Problems with low ESS when using brms::brm_multiple

Hi,
I am experiencing very low ESS when using brm_multiple (some values in the region of 100 when I would expect 2000). Here are details of relevant information, I would be really grateful for some advice. Thank you.

  • Windows 10, R version 4.1.3, BRMS version 2.16.3, MICE version 3.14.0

  • Data with 2700 observations has approximately 14% missing data, all of class numerical. Used mice to create 30 imputed datasets using method=“cart”. Plots used to assess quality of imputations and I was satisfied.

  • Data used in a brm_multiple model with one random effect and 40 fixed effects, horseshoe priors applied to 30 of the fixed effects, default priors applied otherwise. Used 2000 iterations (1000 warmup), 4 chains. Specified a high target acceptance criterion of 0.98.

  • If I select one of the imputed datasets and run as a brm model, the ESS and rhats are all satisfactory. However, if I run the brm_multiple model i.e. to pool all 30 imputed datasets, I am left with very low ESS as described above (NB also convergence warnings and rhats>1.05 but as described in the brm_multiple documentation this is likely false warnings due to the chains across different datasets maybe not overlapping)

I am grateful for any advice, thank you.

1 Like

I have encountered the same issue! But if I look at the individual models using the code provided here, the bulk/tail ESS are much higher than the summary output of brm_multiple suggests (false positive). I think it has to do with the way how brm_multiple computes the pooled ESS and they are not representative of the individual ones like the Rhats. @paul.buerkner How does brm_multiple compute pooled ESS’s and do you have an idea why it is super low sometimes despite the individual ones being high?

The overall ESS are misleading. In the latest brms versions (at least from github), I am no longer showing them in the summary output. Instead, I present a bit of code in the doc of brm_multiple how to compute convergence measures individually for each model.

2 Likes