OBJECTIVES : Partitioned survival analysis (PartSA) is a commonly used modelling technique for economic evaluations in oncology. This method typically involves parametrising and extrapolating progression free survival (PFS) and overall survival (OS) to estimate costs and effects of interventions and their comparators. In probabilistic analyses, the PFS and OS estimates are often sampled independently, which ignores the relationship that exists between these two outcomes. We illustrate a straightforward method to factor in this relationship by bootstrapping the underlying trial data.

METHODS : The trial data is repeatedly resampled with replacement, for many iterations (in this example, 10,000). Within each of the 10,000 resamples, the OS and PFS models are re-fitted and their coefficients recorded. This results in 10,000 sets of OS and PFS model coefficients which are internally consistent with each other. The OS/PFS relationship can be factored into subsequent probabilistic PartSA analyses, by either estimating a covariance matrix across the OS and PFS coefficients, or by simply drawing from a random row of the 10,000 coefficient sets (dependent modelling).

We compare the uncertainty impact of the dependent modelling approach with the conventional independent approach, by using two publicly available cancer datasets (pancreatic and liver cancer) and observing the distribution of probabilistic post-progression survival.

RESULTS : In the pancreatic data, the dependent modelling method considerably reduced the uncertainty of PPS compared to independent sampling (standard deviation of 18 vs 46 days). This is a direct result of factoring in the relationship between PFS and OS (correlation coefficient of 0.84, vs zero). A similar pattern is observed in the liver cancer data.

CONCLUSIONS : It is typical for costs to be strongly associated with PFS (due to treat-until-progression regimens), and QALYs to be strongly associated with OS. Therefore, the relationship between PFS and OS (and hence PPS) is a crucial to the validity of cost-effectiveness results.