David Draper: Model Mis-specification Bias and Model Uncertainty: Bayesian Solutions……………………………………………………………5 Estimation in Cox Regression Revisited……………………………………7 Georg Ferber: the Maximum Mean Difference -statistical Problems in Assessing Cardiac Safety……………………………………..11 Klaas Prins: Modeling and S

In a typical application of the statistical paradigm, there's some quantity Q about which I'm at least partially uncertain, and I wish to quantify my uncertainty about Q, for the purpose of (a) sharing this information with other people or (b) helping myself or others to make a choice in the face of this uncertainty. Uncertainty quantification is usually based on a probability model M, which relates Q to known quantities (such as data values D); M will in turn be based on assumptions and judgments on my part about how Q and D are related, but I'm not always certain about the "right" assumptions and judgments to make. To be completely honest, then, I have to acknowledge two sources of uncertainty: I'm uncertain about Q, and I'm also uncertain about how to quantify my uncertainty about Q. This second source is model uncertainty; and if I get the model "wrong," the result will typically be model mis-specification bias. In this talk I'll speak to two questions: why model uncertainty and model mis-specification bias matter, and what to do about them. Topics to be addressed, in the context of one or more applied examples, will include (1) What is a statistical model (the standard parametric answer; de Finetti's predictive answer based on exchangeability); (2) Where do models come from (in practice many specification details often come from looking at the data); (3) Consequences of failure to acknowledge model uncertainty and model mis-specification bias (if the right price is not paid in (2), the result will often be poor calibration of inferential and predictive statements); and (4) Methods for coping with model uncertainty and model mis-specification bias (Bayesian cross-validation, Bayesian model averaging, Bayesian nonparametric methods). The Bayesian approach to dose finding in clinical trials has a modeling and a decision-making component. The former necessitates a flexible class of dose-response models (parametric or nonparametric), whereas the latter relies on decisions that are based on appropriate inferential summaries and can be either informal (posterior-and/or predictive-based) or formal, i.e., fully decision-analytic using utilities. Dose-response modeling can be done in various ways, ranging from simple models with a small number of parameters (typical for early development trials) to high-dimensional or non-parametric models. Often in early clinical development, little is known about the possible shape of the dose response relationship within the studied dose range (tyical 3 to 4 active doses), except it is monotone. If …