Household expenditure patterns and gender bias evidence from selected Indian states.doc

household expenditure patterns and gender bias: evidence from selected indian states* pushkar maitra and ranjan ray acknowledge funding provided by an australian research council discovery grant. geoff lancaster and ranjan ray acknowledge funding provided by an australian research council large grant. we would like to thank gary deng for his excellent and painstaking research assistance, david stifel, geeta gandhi kingdon, brett inder, seminar participants at the university of sydney and purdue university and participants at esam 2003 meetings held at sydneythe econometric society australasian meeting, the , 32nd australian conference of economists held at canberra, the 2003 neudc conference held at yale university, the empirical development economics workshop at monash university and the european society of population economics annual conferencespe04 meetings at bergen for their comments on earlier versions of the paper. the usual caveat applies. geoffrey lancastera geoffrey lancaster, school of economics, university of tasmania, gpo box 252-85, hobart, tasmania, 7001, australia. email: geoffrey.l.au , pushkar maitrab pushkar maitra, department of economics, monash university, clayton campus, vic 3800, australia. email: pushkar.maitrab.au. corresponding author. and ranjan rayc ranjan ray, school of economics, university of tasmania, gpo box 252-85, hobart, tasmania, 7001, australia. email: ranjan.r.au. corresponding author.1household expenditure patterns and gender bias:evidence from selected indian statesabstractthis paper investigates using indian data the existence and nature of gender bias in the intra household allocation of expenditure. we estimate an extended version of the collective household model where the welfare weights, i.e. the bargaining power of the adult decision makers, are simultaneously determined with the households expenditure outcomes. we are able to detect significant gender bias in case of some items, most notably, education education where previous attempts have failed. we also find that the results of the test of gender bias vary sharply between households at different levels of adult literacy. this is particularly true of household spending on educationeducation. the gender bias in the case of this item is, generally, more likely to prevail in households with low levels of adult educational attainment than in more literate households. not only is this result of considerable policy importance, given the strong role that this item plays in human capital formation, it suggests another possible reason for previous failures to detect gender bias, namely, the overlooking of possible interaction between the adults literacy levels and the gender issue.keywords: expenditure patterns, gender bias, india.181.introductionthe issue of gender bias against women, especially girls, in the distribution of resources within the household has been a topic of much research in recent years (see, for example, deaton, (1989), pitt, rosenzweig and hassan, (1990), ahmad and morduch, (1993), haddad and reardon, (1993), bhalotra and attfield, (1998). udry (1997) and duflo (2005) presents a useful summariesy of the literature. this topicissue, which we revisit in this studypaper, is of particular importance in the context of developing countries since, if true, then gender bias in the intra household allocation of resources could explain what dreze and sen (1989) refer to as the phenomenon of “missing women” in south asia. the difficulty of observing the inner workings of the household, typically because of lack of adequate data, implies that one is restricted to examining this issue using externally observable outcomes like health, education and expenditure. a common approach has been to examine how household expenditure on a particular good changes with the gender composition of the household (deaton, 1989; 1997). the methodology usually adopted in testing for gender bias is based on a quantification of the differential between the reduction of the consumption of “adult goods” associated with the ceteris paribus addition of a boy and a girl child to the household. if this differential is statistically significant, then gender bias is said to exist, with the sign of the differential indicating the direction of the bias. since tthe validity of this procedure relies, however, on the strong assumption of demographic separability between “adult goods” and other items underlying the restrictive rothbarth model (see myles, 1995), ). this general test has been modified we modify it to test for the statistical significance of the gender differential between the marginal impact of a boy and that of a girl child on the expenditure share of an item in the household budget. stated differently, if one replaces a girl in a certain age group by a boy in that same age group, holding everything else constant, then the extent to which the expenditure share changes as a result of this thought experiment gives us a measure of gender bias. however, most of the investigations based on household expenditure data have failed to detect systematic evidence of gender bias even in societies where gender discrimination is known to exist as reflected, for example, in differential mortality rates between boys and girls.add some references: see duflo (2005) khanna et al (2003) using data from india find that in poor neighbourhoods in new delhi, girls are more than twice as likely to die of diahorrea compared to boys. rose (1999) finds that in times of droughts when families cannot feed everyone, they disproportionately sacrifice the welfare of girls. while deaton (1989) calls this a “puzzle”., udry (1997) argues that this might be due to sample truncation: girls have been so discriminated against that they have died, and are missing from the sample. rose (1999) suggests that the failure to find the expected gender bias in resource allocation might be due to differential mortality selection. she argues that if households are willing to sacrifice daughters in order to cope with adverse shocks, one might not observe gender bias in resource allocation since this masks gender bias in mortality selection in the first place. ahmad and morduch (1993) argue that, even if allocations are equal, girls might still be at a disadvantage because their needs might be higher. yet another reason might be that these tests have been conducted on data sets where gender bias is not likely to be particularly strong (cote d ivoire (deaton, 1989) or burkina faso (haddad and reardon, 1993), regions where girls are typically not seen as burden on their parents. this choice of the “appropriate” region might indeed be an important issue. for example, gibson and rozelle (2004) using data from papua new guinea and gong, van soest and zhang (2004) using data from rural china do find significant gender composition effects in expenditure patterns. the importance of region is further underlined by the results of the present studyis paper, which finds evidence of gender bias in education spending in some areas in india, typically the backward ones, though not in others. we adopt and extend the collective household model (bourguignon, et. al., 1993, browning and chiappori, 1998), by relaxing not only the conventional assumption of income pooling between the household income earners in achieving the expenditure outcomes but also relaxing a potentially limiting characteristic of such models, namely, that the welfare weight assigned to each income earner is exogenous to the household decision making process. following basu (20016), who criticised this assumption, we propose and apply a framework where the welfare weight of the adult male vis-vis the adult female income earner, namely, the “bargaining power” variable is jointly determined with the households expenditure outcomes. perhaps because of these departures from the conventional methodology, we are able to detect significant gender bias in case of some items, most notably, education education where previous attempts have failed. for example, kingdon (2005), using data from rural india, finds that, while individual level data exhibits gender bias, aggregate (household level) data fails to pick up any evidence of gender bias in educational expenditures. we also find that the results of the test of gender bias vary sharply between households at different levels of adult literacy. this is particularly true of household spending on educationeducation. the gender bias in the case of this item is, generally, more likely to prevail in households with low levels of adult educational attainment than in more literate households. not only is this result of considerable policy importance, given the strong role that this item plays in human capital formation, it suggests another possible reason for previous failures to detect gender bias, namely, the overlooking of possible interaction between the adults literacy levels and the gender issue.the rest of the paper is organised as follows: section 2 describes the theoretical framework and the estimation methodology adopted in this study. the two indian data sets that have been used here, namely, the “survey of living conditions” (slc) and the national sample survey (nss) are described in section 3. while section 4 reports the results of estimation on the slc data, section 5 reports the corresponding results on the nss data sets from the 3 chosen states, namely, kerala, bihar and maharashtra. section 6 examines the robustness of the principal results of the previous sections by controlling for the adults educational attainments, and allowing for interaction between gender and the education levels of both parents in the estimated budget share equations. given the importance of education in human capital formation, section 7 reports and discusses the tests of gender bias in educational spending in the remaining states of india. section 8 summarises the principal findings and concludes the paper. 2.theoretical framework and estimation methodologyfollowing browning and chiappori (1998), the households objective function can be written as the weighted sum of utilities of the different members. consider a household with two members: a man, and a woman . assuming that utility depends on consumption and leisure , the households problem can be written as:(1)subject to the full income constraint:(2) represents the utility of member , represents a vector of private consumption of individual , the household utility function is assumed to be the weighted sum of the utilities of the different members of the household. represent the wage rate, time endowment and leisure of individual . total household unearned income is i and p represents a vector of prices for x. note that prices and wages are assumed to be exogenous. the variable , which denotes the welfare weight of member m, depends on prices, household income and other variables such as the distribution of income, bargaining strength and so on. let uswe assume that the utility function is separable in consumption and leisure. households allocate income to each individual according to some sharing rule and then each individual maximises his/her sub-utility subject to the income allocated to him/her. if denotes the income share of male member m then denotes the share of female member f. individual m solves the following utility maximisation problem:(3)and individual f solves the following:(4)where m denotes household income.equations (3) and (4) can be solved to obtain the individual members demand equations in budget share form, i.e. as the share of that members allotted expenditure: (5a)(5b)where and denote the stochastic errors.now and are typically not observed: