Challenges in Obtaining Valid Causal Effect Estimates With Machine Learning Algorithms

Unlike parametric regression, machine learning (ML) methods do not generally require precise knowledge of the true data-generating mechanisms. As such, numerous authors have advocated for ML methods to estimate causal effects. Unfortunately, ML algorithms can perform worse than parametric regression. We demonstrate the performance of ML-based singly and doubly robust estimators. We used 100 Monte Carlo samples with sample sizes of 200, 1,200, and 5,000 to investigate bias and confidence-interval coverage under several scenarios. In a simple confounding scenario, confounders were related to the treatment and the outcome via parametric models. In a complex confounding scenario, the simple confounders were transformed to induce complicated nonlinear relationships. In the simple scenario, when ML algorithms were used, double-robust estimators were superior to singly robust estimators. In the complex scenario, single-robust estimators with ML algorithms were at least as biased as estimators using misspecified parametric models. Doubly robust estimators were less biased, but coverage was well below nominal. The use of sample splitting, inclusion of confounder interactions, reliance on a richly specified ML algorithm, and use of doubly robust estimators was the only explored approach that yielded negligible bias and nominal coverage. Our results suggest that ML-based singly robust methods should be avoided.

Automated summary

Unlike parametric regression, machine learning methods do not require precise knowledge of the data-generating mechanisms. However, machine learning algorithms can perform worse than parametric regression. In the investigated scenarios, ML-based singly robust estimators were less biased in simple confounding scenario but as biased as misspecified parametric models in complex confounding scenario. Doubly robust estimators were less biased, but coverage was below nominal. The use of sample splitting, inclusion of confounder interactions, reliance on richly specified ML algorithm, and use of doubly robust estimators yielded negligible bias and nominal coverage.