Decision analytic modelling for economic evaluation

Here, and represent the effectiveness associated with treatments A and B, and and are the costs of treatments A and B, respectively. Presenting the results of CEA and CUA analyses in this way offers significant advantages over ICERs: it provides an easier interpretation of the results of analyses where multiple treatments are compared¹⁵², offers a tractable estimate of the variability in the results¹⁵⁰ and circumvents problems associated with the interpretation of ratios¹⁵³.

3.1.5. Decision analytic modelling for economic evaluation

In relation to the analytic methods they employ, cost-effectiveness and cost-utility analyses are often distinguished between trial-based and modelling studies^154;155. The former studies are carried out to analyse evidence obtained from a single clinical trial^155;156, while the latter synthesise evidence from many sources by using analytic structures¹⁵⁷. Limitations around evidence produced by a single study as well as the recognition that all available evidence needs to be taken into account have led health economists and decision-makers to regard modelling studies as the preferable framework for synthesising evidence and informing decision-making^158-160.

Decision modelling is an integral part of decision analysis¹⁰⁷ and, apart from in health care, it has been successfully used in different disciplines including engineering, environmental risk analysis and operational research^161-163. Brennan et al.164[p.1296]

define a model-based analysis as “a formal comparison of health technologies, synthesising

makers to adopt”. In the context of health care, results of models are routinely used in treatment coverage decisions by the National Institute for Health and Clinical Excellence^160;165, which sees modelling as “an important framework for synthesising available evidence and generating relevant estimates of clinical and cost-effectiveness”^114[p.42].

Different types of decision analytic models exist, the most common of which are decision trees, Markov models and individual sampling models157;166;167

. The main stages in modelling are depicted in Figure 3.1.

Figure 3.1: Steps in decision analytic modelling

The first step in modelling requires defining the question to be answered. This typically involves specifying the clinical area, the population of interest and the treatments to be

Develop model structure

Populate model with evidence and characterise uncertainty Specify decision problem and available actions

Generate and present results

assessed. Once the decision problem and the comparators are specified, a choice is needed on the type and structure of the model. In choosing the appropriate type, pertinent considerations include, amongst other, the existence of interactions between affected individuals¹⁶⁸, the importance of time¹⁶⁴ and the availability of cohort information^164;166. In addition, the model structure should allow a valid representation of the natural clinical progression of the disease in question, one that takes into account the particularities of the decision problem, that is whether the disease is acute or chronic, whether the options under comparison are diagnostic or therapeutic and whether the risk of experiencing an event changes over time^51;157.

The next step involves populating the model with available information. This process requires identifying and synthesising the available evidence, as well as converting this evidence into a form appropriate for use in the decision model. At this stage, uncertainty arising from the methods used in the model, from the model structure itself, as well as from the use of parameter values obtained from samples needs to be characterised and accounted for, preferably by using both probabilistic and deterministic sensitivity analysis techniques^169;170. In deterministic sensitivity analysis, the impact of uncertainty is explored by recalculating the results for different plausible values of one or more uncertain parameters. On the other hand, in probabilistic sensitivity analysis, uncertain parameters are assigned probability distributions, rather than single values. Uncertainty in the results is then propagated by drawing a large number of values from each distribution in subsequent simulations—typically by using Monte Carlo methods^171;172.

The process gives a large number of cost and effect pairs together with the uncertainty associated with them¹⁶⁹.

The final stage in modelling involves presenting results numerically, in terms of ICERs and NMBs, and graphically, typically by plotting the generated cost and effect pairs on cost-effectiveness planes (CE planes)¹⁷³, cost-effectiveness acceptability curves (CEACs)¹⁷⁴ and, more recently, cost-effectiveness acceptability frontiers (CEAFs)¹⁷⁵. In brief, a CE plane plots paired estimates of incremental costs and benefits obtained from Monte Carlo simulations on a four-quadrant plane. Depending on the quadrant on which cost-effectiveness results are located, a treatment may be more effective and more costly (North East quadrant), more effective and less costly (South East quadrant), less effective and less costly (South West quadrant) or less effective and more costly (North West quadrant), as compared to an alternative treatment.

In turn, CEACs can be used to represent the probability of a treatment being cost-effective at different ceiling ratio values^174;175. For different ceiling ratios, a CEAC is produced by counting the proportion of incremental cost and effect pairs with a value less than the ceiling ratio⁵¹. Last, as there is a possibility that the intervention with the highest probability of being cost-effective may not result in the highest NMBs due to non-linearity in the model, results may also be presented as a CEAF. This is formed by a curve representing the interventions associated with the highest NMBs over a range of ceiling ratio values^175;176. A description of the methods involved in structuring and analysing decision models for the case studies used in this project are given in Chapters 4 and 5.