2017/11/19 by Ashley I. Naimi, Alan Mishler, Naimi, Ashley I +3 · 2 citations
Computer Science · Decision Sciences · Mathematics · #Advanced Causal Inference Techniques #Bayesian Modeling and Causal Inference #FOS: Computer and information sciences #Methodology (stat.ME) #Multi-Criteria Decision Making
paper · pdf · doi:10.48550/arxiv.1711.07137
openalex publication_date 2017/11/19 · openalex created_date 2022/08/26 · openalex updated_date 2026/07/28
Unlike parametric regression, machine learning (ML) methods do not generally\nrequire precise knowledge of the true data generating mechanisms. As such,\nnumerous authors have advocated for ML methods to estimate causal effects.\nUnfortunately, ML algorithms can perform worse than parametric regression. We\ndemonstrate the performance of ML-based single- and double-robust estimators.\nWe use 100 Monte Carlo samples with sample sizes of 200, 1200, and 5000 to\ninvestigate bias and confidence interval coverage under several scenarios. In a\nsimple confounding scenario, confounders were related to the treatment and the\noutcome via parametric models. In a complex confounding scenario, the simple\nconfounders were transformed to induce complicated nonlinear relationships. In\nthe simple scenario, when ML algorithms were used, double-robust estimators\nwere superior to single-robust estimators. In the complex scenario,\nsingle-robust estimators with ML algorithms were at least as biased as\nestimators using misspecified parametric models. Double-robust estimators were\nless biased, but coverage was well below nominal. The use of sample splitting,\ninclusion of confounder interactions, reliance on a richly specified ML\nalgorithm, and use of doubly robust estimators was the only explored approach\nthat yielded negligible bias and nominal coverage. Our results suggest that ML\nbased singly robust methods should be avoided.\n