Doubly robust, machine learning effect estimation in real-world clinical sciences: A practical evaluation of performance in molecular epidemiology cohort settings

Modern efficient estimators such as AIPW and TMLE facilitate the application of flexible, non-parametric machine learning algorithms to improve treatment and outcome model fit, allowing for some model misspecification while still maintaining desired bias and variance properties. Recent simulation work has pointed to essential conditions for effective application including: the need for cross-fitting, using of a broad library of well-tuned, flexible learners, and sufficiently large sample sizes. In these settings,cross-fit, doubly robust estimators fit with machine learning appear to be clearly superior to conventional alternatives. However, commonly simulated conditions differ in important ways from settings in which these estimators may be most useful, namely in high-dimensional, observational settings where: costs of measurements limit sample size, high numbers of covariates may only contain a subset of true confounders, and where model misspecification may include the omission of essential biological interactions. In such settings, computationally-intensive and challenging to optimize cross-fit, ensemble learning-based estimators may have less of a practical advantage. We present extensive simulation results drawing data on 331 covariates from 1178 subjects of a multi-omic, longitudinal birth cohort while fixing treatment and outcome effects. We fit models under various conditions including under- and over- (e.g. excess orthogonal covariates) specification, and missing interactions using both state-of-the-art and less-computationally intensive (e.g. singly-fit,parametric) estimators. In real data structures, we find in nearly every scenario (e.g. model misspecification, single- or cross-fit- estimators), that efficient estimators fit with parametric learner out perform those that include non-parametric learners on the basis of bias and coverage.