Adverse Event (AE) count data in healthy populations, such as in vaccine clinical trials, may be difficult to model well, due to an excess of zeroes relative to the parametric distributions assumed. We fitted a variety of models to an integrated safety database for a vaccine.
Trumenba (TM) (Bivalent rLP2086) was the first vaccine approved in the United States to prevent meningococcal serogroup B disease in individuals aged 10-25 years. An integrated analysis of AEs from 8 randomized, controlled trials is presented. Each trial studied bivalent rLP2086 compared to placebo or active controls. The number of AEs occurring from first dose to 30 days after the last dose was analyzed. Six models where compared: standard Poisson and Negative Binomial models, and their corresponding zeroinflation and hurdle models. Models were evaluated in terms of their ability to predict the number of AEs and by goodness of fit statistics.
Models based on the Poisson distribution were a poor fit. A Zero-Inflated Negative Binomial model was found to provide the closest fit. Zero-inflated models may provide better fit to AE data from healthy populations, compared to conventional parametric models.
|