Effect of sample size and data maturity on parametric survival modeling projections in advanced cancer

Graham, C., Davis, K., & Goyal, R. (2014). Effect of sample size and data maturity on parametric survival modeling projections in advanced cancer. Value in Health, 17(7), A566. DOI: 10.1016/j.jval.2014.08.1881

Objectives
Parametric survival modeling (PSM) is often used in cost-effectiveness analyses of oncology treatments to aid in lifetime projections due to right censoring of data. We sought to better understand the effect of sample sizes and data maturity (follow-up time) on PSM projections to aid in the design of clinical trials and the interpretation of cost-effectiveness models.

Methods
We modeled overall survival (OS) for advanced colorectal cancer patients treated with first-line chemotherapy and/or a biologic using SEER-Medicare data (2004-2010). Survival was estimated using Kaplan-Meier (KM) and PSM methods. From the full cohort, we randomly drew patients to match typical sample sizes from Phase II and III clinical trials (n=50,100,200, and 400). Additionally, arbitrary data cutoffs were created to proxy clinical trial follow-up times (t=3,6, 9,12,24, and 36 months). Using PSM methods mean survival from the full cohort was compared with survival from the combinations of sample sizes and follow-up times.

Results
Using the KM method, 6% of patients were alive at the end of the follow-up period (6.5 years). Mean OS from the full cohort was estimated to be 21.9 months using the PSM method (best fit Weibull curve). OS estimates for the sample size and follow-up time combinations ranged from 5.9-28.0 months. Minimum and maximum survival projections represented a 73% underestimation and 28% overestimation of survival compared with the full cohort projection, respectively. Projection accuracy was improved when t?6 months and n?200.

Conclusions
Both sample size and data maturity have a profound effect on survival projections. Care should be taken when interpreting projections in cost-effectiveness models, especially when sample size is low and follow-up time short. In addition to power calculations, clinical trial design should account for these issues. Additional analyses in other cancer types may provide further guidance for optimum trial design.