English

Abstract

Meta-analysis that pools two levels of data, namely, aggregate data (AD) and individual patient data (IPD) is increasingly common. The implications of pooling these data on the overall meta-analysis estimates have not been fully explored. We examined some of the statistical properties of overall estimate of the treatment effects from meta-analysis which combine the AD and IPD studies. We compared these estimates with those from the all-AD and all-IPD meta-analyses in terms of the bias, root-mean-square-error and coverage probability. We used simulated meta-analyses to evaluate the behaviour of these estimates. The results indicated superiority of estimates from IPD meta-analysis, compared to those from combined-level studies and AD studies, in terms of the accuracy and efficiency of the estimates. Additionally, for the same statistical properties examined, the estimates from combined-level studies displayed better results over those from all-AD meta-analysis. In the scenario involving different ratios of AD:IPD, we found that combined-level studies generated estimates with better statistical properties, irrelevant of the composition of the AD:IPD ratio. Therefore, whenever possible, we recommend the inclusion of available IPD studies when conducting traditional AD meta-analysis.