Group Comparisons Involving Zero-Inflated Count Data in Clinical Trials

Abstract

In clinical trials, outcomes of count data sometimes have excess zeros. When a test drug is compared to a control, zero-inflated data may be ignored or interest is taken only in the proportion of zero counts. By applying the two-part model, Lachenbruch (2001a) suggested a test statistic called the two-part statistic that combines the test statistics of the zero part and the non-zero part. The test for the zero part is the chi-square test. The test for the non-zero part may be a Wilcoxon test, a t -test, etc. This article proposes methods for calculating the sample size and power for the two-part statistic with zero-inflated Poisson data. We developed the methods of sample size and power for the two-part statistic using the Wilcoxon test adjusted for ties. The relationship between the non-zero part and zero-truncated Poisson distribution is also described. Furthermore, we examine the power of the two-part statistic, conventional methods, and the zero-inflated Poisson model. in which patients do not recover but have a small value of the outcome that is zero by chance. The zero-inflated Poisson (ZIP) distribution or zero-inflated negative binomial distribution can be applied to count data with excess zeros. This article focuses on the ZIP distribution. The ZIP distribution has two parameters; λ is the Poisson parameter and ω expresses the extent of zero-inflation compared with zero counts that occur from the Poisson distribution.