Assessing two meta-analysis (MA) methods: Individual patient data-based (IPD) versus literature-based abstracted data (AD) in 10 MA including 37,002 patients (pts)-Are there differences of concern?

6054 Background: MA are widely used in evidence-based medicine (EBM); whether the two most frequently used methods (IPD, AD) provide equally reliable results is debated. Here we analyzed the conclusions achieved by each method using the same panel of studies, weighing agreement/discrepancy and evaluating predictivity of AD results with respect to IPD using multiple methods. METHODS 10 different settings (Table) subjected to IPD were selected to also conduct AD. Hazard and risk ratios (Rs) with 95% confidence intervals (CI) were derived. Methods agreement was analyzed (Blant-Altman method) as difference or ratio and compared using variance-ratio (F-test) and concordance correlation coefficient (CCC). Correlations between AD/IPD-Rs were estimated with linear regression [Pearson (r), R2 coefficients, Spearman-Rho/Kendall-Tau coefficients] to derive a predictive estimation of AD on IPD. RESULTS 14 outcomes in 10 MAs (37,002 pts) were analyzed (Table). Graphical comparison with the B-A method did not show meaningful differences between AD and IPD (mean difference 0.0075, 95% CI -0.0052, 0.02), without significant discrepancy, with no significant differences in variance (F-test 1.06, p=0.91), and high CCC (0.98, 95% CI 0.59, 0.99). A highly significant correlation by both parametric and nonparametric tests allowed derivation of a linear regression equation able to predict IPD results from AD data (YAD=[0.04315] + XIPD[0.958]; R2 0.97). CONCLUSIONS Analysis of MAs conducted by AD or IPD methods showed: 1) no significant differences between the 2 MA methods; 2) AD results predict IPD results, without significant risk of overestimation. [Table: see text].