Primary Submission Category: Applicants in Social Sciences
Designing Realistic and Interpretable Optimal Treatment Regimes for Personalized Education
Authors: Chenguang Pan, Youmi Suk,
Presenting Author: Chenguang Pan*
Optimal dynamic treatment regimes (ODTR) have gained popularity in computer science and biostatistics for personalized recommendations, but their application in personalized education has been limited. A critical aspect of real-world ODTR applications is to ensure that the estimated regime is feasible and implementable in practice. This study addresses this challenge by incorporating feasibility constraints into the ODTR estimation process. Given that pre-determined constraints may not always be available, we develop a data-driven, post-determined strategy. Specifically, we incorporate a propensity score-based constraint into the ODTR estimation procedure. Instead of using arbitrarily chosen, one-size-fits-all thresholds (e.g., 0.05), our method dynamically adjusts thresholds at each decision stage based on data. This approach accommodates both binary and multi-categorical treatments over multiple stages. Through simulation studies, we demonstrate the effectiveness of our method, coupled with Targeted Maximum Likelihood Estimation (TMLE), in producing more feasible treatment regimes, albeit with a slight trade-off in utility. Feasibility is evaluated using the coverage rate with a set of observed treatment sequences. Finally, we illustrate the trade-off between the utility and feasibility of ODTRs using data from High School Longitudinal Study of 2009 to recommend the optimal and feasible math course for each student at each stage of their high school education.