5.2: Empirical agent-based models#
Section 5.1 established what an empirical ABM is and why grounding a model in real-life observations matters. The natural next question is where that grounding actually happens: when we say a model is tied to evidence, which parts of it are we tying down? A model is not a single object that is either grounded or not; it is grounded piece by piece, and different pieces demand different kinds of data. This page names the five dimensions along which an ABM can be grounded, and it introduces them in a deliberate order, from the easiest to the hardest to pin down with evidence. Those five dimensions are the spatial environment, socio-economic-demographic attributes, behavioural attributes, rules of actions and interactions, and social networks. Each provides a different lens through which to understand the complex dynamics of the system being modelled, and each comes at a different cost, which is why the order in which we meet them matters. Select each dimension to explore how it is grounded in evidence:
5.2.1: The five dimensions of empirical grounding#
The dimensions distinguish where evidence enters the model. Spatial and demographic data mainly describe the setting and population; behavioural attributes and action rules describe how those agents decide; network data describe who can influence or interact with whom. Grounding one dimension does not automatically ground the others. A model can use a real map while relying on assumed behaviour, or use observed decision rules on an artificial network. Reading the tabs with that distinction in mind makes gaps in the empirical argument easier to identify.
The spatial environment is often the easiest dimension to ground empirically. It involves using real-world geographic data to create a realistic representation of the physical space in which agents operate, including land use patterns, infrastructure, and natural features. By grounding the spatial environment in empirical data, modellers can capture the influence of physical space on agent behaviour and interactions, leading to more accurate simulations of phenomena such as urban development, transportation, and environmental change.
It is easy to get data on the spatial environment, and it is often the first step in creating an empirical ABM. For example, in models of urban development, modellers can use geographic information systems (GIS) data to create a realistic representation of the city, including the location of buildings, roads, and green spaces. In models of environmental change, modellers can use satellite imagery and other remote sensing data to create a realistic representation of the landscape and its changes over time.
Note
GIS data can be found in many places, such as OpenStreetMap, government databases, and satellite imagery providers. This makes it relatively easy to ground the spatial environment in empirical data.
Another important dimension of empirical grounding is the socio-economic-demographic attributes of agents. These include characteristics such as income, age, and education, along with other factors that can influence decision-making and behaviour. For example, in models of consumer behaviour, the income distribution and educational background of agents can help to predict purchasing patterns and responses to marketing strategies. In models of labour markets, attributes such as education and work experience can be crucial for understanding employment dynamics and wage disparities. For firms, attributes such as the capital/labour ratio or the costs of production can be important for modelling competition and market dynamics. By drawing on empirical data, modellers can create more realistic simulations that reflect the diversity of agents in the real world and provide insights into how different socio-economic-demographic factors influence behaviour and outcomes.
Note
These attributes matter because empirically observed differences between agents (on the micro-level) often drive different outcomes at the system level (macro-level).
Behavioural attributes are also critical for empirical grounding. The bounded rationality account of ABM suggests that agents are rarely perfectly rational; it is therefore important to incorporate ‘soft’ values for different socio-behavioural characteristics, such as risk aversion, social preferences, and cognitive biases. By drawing on these empirically observed behavioural attributes, modellers can create more realistic simulations of human behaviour, which in turn yield better insights into phenomena such as consumer choice, social influence, and collective action.
Note
It can be difficult to obtain empirical data on behavioural attributes, as they are often context-dependent and can vary across different populations and situations. However, by combining surveys, experiments, and observational studies, modellers can gain insights into the behavioural attributes of agents and incorporate them into their models.
The rules of actions and interactions are another key dimension of empirical grounding. This involves understanding how agents actually interact with each other and the sequence of steps that lead to particular decisions or actions. Data on these interactions can be collected in various forms, such as qualitative data from interviews and case studies, or dynamic data from role-playing games and controlled laboratory experiments. This way, modellers can capture the complexity of social interactions and the emergent phenomena that arise from them. For example, in a model of firms competing for customers, the rule might specify how a firm decides to lower its price in response to a rival, or how a household chooses which shop to visit; grounding these rules means observing the actual decision steps rather than assuming them. The topology of who is connected to whom, and how influence travels along those connections, is a separate dimension in its own right, treated next.
