Discussion on "Nonparanormal adjusted marginal inference" by Susanne Dandl and Torsten Hothorn.

Abstract

In the context of randomized studies, a number of covariate-adjusted estimators for a variety of marginal treatment effects have been developed that are guaranteed to be consistent and asymptotically normal even when the statistical models they are derived from are misspecified. When employing prognostic baseline covariates, these estimators often produce more precise estimates than corresponding unadjusted estimators do. Dandl and Hothorn's clever NAMI estimators lack this robust-to-misspecification property and hence are likely to be unattractive to regulatory agencies, such as the FDA, that review the methodology of and results from randomized trials.