Robust Machine Learning for Treatment Effects in Multilevel Observational Studies Under Cluster-level Unmeasured Confounding

Designing Optimal, Data-Driven Policies from Multisite Randomized Trials

Optimal treatment regimes (OTRs) have been widely employed in computer science and personalized medicine to provide data-driven, optimal recommendations to individuals. However, previous research on OTRs has primarily focused on settings that are independent and identically distributed, with little attention given to the unique characteristics of educational settings, where students are nested within schools and there are hierarchical dependencies. The goal of this study is to propose a framework for designing OTRs from multisite randomized trials, a commonly used experimental design in education and psychology to evaluate educational programs. We investigate modifications to popular OTR methods, specifically Q-learning and weighting methods, in order to improve their performance in multisite randomized trials. A total of 12 modifications, 6 for Q-learning and 6 for weighting, are proposed by utilizing different multilevel models, moderators, and augmentations. Simulation studies reveal that all Q-learning modifications improve performance in multisite randomized trials and the modifications that incorporate random treatment effects show the most promise in handling cluster-level moderators. Among weighting methods, the modification that incorporates cluster dummies into moderator variables and augmentation terms performs best across simulation conditions. The proposed modifications are demonstrated through an application to estimate an OTR of conditional cash transfer programs using a multisite randomized trial in Colombia to maximize educational attainment.

Tuning Random Forests for Causal Inference under Cluster-Level Unmeasured Confounding

Recently, there has been growing interest in using machine learning methods for causal inference due to their automatic and flexible ability to model the propensity score and the outcome model. However, almost all the machine learning methods for causal inference have been studied under the assumption of no unmeasured confounding and there is little work on handling omitted/unmeasured variable bias. This paper focuses on a machine learning method based on random forests known as Causal Forests and presents five simple modifications for tuning Causal Forests so that they are robust to cluster-level unmeasured confounding. Our simulation study finds that adjusting the default tuning procedure with the propensity score from fixed effects logistic regression or using variables that are centered to their cluster means produces estimates that are more robust to cluster-level unmeasured confounding. Also, when these parametric propensity score models are mis-specified, our modified machine learning methods remain robust to bias from cluster-level unmeasured confounders compared to existing parametric approaches based on propensity score weighting. We conclude by demonstrating our proposals in a real data study concerning the effect of taking an eighth-grade algebra course on math achievement scores from the Early Childhood Longitudinal Study.

Stan and BART for Causal Inference: Estimating Heterogeneous Treatment Effects Using the Power of Stan and the Flexibility of Machine Learning

A wide range of machine-learning-based approaches have been developed in the past decade, increasing our ability to accurately model nonlinear and nonadditive response surfaces. This has improved performance for inferential tasks such as estimating average treatment effects in situations where standard parametric models may not fit the data well. These methods have also shown promise for the related task of identifying heterogeneous treatment effects. However, the estimation of both overall and heterogeneous treatment effects can be hampered when data are structured within groups if we fail to correctly model the dependence between observations. Most machine learning methods do not readily accommodate such structure. This paper introduces a new algorithm, stan4bart, that combines the flexibility of Bayesian Additive Regression Trees (BART) for fitting nonlinear response surfaces with the computational and statistical efficiencies of using Stan for the parametric components of the model. We demonstrate how stan4bart can be used to estimate average, subgroup, and individual-level treatment effects with stronger performance than other flexible approaches that ignore the multilevel structure of the data as well as multilevel approaches that have strict parametric forms.