A Model Selection Paradigm for Modeling Recurrent Adenoma Data in Polyp Prevention Trials

Abstract

Colorectal polyp prevention trials (PPTs) are randomized, placebo-controlled clinical trials that evaluate some chemo-preventive agent and include participants who will be followed for at least 3 years to compare the recurrence rates (counts) of adenomas. A large proportion of zero counts will likely be observed in both groups at the end of the observation period. Poisson general linear models (GLMs) are usually employed for estimation of recurrence in PPTs. Other models, including the negative binomial (NB2), zero-inflated Poisson (ZIP), and zero-inflated negative binomial (ZINB) may be better suited to handle zero-inflation or other forms of overdispersion that are common in count data. A model selection paradigm that determines a statistical approach for choosing the best fitting model for recurrence data is described. An example using a subset from a large Phase III clinical trial indicated that the ZINB model was the best fitting model for the data.