An empirical comparison of methods for predicting net survival

Abstract

Background: Providing accurate predictions of long-term net survival for recently diagnosed cancer patients is challenging due to the lack of follow-up. The aim of this study was to empirically compare predictions of net survival obtained from a flexible parametric excess hazard model to predictions obtained using the period and hybrid approaches.

Methods: The Cancer Registry of Norway was used to identify patients diagnosed with cancer during the period 1953-2008. Patients were then followed up for survival until the end of 2013. Net survival was calculated for 23 different cancer sites at 5, 10 and 15 years after diagnosis for each patient cohort. Observed net survival was estimated using the PP estimator. Predicted net survival was obtained omitting the most recent follow-up years using three approaches: a flexible parametric excess hazard model (FPM), the period approach, and the hybrid approach. All estimates were age-standardized to the age distribution of the cohort for which predictions were made. Prediction errors were calculated as the absolute difference between observed and predicted net survival.

Results: Average absolute prediction error across all cancer sites was smallest for FPM for 5-year, 10-year and 15-year net survival. FPM and the hybrid approach gave better predictions of 10- and 15-year net survival than the period approach. The period and hybrid approaches tended to over-estimate net survival for cancer sites with poor survival, and under-estimate net survival for cancer sites where survival has increased over time. Uncertainty in the predictions was considerably smaller when FPM was used compared to the other approaches.

Conclusions: FPM should be considered for predicting net survival when follow-up is incomplete.

Keywords: Flexible parametric models; Net survival; Non-parametric estimator; Prediction; Prediction error.