Brms non-linear models: Informative priors

Hi there!

So I am using brms non linear models to fit a model to some learning processes.

I have to include a lot of group effects and two predictors, and the data is quite messy.
The only way this works is by setting informative priors on nearly everything. So for example the data has an empirical standard error of roughly one, and the model does not converge anymore if I don’t set the appropriate prior of sigma to a half Student’s t-distribution with scale four and two degrees of freedom. It’s the same with almost all parameters, the scaling has to be done quite restrictive.

I have no good feeling about this as it makes it difficult to do sensitivity analyses and so I was wondering if anybody had helpful remarks or knows about ressources for bayesian non-linear models as specified in brms.

Any feedback is highly appreciated! Thanks and
all the best


1 Like

Well a couple general things come to mind.

Simulate fake data so you know what all your outcomes should be.
Run the model on the fake data with your priors.
For each prior document why it’s set to that particular value (include journal references, references to previous work, field notes, and expert opinion as needed).
Check to see if you can recover the parameters.
With the fake data model you can then play around with other justifiable priors including slightly informative.
Record all these and include in an appendices or part of your main findings.