Abstract
Causal graph discovery and causal effect estimation are two fundamental tasksin causal inference. While many methods have been developed for each taskindividually, statistical challenges arise when applying these methods jointly:estimating causal effects after running causal discovery algorithms on the samedata leads to "double dipping," invalidating coverage guarantees of classicalconfidence intervals. To this end, we develop tools for validpost-causal-discovery inference. One key contribution is a randomized versionof the greedy equivalence search (GES) algorithm, which permits a valid,finite-sample correction of classical confidence intervals. Across empiricalstudies, we show that a naive combination of causal discovery and subsequentinference algorithms typically leads to highly inflated miscoverage rates; atthe same time, our noisy GES method provides reliable coverage control whileachieving more accurate causal graph recovery than data splitting.