Abstract

This presentation focuses on practical applications of machine learning approaches to better understand (a) heterogeneity of adherence, per-protocol causal estimates, and (b) heterogeneity of treatment response causal effects (Weissler et al., 2021). Specifically, discussion on the development of single training models and meta-training models are highlighted for their improved statistical efficiency and less stringent model assumptions compared to more traditional approaches. For differential adherence, classic training approaches such as complier average causal effects (CACE) (Jo et al., 2008) to evolving meta-learning algorithms are presented (Zhong et al., 2022). For differential response, classic training models applied in adaptive intervention designs (DeGarmo & Gewirtz, 2019) to evolving doubly-robust estimator meta-learner algorithms are discussed for conditional average treatment effects (CATE) (Hamaya et al., 2025). Basic applications are illustrated; sample data and R code are provided.