Choix de l’a priori en statistique Bayésienne dans les essais expérimentaux.

Abstract

The aim of this thesis, which was motivated by methodological problems that arise in the context of experimental trials, is to propose a hybrid Bayesian-frequentist approach using sequential Bayesian prediction of the index of satisfaction, which is often done in the case of clinical trial protocol. For an interim analysis of the prediction hypothesis, such as futility monitoring of the experimental treatment with delayed results, the prediction of satisfaction takes into account the amount of data remaining to be observed in a clinical trial and offers the possibility of integrating additional information via auxiliary variables by the a priori distribution. The design of the prediction of satisfaction guarantees the type I error rate and does not require calculation or intensive simulation.