Error with gamma prior

Interfaces brms
1

I am trying to set a gamma prior for one of the variables in Beta regression. In fact my model has been working with no issues under R 3.6.1 with brms 2.10.5, however now after upgrading to the latest versions (R 4.0.2 with brms 2.13.0) I keep getting the following error:

Warning message:
It appears as if you have specified a lower bounded prior on a parameter that has no natural lower bound.
If this is really what you want, please specify argument ‘lb’ of ‘set_prior’ appropriately.
Warning occurred for prior b_variable ~ gamma(2,1)

This error pops up during sampling half way through (not when setting priors or model compilation).
The variable in question is strictly positive. When I try to specify “lb” argument, I get another error:

Error: Argument ‘coef’ may not be specified when using boundaries.

What am I doing wrong?

Currently the prior is specified as follows (and that worked before):

set_prior(“gamma(2,1)”,

  •        class = "b",*

  •        coef = "variable")*
2

Currently lb und ub can only be specified for a whole parameter class. I realise this is unnecessarily restricting and may change that in the future. Feel free to open an issue on github about it.

3

Is there a way to fix the issue without specifying “lb”, but keeping gamma prior?

4

Ignore the warning and hope you don’t get too many divergent transitions because of the missing boundary.

5

Is this still true with brms or are we able to specify bounds for specific coefficients instead of an entire class?

6

This appears to be still the case. As a workaround, you can modify the Stan code within the fit object itself and re-run it to obtain the desired results without having to stray too far from a brms-based workflow.

You can find some instructions on how to do so here. Although I wrote that in the context of setting constraints among parameters, the approach to modifying the model code should work in general.

1 Like
7

You can hack your way around this via the nonlinear syntax. For example:

dat <- data.frame(y = rnorm(10), x1 = rnorm(10), x2 = rnorm(10))

mod <- brm(
  bf(
    y ~ lp1 + lp2,
    lp1 ~ x1,
    lp2 ~ x2,
    nl = T
  ),
  data = dat, 
  prior = set_prior("normal(0, 1)", class = "b", nlpar = "lp1", lb = 0)
)

prior_summary(mod)
1 Like