|
Abstract:
|
Often, adverse events (AE) are summarized as counts and rates per patient time-at-risk. In comparing AE rates between treatment groups, typically the same distributional assumption is applied across all AEs with study protocols specifying the same modeling approach for each AE class despite known differences in dispersion. Using data from a randomized, 2-arm trial evaluating an experimental therapy in advanced heart failure patients, we evaluated differences in performance and fit of Poisson, negative binomial (NB), Poisson Inverse-Gaussian (PIG), zero-inflated Poisson (ZIP), zero-inflated negative binomial (ZINB), Poisson hurdle (PH), and negative binomial hurdle (NBH) models across various AE types. Model estimates, fit, and performance were assessed using repeated sub-sampling cross-validation. For AEs with mild overdispersion, model validation mean square error (MSE) was reasonable and similar across all models. Conversely, AEs with high overdispersion had high MSE across all models; NB, PIG and ZINB models had similar standard error (SE) estimates and better fit compared to Poisson models. Our results indicate a “one-model fits all” approach is likely inadequate.
|
