1.3.2: How much detail? KISS and KIDS#
An early decision in any modelling project is how much detail to include. Two opposing strategies frame the choice:
Keep It Simple, Stupid. Start with the simplest model that could plausibly address the research question, and add complexity only when the simpler model demonstrably fails to produce the behaviour you are trying to explain. The rationale is scientific discipline, not laziness: a simple model forces you to identify the essential mechanism, and a feature that can be removed without changing the conclusion was not needed. KISS suits exploratory and theoretical work, where the goal is to understand a mechanism. The strategy is associated with Axelrod’s account of simulation as a tool for theory-building [1].
Keep It Descriptive, Stupid. Proposed by Edmonds and Moss [2] as a counterpoint to KISS, KIDS argues that you should include the detail the evidence justifies from the outset, and simplify only where a simplification can be shown to be harmless. It suits applied work, where the model must match real behaviour in a real context, rich empirical data are available, or stakeholder credibility requires behavioural realism. In practice, many applied agent-based models follow KIDS.
1.3.2.1: When each strategy fits#
The two strategies are not rivals so much as starting points suited to different goals. KISS is strongest when the purpose is to isolate a mechanism and ask which ingredients are sufficient to generate it. KIDS is stronger when the purpose is to represent a particular case whose observed geography, institutions, or behavioural differences are part of the question. The evidence available matters too: descriptive detail is defensible when data or theory can support it, but becomes an unsupported assumption when they cannot [3].
The important thing is therefore to state which strategy you are following and why, and to let the research question rather than habit decide. A project may also move between them: it can begin with a simple mechanism, then add empirically justified detail as the question becomes more specific. Each of the remaining subsections makes one half of that choice concrete: first the discipline that KISS demands, then the payoff KIDS earns when richer detail is warranted.
1.3.2.2: Making KISS concrete: strip to the essentials#
The rest of this section makes the KISS discipline concrete before returning to KIDS, because KISS is easiest to feel through the act of removing features and watching what survives. The vehicle is Schelling’s segregation model, a classic agent-based model in which agents of two colours relocate whenever too few of their neighbours share their colour; the target behaviour is that the two colours separate into visible clusters. The question the activity poses is which of the model’s features are essential to that behaviour and which can be switched off without losing it.
KISS says: keep only what the question needs. The activity below is a Schelling-style segregation model with several features, only some of which the target behaviour actually requires.
Every feature you could switch off without breaking the target behaviour was a feature the model never needed. That test, removing a part and checking whether the conclusion survives, is the working definition of KISS and the surest guard against needless complexity.
1.3.2.3: How much detail is worth it: diminishing returns#
The activity above asked which features to keep. The next one asks a sharper question: as you keep adding detail, how much do you gain? The key idea it makes visible is that detail pays off unevenly: the first few features carry most of the explanation, and each extra one returns less, so accuracy flattens while the effort of building and running the model keeps climbing.
1.3.2.4: When descriptive detail earns its place (KIDS)#
The diminishing-returns curve is not an argument that detail is always wasteful; it is an argument that detail must earn its place. KIDS is the case in which it does. When the model has to reproduce a specific real system rather than isolate a general mechanism, the features that KISS would strip away, the local geography, the actual policy rules, the observed heterogeneity in how agents behave, are often exactly the ones the evidence demands and the ones a stakeholder will check the model against. A traffic model built to advise on one junction, or an epidemic model calibrated to one region’s contact data, cannot answer its question with the barest mechanism alone; the descriptive detail is what makes its output credible and actionable. In those settings the sweet spot in the curve above simply sits further to the right, because each added feature is anchored to data and so keeps buying accuracy rather than overfitting. The discipline KISS teaches, add a feature only when it earns its keep, still applies; under KIDS the evidence, not the bare mechanism, is what tells you it has.
Key takeaways#
Key takeaways
KISS: start as simple as possible and add detail only when the simpler model fails; best for theory and exploration.
KIDS: include the detail the evidence justifies and simplify only where it is safe; best for applied, data-rich, decision-support work.
State which strategy you are using, and let the research question decide.
References#
Choosing a detail strategy is only half the battle; Section 1.3.3 turns to the common pitfalls that derail models regardless of how much detail they carry.