All Times ET
Keywords: heterogeneous treatment effects, generalized propensity score, learning analytics
It is well known that individuals may respond to treatments with significant heterogeneity. To optimize the treatment effect, it is necessary to recommend treatments based on individual characteristics. Given increasing interests in learning individualized treatment regimes, a number of methods were proposed recently. However, existing methods are usually designed for randomized studies with binary treatments. In this paper, we propose an algorithm to extend the Random Forest of Interaction Trees from a binary treatment to a non-binary treatment. By integrating generalized propensity scores into the interaction tree growing process, the proposed method could handle both randomized or observational study data with general treatment constructs. Moreover, we propose to use linear residuals as new responses to improve the numerical stability of the algorithm, leading to improved prediction accuracy. The performance of the proposed method is evaluated through extensive simulation studies. The application of the proposed method is illustrated in an assessment of multiple voluntary supplemental instruction programs at a large public university.