Biased methods for estimating local and distant failure rates in breast carcinoma and a "commonsense" approach

Abstract

BACKGROUND. Several methods are used to estimate risks of local and distant failure after treatment of breast carcinoma. The authors' purpose was to present a physician-friendly description of the potential bias in these methods, and to suggest an improvement. METHODS. The cumulative incidence based on first event (cumulative incidence [CI]) and Kaplan-Meier method based on first (KM[1st]) or all (KM[any]) events, are applied to a database comprising 2521 women treated for breast carcinoma at the same institution and observed for more than 20 years. The authors relate these estimates to the region containing all possible estimates of failure rate. This region contains the "true" risk (net risk, or risk that would be observed in the absence of competing risks) of local or distant failure. RESULTS. The CI estimate is the lowest possible estimate of the true failure rate. Under certain "commonsense" assumptions, the CI estimate is below the lowest possible estimate of risk of failure. The KM(1st) estimate is higher than the CI estimate and lower than the KM(any) estimate. Under the same commonsense assumptions, the KM(1st) method also underestimates the true failure rate. CONCLUSIONS. Methods based on time to first event such as CI and KM(1st) underestimate the true risk. In the design of clinical trials, consideration should be given to longer follow-up and the KM(any) method of analyzing results because it provides a less biased estimate. © 2001 American Cancer Society.