On Nearly Assumption-Free Tests of Nominal Confidence Interval Coverage for Causal Parameters Estimated by Machine Learning

Abstract

For many causal effect parameters of interest, doubly robust machine learning (DRML) estimators $\hat{\psi}_{1}$ are the state-of-the-art, incorporating the good prediction performance of machine learning; the decreased bias of doubly robust estimators; and the analytic tractability and bias reduction of sample splitting with cross-fitting. Nonetheless, even in the absence of confounding by unmeasured factors, the nominal $(1-\alpha)$ Wald confidence interval $\hat{\psi}_{1}\pm z_{\alpha/2}\widehat{\mathsf{s.e.}}[\hat{\psi}_{1}]$ may still undercover even in large samples, because the bias of $\hat{\psi}_{1}$ may be of the same or even larger order than its standard error of order $n^{-1/2}$.

In this paper, we introduce essentially assumption-free tests that (i) can falsify the null hypothesis that the bias of $\hat{\psi}_{1}$ is of smaller order than its standard error, (ii) can provide a upper confidence bound on the true coverage of the Wald interval, and (iii) are valid under the null under no smoothness/sparsity assumptions on the nuisance parameters. The tests, which we refer to as Assumption Free Empirical Coverage Tests (AFECTs), are based on a U-statistic that estimates part of the bias of $\hat{\psi}_{1}$.

Our claims need to be tempered in several important ways. First no test, including ours, of the null hypothesis that the ratio of the bias to its standard error is smaller than some threshold $\delta$ can be consistent [without additional assumptions (e.g., smoothness or sparsity) that may be incorrect]. Second, the above claims only apply to certain parameters in a particular class. For most of the others, our results are unavoidably less sharp. In particular, for these parameters, we cannot directly test whether the nominal Wald interval $\hat{\psi}_{1}\pm z_{\alpha/2}\widehat{\mathsf{s.e.}}[\hat{\psi}_{1}]$ undercovers. However, we can often test the validity of the smoothness and/or sparsity assumptions used by an analyst to justify a claim that the reported Wald interval’s actual coverage is no less than nominal. Third, in the main text, with the exception of the simulation study in Section 1, we assume we are in the semisupervised data setting (wherein there is a much larger dataset with information only on the covariates), allowing us to regard the covariance matrix of the covariates as known. In the simulation in Section 1, we consider the setting in which estimation of the covariance matrix is required. In the simulation, we used a data adaptive estimator which performs very well in our simulations, but the estimator’s theoretical sampling behavior remains unknown.