PARTITIONED SURVIVAL ANALYSIS: A CRITICAL REVIEW OF THE APPROACH AND APPLICATION TO DECISION MODELLING IN HEALTH CARE

Purpose: Cost-effectiveness models increasingly use a modelling approach called partitioned survival analysis. The appropriateness of this approach has been subject to little critique. The purpose of this work is to describe and critique the partitioned survival analysis approach as a decision modelling tool and to provide formal recommendations to assist different stakeholders in determining it’s appropriateness as a modelling approach and basis for informing policy decisions. 

Method(s): We reviewed the 30 most recent technology appraisals of cancer treatments published by NICE (covering the period May 2013-February 2016). Manufacturer submissions and reports by academic Evidence Review Groups/Assessment Groups were reviewed to establish the way and contexts in which the method is being used, its key assumptions and the prevalence and relative merits of alternative methods.

Result(s): The use of partitioned survival analysis was highly prevalent and used in 23 out of 30 appraisals. The method was generally incorrectly described and its assumptions rarely documented or justified. Critical appraisal of the method identified a central assumption of independence of clinical events. This assumption has implications for the reliability of extrapolations and in particular extrapolation of overall survival, modelling of decision uncertainty and transparency with respect to mechanisms underpinning model predictions. The most common alternative approach was the (semi-) Markov model which was typically used when overall survival data were immature and/or to reflect subsequent lines of therapy. The (semi-) Markov models frequently included an assumption that extensions to time to intermediate endpoints (e.g. progression-free survival) translated directly to extensions to overall survival. These assumptions are not necessitated by the (semi-) Markov modelling approach and should be subject to scrutiny. Appropriately designed (semi-) Markov models may offer advantages over partitioned survival approaches by allowing information on intermediate endpoints to inform survival predictions, providing greater transparency on the mechanisms underpinning these predictions and by accurately reflecting endpoint correlations. They may, however, struggle to incorporate external data such as treatment effects from indirect comparisons or long-term observational data on mortality rates.

Conclusion(s): More consideration should be given to the selection and justification of alternative modelling approaches. Choice of modelling approach is likely to be most important for the extrapolation period. Assumptions underpinning extrapolations should be subject to greater scrutiny and where possible informed by empirical data.