Simple robust procedures for combining risk differences in sets of 2 x 2 tables

Area of Science:

• Biostatistics
• Epidemiology
• Medical Research Methodology

Background:

• Meta-analyses commonly employ random-effects models to address study result heterogeneity.
• Anomalous studies can disproportionately influence meta-analytic findings.
• Risk difference is a key measure in meta-analyses, especially for binary outcomes.

Purpose of the Study:

• To compare the performance of trimmed versus untrimmed meta-analytic estimators for the risk difference.
• To evaluate the robustness of trimmed estimators against anomalous study results.
• To identify the most effective trimmed meta-analytic procedure for handling heterogeneity.

Main Methods:

• A simulation study was conducted to compare four meta-analytic procedures for risk differences.
• Trimmed and untrimmed versions of these procedures were evaluated.
• A modified DerSimonian-Laird estimator and its trimmed adaptation were specifically assessed.
• Winsorization was used to adapt variance component estimation for robustness.

Main Results:

• A modified DerSimonian-Laird estimator is effective when random-effects models capture study variability.
• A 20% trimmed, weighted version of this estimator demonstrated resistance to highly anomalous study results.
• Among the four trimmed procedures, the modified DerSimonian-Laird trimmed version performed best across various simulations.
• No single method (trimmed or untrimmed) was universally superior.

Conclusions:

• Trimmed meta-analytic procedures, particularly a modified DerSimonian-Laird estimator, offer enhanced resistance to outlier studies in meta-analyses.
• The choice of meta-analytic method depends on the specific characteristics of the data and the presence of heterogeneity.
• Further research may be needed to establish optimal trimming strategies for different scenarios.