Interventions that impact upon survival form a high proportion of the treatments appraised by NICE, and in these it is essential to accurately estimate the survival benefit associated with the new intervention. This is made difficult because survival data is often censored, meaning that extrapolation techniques must be used to obtain estimates of the full survival benefit. Where such analyses are not completed estimates of the survival benefit will be restricted to that observed directly in the relevant clinical trial(s) and this is likely to represent an underestimate of the true survival gain. This leads to underestimates of the Quality Adjusted Life Years gained, and therefore results in inaccurate estimates of cost-effectiveness.
There are a number of methods available for performing extrapolation. Exponential, Weibull, Gompertz, log-logistic or log normal parametric models can be used, as well as more complex and flexible models. The different methods have varying functional forms and are likely to result in different survival estimates, with the differences potentially large – particularly when a substantial amount of extrapolation is required. It is therefore very important to justify the particular extrapolation approach chosen, to demonstrate that extrapolation has been undertaken appropriately and so that decision makers can be confident in the results of the associated economic analysis. Statistical tests can be used to compare alternative models and their relative fit to the observed trial data. This is important, particularly when there is only a small amount of censoring in the dataset and thus the extrapolation required is minimal. However it is of even greater importance to justify the plausibility of the extrapolated portion of the survival model chosen, as this is likely to have a very large influence on the estimated mean survival. This is difficult, but may be achieved through the use of external data sources, biological plausbility, or clinical expert opinion.
A review of the survival analyses included in NICE Technology Appraisals (TAs) of metastatic and/or advanced cancer interventions demonstrates that a wide range of methods have been used. This is to be expected, because different methods will be appropriate in different circumstances and contexts. However the review also clearly demonstrates that in the vast majority of TAs a systematic approach to survival analysis has not been taken, and the extent to which chosen methods have been justified differs markedly between TAs and is usually sub-optimal.
In the form of a Survival Model Selection Process algorithm we provide recommendations for how survival analysis can be undertaken more systematically. This involves fitting and testing a range of survival models and comparing these based upon internal validity (how well they fit to the observed trial data) and external validity (how plausible their extrapolated portions are). Following this process should improve the likelihood that appropriate survival models are chosen, leading to more robust economic evaluations.
Keywords: Oncology; survival analysis; overall survival; extrapolation; modeling; cost-utility analysis; technology assessment; resource allocation; pharmacoeconomics; prediction; statistical methods.
Copyright © 2013 National Institute for Health and Clinical Excellence, unless otherwise stated. All rights reserved.