Combining cox regressions across a heterogeneous distributed research network facing small and zero counts

Area of Science:

• Biostatistics
• Epidemiology
• Health Data Science

Background:

• Medical intervention studies increasingly use distributed data from multiple sources like electronic health records.
• Privacy concerns limit data sharing, necessitating cross-network analyses using only summary statistics (e.g., hazard ratios).
• Standard meta-analysis methods combining per-site summary statistics can introduce bias, especially with small outcome counts in Cox proportional hazards models.

Purpose of the Study:

• To address bias in meta-analyses of distributed health data when using Cox proportional hazards models with small outcome counts.
• To propose and evaluate novel methods that avoid normal approximations for improved accuracy.

Main Methods:

• Proposed flexible approximations: skew-normal, one-dimensional grid, and a custom parametric function.
• Evaluated methods using extensive simulation studies for both fixed-effects and Bayesian random-effects models.
• Applied the novel approaches to real-world comparative safety studies of antidepressants across four healthcare databases.

Main Results:

• Standard meta-analysis methods using normal approximations lead to substantial bias with small outcome counts.
• The proposed flexible approximation methods demonstrate reduced bias in simulation studies.
• Bias reduction was observed across both fixed-effects and random-effects model frameworks.

Conclusions:

• Flexible approximations offer a more accurate approach for meta-analyses of distributed health data compared to standard methods relying on normal approximations.
• These methods are crucial for reliable comparative effectiveness and safety research using real-world health data, particularly when dealing with rare outcomes.