Table 2 Summary of key findings
• There is a large risk of error when extrapolating clinical time-to-event data with small sample sizes, which is observed regardless of whether the underlying event distribution has been correctly specified when undertaking extrapolation. Error is more markedly reduced by larger samples than by observing more events with longer follow-up alone. • Uncertainty may not be sufficiently captured within estimated confidence intervals when extrapolating limited clinical data for use in decision models, suggesting that probabilistic analysis is not sufficient to overcome the limitations of small samples or of short follow-up in large samples. • Identifying lifetime time horizon based on the model’s extrapolated output will not reliably estimate mean lifetime survival and its uncertainty. • For data with an exponential event distribution, AIC less frequently correctly identified the true distribution and performed very poorly in estimating outcomes and appropriately capturing their uncertainty compared to selections based on BIC. |