4 Know why you’re building a model
Being clear about the purpose of your model is what lets you narrow down the type of model to build. A few of the common reasons:
Causal inference. You want to know the causes or drivers of something in nature: what makes this happen? These questions usually start from data, and you’re effectively testing a hypothesis about cause.
Prediction. You want to predict: the future (forecasting), places you haven’t observed (often interpolation), or the past (hindcasting). Predictions are also usually built from data, sometimes just by learning patterns well enough to predict to new times and places.
A mix of both. A mechanistic model that tells you about drivers, which you then use to make predictions.
Consequence modelling: testing the logic of your assumptions. You gather a set of assumptions (things people might broadly agree on about a system), put them together, and see whether the conclusions are what you’d have expected. Where there are nonlinearities or Bayesian thinking is at play, our human intuition is often mistaken.
A classic example of consequence modelling is MacArthur in the 1950s arguing that more diverse ecosystems should be more stable (MacArthur 1955), and then Robert May building a model, running the simulations, and showing the opposite: as you add more strongly-interacting species, the system becomes more likely to collapse, not less (May 1972). May didn’t prove MacArthur wrong so much as show that “more diverse” is not on its own a sufficient condition for stability; you need further conditions such as nested interactions and many weak interactions. That single result inspired fifty years of work that is ongoing. This is one of the most powerful things modelling can do: take a set of assumptions and show they don’t logically produce the outcome everyone expected.
An interesting feature of consequence modelling is there is often no direct use of data whatsoever. The model is built purely on our assumptions and reasoning from past studies.
Scenario modelling. Closely related, but now you’re asking what your assumptions imply for the future. This is especially useful for prediction (particularly when it’s empirically well-founded), and also for working out which uncertainties matter: there’s a lot we don’t know about nature, and running scenarios shows whether knowing a given thing would actually change our conclusions.
Guiding empirical work. Modelling is a useful tool to guide field scientists. By seeing whether a prediction is sensitive to a parameter or a structural choice, a model can tell empiricists what’s worth going out and measuring, and, just as usefully, what isn’t.
Decision-making. In decision science you care not about the outcome for its own sake but about the decision it implies. A key and slightly counter-intuitive point: knowing an outcome better often doesn’t change what you’d decide. So instead of asking how sensitive the outcome is to a parameter, you ask how sensitive the decision is. Frequently the answer is “not at all: we already know what to do,” and that itself is worth knowing.
One of my favourite examples is the question of when we need to monitor threatened animal populations. Ecologists often say “we need more monitoring”, because they are coming from a viewpoint of trying to narrow down biological uncertainties. But the decisions we might take, like protecting the species from humans, are often the same regardless of how much more we know about the population (McDonald-Madden et al. 2010). In this case the modelling tells us we don’t need to spend more of our conservation budget on monitoring; we should just throw it all into protecting the species.
4.1 A simple map
It helps to picture these purposes as a space rather than a list. One useful sketch is a set of quadrants: roughly, an axis from causal inference (understanding causes) to prediction (forecasting outcomes), crossed with an axis from data-driven to assumption-driven. Any given piece of modelling sits somewhere in that space rather than in one box:
| Data-driven | Assumption-driven | |
|---|---|---|
| Causal inference (why?) | Fitting mechanistic models to data | Consequence / strategic modelling (May) |
| Prediction (what next?) | Statistical forecasting, interpolation | Scenario modelling, “what-if” exploration |
Most real projects drift across these cells. Knowing which corner you’re aiming for helps you make modelling decisions appropriate to the goal.