I would actually like @paul.buerkner to weigh in on this matter due to his monotonic paper. I asked a similar question (I believe!) in another thread in this forum. In short, depending on what data (e.g., Likert scale 1-5) and if we have nobody answering 4, or 5, should we use, i.e., Dirichlet(c(1,1,1,1)) or Dirichlet(c(1,1,1))?
Question 2
a) If I have ordered categorical predictors from 1-5 and 4 has not been used should I still use ordered categorical in five levels, i.e.,factor(x, c("1", "2", "3", "4", "5"), ordered=TRUE)? My question is due to therlsvariable above which is really 1-8, but the data has recorded only levels 1 and 2.
b) If yes, does the same apply when 5 has not been used above in a) , i.e., should I still usefactor(x, c("1", "2", "3", "4", "5"), ordered=TRUE)?
As an example, if I model this ordered categorical variable:
table(bdf$f6f_rls)
1 2 3 4 5 6 7 8
27 0 0 0 0 0 0 0
as mo(f6f_rls) in brms it bails when starting to sample:
Exception: model159d46253acca_21c393ad993ab34f22af08399084b917_namespace::model159d46253acca_21c393ad993ab34f22af08399084b917: Jmo[i_0__] is 0, but must be greater than or equal to 1 (in 'model159d46253acca_21c393ad993ab34f22af08399084b917' at line 26)
Well, it’s not strange, it’s not monotonic in that sense, but it would be nice from a model specification point of view to be able to write mo() indicating that it should be a 1,...,x ordered categorical var.