgments

This article has benefited from our discussions with Jean Drèze, Marcel Fafchamps and Måns Söderbom and from comments from seminar participants at the Department of Economics, University of Oxford. Any errors are ours.

Notes

¹ If parents have a preference for having at least one (or some desired number of) boys in the household, they will continue child-bearing till that desired number is reached. This sort of behaviour will lead to girls in the population having more siblings, higher average household size and lower per capita resources than boys. Lower per capita resources due to larger household size imply that girls’ outcomes will be worse than boys’ even in the absence of any within-household differential treatment of sons and daughters.

² Studies by Deaton (1997) and Bhalotra and Attfield (1998) focus on food consumption.

³ The conventional application of the Engel curve technique may fail to pick up bias against girls for another reason as well, namely if the distributional assumptions about the dependant variable and thus the specification of the budget-share equation are wrong. For instance, if the education budget-share for households with positive education spending is distributed log-normally but, because the budget-share equation is fitted on all (zero and nonzero education budget-share) households, the researcher has to use absolute budget-share rather than the log budget-share as the dependant variable, leading to incorrect SEs. However, in large samples such as ours, this is not a particularly important worry.

⁴ These age-gender categories are defined as M0TO4, F0TO4, M5TO9, F5TO9, M10TO14, F10TO14 etc. and are the proportion of males (M) and (F) aged 0–4, 5–19, 10–14 and so in a given household.

⁵ These age-groupings are the same as those used in Subramanian and Deaton (1991) and in Kingdon (2005) for India. While regressions were also estimated for the 20–24 age category (corresponding with higher education ages), we do not report the detailed findings for this age group here (see Aslam and Kingdon, 2005, for these results). Sample selection issues are stronger for this age category because in this age, a high proportion of girls are married and do not live in their natal homes.

⁶ The effect of censored observations (zero consumption expenditure on an item) is a well-discussed issue in the Engle curve literature. For instance, see Beneito (2003) and Yen (2005).

⁷ The total educational expenditure (TOTAL_EDU) variable was truncated at Rs 25 000 to exclude outliers Only 0.6% of the sample reported expenditures greater than Rs 25 000.

⁸ See Aslam and Kingdon (2005) for all the disaggregated results.

⁹ The provinces were also disaggregated by region (urban and rural). A total of 54 equations have been estimated. There are four provinces and three territories in Pakistan. We also wish to present results for Pakistan as a whole, thus making eight geographical units. For five of these units, we have broken the unit up into three samples: rural, urban and whole (rural + urban). Thus, in total we have (5 × 3) + 3 = 18 separate samples. For each of these samples 3 different equations have been fitted, implying a total of 18 × 3 = 54 equations using household level data. Table A1 does not report results by province due to space constraints. Tables 4 and 5 also do not report results by regional categorization for the different provinces. Disaggregated results are available in Aslam and Kingdon (2005).

¹⁰ The theoretical literature suggests that at any given level of per capita resources, larger households will be better off because they share household public goods, such as housing, consumer durables etc. Larger households should, therefore, be able to allocate larger shares to private goods such as education provided they do not substitute towards the ‘cheaper’ public goods. In Pakistani households, economies of scale could be especially important given the norm of a ‘joint family’ system. Deaton and Paxton (1998) did not find evidence of such economies of scale across 7 high and low income countries, though they examined food budget shares.

¹¹ Three exclusion restrictions were used in controlling for possible sample selectivity: LAND_OWN (whether household owns any agricultural land), LAND_ACRES (the amount of land owned by the household) and BUSINESS (whether the household is an owner/proprietor of a nonfarm business). A priori, we might have expected a household owning agricultural land or a business to have a higher demand for child labour, i.e. to affect the school enrolment (or positive education expenditure) decision, but not to affect conditional educational expenditure. However, in no case were the exclusion restrictions jointly significant at the 5% level. The F tests revealed that the p-values of the joint significance of the exclusion restrictions in the probit of current enrolment were: 0.14 (age 5–9), 0.53 (age 10–14) and 0.06 (age 15–19). Only in the 20–24 age-group, the exclusion restrictions were jointly significant (at 4%), but the Lambda term was insignificant (t = −1.27).

¹² If girls’ unobserved traits are important in parents’ decisions about their enrolment/education and boys’ traits are not important (or less important) to parents’ decisions about their schooling, then any pro-male bias will be over-estimated because the female demographic variables will suffer from greater downward bias in the conditional education budget share equation than will male demographic variables.

¹³ For example for the full sample Punjab, the coefficients on M5TO9 and F5TO9 in the conditional OLS of LNEDU_SHARE was bm = 0.9426 and bf = 0.7840 respectively. We can obtain the log transformations of these by using the property of the log normal distribution that the conditional expectation of E(w | x, w > 0) equals exp (xβ + σ²/2). The Exp(·) for this sub-sample is 0.1838. Thus the transformed marginal effect for males is bm*Exp(·) = (0.9426) * (0.0720) = 0.0679 and that for females is bf * Exp(.) = (0.7840) * (0.0720) = 0.0565. The difference between the male and female marginal effects is 0.0679 − 0.0565 = 0.0114. In the table all DME are multiplied by 100 and so the reported DME is 1.14.

¹⁴ For instance, tuition fees for males and females aged 5–9, 15–19 and 20–24 are statistically insignificantly different from each other (for the 10–14 age group they are significantly higher for girls). Similarly, the data suggest that expense on transport is significantly greater for girls aged 10–14, 15–19 and 20–24.

¹⁵ Alderman et al . (1996) attribute reduced availability of schools for females in rural Pakistan to lower adult cognitive achievement while Lloyd et al . (2002) suggest that single-sex girls’ school availability is a key determinant of parent's decision to enrol girls in school in rural Pakistan.

¹⁶ Although we estimated a total of 288 equations (aged 20–24 was a separate category and the results were disaggregated by region for all provinces), as before we do not report all results. More detailed results are available in Aslam and Kingdon (2005).

¹⁷ The full results of the individual-level regressions are available from the authors.

¹⁸ As a referee of this journal points out, age gaps between siblings may differ for girls and boys within the household since a new born's gender may affect parental decisions about the spacing of the next birth (Angrist et al ., 2005). The family fixed effects approach does not address this issue or the possibility of time-varying unobserved household heterogeneity.

¹⁹ An Oaxaca decomposition suggests that much of the gender earnings gap is not explained by differences in the observed characteristics between men and women, suggesting a good deal of gender discrimination in the labour market. Studies by Ashraf and Ashraf (1993) and Siddiqui and Siddiqui (1998) also find evidence of gender discrimination in the Pakistan labour market, though they suffer from methodological limitations such as lack of control for sample selectivity in female work participation and for the endogeneity of education.