Are female-headed households less resilient? Evidence from Oxfam’s impact evaluations

ABSTRACT

Given the complexity of measuring resilience, development practitioners are constantly seeking observable proxies that can help target resilience-building initiatives. Evidence suggests that female-headed and male-headed households differ in terms of their vulnerability and resilience, so household head gender may provide policymakers with important information about how best to target their interventions. However, the extent to which these gender differences can be explained by other observable characteristics, such as education and household demographics, remains an open question. Using an index of resilience employed by a large non-governmental organization (Oxfam GB), we provide evidence on this question by comparing the resilience of female- and male-headed households interviewed in a series of 16 evaluations of rural development projects carried out in 12 countries across Africa, Asia, and Latin America. We find that there is a statistically significant difference between female- and male-headed households, on average, in terms of their measured resilience, and that only just over half of this difference can be explained by observable characteristics. However, since the size of this difference is small, using information on household head gender does not significantly improve the accuracy of targeting methods, such as proxy means tests, that aim to identify households for resilience-building initiatives.

KEYWORDS:

• Resilience
• adaptation
• gender
• female-headed households
• targeting

Acknowledgements

The authors are indebted to Alexia Pretari, Simone Lombardini, and the anonymous reviewers of this article for their valuable comments. The authors would also like to thank Oxfam GB for generously making the data available for this research.

Disclosure statement

The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.

Notes on contributors

Rob Fuller is the Monitoring and Evaluation Manager in the Smallholder Development Unit at AgDevCo, and was previously an impact evaluation adviser at Oxfam GB. He is a specialist in impact evaluation for rural and agricultural development programmes, and in particular has worked on developing frameworks for the measurement of resilience.

Jonathan Lain was, at the time this article was written, an Economist in the Poverty and Equity Global Practice at the World Bank in Indonesia, and prior to that was an impact evaluation adviser at Oxfam GB. His research focuses on labour market inequality, urban poverty, and using new data sources – including ‘big data’ – to improve the measurement of key economic outcomes. He has also conducted impact evaluations of labour market programmes and resilience-building initiatives in both developing and developed country settings.

ORCID

Rob Fuller http://orcid.org/0000-0001-9107-4045

Jonathan Lain http://orcid.org/0000-0003-3872-439X

Notes

1 Kumar and Quisumbing (2013) corroborate this finding with data from Ethiopia, showing that the more severe impact of the crisis on female-headed households is maintained even after controlling for other household-level characteristics.

2 Reports describing the full results for each of the Effectiveness Reviews can be found at www.oxfam.org.uk/effectiveness. The results are synthesized in Fuller and Lain (2017).

3 The two evaluations carried out in Nepal are exceptions to this. In these evaluations – carried out in an area in which many men migrate abroad to work – the questionnaire allowed respondents to identify an individual who was not currently resident in the household as the head of household. For the analysis in this paper, in households that were recorded with an absent head, the spouse of that individual, or (if she/he has no spouse) the oldest member currently resident in the household, is assumed to be the actual household head.

4 Under the first alternative definition, households were considered to be female-headed if and only if there were no adult men (aged 16 years or over) living in the household. Under the second definition, households were considered to be female-headed if the oldest female household member was at least 12 years older than the oldest male household member. The correlation coefficients between the three definitions across the pooled dataset range from 0.55 to 0.63.

5 In Oxfam’s evaluations, characteristics were included in the resilience measure only if they were seen as amenable to change as a result of development interventions (though not necessarily as a result of the activities of the specific project being evaluated). Indicators related to household structure and education were considered to be exogenously determined, so were not included in the resilience measure. In addition, households’ socio-economic status was not included, since, in line with Oxfam’s definition, this was considered primarily to be an outcome of household resilience – something that would improve despite shocks, stresses, and uncertainty in resilient households – rather than an indicator of resilience per se. However, the ownership of specific productive assets (a measure closely related to household wealth) was recognized to be an important driver of resilience, and in most cases was included in the index (Fuller & Lain, 2015). Nevertheless, since the boundaries of what may be considered as a characteristic are somewhat blurred, the analysis in Section 4 adds variables representing household structure, education, and wealth as covariates to the regression models, so that their relationship with resilience is accounted for even without them being directly included in the resilience measure itself.

6 Unlike in the original Alkire and Foster (2011) approach, this is the final aggregation step: this index was not compared to some overall threshold to make a binary categorization of households as either resilient or not resilient.

7 The wealth index was constructed using different combinations of housing and asset indicators for each dataset. However, the protocol adopted for constructing the wealth index was consistent across the datasets, following the approach of Filmer and Pritchett (2001).

8 The outcomes in this paper are therefore effectively reported in terms of Cohen’s d, a measure of standardized mean difference between groups (Cohen, 1992).

9 This is analogous to the situation described by Mincer (1958) when constructing a regression model for earnings: including the (logarithm of) the number of hours worked as a dependent variable in the regression model is equivalent to accounting for the number of hours worked in the independent variable by calculating earnings per hour.

10 As an additional check that the results in this paper were not driven by the project interventions, the analysis was repeated after restricting the sample to households in the comparison groups defined in Oxfam’s impact evaluations (that is, households that did not participate in the projects being evaluated). The resulting estimates are all of a similar magnitude to those presented in this paper, though the main effect is no longer statistically significant, even at the 10 percent level, when applying the two-step random effects approach outlined in Section 4.3.

11 As such, it is not possible to distinguish de facto from de jure female-headed households, using the terminology of Klasen et al. (2015).

12 In principle, random effects can also be introduced in a one-step approach like the pooled regressions on which the main results are based. Incorporating random effects in a one-step approach would involve allowing for dataset-specific values of $\text{β}$ and assuming that these follow some pre-defined distribution (Kontopantelis, 2018). We initially tried to use a one-step approach with random effects in this way, but doing so produced virtually the same point estimates for the average $\text{β}$ (from all the datasets) that currently arise from the pooled regressions, but with smaller standard errors. This is because the distributional assumptions required for the one-step approach with random effects give the model more information, increasing efficiency. Allowing for random effects in the two-step approach rather than the one-step approach is therefore likely to serve as a more exacting robustness check. Similar logic also applies to the treatment of the error terms: the two-step approach explicitly allows for dataset-specific standard errors – rather than assuming some common error term as in the pooled regressions (${\epsilon }_i$) – without requiring complicated distributional assumptions. Additionally, the incorporation of random effects into the two-step approach also aligns more closely with much of the existing meta-analysis literature, and offers a ‘tried-and-tested’ way of allowing for heterogeneity in the effect of interest (Fisher, 2015).

13 Reducing the number of indicators comprising the resilience index does not make it systematically more or less likely that households will be given a higher resilience index score. The resilience index is still calculated as the number of indicators on which a household is classified as resilient divided by the total number of resilience index: while the numerator of this fraction is reduced when the number of indicators is reduced, so is the denominator.

14 When splitting the calibration and validation sample, the randomization was stratified by evaluation dataset.

15 The resilience index scores are normalized before pooling the data from each evaluation dataset, so applying a consistent cut-off for the pooled dataset is a tenable approach.

16 We also attempted to use information on different types of female household head, as in Section 4.2, but this reduced the number of datasets that could be included in the analysis and substantially worsened the performance of the PMT models.