Variable Selection for Confounder Control, Flexible Modeling and Collaborative Targeted Minimum Loss-Based Estimation...

Area of Science:

• Causal Inference
• Statistical Modeling
• Machine Learning in Biostatistics

Background:

• Propensity score modeling is crucial for estimating causal effects in observational studies.
• Automated variable selection for propensity scores can introduce bias, particularly when including variables that are causes of the exposure.
• Flexible modeling approaches (nonparametric, machine learning) offer potential but require careful implementation.

Purpose of the Study:

• To investigate the appropriateness of flexible propensity score modeling within semiparametric causal inference frameworks.
• To evaluate the impact of variable selection strategies on the estimation of causal quantities.
• To compare different automated variable selection methods in various dimensional settings.

Main Methods:

• Overview of knowledge-based vs. statistical variable selection in causal inference.
• Demonstration of consequences of adjusting for pure causes of exposure using inverse probability of treatment weighting (IPTW).
• Description of Collaborative Targeted Minimum Loss-based Estimation (C-TMLE) for covariate selection.
• Simulation study comparing automated variable selection approaches in low- and high-dimensional data.

Main Results:

• IPTW with flexible propensity score prediction can yield inferior causal effect estimates.
• Targeted Minimum Loss-based Estimation (TMLE) and C-TMLE demonstrate robustness to highly treatment-correlated variables when using flexible prediction.
• Standard influence function-based variance estimation methods showed underestimation of standard errors, leading to poor coverage in specific scenarios.

Conclusions:

• Flexible propensity score modeling requires careful consideration to avoid biased causal effect estimation.
• TMLE and C-TMLE show promise for robust causal inference, especially in high-dimensional settings with complex treatment-effect relationships.
• Further research is needed to refine variance estimation methods for flexible prediction models in causal inference.