Brms vs rstanarm for ordered logistic

General
1

Hi guys! I am new in Stan and I am little confused about ordinal regression.Here are some questions:

1] I was trying to fit for my data an ordered logistic regression (only fixed effects) with brm and stan_polr function and after many trials I figured out that they produce the same results (using as argument the cumulative “logit” and family=“logistic” respectively). Which is the difference between these functions then? Maybe that in brms someone can add random effects or other facilities to his formula?

2] The brm function is the respective of lrm (library rms) from classical statistics like stan_polr which is respective with polr from MASS package?

3] For the two different versions (brms and rstanarm) which is the form of the model? I mean the form of the model is this logit(p_j)=b_{0j}+(b_{1} X_1+...+b_k X_k) or logit(p_j)=b_{0j}-(b_{1} X_1+...+b_k X_k) ?

Because in MASS package the - is used and the lrm library the +.

4] Finally, how can I change the baseline category in ordered in stan?In multinomial logistic the 1st level is used instead of ordered where the last level is used as baseline. I want to adjust it so that I have the baseline in 1st level my baseline .

Thank you for your time!

2

Yes, brm can estimate several types of ordinal models and has more flexibility with the parameterization of the linear predictor while stan_polr is essentially just a Bayesian version of MASS::polr. They also have different priors.

I don’t know if I would say that. The main functions in brms map a wide class of R syntax to a wide subset of Stan models for Bayesian inference.

Yeah, there are different parameterizations of a proportional odds model and stan_polr sticks with the one in MASS::polr.

Redefine the outcome using the factor function and use whatever ordering you want for the factor levels. The first one will be considered the baseline.

1 Like
3

But in brms when we run a cumulative logit model (ordinal logistic) , it considers as baseline only the last one instead of multinomial logistic which considers as baseline the first one.I would want to define the first one as baseline also in ordered logistic