Using the expected survival to explain differences between the results of randomized trials: a case in advanced ovarian cancer.

Abstract

A meta-analysis of randomized trials in advanced ovarian cancer showed a longer survival with cyclophosphamide, doxorubicin, and cisplatin (CAP) than with cyclophosphamide and cisplatin (CP; P =.009). In contrast, the results of the large International Collaborative Ovarian Neoplasm Study (ICON2) showed no survival difference between CAP and carboplatin (P =.98). In this article, we show how these discrepant results can be reconciled through the estimation of expected survival curves. A proportional hazards model, fitted to the meta-analysis data, was used to construct the expected survival curve for each treatment arm of the ICON2 trial. Expected survival curves were compared with observed survival curves in the ICON2 trial at all time points using a nonparametric test. The prognostic model for survival obtained in the meta-analysis included extent of residual disease, age, histologic grade, and International Federation of Gynecology and Obstetrics stage. When this model was applied to the ICON2 data, there was no difference between the expected and observed curves in the CAP arm. In contrast, the observed survival curve for carboplatin was far superior to the expected survival curve for CP (P <.01). These analyses provide indirect evidence that better results are achieved with carboplatin alone at an optimally tolerated dose, compared with the CP combination at a cisplatin dose of 50 to 60 mg/m2. The expected survival may provide valuable insight when direct comparisons between randomized groups yield discrepant results across different studies.