Analysis and synthesis of time-to-event data from cancer trials in the assessment of cost-effectiveness

All new treatments are subjected to an analysis of both their effectiveness and cost-effectiveness before they can be recommended for use in the NHS, and data from randomised clinical trials forms the critical element in these assessments. In the case of cancer treatments these trials generate time-to-event data, which require special methods of analysis because many patients have not reached the trial ?end-point? (tumour progression, or death) at the time when the study data is analysed. It is extraordinary, but true, that the data analysis used to assess effectiveness is almost invariably different from the one used to assess cost-effectiveness, and that the two analyses cannot both be correct. This has lead to an element of arbitrariness in the cost-effectiveness calculations, which depend enormously in what assumptions are made. At NICE, Appraisals Committees are, as a result, frequently confronted with alternative sets of assumptions that lead to completely different conclusions, and this has in the past lead to an Appeal. But even if this were not the case, it is clear that methods for analysing time-to-event data for cost-effectiveness calculations need to be completely overhauled. Current methods for deriving estimates from data, and for extrapolating treatment effects over a lifetime, can be shown in many respects ? perhaps surprisingly ? to be incoherent and non-sensical. This research sets out to establish new analytic procedures and methods, which use the available data fairly and give the best predictions possible about the gain in life-expectancy and quality of life due to the new cancer treatments. The aim is to make the task of assessing these treatments more transparent and less open to arbitrary, and incorrect, assumptions.

The efficacy of new cancer treatments is assessed in randomised trials with time-to-event outcomes, and methods for survival data analysis based on Cox regression and proportional hazards are pre-specified in trial protocols, and widely accepted. However, cost-effectiveness analysis (CEA) is an increasingly important, and in some jurisdictions the decisive element in decisions about whether new treatments will be recommended for routine use. Cox regression and hazard ratios, designed for statistical inference, are wholly inadequate as a basis for CEA, which relies on estimated of expected time-to-event. Currently, investigators produce separate analyses of the trial data for the purposes of CEA. Because estimates of expected to-to-event require extrapolation beyond the data, they can be highly sensitive to model choice. In the absence of pre-specified protocols for cost-effectiveness analysis, this leaves conclusions on cost-effectiveness vulnerable to more or les arbitrary model selection. The standard modelling technique used in CEA, the Markov model, cannot correctly capture uncertainty in extrapolation, and is more often than not based on incoherent parameter estimates. The purposes of the research proposed here is to provide a new approach to time-to-event data analysis, fit for the purpose of CEA, that correctly reflects extrapolation uncertainty. We will carry out an exploratory analysis of a representative set of cancer trials, to determine how ?complex? cancer survival curves are in practice, looking at a wide range of models. These will include generalised families of parametric distributions and highly flexible spline functions. We will then conduct simulation studies based on realistically complex models, as dictated by the exploratory analysis, to identify analytic methods and procedures that best recover the expected time-to-event statistics of the generating models. We will also develop new methods for synthesis of evidence from multiple trials, comparing methods based on expected time-to-event estimates, with joint modelling of multiple trials with flexible survival techniques. The research has very wide implications for CEAs in all areas of medicine where time-to-event outcomes are used in trials designed for regulatory approval or re-imbursement decisions, particularly for the design of trial data analysis protocols, and synthesis methods for time-to-event outcomes.