A0288
Title: Efficient design and powerful analysis of experiments via digital twins
Authors: Arman Sabbaghi - Purdue University (United States) [presenting]
Abstract: Designed experiments constitute the gold standard for evaluating the causal effects of new treatments, interventions, and therapies across a wide range of domains. However, modern experiments are becoming increasingly difficult to conduct due to enrollment challenges, long durations, and significant costs. How digital twins can facilitate both the efficient design of experiments and the powerful analysis of data are discussed. In particular, digital twins constructed via the application of generative artificial intelligence (AI) algorithms on historical control data containing covariate information can effectively augment the observed information in an experiment while ensuring control of an estimator bias and control of the Type I error rate, when inferring a causal estimand. The integration of AI-generated digital twins with new statistical methods can ultimately power the next generation of designed experiments, as they incorporate the unique advantages of AI algorithms to address the challenges with the design and analysis of modern experiments while maintaining control of desired frequentist properties.
