Measuring poverty of refugees: Can cross-survey imputation methods substitute for data scarcity?

In a world increasingly characterized by mass data generated by the growth of global connectivity and social media, information on the poor and mobile populations remain very scarce. The poorest members of society are the least connected to the internet and social media whereas highly mobile populations such as migrants are very difficult to capture with surveys. As refugees are both poor and mobile, collecting socioeconomic information on refugees continues to be a challenge and good quality microdata on refugee populations are rare.

This is partly changing. Thanks to recent efforts on the part of the World Bank, the UNHCR and other international organizations, data on refugees are now collected more systematically with household surveys that aim to cover both refugees and host communities. Countries such as Jordan, Uganda, Kenya, and Chad have started to consider refugees as part of their regular household surveys, they are developing special methods to cover these types of populations and are rolling out the first experimental surveys. To boost such efforts, the World Bank and the UNHCR have recently joined forces to collect more and better data on refugees by establishing a Joint Data Center while research on refugees is being strengthened. However, as shown by the experience with the collection of poverty data worldwide, collecting global data on refugees’ well-being will be a long-term process and data gaps will continue to exist for most countries in the foreseeable future.

One alternative approach to filling these data gaps is to employ recent advances with cross-survey estimation methods to estimate poverty using proxies of poverty available in the UNHCR Global Registration System (proGres). This is what we have recently experimented with using UNHCR data from Jordan. Combining data from the UNHCR proGres registration system and household survey data, our study has shown that it’s possible to estimate poverty with a good degree of statistical accuracy using only the few proxies of poverty found in the UNHCR registration system. We also found these estimates to be robust to different poverty lines as shown in Figure 1 (see our study for more detailed explanation). These estimates are also more accurate than those based on asset indexes or proxy means tests while performing well according to targeting indicators. We also provide both theoretical and empirical evidence that cross-survey methods can be applied to relatively small household survey samples with positive prospects for reducing survey costs.

While our findings remain preliminary and subject to validation with other data and in other contexts, they provide a first glimpse of hope in producing more systematic poverty estimates for refugees at the global level. The UNHCR global registration system currently maintains over 14 million individual records, essentially all refugees from countries where the UNHCR is responsible for the registration and assistance of refugees. Where refugee survey data are scarce or missing, estimating poverty using registration data could provide a second-best alternative to survey-based estimation. More generally, these results may not be specific to the Syrian refugees living Jordan. As pointed out in a recent review, imputation methods are particularly relevant in the immediate term (when micro survey data are not fully available for all countries) or where there is a need to back-cast consumption data from a more recent survey for better comparison with older surveys.

If these results are validated in other contexts, the proposed cross-survey imputation methodology could lead to significant improvements in the estimation of poverty among refugee populations, the accuracy of assistance programs targeting refugees, and cost savings in the administration of refugee surveys. While there is no substitute for good survey data, these are potentially significant gains in contexts where proper survey data are not available.

Note: estimates are obtained by imputing from Sample 1 into Sample 2