Cancer net survival on registry data: use of the new unbiased Pohar-Perme estimator and magnitude of the bias with the classical methods

Abstract

Net survival, the survival which might occur if cancer was the only cause of death, is a major epidemiological indicator required for international or temporal comparisons. Recent findings have shown that all classical methods used for routine estimation of net survival from cancer-registry data, sometimes called "relative-survival methods," provide biased estimates. Meanwhile, an unbiased estimator, the Pohar-Perme estimator (PPE), was recently proposed. Using real data, we investigated the magnitude of the errors made by four "relative-survival" methods (Ederer I, Hakulinen, Ederer II and a univariable regression model) vs. PPE as reference and examined the influence of time of follow-up, cancer prognosis, and age on the errors made. The data concerned seven cancer sites (2,51,316 cases) collected by FRANCIM cancer registries. Net survivals were estimated at 5, 10 and 15 years postdiagnosis. At 5 years, the errors were generally small. At 10 years, in good-prognosis cancers, the errors made in nonstandardized estimates with all classical methods were generally great (+2.7 to +9% points in prostate cancer) and increased in age-class estimations (vs. 5-year ones). At 15 years, in bad- or average-prognosis cancers, the errors were often substantial whatever the nature of the estimation. In good-prognosis cancers, the errors in nonstandardized estimates of all classical methods were great and sometimes very important. With all classical methods, great errors occurred in age-class estimates resulting in errors in age-standardized estimates (+0.4 to +3.2% points in breast cancer). In estimating net survival, cancer registries should abandon all classical methods and adopt the new Pohar-Perme estimator.