Generating dispersion predictions in beta regression with brms

Interfaces brms
1

Hello,

I have a beta regression where I model both the response and the dispersion, like this:

mean.spec <- as.formula(y ~ ...)
disp.spec <- as.formula(~ ...)
fit <- brm(bf(mean.spec, phi=disp.spec), family=Beta(), ...)

The model fits fine, but when I try to see what dispersion is taken at each observation:

predict(fit, resp="phi")

I get back predictions of the response, i.e. the resp argument is ignored.

This is likely something quite trivial, but I am stumped. Thank you very much in advance.

Versions:

  • Operating System: Windows 10
  • brms: 2.9.0
2

Hi!

The correct way is to use the ‘fitted’ function, using the argument ‘dpar = phi’.

Have a good day!
Lucas

3

Do not forget that phi is a precision parameter, non linearly and negatively related to dispersion :)

4

Thank you for your reply, Lucas. Wouldn’t fitted ignore the random effects and the residual variability though? I care about both in my application.

5

The argument re_formula controls this and by default will include all group level effects. See ?brms:::fitted.brmsfit for details.

6

Thanks, Lucas and Gavin. fitted indeed works and includes the group effects, which is sufficient in my case, actually. But what should I do if I also wanted to include the residual variability? The documentation directs me to predict.

7

Residual variability applies, by definition, only to the response variable itself. Distributional parameters have no residual variability.

8

You’re right. Thanks everyone.