The type of model used in economic evaluations of healthcare interventions is critical to ensuring the features of the health condition and treatment effects are appropriately captured whilst minimising uncertainty. For example,
Partitioned survival models
Partitioned survival models are widely used in oncology, where the outputs of clinical trials – survival curves – can be used directly to estimate the proportion of patients in each health state over time (most commonly progression free, progressed disease, and dead). This can make application of relative treatment effects (e.g. from an indirect treatment comparison) straightforward, and make decisions regarding extrapolation very transparent. A drawback of this approach is that because survival functions are estimated independently, the relationship between progression and mortality is not explicitly modelled. For example, a change in PFS extrapolation does not change the mortality within the model, which is unattractive because PFS includes progression and deaths, and more generally because progression is presumably prognostic of mortality.
Markov models
Markov models are used where it is possible to define the condition by a set of mutually exclusive and exhaustive health states, but here the transitions between each health state are explicitly modelled. Their use is therefore typically in chronic conditions defined by either levels of severity or disease progression. Extending the oncology example above, this might mean modelling the probability of progression, death pre-progression, and death post-progression. Thus, changes to the predicted probability of progression will change the number of deaths within the model, which is more attractive than in the example above.
Because such models often require transition probabilities that are not directly reported in trial publications, additional analyses, external evidence, or patient-level data may be required. If external comparators need to be considered, applying relative treatment effects from indirect treatment comparisons can become more challenging.
The Markov assumption holds that future transitions depend only on the patient’s current state and not on prior history, which can often be a strong or invalid assumption. In these situations, this assumption might be relaxed using Semi-Markov models, which incorporate ‘tunnel states’ representing time since entry into a given state, at the expense of model size and complexity.
Patient-level simulation
Patient-level simulation approaches may be required where there is significant patient heterogeneity and/or risks are a function of patient history, or the disease cannot easily be represented by mutually exclusive health states. However, their use can be limited as such approaches tend to require more data, they can be computationally intensive, and there is a perception that they are difficult for payers to review and validate.
All model structures involve trade-offs between realism, transparency, data requirements, and uncertainty. The most credible model is rarely the most complex one. It is the one that best reflects the decision problem being addressed.
Choosing the right model structure shapes every downstream conclusion in a cost-effectiveness analysis. If you’re weighing up partitioned survival, Markov, or patient-level simulation for your next evaluation, we can help, get in touch to discuss your modelling requirements.
