Whose Behaviour Gets Modelled? ABMs, Data Gaps, and the Risk of Simulating the Easy-to-Reach

@Rhoda Nakhosi This is a critical reflection that touches the heart of modeling ethics and spatial/social justice. You’ve articulated a problem that too often gets buried in the "limitations" section of academic papers rather than driving upfront model design.

Here are perspectives and experiences addressing your three core questions:

1. Data-Driven Exclusion: Handling Missing or Missing-by-Design Populations

Leaving out marginalized or hard-to-reach populations isn't just a data quality issue—it introduces structural bias into model dynamics. When data doesn't exist, modelers typically rely on three primary strategies, each with distinct trade-offs:

• Synthetic Population Generation & Constraint Matching: Rather than relying purely on empirical survey samples, synthetic micro-data can be generated using marginal distributions (e.g., census aggregates) and spatial micro-simulation. Where individual-level data is thin, Bayesian updating or iterative proportional fitting (IPF) can build plausible agent distributions while explicitly acknowledging missing attributes.

• Bounding via Scenario Ensembles (Worst-Case / Best-Case Parameters): When exact behavioral parameters for a group are unknown, instead of assigning them "average" behavior, researchers simulate the full parameter space. Running sensitivity analyses across extreme behavioral bounds reveals whether the missing population's behavior creates tipping points or systemic shifts in the outcomes.

• Participatory Modeling & Qualitative Imputation: Grounding agent rules in qualitative research (focus groups, key informant interviews, expert-elicited heuristics) allows models to represent behaviors that standard quantitative surveys miss (e.g., trust-based informal networks, avoidance of formal services).

How to flag the gap: Transparent documentation via frameworks like ODD+D (Overview, Design Concepts, and Details + Human Decision Making) is essential. Modelers should explicitly document data heritage for every agent archetype, designating agents as Empirically Anchored, Qualitatively Informed, or Stylized/Assumed.

This is a critical issue in agent-based modelling (ABM), particularly in public health research and low- and middle-income countries (LMICs).

When populations are missing from the data, they do not cease to exist; they become invisible in the model. This is especially concerning for marginalized communities, people facing barriers to healthcare, and individuals whose behaviours are poorly captured by routine surveillance systems.

From an epidemiological perspective, I believe three principles are essential:

1. Make exclusion visible. Where data are unavailable, researchers should explicitly document which populations are underrepresented, distinguish observed behaviours from assumptions, and use sensitivity analyses or alternative scenarios to assess how missing populations could affect the findings.

2. Challenge the assumption of universal behaviour. The average behaviour observed among accessible participants may not reflect the behaviour of the wider population. Differences in socioeconomic status, healthcare access, geography, stigma, and social networks can fundamentally influence individual decisions and interactions.

3. Make equity part of validation. An ABM should be assessed not only on its overall predictive accuracy but also on how well it captures outcomes among populations most likely to experience the consequences of policy decisions. A model can perform well on average while systematically misrepresenting vulnerable groups.

Importantly, limited data should not automatically prevent modelling. It should, however, change how confidently we interpret the results. Participatory research, qualitative evidence, transparent assumptions, uncertainty analysis, and targeted data collection can help represent populations for whom granular data are scarce.

My central argument is that data availability should not determine whose experiences count. ABMs should make uncertainty and structural inequality visible rather than inadvertently reproducing them. Equity must be a deliberate component of model design, calibration, validation, and interpretation—not an afterthought.

Thank you Rhoda Nakhosi for raising this. Very timely question.