Optimal treatment regime estimation in practice: challenges and choices in a randomized clinical trial for depression

An important aspect of precision medicine is the tailoring of treatments to specific patient types. Nowadays, various methods are available to estimate for this purpose so-called optimal treatment regimes, that is, decision rules for treatment assignment that map patterns of pretreatment characteristics to treatment alternatives and that maximize the expected patient benefit. However, the application of these methods to real-life data has been limited and comes with nonstandard statistical issues. In search of best practices, we reanalyzed data from a randomized clinical trial for the treatment of dysthymic disorder. While the original objective of this trial was to detect a marginally best treatment alternative, we wanted to estimate an optimal treatment regime using 2 prominent estimation methods: Q-learning and value search estimation. An important obstacle in the dataset under study was the occurrence of missing values. This was handled with multiple imputation, a thoughtful implementation of which, however, implied several challenges. Other challenges were implied by the concrete implementation of value search estimation. In this paper, all the choices we have made in the analysis to handle the aforementioned issues are detailed together with a motivation and a description of possible alternatives. Accordingly, this paper may serve as a guide to apply optimal treatment regime estimation in data-analytic practice.