Presentation (CTI)

OBJECTIVES: Parametric curves are fitted to Kaplan-Meier data informing long-term survival within cost-effectiveness analyses. Uncertainty in parametric curves is commonly captured in probabilistic analyses using covariance matrices of curve parameters varied via Cholesky decomposition. Resampling trial data then refitting curves per sample (bootstrapping), is an alternative. We compared sampling techniques and the resulting uncertainty in survival estimates.
METHODS: We applied the Cholesky decomposition and bootstrapping (frequentist and Bayesian) as sampling methods for six parametric survival curves (Weibull, log-normal, log-logistic, exponential, Gompertz, and generalised gamma). Each method was applied for the lung (n=288, 72.37% dead at day 1,022) and ovarian (n=26, 46.15% dead at day 1,227) cancer datasets in the R survival package and analysed for 1,000 probabilistic iterations. The relative spread in the mean predicted probability of survival across sampling methods (mean_rel) for each curve was compared at two timepoints: midpoint of trial follow-up (lung: n=511 days; ovarian: n=614 days) and end of trial follow-up (lung: n=1,022 days; ovarian: n=1,227 days).
RESULTS: Excluding generalised gamma, average mean_rel across curves was low: midpoint=0.27% (lung) and 1.97% (ovarian); end of trial=1.34% (lung) and 3.34% (ovarian). Mean_rel generally was higher, especially for the end of trial, for the generalised gamma curve: midpoint=1.96% (lung) and 0.94% (ovarian); end of trial=7.86% (lung) and 9.80% (ovarian). Mean_rel was consistently greater for the more immature, smaller dataset (ovarian). There were 44 (Bayesian) and 6 (frequentist) instances of failure to fit the generalised gamma curve in the ovarian dataset, whereas none in the lung dataset. We found no clear over- or under-estimation for the mean survival across sampling methods.
CONCLUSIONS: Choice of sampling method had limited impact on the parameterised uncertainty of survival estimates, except for a generalised gamma curve (a higher parameter flexible curve with more uncertainty). Replicating this study with more datasets could rule out these findings being dataset specific.