OBJECTIVES: Survival extrapolation is required to estimate survival beyond the duration of a clinical trial. The selection of the survival model is partly guided by goodness-of-fit statistics, namely the Akaike information criteria (AIC) and the Bayesian information criteria (BIC). The objective of this simulation study was to assess the performance of AIC and BIC when selecting between six standard parametric survival models to extrapolate survival data with different levels of censoring.

METHODS: Survival data was simulated from six standard survival distributions (exponential, gamma, Weibull, Gompertz, log-normal, and log-logistic). Censoring was simulated using a uniform distribution. The level of censoring was categorized as low (≈30%), medium (≈50%), and high (≈70%). The six distributions were then fit to each of the six simulated data sets to assess the performance of AIC and BIC.

RESULTS: At a low level of censoring, AIC and BIC indicated the correct survival model in approximately 70% and 80% of the simulations, respectively. With increasing censoring, the accuracy of AIC and BIC decreased. At a high level of censoring, AIC and BIC indicated the correct model in 20% to 40% of the simulations. The exponential and the log-normal distribution were the least sensitive to high censoring.

CONCLUSIONS: At high levels of censoring neither AIC nor BIC can guide the choice of a survival model. At 70% or more censoring, weight should be given to the clinical plausibility of the different survival models rather than AIC and BIC. The simulation study showed that BIC was more often able to correctly identify the underlying survival distribution compared to AIC.