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

Area of Science:

• Machine Learning
• Causal Inference
• Econometrics

Background:

• Machine learning offers flexible modeling for causal inference, particularly for propensity scores and outcome models.
• Existing machine learning methods often assume no unmeasured confounding, limiting their application in real-world scenarios with potential omitted variable bias.

Purpose of the Study:

• To propose and evaluate modifications to Causal Forests to enhance robustness against cluster-level unmeasured confounding.
• To assess the performance of modified Causal Forests compared to traditional methods under unmeasured confounding.

Main Methods:

• Introduced five modifications to the tuning procedure of Causal Forests.
• Utilized simulation studies to compare modified Causal Forests with existing methods.
• Applied the modified methods to real-world data from the Early Childhood Longitudinal Study.

Main Results:

• Adjusting Causal Forests with propensity scores from fixed effects logistic regression or cluster-mean centered variables improved robustness to unmeasured confounding.
• Modified machine learning methods demonstrated resilience to bias from cluster-level unmeasured confounders, even with mis-specified parametric propensity score models.

Conclusions:

• The proposed modifications offer a more reliable approach to causal inference using machine learning when unmeasured confounding is present.
• These methods provide a valuable tool for researchers dealing with complex observational data and potential biases.