Introduction
-
long-term health impacts (e.g., the effects on cardiovascular disease burden and health expenditure of increasing the price of unhealthy foods may manifest and persist for many years after the intervention is introduced);
-
wider societal costs and consequences (e.g., transport policies to improve physical activity have social costs and outcomes beyond health; and different social costs and outcomes are relevant to different modeled scenarios, for example, influencing alcohol use has an important effect on crime whereas reducing stroke incidence has a more important effect on social care);
-
impact on inequalities (e.g., more deprived population groups may respond differently to those who are less deprived following a price increase on unhealthy food);
-
multicomponent interventions (e.g., a policy aimed at increasing physical activity may include both additional bicycle lanes and subsidized gym membership); and
-
interactions within complex non-health sector systems (e.g., if simulating the effect of upgrading home insulation to reduce winter deaths, it may be important to consider interactions with the housing sector, energy sector, and social care).
A taxonomy of epidemiological modeling structures for the economic evaluation of public health interventions
A | B | C | D | |||
---|---|---|---|---|---|---|
Cohort/aggregate-level/counts | Individual-level | |||||
Expected value, continuous state, deterministic | Markovian, discrete state, stochastic | Markovian, discrete state | Non-Markovian, discrete state | |||
1 | No interaction | Untimed | Decision tree rollback or comparative risk assessment | Simulation decision tree or comparative risk assessment | Individual sampling model: Simulated patient-level decision tree or comparative risk assessment | |
2 | Timed | Markov model (deterministic) | Simulation Markov model | Individual sampling model: Simulated patient-level Markov model | ||
3 | Interaction between entity and environment | Discrete time | System dynamics (finite difference equations) | Discrete time Markov chain model | Discrete-time individual event history model | Discrete-time discrete event simulation |
4 | Continuous time | Systems dynamics (ordinary differential equations) | Continuous time Markov chain model | Continuous time individual event history model | Continuous-time discrete event simulation | |
5 | Interaction between heterogeneous entities/spatial aspects important | x | x | x | Agent-based simulation |
Corresponding section of review and table 1 | Modeling method | Advantages | Disadvantages | Public health examples |
---|---|---|---|---|
Section: Decision trees | Decision tree | Can be easy to construct. | No explicit time component. | Comparing exercise referral schemes with usual care to increase physical activity [29]. |
Relatively easy to interpret. | Exponentially more complex with additional disease states. | |||
Table 1: A1, B1, C1, D1 | Can be adapted for cohorts and individuals. | No looping/recurring. | ||
Poorly suited to complex scenarios. | ||||
Section: Comparative riskassessment | Comparative risk assessment | Can model multiple diseases and risk factors simultaneously. | More complex to build than decision trees. | Return on investment of workplace interventions to improve physical activity [32]. |
Can be used for individuals or cohorts. | No explicit time component. | |||
No looping/recurring. | ||||
Table 1: A1, B1, C1, D1 | Unable to model interactions between individuals, populations, or their environment. | |||
Section: Markov models without interaction | Markov models without interaction | Relatively straightforward to construct and to communicate. | The Markovian assumption-individuals have no memory of (are independent of) previous disease states. | |
Can model populations or individuals. | ||||
Table 1: A2, B2, C2, D2 | Has time component. | Can only exist in one disease state. | ||
Allows looping/recurring. | Exponential increase in complexity with increasing numbers of disease states. | |||
Section: System dynamics models | System dynamics models | Allows for interactions between populations and the environment. | Models populations rather than individuals. | Modeling the effects of policies aimed at increasing bicycle commuting rather than travelling by car [63]. |
Table 1: A3, A4 | Allows for feedback and recurring. | |||
Section: Markov chain models and individual-level Markov models with interaction | Markov chain models and Markov individual event history models | Can model individuals or populations. | Markovian assumption still exists (although its impact can be reduced-see main text). | |
Table 1: B3, B4, C3, C4 | Allows for interaction between populations or individuals within the model. | Becomes rapidly more complex with added disease states. | ||
Section: Discrete event simulation | Discrete event simulation | Allows for interaction between individuals and between individuals, populations, and their environment, governed by system rules. | Model structure can be difficult to communicate and interpret. | Evaluating the cost-effectiveness of screening programs [67]. |
Table 1: D3, D4 | Computationally challenging both in terms of designing the model and running it. | |||
Allows for modeling of complex scenarios. | ||||
Section: Agent-based simulation | Agent-based simulation | Allow for interactions within and between individuals, populations, and the environment, governed by rules applied to individuals. | More complex than discrete event simulation. | The Archimedes model for modeling the outcomes of changing health care systems, such as investigating diabetes care [70]. |
Table 1: D5 | Requires large computational power. | |||
Allows for individuals to learn. | Difficult to communicate and interpret model structure. | |||
Allows modeling of complicated systems. | ||||
Table 1: adjunct to A1, B1, C1, D1, A2, B2, C2, D2 | Multistate life tables | Can be used with comparative risk assessment and decision tree models to add a time component. | Assumes diseases are independent of each other. | |
Can be combined with Markov models to increase the numbers of possible disease states without exponentially increasing model complexity. | Model limited by underlying model structure, for example, if combined with a Markov model, the Markovian assumption remains. | |||
Table 1: adjunct to C1, C2, C3, C4, D1, D2 | Microsimulation | Can be combined with decision tree, comparative risk assessment, and Markov models to make it easier to model heterogeneous populations or multiple disease states. | Data requirements and simulations can become computationally challenging with complex models. | The NICE obesity health economic model used by Trueman et al. to estimate the cost-effectiveness of primary care weight management programs [83]. |
Model limited by underlying model structure, for example, if combined with a Markov model, the Markovian assumption remains. |