Is modelling sigma for ordinal data currently possible?

I know that BRMS can do two interesting things, but I have thus far not been able to do both at once.

I have a bunch of images rated 1-5. I tried to create a model that includes sigma like so:

bform <- bf(rate ~ s(b,s),
         sigma~ s(b,s))
brm(bform, data = outsub, chains = 2, cores = 2, warmup = 2000, iter=4000)

This model assumes that the data is gaussian distributed, which is not true, but the model that also predicts sigma is about 20SEs better when I compare it to one that is simply specified:

bform <- bf(rate ~ s(b,s)

It thus seems like it could be valuable to include this in an ordinal model specified as the following:

brm(rating ~ category+distortion + (1|id), 
family = cumulative(threshold = "flexible")

I can specify this model too without issues, but when I try to combine them in

brm(bform,family = cumulative(threshold = "flexible")

… I get the error:

Error: The parameter 'sigma' is not a valid distributional or non-linear parameter. 
Did you forget to set 'nl = TRUE'?

I tried to read through some documentation such as Estimating Distributional Models with brms and it seems like this may only be a feature that works for gaussian and a few other distributions. Is this the case and if so should I simply give up on my attempt to model this as ordinal data for the time being? I am happy to provide data if needed, but given the question’s conceptual nature I chose not to do so for the time being.

Thanks for reading though my question

A quick answer which I hope will help you. The most similar thing to what you want is the disc parameter. I think it is defined as the inverse of the \sigma for the underlying distribution. If I recall correctly it is defined to be 1 at the reference level for all covariates (set to zero)

1 Like