OBJECTIVES: When analyzing the cost-effectiveness of new pharmaceuticals to treat metastatic cancer, a common methodological issue is the extrapolation of costs and benefits beyond those observed in randomized controlled trials (RCTs). A general approach is to predict survival using parametric models. The Surveillance, Epidemiology, and End Results (SEER) Cancer Incidence Research Database provides an opportunity to investigate the validity of model predictions in a cohort of patients with known long-term mortality outcomes. METHODS: Adult women with metastatic breast cancer diagnosed between 1973 and 2004 were extracted from the 1973–2014 SEER database (n=22,357). All had at least 10 years of follow-up, enabling the calculation of 10-year restricted mean survival. To mimic the recruitment period and duration of a typical RCT, patients were allocated a pseudo-randomization date over a two year period. Follow-up times were truncated at three, four and five years from the first pseudo-randomization date. Six common parametric models were fitted to each truncated data set from baseline and from a series of cut points. The accuracy of the predicted 10-year restricted mean survival estimates and the likelihood of optimal model selection based on alternative selection strategies were examined. RESULTS: When models were fitted from baseline, predicted 10-year restricted mean survival was within ±1 month of the observed value (36.75 months) for one model fitted to three years of data (Gompertz: 35.86 months), for one model fitted to four years of data (Gompertz: 36.03 months) and for two models fitted to five years of data (Gompertz: 36.60; generalized gamma: 36.88 months). The Gompertz provided the closest predictions at all time points, however the generalized gamma had the lowest AIC and BIC at all time points. CONCLUSIONS: Investigating the fit of parametric models to truncated registry data can provide insight into model selection strategies for the extrapolation of survival data.