658 – Statistical Learning in Various Areas of Clinical Trials
Is Q-Learning a Valid Method of Knowing?
Francisco J. Diaz
The University of Kansas Medical Center
A great deal of statistical research on Dynamic Treatment Regimes focuses on Q-learning. However, the mathematical coherence and statistical fundamentation of Q-learning are still very poor. In fact, in Q-learning, it is impossible to distinguish between the model explaining or describing the illness phenomenon and the clinical algorithm for treatment individualization. In addition to this epistemological conundrum, Q-learning is mathematically intractable using standard asymptotic or decision theories. Standard theory cannot be used to test the null hypothesis that a treatment has no effect, or to construct confidence intervals. Incoherent definition of covariates is also common. Researchers have attempted to remedy some of these issues, but questions arise about how should we build models in personalized medicine (PM). We discuss here about these issues. As an alternative, Generalized Linear Mixed Effects Models and Empirical Bayesian Feedback can be used to establish a solid paradigm for the construction of the mathematics and statistics of PM research and practice. In fact, there is a long tradition of mixed modeling for treatment individualization in pharmacological literature.