Constructing an optimal adaptive treatment strategy becomes complex when there are a large number of potential tailoring variables. In such scenarios, many of these extraneous variables may contribute little or no benefit to an adaptive strategy while increasing implementation costs and putting an undue burden on patients. Although existing methods allow selection of the informative prognostic factors, statistical inference is complicated by the data-driven selection process. To remedy this deficiency, we adapt the Universal Post-Selection Inference procedure to the semiparametric Robust Q-learning method and the unique challenges encountered in such multistage decision methods. In the process, we also identify a uniform improvement to confidence intervals constructed in this post-selection inference framework. Under certain rate assumptions, we provide theoretical results that demonstrate the validity of confidence regions and tests constructed from our proposed procedure. The performance of our method is compared to the Selective Inference framework through simulation studies, demonstrating the strengths of our procedure and its applicability to multiple selection mechanisms.
View on arXiv