vvery few surveys collect data on gender specific expenditure on specific commodities so and are typically not observed. instead, . wwe assume that the household level budget share of good is the weighted average of the budget share of that good for the spouses , i.e. (6)substituting for and obtained from equations (5a) and (5b) in equation (6) and including standard demographic variables (household size and composition variables) as additional explanatory variables, we can write the estimating equation as follows: see lancaster, maitra and ray (2003) for details on the derivation of equation (7). (7)we estimate equation (7) as a system of equations. denotes household size, and denotes the number of individuals in the age category and gender class, with k denoting age group and denoting gender. note that iif household size and expenditures are sufficient to explain demand, then the coefficients will all be zero. however, in general, household composition will matter and the coefficients measure the effect of changing household composition (in proportion terms) on budget shares. a test of the statistical significance of the estimated difference constitutes a test of gender bias in the age group k in the expenditure allocation of item g. in other words, the null hypothesis, maintains that there is no gender bias in case of age group k and item g. in this paper, we have chosen the following age groups: 0 5 years, 6 10 years, 11 16 years, 17 60 years, and 61 years and above. females in the age group, 61 years and above, were chosen as the omitted category in equation system (7). if it is the case that the coefficients for boys and girls are different, then we can conclude that everything else constant, expenditure on a particular commodity depends on the gender composition of children. the choice of the age groups, 6 - 10 years, 11 - 16 years is justified by the fact that in india, while the former includes children enrolled in the primary school (grade 1 grade v), the latter includes children enrolled in the middle and high school.a significant feature of this study is that the male welfare weight, , which is a determinant of budget shares (see equation (7), is jointly estimated with household expenditure and the budget shares, . , measured by the males share of household income, denotes the male income earners bargaining power within the household, so that denotes that of the female income earner., as defined here, there is a fairly large literature in sociology and anthropology that argues that the males share of household earnings is a good measure of his bargaining power within the household (see blumberg and coleman, 1989, desai and jain, 1994 and riley, 1997). could be correlated with the unobserved determinants of budget shares, i.e., is potentially endogenous in the budget share equations - see lancaster, maitra and ray (2006) for details on the estimation of , the relative bargaining power within the household. per capita household expenditure is used as a proxy for household permanent income. household expenditure is easier to measure compared to household income and is measured with less error. moreover, household expenditure is a better proxy than current income because while income might be subject to transitory fluctuations, households use a variety of mechanisms to smooth consumption over time. however, household expenditure is also likely to be correlated with unobserved determinants of the household budget shares and failure to account for this potential endogeneity could result in inconsistent estimates.to account for the potential endogeneity of male bargaining power and per capita household expenditure, we jointly estimate the share of male earnings , per capita household expenditure and the budget shares as a set of simultaneous equations using 3sls estimation. the advantage of the 3sls estimation methodology is that it takes into account not only the potential endogeneity of the regressors but also allows for a non-diagonal covariance matrix of the residuals of the estimating equations. to focus our mind on theour empirical analysis that is conducted here, let us note that it is based on the 3sls estimation of the following system of equations:(8a)(8b)(8c)where is per capita household expenditure, are the vectors of exogenous determinants in the three equations, and the other variables are as defined earlier. while the estimable form for the budget share equations, , (equation (8c) has been presented previously in equation (7), we assume linear functional forms for equations (8a) and (8b) to keep the estimation simple.3.data and descriptive statisticstwo different data sets are used in the estimation. the first is the uttar pradesh and bihar “survey of living conditions” (henceforth referred to as the slc data set), which is a two-part study of rural poverty carried out in 1997 98 in south and eastern uttar pradesh and north and central bihar. the data set used is from the quantitative component of the study, which draws on data collected from household and community surveys modelled after the world banks living standard measurement surveys (lsms). the data was collected between december 1997 and march 1998 from 120 villages drawn from a sample of 25 districts in the states of uttar pradesh and bihar. 2250 households were interviewed for this survey. information was collected on household demographics, economic activities, housing, education, health, marriage and maternity history expenditures etc. for the purposes of this paper we use household level data from 2248 households. this data set is available from the world bank from /lsms/country/india/upbhhome.html the second data set used here is that on household expenditure collected in the 50th round (1993/94) of the national sample survey (nss). the nss, because of its wider coverage involving all regions/states in india, has been ext