Between-study heterogeneity poses challenges to the generalisability of meta-analytical results, which can influence their ability to predict outcomes in future studies. Prediction intervals have been proposed to account for both uncertainty and heterogeneity, yet their real-world performance in predicting future studies has not been systematically evaluated.
This study aims to assess the prediction performance of meta-analyses, focusing on how effectively they predict later study results based on meta-analyses of earlier studies.
This empirical study used a comprehensive collection of meta-analyses from the Cochrane Database of Systematic Reviews. Through in-sample evaluation, the success of predicting later study results was assessed based on meta-analyses of earlier studies in Cochrane reviews. The impact of factors such as the number of studies in the meta-analysis and uncertainties in heterogeneity estimation was also analysed.
The findings reveal that prediction failures are common, particularly as the number of studies in the meta-analysis increases. This may be attributed to uncertainties in estimating between-study heterogeneity. Conversely, when the number of studies is small, the proportion of successful predictions is high. However, this is likely due to large uncertainties in predictions and the limited information provided by fewer studies, which may reduce their utility in providing valuable evidence for future studies.
These results underscore the importance of cautious interpretation and further investigation when applying meta-analytical findings to future studies. Our findings suggest several potential strategies for predicting future study results through evidence synthesis, with particular emphasis on carefully considering between-study heterogeneity, the number of studies included in a meta-analysis, and the temporal trends in individual study results.