The uncertainty associated with an economic evaluation should be appropriately characterised.
Importance to decision making
Decisions regarding resource allocation in health are unavoidable. All decisions carry a risk that a more optimal course of action could have been selected, and so in making a decision uncertainty must be acknowledged and measured.
For the chances of reaching a good decision to be optimised, the decision-maker needs to be aware of the magnitude of uncertainty in the results of an economic evaluation. All economic evaluations contain uncertainty, and it is important that all types of uncertainty are appropriately presented to the decision-maker. The types of uncertainty include uncertainty about the source of parameters used in the analysis, the precision of those parameters, and whether models or simulations of how the costs and effects of the intervention and comparators will behave are accurate. The characterisation of this uncertainty enables the decision-maker to make a judgement based not only on a likely estimate of the incremental costs and effects of an intervention, but also on the confidence that those costs and effects reflect reality.
Characterising the uncertainty will also enable the decision-maker to be informed about courses of action that could reduce this uncertainty. This could involve delaying implementation to allow for more evidence to be garnered. In this situation, appropriately characterising uncertainty will allow the decision-maker to make an informed trade-off of the value of new information, the implications of potentially delaying treatment to patients or individuals, and irrecoverable costs associated with implementing funding for an intervention.
There are a number of potential selection biases and uncertainties in any economic evaluation, and these should be identified and quantified where possible. There are three types of bias or uncertainty to consider:
Structural uncertainty – for example in relation to the categorisation of different states of health and the representation of different pathways of care. These structural assumptions should be clearly documented and the evidence and rationale to support them provided. The impact of structural uncertainty on estimates of cost effectiveness should be explored by separate analyses of a representative range of plausible scenarios.
Source of values to inform parameter estimates – the implications of different estimates of key parameters (such as estimates of relative effectiveness) must be reflected in sensitivity analyses (for example, through the inclusion of alternative sources of parameter estimates). Inputs must be fully justified, and uncertainty explored by sensitivity analyses using alternative input values.
Parameter precision – uncertainty around the mean health and cost inputs in the model. Distributions should be assigned to characterise the uncertainty associated with the (precision of) mean parameter values. Probabilistic sensitivity analysis (PSA) is preferred, as this enables the uncertainty associated with parameters to be simultaneously reflected in the results of the model. In non-linear decision-models – where there isn’t a straight-line relationship between inputs and outputs of a model (such as Markov models) – probabilistic methods provide the best estimates of mean costs and outcomes. Simple decision trees are usually linear. The mean value, distribution around the mean, and the source and rationale for the supporting evidence should be clearly described for each parameter included in the model. Evidence about the extent of correlation between individual parameters should be considered carefully. Assumptions made about the correlations should be clearly presented.
Where lack of evidence restricts reliable estimations of mean values and their distributions, unsupported assumptions or exclusion of parameters in a PSA will limit its usefulness to characterise uncertainty, and may give a false impression of the degree of uncertainty. For this reason, PSA is not explicitly required in all economic evaluations at this time; however any decision not to conduct PSA should be clearly and transparently explained in the analysis. Future iterations of the Gates-RC will provide further specification on the application of PSA.