11: Balanced Subgroup Discovery Via Matching, Decision Trees, and Randomization Inference

Music City Center 

We recently developed a method for data-driven heterogeneous treatment effect subgroup discovery that combines matching and decision trees (mTree). Matching leads to balance within discovered subgroups, overcoming the limitation that insufficient balance in subgroups may lead to findings that cannot be replicated. Decision trees are popular in medicine because they are an effective decision-making technique providing high classification accuracy with a simple representation of gathered knowledge, i.e., they are not “Black Boxes”.

Our previous work did not propose an approach for statistical inference within subgroups. In the typical superpopulation inference framework, re-using data for hypothesis generation and hypothesis testing creates type 1 error issues. To overcome this challenge, we adopt the randomization inference framework wherein the goal is to make inferences about treatment effects in the sample alone. In this work we extend the mTree method to accommodate time-to-event outcomes and develop new randomization inference estimators of within-subgroup additive and multiplicate treatment effects. The methods are applied to a systolic blood pressure intervention trial.

Heterogeneous treatment effects

Subgroup discovery

Matching

Decision trees

Randomization inference 

Section on Statistics in Epidemiology