19) Title: Optimal and adaptive designs for modern medical experimentation

Organizers: Martin Forster, University of Bologna, Italy and Rosamarie Frieri, University of Bologna, Italy

Emails: martin.forster@unibo.it, rosamarie.frieri2@unibo.it

Optimal designs represent the main statistical tool for maximizing experimental information. Combined with sequential adaptation, they permit experiments to be modified and/or stopped on the basis of the accrued data, so as to draw correct inferential conclusions as quickly as possible. The methodology underlying such designs is attractive for the ethical conduct of later stage clinical trials for treatment comparisons in the presence of several patient characteristics, such as those measuring genetic, biological and clinical characteristics. We will present a series of papers covering theoretical innovations in frequentist and Bayesian optimal adaptive experimental methods and apply them to the question of how to conduct clinical trials in a more ethical and efficient way, thereby bringing the best treatments to patients more quickly.
The principal topics that we will present are the following:
*** Adaptive experimental designs that vary the allocation probabilities to the treatment arms for ethical reasons, according to earlier responses, covariates, previous assignments, and/or the profile of the current patient, with the goal of skewing the allocations towards the superior treatment, or to identify subpopulations which will benefit from a treatment.
*** Covariate-Adaptive designs, which modify the allocations to force the balance between the treatment groups with respect to a set of covariates. Since most existing procedures can only handle a few factors, we will introduce novel designs that are able to handle a vast number of covariates that are available in large datasets.
*** New selection methods to draw optimal subsamples from the original large dataset, with the aim of identifying a subset of patient characteristics which influence their health outcomes (a task which may be computationally prohibitive with big-data)
*** Identification of the doses of new drugs which maximize the estimation precision of the parameters of interest. We will propose new optimality criteria based on modifications of the likelihood approach to avoid possible degeneracy of standard inference. In addition, since optimal doses for estimation do not generally allow model validation, we also face the problem of determining optimal doses to identify the right curve among a class of widely used dose-response models, such as fractional polynomials.
*** Investigation of the relationship between classical optimality criteria and the power of tests, from both frequentist and Bayesian perspectives.
*** Optimal sequential experimentation which accounts for the cost-effectiveness of both the treatments and research process, taking frequentist and Bayesian perspectives.
The event is organized as part of the PRIN2022PNRR project (https://sites.google.com/view/prin-project/home)

Scroll to Top