|
Abstract:
|
Evaluating treatment effect heterogeneity is at the core of treatment decision in many areas. Existing methods focus on estimating conditional average treatment effect via flexible machine learning algorithms. Despite the theoretical appeal in terms of consistency and convergence rate, they generally perform poorly in uncertainty quantification, which is crucial for reliable decision making in sensitive and noisy environments. In this work, we propose a conformal inference-based approach that can produce reliable interval estimates for counterfactuals and individual treatment effects under the potential outcome framework. Precisely, for randomized experiments of which the treatment assignment mechanism is fully known, the intervals have guaranteed average coverage (say 90%) without any assumption other than exchangeability of data. In observational studies, the intervals satisfy a "doubly robust" property that the average coverage is approximately controlled if either the treatment assignment mechanism or the outcome generating process can be estimated well. This is analogous to the double robustness in causal inference of average effects though it is fundamentally different.
|
