Projecting overall survival in health-economic models: uncertainty and maturity of data

Abstract

Objective As lifetime horizons are considered for economic evaluations, the Kaplan–Meier (KM) estimate is used to extrapolate survival in cases of immature overall survival (OS) data. This study estimated the error induced by the choice of distribution when extrapolating different levels of OS maturity. Methods Fifteen phase 3 trials reporting KM estimates of OS where at least 70% maturity (i.e. 70% of the population had died during follow-up) were included and compared to artificially created truncated data (30 and 50% maturity). Individual patient-data were reproduced using the Guyot algorithm based on digitized KM curves. Parametric survival distributions were fit for each arm in each study, for each maturity level, using the same time horizon (equal to the maximum follow-up). For each KM curve, the best distribution was chosen based on visual inspection, Akaike/Bayesian information criteria, and external validity. Outcomes were measured as life expectancy in months (LM) and life months gained (LMG). Results The Weibull (33%), log-logistic (32%) and log-normal (27%) were most often selected as the best fitting distribution. Compared to LM at full maturity, LM was overestimated in 23 and 40% of cases, at 30 and 50% maturity, respectively. Mean absolute error was at 30% maturity, and decreased to at 50% maturity. When comparing to mature data, the mean percentage of error in LMG was and 6 at 30 and 50% maturity, respectively. Conclusion The extent of OS maturity increases the risk of error when projecting long-term life expectancy for economic models. Even marginal gains in OS maturity result in more accurate estimations and should be considered when developing models.