OBJECTIVES: We sought to explore the impact of short life expectancy, which is a standard exclusion criterion in clinical trials, on choice of model, predicted mean survival, and maximum survival (time to reach 1% survival).

METHODS: Survival data were simulated using exponential distributions with 0-100% of patients with <3 months survival excluded at random. A recent study reported 80% accuracy by physicians at predicting <3 months life expectancy. Patients were entered at random times at 0-21 months with a maximum follow-up of 24 months and compared to complete uncensored data. In total, 7 standard parametric models, and 7 models with a turning point at 3 months were fitted to the simulated data and survival estimates compared with uncensored data. Best fitting models were selected using Akaike, and Bayesian information criteria (AIC, BIC).

RESULTS: As the proportion of patients removed increased, the best fitting standard models changed from exponential to Weibull/gamma, to log-logistic, to log-normal/generalized gamma. With 80% of patients with < 3 months survival excluded, the uncensored data gave a mean survival of 1.58 years (95% prediction intervals (pi): 1.48, 1.69), and maximum survival of 6.51 years (95% pi: 5.57, 7.71). The standard parametric models selected on AIC and BIC gave mean survival predictions of 1.95 years (overestimation of 23%) (95% pi: 1.39, 2.42), and maximum survival of 15.07-15.21 years (overestimation of ~133%) (95% pi: 4.41, 23.05). The models which included a turning point at 3 months gave a mean survival of 1.58 years (0% overestimation) (95% pi: 1.43, 1.77) and maximum survival of 6.59-6.60 years (1% overestimation) (95% pi: 5.87, 7.73).

CONCLUSIONS: Exclusion criteria violate the random sampling assumption made by standard survival models, which may result in substantial overestimation and uncertainty in survival estimates. When exclusion criteria are present, only models with a turning point should be fitted to the survival data.