Conformal prediction intervals for individual treatment effects (ITEs) usually guarantee marginal coverage, averaged over the covariate distribution, and can undercover badly where effects are most heterogeneous: in our simulations, a globally calibrated conformal meta-learner reaches 94.5% marginal coverage at the 90% level but only 62.3% in the worst covariate slab. We propose a localized split-conformal procedure that calibrates counterfactual intervals with kernel weights centered at the covariate value of the query, using a kernel built from causal-forest variable importance. The intervals have exact finite-sample coverage for units drawn from a kernel neighborhood of the query, need no propensity model, and cover the query itself with probability at least the nominal level minus an explicit localization error, which makes the bandwidth trade-off interpretable. We give a bandwidth-selection rule that uses only the training fold and a variant that calibrates doubly robust pseudo-outcomes. In simulations with 20 covariates, localization raises the worst-slab coverage of out-of-sample ITE intervals from 0.58-0.87 for global methods to 0.98 at a similar length, with 4.7% of intervals infinite. On the IHDP benchmark, where local information is scarce, the procedure reduces to global calibration.
翻译:暂无翻译