Abstract
Background
Flexible, data-adaptive algorithms (machine learning; ML) for nuisance parameter estimation in epidemiologic causal inference have promising asymptotic properties for complex, high-dimensional data. However, recently proposed applications (e.g. targeted maximum likelihood estimation; TMLE) may produce biases parameter and standard error estimates in common real-world cohort settings. The relative performance of these novel estimators over simpler approaches in such settings is unclear.
Methods
We apply double-crossfit TMLE, augmented inverse probability weighting (AIPW), and standard IPW to simple simulations (5 covariates) and “real-world” data using covariate-structure-preserving (“plasmode”) simulations of 1,178 subjects and 331 covariates from a longitudinal birth cohort. We evaluate various data generating and estimation scenarios including: under- and over- (e.g. excess orthogonal covariates) identification, poor data support, near-instruments, and mis-specified biological interactions. We also track representative computation times.
Results
We replicate optimal performance of cross-fit, doubly robust estimators in simple data generating processes. However, in nearly every real world-based scenario, estimators fit with parametric learners outperform those that include non-parametric learners in terms of mean bias and confidence interval coverage. Even when correctly specified, estimators fit with non-parametric algorithms (xgboost, random forest) performed poorly (e.g. 24% bias, 57% coverage vs. 10% bias, 79% coverage for parametric fit), at times underperforming simple IPW.
Conclusions
In typical epidemiologic data sets, double-crossfit estimators fit with simple smooth, parametric learners may be the optimal solution, taking 2-5 times less computation time than flexible non-parametric models, while having equal or better performance. No approaches are optimal, and estimators should be compared on simulations close to the source data.
Key messages
In epidemiologic studies, use of flexible non-parametric algorithms for effect estimation should be strongly justified (i.e. high-dimensional covariates) and performed with care. Parametric learners may be a safer option with few drawbacks.