Tuning LASSO Models for Propensity Score Weighting and Using Synthetic Negative Control Exposures for Residual Bias...

Area of Science:

• Causal inference
• Epidemiology
• Health data science

Background:

• Propensity score (PS) methods are crucial for controlling confounding in high-dimensional healthcare databases.
• Least absolute shrinkage and selection operator (LASSO) is common for PS estimation, with cross-validation typically selecting regularization parameters.
• Standard prediction-based tuning may over-regularize PS models, increasing bias in PS-weighted estimators.

Purpose of the Study:

• To evaluate balance metrics for selecting propensity score model undersmoothing when efficient influence functions are unavailable.
• To introduce synthetic negative control exposures for detecting residual confounding in PS weighting analyses.
• To provide practical methods for improving the reliability of propensity score weighting.

Main Methods:

• Investigated balance metrics for tuning propensity score model regularization (undersmoothing).
• Developed a framework using synthetic negative control exposures to detect bias from unmeasured confounding.
• Conducted numerical studies to compare methods against standard cross-validation.

Main Results:

• Balance-based undersmoothing consistently reduced bias compared to standard cross-validation.
• Synthetic negative control exposures effectively identified analyses with residual confounding.
• The proposed methods offer practical improvements for propensity score weighting.

Conclusions:

• Balance metrics provide a practical alternative for tuning propensity score models when theory-driven approaches are infeasible.
• Synthetic negative controls are valuable tools for assessing the validity of propensity score analyses.
• These methods enhance the robustness and interpretability of causal inference from observational healthcare data.