Problematic sampiling with brms after updating cmdstan

Interfaces brms
1

Hi All,

I can really use your help!

I updated cmdstan and brms. I get the following error when I start sampling, and then it ‘settles’ for a while, but at the end Rhats are very high. I tried y ~ x + (x|group) with two different datasets and different variables that previously had no issues and gave good results. This also happens for logistic…

We have a deadline and the main workhouse is out of service… :(
Any idea whats going on??

Thanks
Nitzan

model = brm(y~ x+ (x | group), data = df ,chains = 4, warmup  = 2000, iter    = 2250,  cores   = 4,backend='cmdstan')


library(cmdstanr)
This is cmdstanr version 0.6.1
- CmdStanR documentation and vignettes: mc-stan.org/cmdstanr
- CmdStan path: C:/Users/user/Documents/.cmdstan/cmdstan-2.32.2
- CmdStan version: 2.32.2

> packageVersion("brms")
[1] ‘2.20.3’

image
image1340×177 12.9 KB

2

No idea what’s going on yet unfortunately, but cmdstanr::install_cmdstan() has a version argument, so you could install 2.31.0 or an older version of those worked for you.

1 Like
3

Thanks. Do I need to uninstall anything, or just re-install a previous version?

4

Do the warnings occur only a handful of times at the start of sampling, or do they occur frequently throughout sampling or error?

1 Like
5

Only a handful of times at the start of sampling

6

Then like the warning says, it’s likely safe to ignore - it’s just the sampler initially exploring the parameter space.

If you’re concerned or would like someone to investigate more, then you’ll need to share a reproducible example so that we can debug locally

2 Likes
7

I agree with everything @andrjohns said, but if you are really running into convergence issues (e.g. high R-hats) with the same model and data that ran fine before then that is somewhat alarming (I assume that’s what you mean by “previously had no issues”). If you could share a reproducible example where the behavior is fine with an older version of the software (cmdstan and/or brms) and problematic with a newer version of the software that would be really helpful.

2 Likes
8

I think this is just cmdstan being more verbose with warnings. I.e. the “problem” existed before but wasn’t communicated in bright red letters. And, as @andrjohns said, if it only happens a few times early on, you can do as the warning says and ignore it.

1 Like
9

Many thanks everyone. I hope that indeed this is just a verbose warning… I
If you can teach me how to move from one cmdstan to another on the same computer, I can try generating some replicated example to compare rhats with the same model and data.

  1. When installing cmdstan, do you have to remove the previous version?
  2. Can I switch from one version of cmdstand to another with setting PATH or is there another way?

Nitzan.

10

You should be able to tell cmdstanr which cmdstan to use via cmdstanr::set_cmdstan_path().

You don’t have to remove old versions, they’re installed in their own directories.

1 Like