Hurdle Negative Binomial Regression Model
Keywords:Bayesian, Gibbs Sampling, Markov Chain Monte Carlo, Motoric Complication, Parkinson
Poisson regression is a common regression model used for count data with equidispersion. However, in real data application, overdispersion often encountered, suggesting the seek for alternative model to the Poisson regression. In overdispersion data due to excess zeros and additional overdispersion in positive values, one of alternative model that can be used is hurdle negative binomial model. Hurdle negative binomial model is a two-part model consists of binary model and zero-truncated negative binomial model. In this study we discuss hurdle negative binomial model and Bayesian approach for the model’s parameter estimation, then apply the method for modelling frequency of motoric complication in people with early Parkinson’s disease. Markov Chain Monte Carlo with Gibbs Sampling (MCMC-GS) was implemented to sample the regression parameters from their posterior distribution. The result showed that hurdle negative binomial model fit the data satisfactorily, as implied by the convergence and unimodality of posterior density of the parameters of interest. We also identified risk factors for motoric complications.