Generalized meta-analysis for multiple regression models across studies with disparate covariate information

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Modeling

Background:

• Meta-analysis synthesizes data from multiple studies for logistical convenience and statistical efficiency.
• Existing methods often struggle with combining multivariate regression parameters across studies with differing covariate information.

Purpose of the Study:

• To develop a generalized meta-analysis approach for combining multivariate regression parameters from studies with varying covariate data.
• To provide a robust framework for estimating parameters of a maximal model using information from reduced models across studies.
• To introduce a diagnostic test for detecting assumption violations, such as population heterogeneity.

Main Methods:

• Utilizing algebraic relationships among regression parameters to specify moment equations for parameter estimation.
• Employing the generalized method of moments (GMM) for solving these equations, with optimal weighting to account for parameter estimate uncertainty.
• Extending the iterated reweighted least-squares algorithm for fitting generalized linear regression models within the proposed framework.

Main Results:

• The proposed generalized meta-analysis approach effectively combines multivariate regression parameters from studies with disparate covariate information.
• The method allows for the estimation of a maximal model's parameters by leveraging information from reduced models.
• A novel diagnostic test was developed to detect heterogeneity and other model assumption violations.

Conclusions:

• The developed generalized meta-analysis framework offers a statistically efficient and flexible approach for synthesizing complex data from multiple studies.
• This method facilitates the creation of robust prediction models, as demonstrated in a breast cancer risk prediction example.
• The approach enhances the ability to handle varying covariate information and assess model assumptions in meta-analytic settings.