Assessing Treatment Effects in Observational Data With Missing Confounders: A Comparative Study of Practical...

Area of Science:

• Pharmacoepidemiology
• Biostatistics
• Health Data Science

Background:

• Pharmacoepidemiology relies on administrative and electronic health records (EHR) for safety and effectiveness studies.
• Missing confounder data is common in these real-world datasets, necessitating robust analytical approaches.
• Multiple imputation (MI) and inverse-probability weighting (IPW) are standard but may not be optimal for missing data.

Purpose of the Study:

• To evaluate and compare the performance of under-utilized doubly-robust methods against traditional MI and IPW for handling missing data in pharmacoepidemiology.
• To provide guidance on selecting appropriate missing data analysis methods based on performance across various scenarios.

Main Methods:

• Investigated two doubly-robust estimators: generalized raking and inverse probability-weighted targeted maximum likelihood estimation (TMLE).
• Conducted extensive numerical studies with synthetic data across diverse missingness and data-generating scenarios, including rare outcomes and high missingness proportions.
• Utilized plasmode simulation studies emulating large EHR cohort data to assess performance in a rare-outcome setting with substantial missing confounder data (>50%).

Main Results:

• Doubly-robust methods demonstrated superior performance, particularly in terms of the bias-variance trade-off, compared to MI and IPW across various scenarios.
• Generalized raking and TMLE showed greater efficiency and robustness, especially under conditions of high missingness and rare outcomes.
• The study identified specific scenarios where doubly-robust methods significantly reduce bias and improve precision in effect estimation.

Conclusions:

• Doubly-robust methods, specifically generalized raking and TMLE, represent valuable, yet under-utilized, tools for missing data analysis in pharmacoepidemiology.
• These methods offer improved statistical efficiency and robustness, leading to more reliable safety and effectiveness estimates from real-world data.
• Researchers are encouraged to adopt doubly-robust approaches for handling missing confounder data to enhance the validity of pharmacoepidemiological findings.