Analysis of Zero-Inated Count Data from Longitudinal Trials with Dropouts

Jingyuan Yang1, Xiaoming Li2 and Frank Liu2
1Ohio State University, OH
2Merck Research Laboratories, PA

Abstract

This research is motivated by vaccine clinical trials, in which binary outcomes are recorded at sequential time points post vaccination. The primary endpoint is the count of events during the follow-up time, and the parameter of interest is the proportion of subjects with at least one event. Potential issues on the analysis of these count data may include 1) over-dispersion arising from the intra-subject correlation among binary outcomes, 2) zero-inflation caused by low incidence rate, and 3) the missing data due to early dropouts. The conventional methods such as Poisson regression may not be appropriate because of these issues. We evaluated several alternatives including Poisson hurdle (PH), negative binomial (NB), and negative binomial hurdle (NBH) models. A new link function is proposed for the PH and NBH models in an effort to improve the model fitting and interpretability. When there are missing outcomes due to early dropout, the NBH model with offset and the multiple imputation (MI) method are evaluated in terms of their ability to adjust the bias. The characteristics of different methods are compared via simulations.