Discrepancy between summary() and tab_model() for brms models

Question

Discrepancy between summary() and tab_model() for brms models

maxfarrell opened this issue 7 months ago · 5 comments

I'm having trouble formatting the summary of my hierachical brms model via tab_model()

I would like to essentially re-create what I see with summary(model) via brms.

I see that the "fixed-effect" coefficients differ slightly compared to the table, but more concerning is that the standard deviation parameters for my hierarchical components are not shown. Instead, there is a Tau estimate, which is not reflected in the summary() output.

Any advice or clarifcation on how to show the hierarchical group varances, rather than Taus would be greatly appreciated.

Answer 1 · 2024-04-23T10:31:36.000Z

Coefficient may differ, depending on whether mean or median (or another centrality measure) is reported. tab_model() reports the variances, while summary() usually reports SD of the random effects parameter. You cannot change this part of the table output in tab_model(), unfortunately.
Alternatively, you may try following packages:

parameters (using model_parameters() and print_html()
modelsummary
gtsummary

(There are probably more, but these are the packages I usually use)

Answer 2 · 2024-04-23T11:05:40.000Z

Hi Daneil,

Thanks for the clarification. In this case, what is the tab_model() sigma^2 for the random effect denoting?

Answer 3 · 2024-04-23T11:54:07.000Z

Sigma^2 is the resiudal variance (you also get this from the summary). See also https://easystats.github.io/insight/reference/get_variance.html, where Tau etc. are also explained.

You actually find most values directly in the summary output:

Answer 4 · 2024-04-23T11:54:32.000Z

(just realized, you're talking about brms - the summary output may differ)

Answer 5 · 2024-04-23T13:19:03.000Z

Hi Daniel,

Thanks - yes I'm fitting a hierarchical bernoulli model via brms so the output is different, and I don't think there should be an estimated sigma^2 term as that's not a parameter in the model (other than the sigmas for the hierarchical effects). Just wondering if this something coming out because I'm using brms when this function isn't supported for these types of models.

I'll look into the other pacakges you mentioned in the mean time