Abstract:Doubly robust estimators with cross-fitting have gained popularity in causal inference due to their favorable structure-agnostic error guarantees. However, when additional structure, such as Hölder smoothness, is available then more accurate "double cross-fit doubly robust" (DCDR) estimators can be constructed by splitting the training data and undersmoothing nuisance function estimators on independent samples. We study a DCDR estimator of the Expected Conditional Covariance, a functional of interest in causal inference and conditional independence testing, and derive a series of increasingly powerful results with progressively stronger assumptions. We first provide a structure-agnostic error analysis for the DCDR estimator with no assumptions on the nuisance functions or their estimators. Then, assuming the nuisance functions are Hölder smooth, but without assuming knowledge of the true smoothness level or the covariate density, we establish that DCDR estimators with several linear smoothers are semiparametric efficient under minimal conditions and achieve fast convergence rates in the non-$\sqrt{n}$ regime. When the covariate density and smoothnesses are known, we propose a minimax rate-optimal DCDR estimator based on undersmoothed kernel regression. Moreover, we show an undersmoothed DCDR estimator satisfies a slower-than-$\sqrt{n}$ central limit theorem, and that inference is possible even in the non-$\sqrt{n}$ regime. Finally, we support our theoretical results with simulations, providing intuition for double cross-fitting and undersmoothing, demonstrating where our estimator achieves semiparametric efficiency while the usual "single cross-fit" estimator fails, and illustrating asymptotic normality for the undersmoothed DCDR estimator.

Doubly robust estimation and inference for a log-concave counterfactual density

Doubly Robust Inference in Causal Latent Factor Models

Conservative Inference for Counterfactuals

Causal effects based on distributional distances

Doubly Robust Kernel Statistics for Testing Distributional Treatment Effects

Causal Inference with High-dimensional Discrete Covariates

Doubly-Valid/Doubly-Sharp Sensitivity Analysis for Causal Inference with Unmeasured Confounding

Scalable Counterfactual Distribution Estimation in Multivariate Causal Models

Relaxed Doubly Robust Estimation in Causal Inference

Estimating individual treatment effect: generalization bounds and algorithms

Enhanced doubly robust estimation with concave link functions for estimands in clinical trials

Bayesian Inference for Multivariate Monotone Densities

Double Cross-fit Doubly Robust Estimators: Beyond Series Regression

Learning to Bound Counterfactual Inference from Observational, Biased and Randomised Data

Doubly Robust Estimation in Causal Inference with Missing Outcomes: with an Application to the Aerobics Center Longitudinal Study

Counterfactual Sensitivity and Robustness

Estimation and inference for the counterfactual distribution and quantile functions in continuous treatment models

Conformal Inference of Counterfactuals and Individual Treatment Effects

Conformal Counterfactual Inference under Hidden Confounding

Robust Estimation of Causal Effects of Binary Treatments in Unconfounded Studies with Dichotomous Outcomes

Universal Inference Meets Random Projections: A Scalable Test for Log-concavity