Innovative estimation of survival using log-normal survival modelling on ACCENT database

BACKGROUND: The ACCENT database, with individual patient data for 20 898 patients from 18 colon cancer clinical trials, was used to support Food and Drug Administration (FDA) approval of 3-year disease-free survival as a surrogate for 5-year overall survival. We hypothesised substantive differences in survival estimation with log-normal modelling rather than standard Kaplan–Meier or Cox approaches. METHODS: Time to relapse, disease-free survival, and overall survival were estimated using Kaplan–Meier, Cox, and log-normal approaches for male subjects aged 60–65 years, with stage III colon cancer, treated with 5-fluorouracil-based chemotherapy regimens (with 5FU), or with surgery alone (without 5FU). RESULTS: Absolute differences between Cox and log-normal estimates with (without) 5FU varied by end point. The log-normal model had 5.8 (6.3)% higher estimated 3-year time to relapse than the Cox model; 4.8 (5.1)% higher 3-year disease-free survival; and 3.2 (2.2)% higher 5-year overall survival. Model checking indicated greater data support for the log-normal than the Cox model, with Cox and Kaplan–Meier estimates being more similar. All three model types indicate consistent evidence of treatment benefit on both 3-year disease-free survival and 5-year overall survival; patients allocated to 5FU had 5.0–6.7% higher 3-year disease-free survival and 5.3–6.8% higher 5-year overall survival. CONCLUSION: Substantive absolute differences between estimates of 3-year disease-free survival and 5-year overall survival with log-normal and Cox models were large enough to be clinically relevant, and warrant further consideration.