Dose-response modeling with bivariate binary data under model uncertainty

When modeling a dose-response for a drug based on bivariate binary data such as two co-primary efficacy endpoints in early stages of development, there is usually uncertainty about the from of the true underlying dose-response shape. Often, investigators fit several different models that are deemed plausible, but later fail to acknowledge this uncertainty in inference that is based on a single model selected via e.g. the minimum AIC criterion. This leads to an inflation in the error of the proof of activity decision and may also result in poor estimation of a target dose that is used in future trials. In this article we acknowledge model uncertainty by fitting several candidate models for a bivariate binary response and develop a principled approach to establish proof of activity and a target dose.