Appendix 3 Subsample estimates
6.3 Sample, measurement, and empirical strategy
6.3.5 Problem of endogeneity
At this stage, it may be worth pointing out the fact that estimating the impact of empowerment on life satisfaction can be biased either because of reverse causality or omitted unobservable characteristics (e.g. personal characteristics) (Fernandez et al., 2015). In addition there may be some omitted variables at the community, household and personal level. In particular, it is possible that communities might have different gender norms that could influence women’s well-being. These could include the imposition of patriarchal norms on women such as purdah, the custom of early marriage or even practices like female genital mutilation in some parts of the world (e.g. Africa). These norms might vary across communities and influence our estimates of the effect of empowerment on the LS. In addition, there could be household level variation in norms that are not captured in our empowerment indices. Again, this could bias the coefficient of empowerment in our model.
Correcting for omitted variables of this kind is not always easy because such factors are not easy to measure. However, we have attempted to correct for them by including household and regional fixed effects, which should capture the unobserved effect of household and regional norms that our model fails to control explicitly. First, we estimate a community fixed effects model which allows us to capture omitted locality specific gender and cultural norms (e.g. the custom of purdah governing the outside engagement of women and hence their threat points in intra-household bargaining) that can bias our estimate of the empowerment variable. Second, we exploit the fact that BIHS collected data on both husband and wife to estimate a household fixed effects model. Household levels fixed effects can help control for time- invariant household specific factors that affect well-being and are correlated with women’s empowerment. By including both community and household level fixed effects, we are able to control for unobservable factors that might influence life satisfaction at both these levels.
The other cause of potential endogeneity is reverse causality or feedback effect from life satisfaction to empowerment. It is possible that extrovert and happier women proactively take part in household decisions, while at the same time, being involved in these decisions makes them happier and more fulfilled. In order to address this particular issue which undermines the exogeneity of the empowerment variable in the well-being function, we have estimated Equation 6.1 using the instrumental variables (IV) approach as well. To do so, we use information on the average number of community activities that the villagers participated over the past 12 months. The instrument is computed at village level for women and men separately, after excluding the respondent. More explicitly, the construction of the instrument is expressed in the Equation 6.2.
Average number of community activities participated by the villagers,
1 1 1 , , ,
n ncap ancap n j j g v g v Equation 6.2Where ancapv,g is the number of community activities participated by n-1 individuals of gender g in the village v; ncapv,g,j is the number of community activities j-individual of g- gender in v-village participated in the past 12 months. The number of villagers of g-gender is n and g takes account of the gender of the villagers i.e. male or female. Eventually ancapv,gis used to predict the i-th individual’s empowerment index in the first stage. The IV process is delineated as the following Equations 6.3 and 6.4.
i i i i i i i v i ancap female DE HL SE HH D v d em5 0 1 2 3 4 5 6 7 Equation 6.3 i i i i i i i i i em d female DE HL SE HH D u LS
0
1 ˆ5
2
3
4
5
6
7 Equation 6.4 From Table 6.1, we can see that on an average the female villagers had participated in less than one (0.817) activity, male villagers participated in more than one but less than two (1.6) community activities. Although we have discussed the rationale for using this instrument in the previous chapter, it may be useful to point out that our current instrument is slightly different from the one used in the previous chapter.In chapter 5, we have used the information on how many community activities a mother had participated in the previous 12 months as an instrument for her empowerment index. On the other hand in this chapter, we use information on how many community activities the
villagers had participated in the previous 12 months. Clearly we use two different instruments to deal with the potential endogeneity of the empowerment variable; however, the following rationale may be useful to justify their usage.
While estimating the effect of mothers’ empowerment on food security indicators, we suspect empowerment variable is endogenous because the empowerment index possibly does not account for household-specific gender norms that may have a bearing on women’s ability to influence households’ food consumption bundle and on their ability to influence the allocation of nutrients among household members. That being said, on the other hand, we suspect that the source of endogeneity of empowerment index in the life satisfaction model may originate from a personal trait such that a happier individual may proactively exercise agency and thereby becomes empowered. This indicates to the fact that the source of endogeneity is different across the models.
The criteria of a suitable instrument suggest us that an instrument should not be correlated with the source of bias and that it is correlated with the endogenous variable as highly as possible. Participation in the community activities in the past year may be relevant to the current state of empowerment because through participation individuals are likely to have an exposure to different ideas, likely to be aware of their rights, likely to be confident to raise voices against concerns as well as to seek support in case of an issue. Thus women’s participation in the community activities in the previous year may well suggest the extent of current empowerment. However what if the women had already been empowered in the previous year and hence participated in the community activities. Since in the dataset there is no information on the extent of past year’s empowerment, we cannot examine if there has been any improvement in the level of empowerment because of the participation. This we recognise as a potential weakness of our strategy. That being said, we have also argued that women’s participation in the past year is not directly related to household-specific gender norms concerning intra-household allocation of nutrients in the current period. This means that we expect that women’s participation in the previous year has no direct effect on individuals’ nutrient intake in the current period.
We could have used the same instrument in the life satisfaction model if we did not suspect the source of endogeneity is arising from a reverse causation because of personal traits. In other words, using individual-level participation information would not guarantee that the instrument is uncorrelated with the source of bias. Hence we prefer using the community level
variable that computes an average of the number of community activities in which the villagers except the individual being studied had participated in the past year. The idea here is that the community average would still give an approximate estimate of individual-level information on participation, which can be used to predict current empowerment. There may still be a concern that a more engaged community might create a happier atmosphere to live in and thereby influences people’s life satisfaction, which hints at community-specific characteristics. However we have attempted to address such concern by estimating the life satisfaction model at community fixed-effects level and have found evidence to suggest that the relationship between empowerment and life satisfaction is not influenced by community- specific characteristics.
Following to the above argument for using two different instruments in two different models, we would also like to stress that community-level information on participation may not be truly exogenous in predicting empowerment in the case of food security model. If a higher participation by the villagers is meant to be symptomatic of a prosperous, advanced or well- off community then it is very much likely that the food security of that community’s members is also enhanced. Moreover household-specific gender norm may also be associated with community-level characteristics. Hence in the food security chapter we prefer mother-specific participation information rather than community level variable. All things considered, we prefer to use two different instruments in our two separate models with appropriate caution. In what follows we will also perform the formal IV diagnostic tests.