An empirical comparison of methods for predicting net survival

Area of Science:

• Epidemiology
• Biostatistics
• Cancer Research

Background:

• Accurate long-term net survival prediction for cancer patients is challenging due to incomplete follow-up data.
• Existing prediction methods may not adequately account for evolving survival trends and data limitations.

Purpose of the Study:

• To empirically compare the predictive accuracy of a flexible parametric excess hazard model (FPM) against period and hybrid approaches for net survival.
• To evaluate prediction errors and uncertainty associated with different methods when predicting long-term cancer survival.

Main Methods:

• Utilized Norwegian Cancer Registry data (1953-2008 diagnoses, followed until 2013) for 23 cancer sites.
• Calculated observed and predicted net survival at 5, 10, and 15 years using PP estimator.
• Compared FPM, period, and hybrid approaches for predicting net survival, omitting recent follow-up data; calculated absolute prediction errors.

Main Results:

• Flexible parametric excess hazard model (FPM) demonstrated the smallest average absolute prediction error across all time points (5, 10, 15 years).
• FPM and the hybrid approach outperformed the period approach for 10- and 15-year net survival predictions.
• Period and hybrid methods showed biases, overestimating survival for poor-prognosis cancers and underestimating for those with improving survival trends; FPM had lower prediction uncertainty.

Conclusions:

• Flexible parametric excess hazard model (FPM) is recommended for predicting net survival, especially when patient follow-up data is incomplete.
• FPM offers a more accurate and less uncertain approach to long-term cancer survival prediction compared to period and hybrid methods.