A0308
Title: A flexible framework for adaptive designs based on the simulated annealing algorithm
Authors: Francesco Mariani - University of Bologna (Italy) [presenting]
Rosamarie Frieri - University of Bologna (Italy)
Marco Novelli - University of Bologna (Italy)
Abstract: In adaptive randomization designs, new trial participants are sequentially assigned to a particular treatment based on observed covariates and/or responses. While accounting for covariate profiles can increase statistical power and efficiency, solely relying on them may cause ethical concerns since patients are potentially allocated to the inferior treatment. A sequential procedure is proposed based on the simulated annealing (SA) algorithm that allows combining different research needs by allocating patients in a covariate-adjusted (CA) and/or response-adaptive (RA) manner. Thanks to SA flexibility, the resulting procedure can be: (i) entirely RA, when patients are assigned to the best-performing treatment; (ii) RA + CA, where allocation also aims at balancing covariates across treatments; (iii) CARA, assigning each patient to the most effective treatment based on their covariate profile, thus reflecting a personalized medicine approach. The procedure can be either model-free or model-based, so as to work at a population or patient-specific level, and thus depending on the assumptions over treatment-covariate interactions. In complex scenarios where treatment effect varies with covariates, Bayesian additive regression trees (BART) are incorporated to guide SA toward optimal allocations. The resulting class of CARA designs has appealing statistical and ethical properties, as illustrated through simulation studies.
