Computationally efficient methods for fitting mixed models to electronic health records data.

Area of Science:

• Biostatistics
• Health Informatics
• Epidemiology

Background:

• Analysis of large healthcare datasets presents computational challenges.
• Mixed-effects models are suitable for clustered data, like patient records within general practices.
• Existing methods for large datasets, such as subsampling, may impact statistical power.

Purpose of the Study:

• To describe novel statistical methods for analyzing large primary care datasets.
• To investigate associations between patient characteristics and health outcomes while accounting for practice-level variation.
• To compare the efficiency and accuracy of new methods against existing approaches.

Main Methods:

• Development and application of weighted regression for large datasets.
• Utilizing meta-analysis of practice-specific regression coefficients.
• Comparison with a subsampling approach for mixed-effects model fitting.
• Inclusion of random intercepts to model heterogeneity among general practices.

Main Results:

• Weighted regression and meta-analysis significantly reduce dataset size and analysis time.
• Both methods yield point estimates comparable to analyzing the entire dataset.
• Weighted regression and meta-analysis provide standard errors similar to full dataset analysis.
• Subsampling resulted in larger standard errors compared to the proposed methods.

Conclusions:

• Weighted regression and meta-analysis are efficient and effective for analyzing large primary care datasets.
• These methods provide accurate estimates and standard errors, facilitating research on patient characteristics and outcomes.
• The described statistical techniques are readily implementable in standard statistical software.