POISSON MIXED MODELS WITH A BOOSTING APPROACH FOR THE ANALYSIS OF COUNT DATA

Abstract

Boosting is a powerful technique for enhancing predictive accuracy by iteratively reweighting observations, and is particularly effective in high-dimensional settings and for variable selection. While previous studies have demonstrated the advantages of integrating boosting with generalized linear mixed models (GLMMs) for binary outcomes, its application to count data within hierarchical frameworks remains limited. This study addresses that gap by extending boosting methods to count data through the development of a boosted Poisson mixed model (bPMM), a novel approach for small area estimation and variable selection in complex survey designs. The proposed model is applied to fertility data in the Indonesian provinces of Bali and East Nusa Tenggara, where the response variable is the number of live births and the predictors include twenty-eight socio-demographic covariates. Using the Akaike Information Criterion (AIC) for model selection, three significant variables were identified in Bali (Model 1), and one in East Nusa Tenggara (Model 2). The results demonstrate that bPMM not only improves variable selection in high-dimensional settings but also accommodates hierarchical structure in count data.