Models of Learning
Developed by
Jill O'Reilly
Hanneke den Ouden
August 2015

A very brief recipe for modelling your data
To finish this section, below is a very brief overview of the steps you would want to take when trying a modeling approach to your data.
This is by no means exhaustive nor a set of hard rules, and definitely doesn't replace
you having to use your own brain, but we hope it will give you a good basis to start from.

Define your model: this is your quantitative hypothesis of how your subjects do your task.

Do a generative model check, a.k.a. Does your model do what it's supposed to do?:

Simulate data using different parameter settings and let the model play your task.

Does your model generate the behaviour you (expect to) see in your subjects?

If not, it won't be a very good model for your data!

See whether your model can recover parameters: plot the correlation of simulated and recovered parameters. These should be closely correlated.

If the estimated parameters are wildly off what you put in, you have a problem. You need to reconsider the parameterisation.

It's likely that some of the parameters are correlated

Optional (more advanced): When you generate data from your model, do other simpler models have higher model evidence? Are they a better model for explaining your data?

If this is the case, your model is likely overparameterised,i.e. you have correlations between your parameters or some of your parameters do not capture any variance

Fit your data, and see whether the evidence 'landscape' looks like it has a
nice peak in it, meaning that you can be relatively confident in your parameter estimates.

If you have multiple models (usually you will):
do model comparison to see which model is the best and most parsimonious explanation of your data.

Do inference on the parameters of your model.
Back to top page ►►►
