OBJECTIVES

Oncologic/hematologic economic evaluations require estimation of mean survival benefit, which requires extrapolation of survival beyond the observed clinical trial data. In this analysis the impact of restricted data availability (i.e. small sample sizes and short follow-up) on parametric estimates of survival was compared between three cancer populations with differing prognosis (chronic myeloid leukemia (CML), stage IV lung cancer and stage IV prostate cancer).

METHODS

Data from the National Cancer Institute's Surveillance, Epidemiology, and End Results (SEER) registry were employed. Lung and prostate cancer patients diagnosed between 1998-2003 and CML patients diagnosed from 1988-1998, with follow-up data available until 2016 were included. Overall survival was estimated using standard parametric models (exponential, Weibull, log-logistic, log-normal, and Gompertz). For each population, survival analyses were run for 4 sample sizes (n= 50, 100, 250, 500) and 5 follow-up durations (follow-up months = 12, 24, 60, 120, 240) yielding 20 permuted scenarios. Using bootstrap techniques, mean survival estimates and root mean square error (RMSE) relative to the observed mean survival for the whole SEER sample were calculated for each scenario in each population.

RESULTS

Mean overall survival was 7.08, 7, and 0.08 years for patients with CML, prostate, and lung cancer, respectively. Across the populations, exponential and Weibull distributions were relatively robust to data limitations. Log-logistic and log-normal distributions were sensitive (large RMSE) to small sample sizes. Gompertz distributions produced the largest overestimates and RMSE across the scenarios. RMSE was higher in populations with a better prognosis (CML and prostate), especially for the restricted follow-up time scenarios.

CONCLUSIONS

This analysis showed that log-normal and log-logistic distributions are more sensitive to sample size restrictions and that restricted follow-up times have the biggest impact in populations with a better prognosis. Results highlight the importance of exercising caution in survival model selection.