Developing Poverty Assessment Tools Based on Principal Component Analysis: Results from Bangladesh, Kazakhstan, Uganda, and Peru

Developing accurate, yet operational poverty assessment tools to target the poorest households remains a challenge for applied policy research. This paper aims to develop poverty assessment tools for four countries: Bangladesh, Peru, Uganda, and Kazakhstan. The research applies the Principal Component Analysis (PCA) to seek the best set of variables that predict the household poverty status using easily measurable socio-economic indicators. Out of sample validations tests are performed to assess the prediction power of a tool. Finally, the PCA results are compared with those obtained from regressions models. In-sample estimation results suggest that the Quantile regression technique is the first best method in all four countries, except Kazakhstan. The PCA method is the second best technique for two of the countries. In comparison with regression techniques, PCA models accurately predict a large percentage of households. With regard to out-of sample validations, there is no clear trend; neither the PCA method nor the Quantile regression consistently yields the most robust results. The results highlight the need to assess the out-of-sample performance and thereby the robustness of a poverty assessment tool in estimating the poverty status of a new sample. We conclude that measures of relative poverty estimated with PCA method can yield fairly accurate, but not so robust predictions of absolute poverty as compared to more complex regression models