Zero-cell corrections in random-effects meta-analyses.

Area of Science:

• Biostatistics
• Meta-analysis
• Epidemiology

Background:

• Standard estimators for log odds ratio (unconditional maximum likelihood estimator) and its standard error (delta-method) are undefined when 2x2 tables have zero cells.
• Zero-cell issues persist in meta-analysis for overall log odds ratio estimation.
• Zero-cell corrections, while known to be problematic, are still frequently employed.

Purpose of the Study:

• To compare the performance of two distinct zero-cell correction methods for log odds ratio estimation in meta-analysis.
• To identify and discourage a particularly detrimental zero-cell correction strategy.

Main Methods:

• A simulation study was conducted to evaluate two zero-cell correction methods under an ordinary random-effects model.
• Method (a): Adding 0.5 to all cells in all 2x2 tables, irrespective of zero cells.
• Method (b): Adding 0.5 only to cells in 2x2 tables containing at least one zero cell.

Main Results:

• Correction method (a) demonstrated poorer performance compared to method (b).
• The simulation results indicate significant negative impacts from applying a universal increment.

Conclusions:

• The study strongly advises against the use of adding a constant increment to all cells in all 2x2 tables, regardless of zero-cell presence.
• Correction method (b), which selectively targets tables with zero cells, is comparatively superior.