Adjusting for indirectly measured confounding using large-scale propensity score.

Area of Science:

• Causal inference
• Observational data analysis
• Medical informatics

Background:

• Confounding is a significant challenge in drawing causal conclusions from observational medical data.
• Electronic health records (EHRs) and administrative claims offer vast datasets but are prone to confounding.
• The presence of numerous covariates in modern medical data presents an opportunity to improve causal inference.

Purpose of the Study:

• To investigate the effectiveness of the large-scale propensity score (LSPS) approach for causal analysis in medical data.
• To determine if LSPS can adjust for indirectly measured confounders using extensive covariate information.
• To assess LSPS's ability to mitigate bias from colliders and other problematic variables.

Main Methods:

• Evaluation of the large-scale propensity score (LSPS) method.
• Application of LSPS to simulated medical datasets.
• Validation of LSPS performance on real-world medical data, including EHRs and administrative claims.

Main Results:

• LSPS demonstrates the capability to adjust for indirectly measured confounders by incorporating tens of thousands of covariates.
• The study identifies conditions under which LSPS successfully removes bias stemming from unmeasured confounders.
• LSPS shows potential to avoid introducing bias from inadvertently adjusted variables, such as colliders.

Conclusions:

• The large-scale propensity score (LSPS) approach is a promising method for causal inference with large observational medical datasets.
• LSPS offers a robust strategy for handling complex confounding, including indirectly measured factors and colliders.
• Findings support the utility of LSPS in improving the reliability of causal conclusions drawn from EHR and claims data.