Clinical researchers wanting to make principled (evidence based) rules for tailoring treatment have run multi-stage randomized clinical trials in order to evaluate and compare different long-term treatment strategies. Q-learning and other extensions of regression to multi-stage data can be used effectively to construct decision rules that lead to a favorable clinical outcome. However, in order for these methods to be more widely adopted, one must be able to conduct statistical inference. In particular, one must be able to able to answer the following types of questions: Do two treatment strategies result in significantly different clinical outcomes? Which patient variables are relevant for tailoring treatment? The need to address these questions has led us to develop new statistical methodology that allows for the construction of valid confidence intervals for parameters in the regression models used in Q-learning. The focus of this talk will be on the use of this new methodology to construct meaningful and interpretable regression models via Q-learning and the application of confidence intervals to glean relevant scientific knowledge. We illustrate these ideas by means of a case study of the Study of the Adaptive Interventions for Children with ADHD Trial.
Return to Biostatistics Working Group