PNS336 NETWORK META-ANALYSIS (NMA) WITH TIME-VARYING HAZARD: A COMPARATIVE ANALYSIS OF SURVIVAL OUTCOMES

Abstract

In oncology, the standard approach to NMA of survival outcomes is to assume proportional hazards (PH), i.e. to assume that the hazard ratio (HR, or treatment effect) remains constant over time. This, however, may not always hold true. The objective of this analysis was to perform a NMA with alternative methodologies and compare the results to understand how the selected method might influence survival estimates. Using a Cochrane review as a basis, the studies included in the NMA assessed progression-free and overall survival (PFS/OS) in metastatic colorectal cancer patients (KRAS exon 2 wild-type) receiving first-line epidermal growth factor receptor inhibitors. Kaplan-Meier curves were digitised, individual patient-level data (IPD) reconstructed and the PH assumption tested. The NMA was conducted with the four following methods: 1) using HRs reported in the publications; 2) considering treatment effects on shape and scale of parametric survival curves; 3) fractional polynomials; 4) restricted cubic splines. Methods 2–4 required IPD. Six randomised control trials (RCTs) were included in the analyses, of which five investigated cetuximab/chemotherapy vs chemotherapy alone and one panitumumab/chemotherapy vs chemotherapy alone. The PH assumption did not hold for PFS and/or OS in all RCTs. Substantial differences were observed in hazards, over time, with each of the four methods. Comparing the four NMA methods, differences were observed in the estimated median PFS and OS: for cetuximab/chemotherapy, the standard approach yielded median PFS and OS of 10.1 and 24.2 months, respectively, with up to a 15% variation with the alternative methods. This study highlights the importance of using an appropriate approach to model selection in survival analysis, as model selection may have a substantial impact on the estimated relative treatment effects and consequently on the long-term cost-effectiveness of health technologies.