Logistic_normal() family link function throws 'response category' error

Modeling
1

Hello, I’m running into a problem with the logistic normal family function in brms. I’m trying to compare logistic-normal and beta regression for modeling skin-conductance ratios across an experimental session. I have 6 explanatory variables to consider from, with 2 that are factors (sex and race). I could treat the latter as non-factor variables, but I’ll consider that afterwards.

Here’s a glimpse of my data.

glimpse(fh.sfdata[,-c(1,2)])

Rows: 243
Columns: 7
$ PCL5 <dbl> 12, 0, 3, 2, 0, 0, 5, 6, 3, 2, 3, 21, 0, 4, 13, 6, 15, 7, 4, 18, 1, 0, 10, 8, 0, 0, 3, 9, 2…
$ BDII <dbl> 9, 0, 3, 0, 21, 1, 1, 4, 4, 2, 14, 12, 2, 4, 5, 6, 7, 1, 9, 3, 11, 2, 10, 7, 3, 5, 0, 0, 25…
$ STAI <dbl> 42, 25, 37, 20, 62, 23, 26, 22, 35, 23, 32, 38, 29, 32, 35, 33, 45, 34, 36, 25, 39, 27, 27,…
$ Sex  <fct> 2, 1, 2, 2, 2, 2, 1, 1, 1, 2, 2, 2, 1, 2, 2, 2, 1, 2, 2, 2, 1, 2, 2, 2, 2, 1, 1, 2, 2, 2, 2…
$ Race <fct> 1, 3, 2, 1, 3, 2, 3, 6, 3, 4, 1, 2, 3, 3, 3, 1, 3, 1, 1, 2, 3, 1, 3, 3, 2, 1, 4, 1, 1, 3, 1…
$ Age  <dbl> 21, 23, 22, 29, 18, 23, 22, 25, 24, 33, 19, 21, 25, 19, 18, 22, 23, 27, 29, 22, 20, 25, 21,…
$ SF   <dbl> 0.44327570, 0.45485870, 0.47522750, 0.47923320, 0.42417820, 0.17689910, 0.47311830, 0.49595…

When I go to examine what the priors are for logistic_normal, I run into this issue and similarly if I run a brm fit.

default_prior(bf(SF ~ PCL5 + BDII + STAI + Age + Sex + Race,
       sigma ~ PCL5 + BDII + STAI + Age + Sex + Race),
       data = fh.sfdata, family = logistic_normal(),
       chains = 1, warmup = 500, iter = 2000, cores = 2, seed = 822025)

Error: At least 2 response categories are required.

I’m completely lost on this issue. Even if I remove the factor variables, the issue is the same. Running a simpler model with one predictor also caused this issue. Has anyone encountered this issue with logit-normal models?

2

For anyone else who runs into this issue, I’ve realized the issue is that the logistic_normal() doesn’t allow univariate response variables. I was expecting something similar to pymc’s logitnormal. There’s a demonstration of how to use it here and reasoning for why the multivariate and not univariate implementation is allowed as well.