Modeling over dispersed data: double Poisson regression and zero-inflated Poisson regression with application to neonatal mortality

Abstract

Neonatal mortality remains a major public health issue, particularly in developing countries, requiring robust statistical approaches to identify its determinants. This study compares Poisson Regression (PR), Zero-Inflated Poisson Regression (ZIP), and Double Poisson Regression (DPR) in modeling neonatal mortality data. The dependent variable measures neonatal deaths, while five independent variables capture patient diagnoses.Although PR assumes equidispersion, diagnostics revealed overdispersion and excess zeros, necessitating ZIP and DPR. ZIP accounts for zero inflation, while DPR flexibly handles both overdispersion and underdispersion. Model fit was evaluated using AIC and BIC, with DPR outperforming the other models by achieving the lowest values, indicating superior performance.