Agent Based Model: Essential Tool or Unnecessary Complexity?
Rhoda, I think the key question is not “Can we use an Agent-Based Model?” but rather “What question can an ABM answer that a simpler method cannot answer adequately?”
For me, an ABM becomes justified when individual heterogeneity, interactions, adaptation, networks, feedback loops, and emergent behaviour are central to the decision problem. ABMs are particularly useful in public health when people or organisations behave differently and their interactions can change the trajectory of the system.
I would use a simple decision test:
1. Is the research question inherently individual-level?
If individuals differ meaningfully in risk, behaviour, access to services, treatment adherence, mobility, or decision-making, ABM may add value.
2. Do interactions between individuals matter?
For example, infectious disease transmission, social networks, health-seeking behaviour, peer influence, or household/community effects. If individuals can influence one another, an aggregate model may hide an important mechanism.
3. Are there feedback loops or emergent outcomes?
If changing one part of the system changes behaviour elsewhere, creating effects that cannot easily be predicted by simply adding individual effects, ABM becomes more attractive. Complex-systems guidance specifically highlights dynamics, feedback, non-linearity and interactions as reasons complexity may be warranted.
4. Is there a meaningful counterfactual that we cannot test empirically?
This is particularly important in public health. If we want to ask “What would happen if we changed this policy, behaviour or service configuration?” but a real-world experiment is impossible or unethical, simulation can help explore plausible scenarios.
But I would also apply the “complexity penalty.”
If a regression model, cohort analysis, compartmental model, microsimulation, or discrete-event simulation can answer the question adequately, I would hesitate to build an ABM simply because it is more sophisticated.
ABMs require additional assumptions, parameters, programming, calibration and validation. As complexity increases, uncertainty and the difficulty of external validation can also increase.
So my rule of thumb would be:
Use ABM when the mechanism of interaction is part of the research question—not merely because the population is large or the problem sounds complex.
For example, in TB research, I could justify ABM if the question involves interactions among patients, households, communities, health facilities, providers, diagnostic pathways and treatment behaviours, and we want to understand how these interactions collectively produce diagnostic or treatment delays.
But if my question is simply “Which factors are associated with patient delay?”, a well-designed epidemiological analysis may be much more appropriate. An ABM there could become complexity for complexity’s sake.
Ultimately, I would ask three questions before building one:
What decision are we trying to inform?
What mechanism must be represented?
What does ABM allow us to learn that a simpler model cannot?
If we cannot answer the third question convincingly, I probably wouldn't build the ABM.
And perhaps the most important principle is: the model should be as complex as necessary, but no more complex than the decision problem requires.
@Henry Magoba, your decision test is the clearest thing in this thread, and the fourth filter you and Charles landed on together is the piece I suspect most of us will actually use. What I want to press on is the calibration burden you mentioned. You noted that ABMs require additional assumptions, parameters, calibration and validation and that uncertainty grows with complexity. In practice, how do you decide when the calibration data you have is good enough to justify the build, versus when the absence of that data is itself the signal that the ABM is premature? I ask because in public health settings we often have rich individual-level data but poor network data, and I wonder whether that asymmetry should be part of the decision test.
I’d say the key question is whether individual behaviour and interactions actually drive the outcome you’re trying to understand.
If the problem involves people making different decisions, influencing each other, adapting to their environment, or producing unexpected “bottom-up” effects, then an Agent-Based Model (ABM) can be really useful.
But if a simpler statistical model, system dynamics model, or straightforward simulation can answer the question, then ABM may just add unnecessary complexity.
A simple rule I’d use is: don’t use ABM because the problem is complex; use it when the complexity comes from interactions between individual agents.
@Desmond Angira, your rule complexity from interactions, not from scale is the one I have already written down. What I want to add is a caution about the opposite failure mode: sometimes we avoid ABM not because the interactions do not matter but because we have never modelled interactions before and it feels like a leap. The rule protects against over-engineering, but it can also become a reason to stay with methods we are comfortable with. How do you tell the difference between "a simpler model is adequate" and "a simpler model is familiar"? That distinction feels important for a community like this one, where the tools we reach for are partly shaped by what we already know how to use.
@Charles Bawate, your Purpose Rwanda example is the most concrete thing in this thread, and the honesty of "not yet, but possibly soon" is what makes it useful. The piece I want to pull on is your point about recovery spreading through networks beyond the formal mentorship relationship, siblings, friends, people never enrolled. That is a spillover question, and spillovers are notoriously hard to estimate empirically because the control group is often contaminated by exactly the mechanism you want to measure. So the ABM is not just adding complexity; it may be the only way to get at a quantity the RCT structurally cannot identify. Does that change how you think about the sequencing or does the spillover question become more urgent once you realize the trial itself may be affected by it?
Charles, thank you for engaging with the framework so thoughtfully. I think you have identified an important extension to the original decision test: ABM should ultimately be justified not only by whether it can explain a mechanism, but by whether understanding that mechanism can change a decision, intervention, or policy.
Your Purpose Rwanda example illustrates this distinction very well. The current RCT can answer the effectiveness question without introducing unnecessary modelling complexity. However, if the subsequent question becomes how recovery behaviours diffuse through formal and informal peer networks, then the network itself becomes part of the causal mechanism. Heterogeneity in receptivity, network structure, peer influence, spillovers and feedback could generate emergent patterns that conventional aggregate approaches may struggle to represent adequately.
