Variable selection-combined causal mediation analysis for continuous treatments with application to large-dimensional...

Area of Science:

• Causal Inference
• Statistical Modeling
• Biostatistics

Background:

• Causal mediation effects are crucial for understanding complex relationships in large-scale data.
• Existing methods often focus on binary variables, limiting applications with continuous treatments and mediators.
• Large-scale covariate selection remains a challenge in continuous variable settings.

Purpose of the Study:

• To develop a novel weighted semiparametric estimation framework for causal mediation effects in continuous variable settings.
• To address limitations in large-scale covariate selection for continuous treatments and mediators.
• To enhance estimation accuracy and efficiency in high-dimensional causal inference.

Main Methods:

• Proposed a weighted semiparametric estimation framework.
• Integrated generalized outcome-adaptive LASSO with generalized propensity score modeling.
• Applied the method to continuous treatment and mediator variables for causal mediation analysis.

Main Results:

• The proposed method demonstrated superior selection accuracy and estimation efficiency compared to existing regularization-based methods.
• Successfully incorporated outcome-related key variables while excluding noise covariates.
• The approach offers a balance between efficiency, bias reduction, and high-dimensional data filtering.

Conclusions:

• The developed method provides a robust and interpretable alternative for estimating causal mediation effects in continuous, high-dimensional settings.
• It enhances inferential robustness by effectively managing complex covariate structures.
• Real-world application in the UK Biobank demonstrated its utility in quantifying mediation effects in health research.