Time-dependent summary receiver operating characteristics for meta-analysis of prognostic studies

Area of Science:

• Biostatistics
• Medical Informatics
• Epidemiology

Background:

• Prognostic studies are crucial for assessing biomarker associations with patient outcomes and informing medical decisions.
• Standard meta-analysis techniques face challenges with prognostic studies due to reliance on study-specific cut-off values.
• Existing methods struggle to synthesize evidence from prognostic studies with time-to-event outcomes.

Purpose of the Study:

• To propose novel methods for estimating time-dependent summary receiver operating characteristic (ROC) curves in meta-analyses of prognostic studies.
• To address the complexities of analyzing prognostic studies with right-censored time-to-event data.
• To provide robust statistical frameworks for synthesizing evidence from multiple prognostic studies.

Main Methods:

• Introduction of a bivariate normal model for time-dependent sensitivity and specificity, utilizing summary statistics from published studies.
• Development of a bivariate binomial model incorporating a multiple imputation method for inference.
• Application of Rubin's variance formula justified from a Bayesian perspective for the bivariate binomial model.

Main Results:

• Both proposed methods demonstrate effectiveness in meta-analyses with a moderate to large number of studies.
• The bivariate binomial model with multiple imputation shows superior performance compared to the bivariate normal model when the number of studies is small.
• Simulation studies and real-data application validate the proposed methodologies.

Conclusions:

• The developed methods offer valid and asymptotically sound inferences for meta-analyses of prognostic studies.
• The bivariate binomial model with multiple imputation is recommended for its robustness, especially with fewer studies.
• These approaches improve the synthesis of evidence from prognostic biomarker research, aiding clinical decision-making.