OBJECTIVES : Different modelling approaches are used inconsistently to estimate health state sojourn time and resulting quality-adjusted life years (QALY) gains within health technology appraisals of cancer therapies, despite models following the same three state structure. This study assesses the performance of methods to estimate health state sojourn time/QALYs when developing a model based on information collected in a randomised control trial (RCT).

METHODS : Common methods used in health economics (partitioned-survival model[PSM]; state-transition model [STM]; multi-state) alongside additional methods to jointly model progression-free survival (PFS) and overall survival (OS) are applied to a series of simulated datasets in which a range of trial data have been generated covering a wide range of possible scenarios (54 scenarios), to test the appropriateness and performance of each method subject to the nature of the data available.

RESULTS : Depending on the scenario examined, the PSM was typically less biased compared with the STM when estimating LY, but less precise. However, in some scenarios, the STM was less biased compared with the PSM. Larger biases were observed for the STM when the trial data included dependence between the time to progression (TTP) and the time to death following progression (post-progression survival [PPS]). This is because not all randomised patients are included in the dataset used to estimate the transition from progression to death, but only the subset of patient who progressed. Making PPS conditional on TTP did not always improve the performance of the STM. Jointly modelling PFS and OS using a copula model tended to slightly improve the performance compared with separately fitted models.

CONCLUSIONS : When developing a model based on information collected in an RCT, the choice of analytic approach needs to be informed by the characteristics of the data available. Whilst PSMs do not model the underlying process, the STM results may be affected by selection bias.