3A.7 Model validation

Page last updated: September 2016

Validation of an economic model to demonstrate that the generated results represent what they are intended to represent is best practice. It helps to reduce some of the uncertainty associated with modelling, and a more thoroughly validated model allows more confidence in its predictions.

3A.7.1 Operational validation of the economic model

Model traces for the proposed medicine and its comparator provide a clear depiction of the implications of the model. They can inform the face validity of the model logic, computerisation and external validity.

Use traces to track patients through the model and demonstrate that the logic of the model is correct. Present traces representing the proportions of the cohorts in each health state over time, and the cumulative sum of the undiscounted costs and outcomes (eg QALYs) over time. If applicable, state the number of events over time where patient-relevant events occur within a health state. Comment on whether each of the model traces is logical – for example, ensure that any traces of overall survival converge to zero at or before the time horizon of the model (see Subsections 3A.2 and 3A.4).

Compare model traces with corresponding empirical data, where possible, to identify whether outcomes are consistent. Consider both data sources used in the model (dependent validation) and data sources not used in the model (independent validation). For example, compare predicted clinical events with observed data on the natural history of the medical condition. Comment on and explain any differences indicated by these comparisons.

In addition, compare modelled outcomes against outcomes from similar models as a cross-validation tool to identify consistencies (or differences that can be explained).

3A.7.2 Other validation techniques

Present or cross-reference any other completed model validation exercises. The Assessment of the Validation Status of Health-Economic Decision Models (AdViSHE) Study Group describe a range of validation processes, and these should be considered.⁴⁸