An Assessment of Methods to Combine Published Survival Curves

Abstract

Purpose. To assess the accuracies of different techniques for combining published survival curves, for use in disease modeling applications. Methods. Five methods were identified: 1) iterative generalized least-squares (IGLS), 2) meta-analysis of failure-time data with adjustment for covariates (MFD), 3) nonlinear regression (NLR), 4) log relative risk (LRR), and 5) weighted LRR (w-LRR). Each method was used to combine the survival curves from eight single-arm Phase II trials of chemotherapy in 918 patients with advanced non-small-cell lung cancer (NSCLC). The resulting summary curves were compared with the curve calculated from the corresponding individual patient data (IPD). Results. All methods were able to produce accurate summary survival curves statistically similar to the IPD-derived curve. Maximum discrepancies ranged from 1.8% to 4.7%. MFD appeared to be the most accurate when censoring information was complete. Characteristics of the component trials that adversely affected the accuracies of the different techniques were 1) a high proportion of censored observations (MFD); 2) variability in the length of follow-up (IGLS, NLR, LRR, w-LRR); and 3) the heterogeneity of the treatment results (NLR, w-LRR). Conclusions. All methods were able to accurately reproduce summary survival curves from the published literature. The best method depends on characteristics of the data and the purpose of the analysis. Key words: survival analysis; meta-analysis; life tables; proportional hazards models. (Med Decis Making 2000;20:104-111)