Consequences of multiple imputation of missing standard deviations and sample sizes in meta-analysis

Area of Science:

• Ecology
• Biostatistics
• Scientific Research Methods

Background:

• Meta-analyses require complete variance measures (e.g., standard deviation values) and sample sizes for accurate weighting.
• Incompletely reported data is a common issue in published ecological meta-analyses, affecting a majority of studies.
• Omitting studies with missing data is not a viable solution for meta-analysis.

Purpose of the Study:

• To systematically survey the frequency and treatment of missing data in ecological meta-analyses.
• To investigate the performance of 14 different methods for treating or imputing missing standard deviations (SDs) and sample sizes (SSs).
• To assess the impact of missing data imputation on the accuracy of meta-analysis results.

Main Methods:

• Systematic literature survey of ecological meta-analyses to identify missing data patterns.
• Simulation of meta-analysis datasets with varying degrees of missing SDs and SSs.
• Performance assessment of 14 imputation and treatment options against complete data benchmarks.

Main Results:

• Unweighted and sample size-based variance approximations can provide unbiased results when effect sizes are independent of SDs and SSs.
• The effectiveness of imputation methods varies with data structure, particularly when effect sizes, SDs, and SSs are correlated.
• Multiple imputation of up to 90% missing SDs and SSs yielded results comparable to complete data analyses under ideal conditions (missing at random, unrelated to effect size).

Conclusions:

• Multiple imputation is a promising approach to address incompletely reported data in meta-analyses, extending beyond ecological studies.
• Careful consideration of missing data patterns and potential correlations between variables is crucial for reliable imputation.
• Imputation methods can significantly improve the robustness of meta-analyses when faced with missing variance measures and sample sizes.