Abstract
6584 Background: The Cox proportional hazards model allows one to estimate the effect of factors influencing a time-to- event or survival time. This model, however, relies on a fundamental assumption, the proportionality of the hazards (PH): given two individuals with distinct values for a specific factor, the ratio of the estimated hazards is constant over time. A typical example in the genral population is the effect of age on the survival time following surgery. If age is more likely to impact the survival immediately after surgery than weeks or months later, then the PH hypothesis would not hold, and thus misleading effect estimates could be derived. Altman et al. showed that the majority of cancer publications do not report verifying this crucial assumption. Methods: We emphasize the importance of the assumption of the Cox model, and the misleading conclusions that can be inferred if it is violated. We recall that the hazard ratio (HR) provided by the Cox model is a time-averaged hazard ratio. If the HR changes over time, this time-averaged HR might not be representative of the true effect. Results: We illustrate our discussion by analyzing time to metastases in 855 women treated with surgery for a primary breast cancer, with the following prognostic factors: age, tumor size, SBR grade, vascular emboli, lymph node involvement, estrogen (ER) and progesterone receptor status, Her2 status, Mib1 proliferation index. The median follow-up time was 11 years, and there were 216 relapses. While ER was not significantly associated with the time to metastases in the initial Cox model, further analyses showed that ER effect was not constant over time. On average over the study follow-up, ER is not associated with time to metastases, but given the non constant HR, ER may have an impact over some specific periods of time. Conclusion : Our example illustrates that misleading conclusions can be inferred when one ignores the PH assumption. We provide statistical elements the reader should look for when reviewing clinical publications relying on a Cox model; this is particularly important when the follow-up period is long. No significant financial relationships to disclose.