Title:Assumption-robust Causal Inference

Abstract:In observational causal inference, it is common to encounter multiple adjustment sets that appear equally plausible. It is often untestable which of these adjustment sets are valid to adjust for (i.e., satisfies ignorability). This discrepancy can pose practical challenges as it is typically unclear how to reconcile multiple, possibly conflicting estimates of the average treatment effect (ATE). A naive approach is to report the whole range (convex hull of the union) of the resulting confidence intervals. However, the width of this interval might not shrink to zero in large samples and can be unnecessarily wide in real applications. To address this issue, we propose a summary procedure that generates a single estimate, one confidence interval, and identifies a set of units for which the causal effect estimate remains valid, provided at least one adjustment set is valid. The width of our proposed confidence interval shrinks to zero with sample size at $n^{-1/2}$ rate, unlike the original range which is of constant order. Thus, our assumption-robust approach enables reliable causal inference on the ATE even in scenarios where most of the adjustment sets are invalid. Admittedly, this robustness comes at a cost: our inferential guarantees apply to a target population close to, but different from, the one originally intended. We use synthetic and real-data examples to demonstrate that our proposed procedure provides substantially tighter confidence intervals for the ATE as compared to the whole range. In particular, for a real-world dataset on 401(k) retirement plans our method produces a confidence interval 50\% shorter than the whole range of confidence intervals based on multiple adjustment sets.