gms | German Medical Science

Due to concerns about parametric model misspecification, there is interest in using machine learning to adjust for confounding when evaluating the causal effect of an exposure on an outcome. Unfortunately, exposure effect estimators that rely on machine learning predictions are generally subject to so-called plug-in bias, which can render naive p-values and confidence intervals invalid. Progress has been made via proposals like Targeted Maximum Likelihood Estimation and more recently Double Machine Learning, which rely on learning the conditional mean of both the outcome and exposure. Valid inference can then be obtained so long as both algorithms converge (sufficiently fast) to the truth. We will show that by implementing the machine learning techniques in a specific way, we can develop exposure effect estimators that have good properties even when one of the first-stage algorithms does not converge to the truth, along with honest tests and confidence intervals. Our proposal leads to reduced bias and improved confidence interval coverage in moderate-samples, as we observe in simulations studies. We illustrate the proposal in a case study looking at the effect of obesity on the probability of survival within patients in the Ghent University Hospital Intensive Care Unit.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.