3616 Background: The ACCENT database contains individual patient data from over 32,000 patients enrolled onto 18 international clinical trials. Based on recent breast cancer literature, we hypothesized that there could be substantive differences in apparent efficacy estimates using a log-normal (L-N) survival model rather than with standard Kaplan-Meier (K-M) or Cox model methods. While both Cox and L-N survival analyses offer greater specification by individual patient characteristics, the L-N model may more robustly estimate survival under model misspecification.
METHODS
We compared DFS (N=1540) and OS (N=1545) K-M survival estimates for males with stage III disease who were aged 60-65 and received 1) 5FU+LV+/-LEV or 2) other systemic therapy. Cox and L-N modeling with 9,961 patients for DFS and 9,981 patients for OS was for males (all ages: estimation at 60, 62, and 65 yr) with stage II or III disease (estimation for stage III) by the same therapies. Comparisons were at 3, 5, and 8 yr.
RESULTS
There was evidence that the Cox assumption of proportional hazards was violated, and of minor departures from a L-N distribution in the tails of the DFS and OS distributions. Absolute differences between Cox and L-N estimates varied by endpoint and time of assessment: DFS at 3 yr 3.3-4.2%, at 5 yr 1.4-2.7%, and at 8 yr 0.1-1.8%; OS at 3 yr 0.7-1.3%, at 5 yr 1.6%-2.3%, and at 8 yr 0.1-0.8%. While K-M and Cox estimates were generally more similar, than K-M and L-N, this would in part be attributable to the step-wise adjustments at events for K-M and Cox, rather than smooth modeling with the L-N. The need of the Cox model for a baseline hazard function is a disadvantage for extrapolation, compared with the L-N.
CONCLUSIONS
Even with reasonably large population subgroups, there were substantive differences in apparent survival (3.3-4.2%) between Cox and L-N model types, particularly for 3-yr DFS. Smaller differences (0.1-0.8%) between the methods were observed near the truncation of follow-up at 8 yr. The magnitude of differences in survival estimates with a L-N versus a Cox model were large enough to be clinically relevant and warrant further consideration as we evaluate new therapies and prognostic/predictive factors.