STATISTICAL CONSIDERATIONS IN ESTIMATING SURVIVAL FOR ECONOMIC EVALUATIONS IN ONCOLOGY

OBJECTIVES: Economic evaluations in oncology require estimating survival benefits which is used to obtain quality adjusted life years (QALYs) and incremental cost-effectiveness ratios (ICERs).  However, few guidelines exist on how survival data should be analyzed and extrapolated to obtain full survival benefit for economic evaluation. A recent NICE Decision Support Unit document details an algorithm for selecting survival models for economic evaluations alongside clinical trials.  We use this algorithm and other published literature to demonstrate how different models lead to varying survival estimates and how survival data can be systematically assessed in a patient registry using patient-level data. METHODS: Data from the National Cancer Institute's Surveillance Epidemiology and End Results (SEER) were used.  Surgical treatment for prostate cancer was used to illustrate the methods, but the approach is transferrable to other cancers and treatment strategies.  Patients diagnosed with prostate cancer  (PC) between 1991 and 2001 were included, the sample was limited to stage IV PC patients.  Survival between surgery and non-surgery group was estimated via Kaplan Meier, parametric and semi-parametric methods.  Several model fit criteria’s such as visual inspection, log-cumulative hazard plots, Cox-Snell residuals, Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC) along with proportionality assumption tests were used to select appropriate method and distribution.  Observed and extrapolated mean estimates were calculated and compared.  RESULTS: Analysis indicated that survival time and benefit differed based on the model selected.  Our case example demonstrated the best fit was with Weibull and exponential distributions – however, consideration must also be given to the tail in any extrapolation of the parametric distributions selected.  CONCLUSIONS: Systematic analysis of survival data is an important evaluation criterion by health technology assessments.  Selection of survival models must be justified using appropriate steps as different models can yield varying estimates, and improper selection can translate to incorrect estimation of QALYs and the resulting ICERs.