I particularly like your distinction between “a real pending decision” and “mechanistic curiosity.” I would probably make that an explicit fourth filter:
1. What decision are we informing?
2. What mechanism must be represented?
3. What does ABM reveal that a simpler approach cannot?
4. Will that additional knowledge change what someone actually does?
That final question protects us from what I would call methodological over-engineering. A model can be scientifically fascinating and technically sophisticated while still having limited decision value.
There is also an interesting possibility for your setting: the RCT could establish the intervention effect, while a subsequent network-informed ABM could explore spillovers, indirect effects, reach, diffusion and alternative network configurations that the trial itself cannot feasibly test. In that situation, the ABM is not competing with the RCT; it is answering a different, complementary question.
So perhaps the strongest principle is not simply “use ABM when complexity comes from interactions,” but:
“Use ABM when interaction-driven complexity is both substantively necessary to answer the question and consequential for a decision.”
That, to me, is when the complexity is genuinely earned.
Dr. Magoba, Desmond, and this whole exchange
deserve a closing synthesis, since together you've built something sharper than
any single post.
Desmond's rule — "complexity from
interactions, not complexity from scale" — is the cleanest single-sentence
filter here. It cuts through the instinct to reach for ABM just because a
problem sounds big.
Dr. Magoba's original three-part test gave that
rule structure: is the question individual-level, do interactions matter, are
there feedback loops or a real counterfactual. But his response to my Purpose
Rwanda example added the piece that actually made the framework complete — the
fourth filter: does the mechanism, once understood, change what someone does?
That's the difference between science and decision-relevant science, and it's
the filter I'll carry forward into our own research planning.
His final reframe is the one I'd want stamped at
the top of this thread for anyone finding it later: "Use ABM when
interaction-driven complexity is both substantively necessary to answer the
question and consequential for a decision." Necessary and
consequential — not one or the other.
Where this leaves Purpose Rwanda concretely: the
RCT question stays simple by design. The network-diffusion question stays open,
genuinely interaction-driven, but still short one thing — an actual pending
decision, not just mechanistic curiosity. Dr. Magoba's complementary-models
point (RCT for effect, ABM for spillover, once earned) is the right sequencing,
and it's now the plan.
@Charles Bawate, the synthesis is generous and I think it is right. The one thing I would add, as the person who asked the original question, is that the thread itself answered it in a way I did not expect. I asked whether ABM is an essential tool or unnecessary complexity, and the answer that emerged is that the question is malformed, it is not a property of the tool, it is a property of the match between tool, question, and decision. That is a more useful answer than either "essential" or "unnecessary," and it came from the three of you pushing on each other's framing rather than from anyone defending a position. So thank you for that. I would like to keep the Purpose Rwanda case visible as it develops. If the network question acquires a decision, this thread should be the place where that gets recorded.
I would not choose an Agent-Based Model simply because the system is complex. I would ask whether the micro-level decision processes and interactions are themselves essential to explaining the macro-level outcome.
For me, ABM becomes justified when three conditions are present:
-
Heterogeneity matters : agents differ meaningfully in characteristics, information, preferences, constraints, or behavioural rules, and aggregating them would remove important information.
-
Interactions matter: outcomes depend on who interacts with whom, how those interactions change behaviour, and how those changes propagate through the system.
-
Emergence matters : the phenomenon we are trying to understand cannot be adequately represented by directly specifying the aggregate relationship.
If those conditions are weak, I would prefer a simpler statistical, econometric, system-dynamics, or machine-learning model. There is little value in introducing agents, interaction networks, behavioural rules, and simulation uncertainty if they do not materially improve the research question or the validity of the inference.
I also think there should be a complexity budget. Every additional layer, adaptive behaviour, RL, LLM-based agents, network structure, endogenous learning, introduces assumptions and additional sources of uncertainty. The question should be:
What does this layer allow us to explain or identify that a simpler model cannot?
A practical test I use is to build the simplest credible model first and then add complexity incrementally. If adding agent heterogeneity or interaction mechanisms produces substantively different and empirically better-supported insights, the complexity is justified. If the conclusions are essentially unchanged, the additional machinery may be unnecessary.
So I don't see ABM as being "better" than simpler approaches. The right ABM is one where the agents are part of the explanation, not merely part of the implementation.
@Rhoda Nakhosi Deciding when to use an Agent-Based Model (ABM) versus a simpler approach comes down to whether system-wide outcomes emerge unpredictably from individual-level interactions. ABM adds genuine value when dealing with heterogeneous actors, localized network dynamics, spatial constraints, and adaptive behaviors such as predicting how individual commute habits trigger office tipping points or how word-of-mouth creates viral adoption curves. However, ABM introduces unnecessary complexity when individuals behave uniformly, aggregate averages reliably capture the system's dynamics, or fine-grained behavioral data is unavailable, turning the simulation into a black box of arbitrary assumptions. Ultimately, an ABM is worth its computational and modeling overhead only when ignoring individual interaction rules causes traditional statistical or equation-based models to fail; if an aggregate approach or standard optimization formula answers the core business question with acceptable accuracy, the simpler tool should always be preferred.
@Rhoda Nakhosi The provided comment makes two primary arguments about model selection in complex systems research. First, it warns that when individual-level data is sparse, complex tools like Agent-Based Modeling (ABM) suffer from "false precision," cloaking arbitrary assumptions in an air of sophistication that discourages critical scrutiny. Second, it cautions against defaulting to simpler aggregate models purely out of familiarity or ease; while simpler models can predict high-level outcomes (such as whether a recovery happens), they are often structurally blind to underlying mechanisms (such as how it spreads) that carry real decision value. Ultimately, the author advocates for preferring simpler tools only if researchers are transparent about the specific, structurally vital questions they sacrifice by ignoring finer-grained dynamics.