Elsawah et al. (2015) [1] provide a comprehensive review of how rules of actions and interactions can be empirically grounded in ABMs, highlighting the importance of using data from surveys, experiments, and observational studies to inform the design of agent behaviours. They present a step-wise methodology for integrating perceptions of stakeholders (qualitative) into formal simulation models (quantitative) with the ultimate goal of improving understanding and communication about decision-making in complex socio-ecological systems.
Tip
Map out the step-by-step choices an agent makes using simple rules, interviews, or logic flowcharts.
The social network dimension is often the hardest to ground empirically. It involves understanding the relationships and interactions between agents, which can be complex and dynamic. Data on social networks can be collected through surveys, social media analysis, or ethnographic studies, but the full extent of social interactions, and the influence they exert on agent behaviour, can be difficult to capture. Despite these challenges, empirical data on social networks is crucial for realistic ABMs. In models of disease spread, for example, the social network can help to predict how a disease moves through a population and to identify key individuals or groups for intervention strategies. In models of social influence, it can reveal how behaviours or opinions propagate and can pinpoint key influencers. The social network is therefore an essential ingredient in ABMs that aim to provide valuable insights into social dynamics.
As Jager et al. (2024) [2] highlight, social innovation depends heavily on these specific network properties. Empirical ABMs must increasingly account for the reality that interaction occurs across multiple layers: professional networks, informal social circles, and digital connectivity. People within these networks are influenced by a complex web of relationships that shapes their decisions and behaviours. Capturing these links, even with partial data, is essential for understanding tipping points, because the success or failure of a social process often hinges on the precise topology of influence.
Important
Network structure is often the hardest dimension to measure, but it can be very influential for emergent dynamics.
5.2.2: Grounding is a matter of priorities#
Knowing that a model can be grounded along five dimensions is only half the picture, because grounding each one costs time, data, and effort, and those resources are always limited. The dimensions also differ in how easily they can be measured, so a realistic project has to decide which ones are worth the cost and which can be left more abstract. Making that choice well means asking which dimensions actually drive the outcomes you care about, rather than simply grounding whatever data happens to be nearest to hand. In the activity below you are given a fixed budget of effort and asked to spend it across the five dimensions. Before you begin, consider the social network in particular, which is the most expensive dimension to ground yet often the most decisive for the behaviour that emerges.
The lesson to carry away is that empirical grounding is a matter of priorities, not completeness. With a limited budget you gain the most by grounding the dimensions that matter most to your question, even when they are the hardest and most expensive to measure.
Prioritising well, however, presupposes that the effort is worth spending in the first place. Before deciding how much to ground, it helps to be clear about what grounding actually buys you, so that the budget above is spent against concrete gains rather than a vague sense that realism is good.
5.2.3: What empirical grounding buys you#
The four cards above name distinct payoffs of grounding a model in evidence. First, it makes the parameters easier to calibrate. A parameter is simply a setting you can adjust in the model, such as how strongly one agent influences another, and to calibrate means to tune those settings until the model’s output matches the real data (calibration is treated in depth in Section 5.4). Grounding also makes scenarios more credible to the people who rely on them, makes the model clearer to explain, and sharpens the theory behind it. These benefits are worth naming because grounding a model takes real effort, and being able to point to the specific gain you are after helps you decide how much of that effort is justified and how to defend it to others. Each payoff also answers a different kind of problem, so recognising which one a situation calls for is a practical skill in itself. The activity below presents four situations a modeller might face, each resolved by one of these payoffs.
The point to keep in mind is that the value of empirical grounding is not vague. It shows up as calibration, credibility, communication, and theory, and being able to name which of these you are buying is what turns the effort into a defensible choice.
The five dimensions describe what in a model can be grounded, but they say nothing about what does the grounding. Each dimension is tied to evidence by a data source or method, a census for demographic attributes, GIS layers for the spatial environment, interviews or laboratory experiments for interaction rules, and so on. Section 5.3 turns to exactly these sources and methods, so that the dimensions of this page and the data of the next read as a means-to-end pair: the five dimensions tell you what to ground, and Section 5.3 shows you what to ground each one with.
Key takeaways#
Key takeaways
Empirical ABMs are grounded in real-life observations, letting the modeller represent components for which evidence exists rather than relying on abstract assumptions.
Grounding happens along five dimensions: the spatial environment, socio-economic-demographic attributes, behavioural attributes, rules of actions and interactions, and social networks.
These dimensions differ in how easily they can be grounded: the spatial environment is usually the easiest, while social networks are often the hardest, yet the network structure can be decisive for emergent dynamics.