Title: Bayesian and frequentist models for extrapolation of event rates for safety and efficacy data in clinical trials
Authors: Daniel Bonzo - LFB USA (United States) [presenting]
Marek Ancukiewicz - LFB-USA (United States)
Abstract: Event rates are one type of endpoint commonly used for evaluation of safety data (e.g., adverse events) and efficacy data (e.g., bleeding episodes). The data arise when one counts, for a patient, a random number of events observed during study follow-up of a random duration. We assume that event counts are overdispersed. We consider the problem of extrapolation of such data. Under certain circumstances, data can be extrapolated to a different population (e.g., from adult to pediatric population), a different but related indication, and different but similar product (e.g., from original drug to generic drug or biosimilar). As the concept of estimand captures population, endpoint, and a measure of effect in general, one can think about extrapolation of historical data from one estimand to another closely related estimand. We propose and evaluate two models for this task: random effect model using a non-parametric estimate of the variance of event rates (frequentist) and power model assuming partial exchangeability and negative-binomial likelihoods (Bayesian). We demonstrate application of the method using clinical trial data. In conclusion, both models require clinical and scientific inputs that can be quantified as weighting of different sources of data. The models can be useful tools for extrapolation of event rates data for medicine development, including extrapolation from adult to pediatric population, from one indication to another, and from one product to another.