OBJECTIVES: For health economic modelling purposes, progression free survival (PFS) and overall survival (OS) from trials typically need to be extrapolated with parametric survival models. These extrapolations increase uncertainty over the decision problem, which may be addressed in the probabilistic sensitivity analysis (PSA). When extrapolations are varied, covariance between coefficients within one parametric model is taken into account by the Cholesky decomposition. However, usually correlation between OS and PFS is not considered, even though correlation is likely, e.g., death is captured by both OS and PFS. This study assesses the impact of including PFS and OS correlations by fitting parametric models on bootstrapped PFS and OS trial data for each PSA iteration.

METHODS: A three-state (PFS, progression and death) partition survival model was constructed. PFS and OS individual patient-level data (IPD) was taken from a publicly available oncology dataset. The PSA with 10,000 iterations was run for two scenarios; 1) Cholesky decomposition was applied separately on the PFS and OS parametric model coefficients; 2) PFS and OS IPD were bootstrapped 10,000 times simultaneously; over each bootstrap PFS and OS parametric models were fitted, and the corresponding coefficients were applied in the PSA. Uncertainty over the incremental mean costs and QALYs was compared.

RESULTS: Incremental costs and QALYs are 79,518 [95% percentiles: -8,341 – 231,752] and 1.62 [95% percentiles: 0.76 – 2.62] for scenario 1 and 81,354 [95% percentiles: 14,224 – 210,126] and 1.63 [95% percentiles: 0.69 – 2.92] for scenario 2, respectively. The probabilities of being cost-effective at thresholds of 30,000, 50,000 and 80,000 are 25%, 50%, and 78% for scenario 1 and 16%, 50%, and 88% for scenario 2.

CONCLUSIONS: We show that fitting parametric models on bootstrapped PFS and OS may impact the initial decision uncertainty and thus suggest its use in future assessments.