Explaining the Decline of Child Mortality in 44 Developing Countries: A Bayesian Extension of Oaxaca Decomposition for Probit Random Effects Models

PWP-CCPR-2020-004

Abstract

We develop a novel extension of Oaxaca decomposition methods for non-linear random effects models to investigate the decline of infant mortality in 42 low and middle income countries. We analyze micro data from 84 Demographic and Health Surveys where surveys from two time periods were available. We predict mortality at the birth level with a Bayesian hierarchical probit regression models. We use the predictions from these models as input for our new Oaxaca method. Our novel approach accounts for uncertainty in the decompostion results, and allows for point estimates, stan- dard deviations, and posterior distributions of the Oaxaca conclusions. Further, our approach does not depend on assumptions such as matched samples between two surveys and and marginalizes ran- dom effects for variables that are not comparable between surveys, such as location effects. For most countries, declines in infant mortality are due to changes in the regression coefficients, not on covari- ate distributions. However, our decomposition results show that there is considerable heterogeneity between countries and uncertainty on which variable matter the most within countries.