Selecting Treatment Effects Models for Domain Adaptation Using Causal Knowledge

While a large number of causal inference models for estimating individualized treatment effects (ITE) have been developed, selecting the best one poses unique challenge, since counterfactuals are never observed. The problem is challenged further in unsupervised domain adaptation (UDA) setting where we access to labeled samples source but desire an ITE model that achieves good performance on target only unlabeled available. Existing selection techniques UDA designed predictive and sub-optimal because they (1) do not account missing (2) examine discriminative density ratios between input covariates factor model’s predictions domain. We leverage invariance structures across domains introduce novel metric specifically under UDA. propose whose interventions satisfy invariant Experimentally, our method selects more robust covariate shifts variety datasets, including effect ventilation COVID-19 patients.