Methods for using clinical laboratory test results as baseline confounders in multi-site observational database...

Area of Science:

• Health research methodology
• Biostatistics
• Epidemiology

Background:

• Missing baseline laboratory results are common in multi-site health studies.
• This missingness can potentially bias study outcomes and affect the reliability of findings.
• Understanding predictors of missing data and evaluating methods to handle it is crucial for robust research.

Purpose of the Study:

• To quantify the extent of missing baseline laboratory data in multi-site studies.
• To identify factors that predict the missingness of these laboratory results.
• To evaluate the performance of different statistical methods for handling missing data.

Main Methods:

• Utilized the Mini-Sentinel Distributed Database across three study sites.
• Examined three exposure-outcome scenarios involving laboratory confounders (glucose, creatinine, INR).
• Compared results from models omitting data, using complete cases, and applying multiple imputation (MI) regression and MI predictive mean matching (PMM).

Main Results:

• Laboratory data availability varied significantly across sites and scenarios (e.g., glucose 27.7-58.9%, creatinine 44.5-79.0%, INR 20.0-92.9%).
• Results differed between complete case analysis and methods accounting for missing data.
• Multiple imputation methods generally yielded similar estimates, with minor variations observed between site-specific and across-site models.

Conclusions:

• Multi-site studies require careful consideration of site-specific variability in missing data patterns.
• Various missing data methods, including multiple imputation, demonstrated comparable performance in the evaluated scenarios.
• Addressing missing data is essential for accurate effect estimation in large-scale observational studies.