Use of emmeans

[nahorp]

Hello,

Firstly, thank you for this package, very helpful!

I am trying to use emmeans on the model fit after running bfrms but I get this error,

Error in vcov(object)[nm, nm, drop = FALSE] : subscript out of bounds

Considering that I can run emmeans on a brmsfit object and the output of bfrms also seems to be a brmsfit object, I am not sure why it wouldn't work...

[singmann]

Sadly, that is not implemented yet and I am unsure we have the capacity to do so. This package mainly serves as a proof of concept for now.

As an alternative, you might check out stanova. It uses the same factor coding, but does not standardise the response variable (i.e., you might want to divide it by the observed SD before fitting). It also works with brms and supports emmeans.

I will leave this issue open in case someone else has this question.

[nahorp]

Thanks for that @singmann, stanova looks good and it does what I need to in emmeans... but is there an easy way to implement the JZS prior in stanova?

I found this website, and it says it is Cauchy(0, 0.707) prior on the Intercept (see screenshot below).

However, I am unsure whether this would be the prior on population-level effects too (i.e., class = "b")? Also, how would I specify the Jeffreys prior on Sigma to properly implement the JZS prior in stanova/brms?

Thank you for your time, appreciate it.

[singmann]

I do not have the time to check the website in detail, but it is unlikely that choosing the correct prior alone is sufficient to exactly replicate the JZS prior. Because this is what bfrms does and to do so it needs to jump through a few hoops.

What stanova does is that it provides you with the correct factor coding also used in the BayesFactor package. If you also normalise your DV by hand (i.e., before fitting the data) and use the Cauchy prior you get something that comes close to the JZS prior in the case of a factor with only two levels. However, it is still not identical (e.g., no Jeffrey's prior on mu and sigma). For a factor with more than two levels it is surely not identical because the Rouder JZS priors use a multivariate Cauchy prior (which one might also be able to implement somehow).

[nahorp]

Thanks for that informative reply @singmann, that's very helpful for me to decide what route to take.