OBJECTIVES: Funding decisions for health technologies often require estimates of future survival, and hence require accurate extrapolations. There is guidance on the use of standard survival models for extrapolation, and growing interest in the use of more flexible models. However, it is unclear if more flexible models provide more accurate extrapolations than standard models. The aim of this study was to populate this evidence-gap using a representative simulation study.

METHODS: We simulated data from a two-mixture-Weibull model, representing patients with either a high hazard (short survival) or a low hazard (long survival). We considered nine scenarios: three different lengths of follow-up and three different sample sizes. Standard survival models were the exponential, Weibull, Gompertz, log-logistic, lognormal, gamma and generalised gamma. Flexible models were fractional polynomials, two spline models (one extrapolated a cubic polynomial trend, the other a linear trend), and a dynamic survival model (DSM) which extrapolates a linear trend. The primary outcome was the mean-squared-error (MSE) which combines bias and variance.

RESULTS: Extrapolations from the fractional polynomials were very unstable with many hazards becoming infinitely large. Spline models with cubic extrapolations were highly variable, with MSE values worse than standard models for all scenarios. Both the DSM and spline model with a linear extrapolation had lower MSE than standard models for all of the scenarios considered, although this improvement was only statistically significant for the DSM.

CONCLUSIONS: Extrapolations from DSMs provided the best trade-off between bias and variance, followed by spline models extrapolating a linear trend. In general, there was considerable uncertainty in the results; a significant improvement compared with standard models was only seen for DSMs. Estimates from the remaining models showed a very high variance, and do not appear suitable for routine generation of extrapolations under the scenarios considered.