5. agent based modelling

Agent-based modelling (ABM) is an individual-level simulation approach where the agents are given rules about their interactions with each other and their environment, which may depend on their individual characteristics (Gilbert, 2019). It differs from typical microsimulation modelling because it is possible to incorporate interactions and feedback by explicitly modelling social networks and/ or the physical (built and natural) environment (Breeze et al., 2023). These interactions enable the analysis of emergent properties and tipping points where a new behaviour becomes the social norm over time. Whilst it is possible to model proximity to place within typical microsimulation models, more complex rules about the influence of social networks and the physical environment upon behaviour can be incorporated, as well as the possibility of changes to the environment as the behaviour of individuals changes. Individuals can be given different rules according to their psychological attributes or individual characteristics, which enables the impact of influences upon people’s behaviour to vary. In turn, this allows for more nuanced understanding of the equity impact of different interventions, including potential for disaggregation of results by intersectional subgroups. It is also possible to model multiple types of agents, for example, consumers and tobacco companies. There may be interactions between the behaviours of these different types of actors which lead to unexpected outcomes. It may be feasible to validate emergent population level impacts of the interventions with data.

If feasible given the resources available, ABMs are preferable over other modelling approaches when one or more of the following holds: (i) It would be useful to explore the impact of interactions between the behaviours of different stakeholders, such as the public and the tobacco industry; (ii) The model aims to assess the cost-effectiveness of interventions about access to places affecting public health, such as green spaces or food outlets; AND there is evidence of substantial interaction between the environment and behaviour; or (iii) There is evidence that social networks will substantially affect relevant outcomes (beyond what was reported in the intervention studies), which may include impacts on health inequalities AND it is unclear whether the interventions would be cost-effective without accounting for these additional impacts.

Models should be as simple as possible to capture the key drivers of the outcomes, and hence it is important to weigh up the benefits of developing an ABM and including the influences on behaviour within the model versus the time and resources required to build such a model. In some cases, there is very little evidence about the costs and/or effects of public health interventions in changing short term outcomes even at an aggregate level (NICE Published Guidance), hence more evidence needs to be collected before more complex models could be usefully parameterised for prediction. In addition, models are only useful if they are credible to stakeholders and policy makers. It may be that increased complexity may decrease credibility, and hence it will be beneficial to co-design models with stakeholders, discussing alternative modelling options, as well as reporting modelling methods and assumptions transparently (Grimm et al., 2020).

ABMs require individual-level data about the population containing the key variables of interest and some evidence to inform the rules of the agents. One of the major advantages of ABM is its flexibility; any of the influences on behaviour outlined within Figure 2 could be incorporated in an ABM if required, using a range of data, from qualitative to quantitative. Rules for the agents could be informed by (a combination of) behavioural systems mapping, qualitative research, heuristics, behavioural theory, statistical or econometric analyses, or secondary literature.

A few attempts have been made to incorporate psychological theory within ABMs of public health interventions (Squires et al., 2023). These studies demonstrate the potential, but also the substantial time, skill and data requirements for such evaluations. Little justification is usually provided for the theory chosen, though some studies have undertaken systematic reviews (Boyd et al., 2022) and conceptual modelling with expert input to identify the most appropriate theorical basis (Garcia et al., 2017). The only behavioural theory which has been quantified within a simulation model to date that considers behaviour maintenance explicitly is the Transtheoretical Model of Change (Garcia et al., 2017; Ernecoff et al., 2016). This involves the stages: Precontemplation; Contemplation; Preparation; Action; and Maintenance, and acknowledges the possibility of relapse. It has, however, been heavily criticised in the health psychology literature because it ignores habits and situational determinants of behaviour (West, 2005). Buckley et al. (2022) used dual process theory within a simulation model of alcohol consumption, with a habitual pathway and an intentional pathway, and a recognition that new habits in terms of alcohol consumption could be formed - a key determinant of behaviour maintenance (Kwasnicka et al., 2016). Other simulation models that have quantified behavioural theory have updated the parameters at each time step but assumed the same mechanisms of behaviour maintenance as behaviour change (Squires et al., 2023). The developers of the COM-B model have considered sustained behaviour change and suggest that changes to capability, opportunity and motivation must be mutually reinforcing for behaviour to be maintained (Michie et al., 2023).

It may be possible to utilise (or adapt) existing agent-based models of public health behaviours which use empirical data, and then incorporate the costs and effects of the interventions being assessed; however, many ABMs are developed to explore a population-level phenomenon over a short timeframe, so may not be easily modifiable for the purposes of health economic modelling. Models built in modular form can be combined and reused, particularly if shared using open-source software. This would make it more feasible to build ABMs within the constraints of a policy making process. It would be possible to modify an existing individual level health economic model to incorporate the effects of social networks or the environment. A software architecture has recently been developed for mechanism-based social system modelling to incorporate behavioural and social theories within ABM (Vu et al., 2020). Changes in macro level behaviours are generally determined by the interactions between micro level behaviours and macro level behaviours; however, it is also possible to use differential equations to represent macro level (behavioural) changes if they are not explained by the included micro level variables.