Variable selection for causal mediation analysis using LASSO-based methods.

Area of Science:

• Statistics
• Epidemiology
• Machine Learning

Background:

• Causal mediation analysis is crucial for understanding complex relationships.
• Estimating causal mediation effects often relies on marginal structural models and inverse probability weighting.
• High-dimensional covariates pose challenges for fitting accurate propensity score models.

Purpose of the Study:

• To investigate the impact of regularization methods on causal mediation effect estimation.
• To evaluate the performance of outcome-adaptive LASSO in high-dimensional mediation analysis.
• To assess bias reduction and efficiency improvements in natural direct and indirect effect estimation.

Main Methods:

• Utilized marginal structural models with inverse probability weighting.
• Employed regularization techniques, including LASSO and its variants, for parsimonious propensity score models.
• Implemented outcome-adaptive LASSO to incorporate outcome information for variable selection in propensity score models.

Main Results:

• Simulation studies demonstrated that outcome-adaptive LASSO enhances the efficiency of natural effect estimators.
• Regularizing propensity score models improved covariate balance, leading to bias reduction in most scenarios.
• The proposed regularization methods were successfully applied to the MIMIC-III ICU database.

Conclusions:

• Outcome-adaptive LASSO offers an efficient approach for variable selection in high-dimensional propensity score models for causal mediation analysis.
• This method can improve the precision and reduce bias of causal effect estimates.
• The findings have implications for analyzing complex health data, such as that from intensive care units.