THE IMPORTANCE OF CLASS PREDICTION IN ZERO-INFLATED MODELS

In a variety of research domains, data are generated as a consequence of the count process and may possess an ‘excess ’ of zeros. There have been many attempts to analyse such data using different statistical methods, including the zero-inflated Poisson (ZiP) and zero-inflated binomial (ZiB) models. The interpretation of these models is however problematic if the covariates considered for the non-zero distribution part of the model are omitted in the conditional part as class predictors. Although the practice of including covariates as both outcome predictors and class predictors is suggested as an option in the literature (not as a necessity), the argument for this practice to become the norm needs to be made more firmly. This argument is illustrated by a hypothetical example of a ZiP model using simulated data based on a genuine clinical dataset in the public domai