As far as I know, there's no good implementation of this (or at least there wasn't several years ago when I had a similar problem); the closest I've seen when I looked into this was with MCMCglmm, but I was in a similar situation as you, I didn't know enough of the Bayesian side to know what I was doing when I used it.

Another alternative might be to just break your analysis down into multiple binary logistic regressions comparing pairs of outcomes (for example, if the outcome variable has three levels A, B, and C, then you could do one binary logit looking at how often they choose A [as opposed to B or C], and another binary logit on just the non-A responses looking at how often they choose B [as opposed to C]). As far as I understand, that's essentially what multinomial logit is doing under the hood anyway, and forcing your analysis to be like this is probably also a good way to get clear about what the predictions are.

When I encountered this, the only option I had was Bayesian. I recommend the brms package because the syntax will look much more familiar (like lme4) and you can rely on defaults (flat priors). That will help you prevent major missteps. However, it may take a while to run and you may have to increase the default number of iterations. Let us know if it works for you. Good luck!

Thank you both for your responses!

Stephen Politzer-Ahles - That, to my understanding, is indeed what the multinomial logit is in effect accomplishing, though the formal presentation of results would obviously be a bit different.

Michael E Young - I'm quite comfortable with lme4 and had read up a bit previously on brms. The only reason I didn't include it in my list is that I ave no experience with Stan, which seems required for the package. Am I mistaken (that'd be great!), or do you know of a similar non-Stan approach?

Stan runs behind the scenes, James. The Installation will install a compiler to create the Stan code from the brms syntax - you don’t have to write Stan nor really understand it.

I've used brms and haven't ever used stan. brms puts all the stan stuff under the hood; you might have to jump through some hoops to install it, but once you get to actually running models you never need to work with stan directly. So far I've found it super easy to use (if anything, my concerns with brms [if any] are that it could make things too easy, allowing me to do stuff I don't really understand...).

brms package is good for this analysis.

Thanks, everyone, for your feedback. I was finally able to check out brms package this weekend. My experience was quite similar to Stephen's - straightforward and can get the job done, but I'm not entirely sure exactly what is happening. Considering my intended audience and my relative lack of understanding of the Bayesian elements at work, I'm redoing the analysis now as a two separate binomial logits... but intend to revisit brms later after brushing up a bit more on Bayesian. Thanks again!

James, that’s precisely what I did the first time I tried it - as two separate logits and as a multinomial Baeyesian. I ultimately published the former due to the audience, but the Bayesian analysis confirmed that I wasn’t misrepresenting anything.

Hi all. Thanks again for your previous guidance. Revisiting this analysis, I realize now that the three alternative outcomes can arguably be conceptualized as ordinal in nature. If going that route, it seems it might be feasible to run a cumulative link mixed model via the ordinal package. Does that seems sensible for an ordinal categorical DV (choice) with a single categorial IV predictor (group) + random effect (SubID)? If so, is there a definitive set of considerations for employing clmm vs. clmm2? Any thoughts or opinions would be appreciated!