Distribution-Free Causal Inference via Counterfactual Prediction

We develop a causal inference framework for multiple groups based on assessing counterfactual predictions and their confidence intervals. This is achieved in a distribution-free manner, which enables valid inferences without relying on a correct specification of the data model. The framework makes use of a tuning-free prediction method, which can be executed with a runtime that scales linearly in the number of datapoints and is operational in high-dimensional covariate scenarios. The approach is illustrated using both real and synthetic datasets.