Is UWLS Really Better for Medical Research?

Area of Science:

• Biostatistics
• Medical Research Methodology
• Data Science

Background:

• Meta-analysis is crucial in medical research for synthesizing evidence.
• Estimator performance can vary, especially in small-sample contexts.
• Model selection criteria like AIC and BIC may be unreliable with small sample sizes and weak signals.

Purpose of the Study:

• To evaluate the performance of the Unrestricted Weighted Least Squares (UWLS) estimator in medical meta-analyses.
• To address limitations of model selection criteria in small-sample settings.
• To compare UWLS with Random Effects (RE) and Fixed Effect (FE) estimators.

Main Methods:

• A large-scale simulation approach was used, generating 108,000 datasets.
• Simulated datasets mirrored characteristics of the Cochrane Database of Systematic Reviews (CDSR).
• Evaluated estimators based on bias, efficiency, and standard error accuracy.

Main Results:

• UWLS, RE, and FE estimators showed similar performance regarding bias and efficiency.
• The RE estimator consistently produced more accurate standard errors than UWLS.
• AIC and BIC were found to be unreliable for model selection in the simulated CDSR-like datasets.

Conclusions:

• The Random Effects (RE) estimator is recommended for general use in medical meta-analyses due to reliable standard errors.
• The choice between UWLS and FE should consider the anticipated level of effect heterogeneity.
• Caution is advised when using AIC and BIC for model selection in small-sample meta-analyses.