Title: Robust double machine learning inference for conditional exposure effects
Authors: Stijn Vansteelandt - Ghent University and London School of Hygiene and Tropical Medicine (Belgium) [presenting]
Oliver Dukes - Ghent University (Belgium)
Abstract: The evaluation of exposure effects from observational studies typically requires adjustment for high-dimensional confounding. This makes standard parametric inferences not entirely satisfactory as model misspecification is likely, and even relatively minor misspecifications over the observed data range may induce large bias in the exposure effect estimate. Over the past 2 decades, there has therefore been growing interest in the use of machine learning methods to assist this task. Naive use of machine learning is itself problematic as the resulting exposure effect estimate is prone to a so-called plug-in bias, and the bootstrap is not guaranteed to deliver valid confidence intervals. Pioneering work on Targeted Maximum Likelihood Estimation, and more recently on Double Machine Learning, has shown how this can be overcome by relying on double robust estimators. In particular, valid inference can be obtained by cleverly relying on machine learning predictions of both exposure and outcome, provided that both algorithms converge sufficiently fast to the truth. We will adapt so-called bias-reduced double-robust estimators to ensure valid inference even when one of the machine learning algorithms does not converge (fast) to the truth, thereby yielding results that deliver better approximations in moderate sample sizes.