PRE-TRIAL MODELING FOR COST-EFFECTIVENESS: ISSUES OF DESIGN AND POPULATION. AN EXAMPLE IN HEART FAILURE

   Purpose: To investigate the challenges in designing a pre-trial cost-effectiveness model to be populated with clinical trial data. The issues are examined through use of the clinical example of statins for heart failure.
   Methods: Decision analytic models are employed to assess the longer-term impact (in terms of quality adjusted life years, costs and cost-effectiveness) of interventions being compared within an analysis. A fully Bayesian decision theoretic approach involves constructing these models early in the lifetime of a technology (when trial evidence is limited) and then updating them following the collection of additional data and re-analysing to estimate cost-effectiveness, direct future research and data collection through the continuous process of evaluation. By definition, these early, pre-trial, models involve issues with model design and population. We constructed a markov model to examine the long term cost-effectiveness of statins in heart failure, and examined the use of a large epidemiological dataset for designing and populating the model.
   Results: Individuals with a 1^st hospitalisation for heart failure (1993-2003) were identified from the linked Scottish Morbidity Record, which records all hospitalisations and subsequent deaths in Scotland. This provided data on subsequent major cardiovascular events (myocardial infarction, stroke, angina and other heart failure events) as well as deaths from cardiovascular and non-cardiovascular causes, which was used to estimate a series of time to event equations. Of 39307 identified individuals, 19% had a primary fatal cardiovascular event, and 59% had a primary non fatal cardiovascular event. Of these, 6233 went on to have a subsequent non fatal event, and 13662 a subsequent CV death. A total of 10326 people died from a non CV cause. Data from the Scottish Health Survey and on the unit costs of hospitalisation were used to predict the utilities and costs associated with each event. AIC/BIC statistics and cox snell residuals were examined to identify the best distributional fit for the survival analysis for prediction and extrapolation. 
  Conclusions: Large epidemiological datasets are helpful tools for developing pre-trial models and extrapolating within-trial results, although they can be limited in terms of the data they contain.  In addition, careful consideration needs to be given to the functional form of the estimating equations, the distribution of the survival analysis, and the ability of survival models to predict beyond the data.