A Cohort-Level Evaluation of Thompson Sampling for Reducing Asthma Risk

Abstract

Public-health programs are continually deciding which outreach action will most reduce risk in the populations they serve, often with limited data and changing conditions. In this study, I apply Thompson Sampling (TS) to cohort-level decision making for asthma-related interventions using the Behavioral Risk Factor Surveillance System (BRFSS) 2015 data, I defined seven intervention “arms” that plausibly affect asthma outcomes including baseline, smoking cessation, vaccinations, inhaler adherence education, preventive check-ups, weight control, and air-quality awareness. I also compared two algorithms: (i) fixed Thompson Sampling (TS) with a Beta(1,1) prior per arm and (ii) Empirical-Bayes Thompson Sampling (EB-TS) that fits Beta priors from offline estimates. Over 10 replications of 30,000 rounds, fixed TS achieves mean expected reward 0.2660 and mean regret 0.002065 per round, concentrating ~99% of pulls on the best arm. EB-TS increases the overall reward to 0.2681 and lower the mean regret to 0.000046 by stabilizing early decisions with data-informed priors. Results suggest fixed TS is a strong, hyperparameter-free baseline for cohort-level outreach, while EB-TS helps when offline reward estimates are reliable and improved reward with dense value feedback.