Estimating Between-Trial Heterogeneity in Meta-Analyses Based on Two-Arm Clinical Trials with Outcomes Reported as Kaplan-Meier Curves

Background Between-trial heterogeneity is a key element in meta-analysis, traditionally quantified using the I² statistic in studies with binary outcomes. However, survival meta-analyses present additional challenges, as outcomes are usually reported through Kaplan–Meier curves and hazard ratios (HRs). Standard methods for heterogeneity estimation in this context remain poorly established, and consensus is lacking. Methods We propose a standardized approach for estimating between-trial heterogeneity in survival meta-analyses using I². The method is applicable both when individual patient data (IPD) are available (collaborative meta-analyses) and when IPD must be reconstructed from published Kaplan–Meier curves (IPDfromKM algorithm). To illustrate the approach, we re-analyzed a published meta-analysis of randomized controlled trials (RCTs) evaluating PARP inhibitor maintenance therapy in extensive-stage small-cell lung cancer. Five RCTs were included, and overall survival was the endpoint. Results The binary meta-analysis of crude survival rates yielded no significant heterogeneity (I² = 0%). By contrast, re-analysis based on reconstructed IPD and log-transformed HRs indicated moderate heterogeneity (I² = 36.3%, τ² = 0.0233, p = 0.179). Estimates of the overall treatment effect were similar between approaches (HR ≈ 1.03–1.04), though confidence intervals differed due to model specifications. Comparative evaluation with other methods (Wald test, likelihood ratio, concordance index) highlighted the unique interpretative advantages of I² in this setting. Discussion Our findings suggest that crude binary analyses may underestimate heterogeneity in survival meta-analyses. The I² statistic provides an intuitive and flexible measure of between-trial variability when survival data are expressed as HRs. While promising, this approach requires further validation across diverse clinical settings.