Category | Description |
---|---|
Aims | The aim of this study was to assess the performance of standard parametric survival analysis techniques for analysis of time-to-event data from clinical trials under conditions of limited data due to small samples or short follow-up. |
Data generating mechanism | Data were generated for the event of interest from an exponential survival distribution, characterized by a constant hazard rate, λ. |
Estimands and population targets | - Exponential distribution of event times - Median survival time, t where S(t) = 0.5 - One-year landmark survival probability, S(t) where t = 365 days - Population time horizon, THpop, defined at 1% survival time, t where S(t) = 0.01 - Restricted mean survival time (RMST) estimated at time horizon THpop |
Methods | Simulated populations were created and nsim = 5000 repetitions drawn. Each repetition included six levels of sample size, nobs = {30, 60, 90, 120, 250, 500}. Within each repetition and sample size, artificially censored datasets were created based on deciles of proportions of events, pe = {10%, 20%, …, 100%}, creating ten levels of follow-up. Standard parametric distributions (exponential, Weibull, log-normal, log-logistic, generalized gamma and Gompertz) were fitted to each grouping for each repetition, nonconverging or implausible fits removed, and estimated model parameters (estimators) collected from extrapolated survival curves: - Information criteria (IC) to determine the best-fitting distribution - Median survival time, t where S(t) = 0.5 - One-year landmark survival probability, S(t) where t = 365 days - Sample time horizon THi, (1% survival time), t where S(t) = 0.01 - Population time horizon RMST (RMST estimated at THpop) - Sample time horizon RMST (RMST estimated at THi) |
Performance measures | - Proportion identifying the true distribution as best fitting - Coverage - Error  ◦ Mean absolute error (MAE)  ◦ Mean absolute percentage error (MAPE)  ◦ Root mean squared error (RMSE)  ◦ Probability of 20% error |