Abstract:
|
In observational studies, treatment is not randomly assigned to subjects. To estimate the causal effect, researchers need to account for the pretreatment covariate differences between treated and control groups. One popular adjustment strategy is through matching, which does not rely on parametric outcome modeling assumptions. When the treatment effects are heterogeneous among subpopulations, it is challenging to identify the subgroups (defined by covariates) with different effects. Ideally, the identification should be based on individual treatment effects, rather than the observed outcomes. We propose a matching-based classification tree strategy, which fits a tree model to the matched pair outcome differences. The simulation study demonstrates that our strategy outperforms other tree-based methods using observed outcomes in many scenarios.
|