Inference for Treatment Regime Models in Personalized Medicine

In medical practice, when more than one treatment option is viable, there is little systematic use of individual patient characteristics to estimate which treatment option is most likely to result in a better outcome for the patient. We introduce a new framework for using statistical models for personalized medicine. Our framework exploits (1) data from a randomized comparative trial, and (2) a regression model for the outcome constructed from domain knowledge and no requirement of correct model specification. We introduce a new "improvement" measure summarizing the extent to which the model's treatment allocations improve future subject outcomes on average compared to a business-as-usual treatment allocation approach. Procedures are provided for estimating this measure as well as asymptotically valid confidence intervals. One may also test a null scenario in which the hypothesized model's treatment allocations are not more useful than the business-as-usual treatment allocation approach. We demonstrate our method's promise on simulated data as well as on data from a randomized experiment testing treatments for depression. An open-source software implementation of our procedures is available within the R package "Personalized Treatment Evaluator" currently available on CRAN as PTE.