Censored Quantile Regressions of Chronic and Transient Seasonal Poverty in Rwanda

Abstract

It is crucial for social policy in Less Developed Countries to identify correlates of poverty at the household level. This has been done in the literature by estimating household poverty equations typically with Tobit and Probit models. However, when the errors in these equations are non‐normal and heteroscedastic, which is usually expected, these models deliver biased estimates. Using quarterly data from Rwanda in 1983, we reject the normality and homoscedasticity assumptions for household chronic and transient latent poverty equations. We treat this problem by estimating censored quantile regressions. Our results of censored quantile regressions and of inconsistent Tobit regressions are substantially different. However, in the case of chronic poverty the signs of the apparently significant coefficients are generally in agreement, while for seasonal transient poverty different variables have significant effects for the two estimation methods. Our second contribution is to study, for the first time, correlates of poverty indicators based on quarterly consumptions. Our results show that in Rwanda different correlates are significant for chronic poverty and for transient seasonal poverty. The effects of the main inputs (land and labour) are more important for the chronic component of poverty than for the transient one. Household location and socio‐demographic characteristics play important roles that are consistent with usual explanations of poverty in the literature.