ABM FORUM

@Henry Magoba Here is my response to this questions.

Dr. Magoba, useful list — I want to sit with just #7, since I think it's the
gatekeeping question the other six depend on.

The honest test I've found useful: ABM earns its complexity only when (a)
interaction between agents is part of the actual research question, not
incidental to it, (b) individual heterogeneity and network structure plausibly
change the outcome, and (c) someone will act differently based on the answer.
If all three aren't true, a compartmental or regression model usually answers
the question more cheaply and more defensibly.

Applying that to Purpose Rwanda's own peer-led recovery work: our core
evaluation question — does the PAT model produce sustained recovery — doesn't
need ABM. A simple outcomes comparison across arms answers it directly. But an
adjacent question keeps surfacing that does pass all three criteria: how does
recovery behaviour itself diffuse through an informal peer network once it
begins, given that mentees often go on to influence siblings and peers never
formally enrolled in the programme at all? That's a genuine interaction effect
with emergent, second-order outcomes a pre-post comparison would miss entirely
— closer to the TB transmission dynamics in your question #1 than to a standard
treatment-effect question.

Where I'd add to your framing: the complexity penalty of ABM isn't just
computational cost or data scarcity, as Nonvicks noted for #4. It's that ABM
answers mechanistic curiosity beautifully but only justifies its cost when a
real pending decision hangs on the mechanism — otherwise the model is
sophisticated but inert. For us, that decision isn't there yet.

So my honest answer to #7: ABM adds genuine value the moment social
diffusion, not just individual treatment response, is the thing you're trying
to predict. Dr. Magoba, curious whether TB surveillance data actually captures
that kind of informal, non-clinical transmission pathway well enough to
parameterize an ABM honestly, or whether that data gap is itself the binding
constraint on #3.

I think ABM is particularly useful for TB because transmission and treatment are influenced by so many individual and community-level factors. It can help us move beyond simply asking “what happens on average?” and explore how different behaviours, social interactions, health-seeking patterns, and health-system factors may affect outcomes.

The challenge, however, is making sure the model reflects reality. Linking ABMs with good surveillance and routine health data could improve their usefulness, but validation is critical. A model is only as helpful as the assumptions and data behind it.

I also see potential for ABM in testing different strategies for screening, referral, diagnosis, and treatment before implementing them at scale. The interesting question for me is when the added complexity of ABM gives us insights that conventional epidemiological models would miss.