A Bayesian approach for Generalized Linear Model Using Non-local Prior (Case Study: Poverty Status in East Java)

Abstract

Generalized Linear Model (GLM) is an extension of the linear regression model. GLM is useful for determining the effect of predictor variables to the response variable that is a member of the exponential distribution family. This study uses a binary response variable having a Binomial distribution. This modelling uses the logit link function. This study aims to evaluate the performance of the Bayesian GLM modelling using local prior distribution and non-local prior distribution. The non-local prior density in this study was pi-MOM. The results of the research on modelling poverty status in East Java using the local prior distribution produced a predictor variable that had a significant effect on poverty status in East Java. Meanwhile, modelling using non-local prior distribution produced three predictor variables that significantly affected the status of poverty in East Java. The goodness criteria of the model showed that the Bayesian GLM modelling using the prior non-local distribution was the best model with the smallest misclassification and the lowest AIC value.