I’ve been fitting a couple of different cumulative probit models to some ordinal (e.g., Likert-style) response data. In both models, I have two group-level predictors (respondent ID and question number). The sole difference between the models is that, for one (let’s call it Model A), the mean of the…

Have you looked at pairs plots for group, the sd’s of group-level effects, and the thresholds (intercepts)? Off the top of my head and with no data, I can’t think of why this would be, but there is shifts in where the variation is (in the thresholds vs the varying intercepts) when group is added vs …

Thanks for your response. I hadn’t looked at the pairs plots, but having now checked them for both models: In Model A (with the effect of group) the only parameters that are obviously related are the intercepts (unsurprisingly). In Model B, the intercepts are also correlated, but much more weakly fo…

[image] ps2: In Model B, the intercepts are also correlated, but much more weakly for some reason. I think actually what you are seeing is the difference in scale on the x and y axis for the two plots. There is a much larger SE (as well as different values) for Intercepts 4 and 3 for Model A …

[image] jd_c: [image] ps2: In Model B, the intercepts are also correlated, but much more weakly for some reason. I think actually what you are seeing is the difference in scale on the x and y axis for the two plots. There is a much larger SE (as well as different values) for Intercepts 4 …

Yeah, that is just wacky looking. I don’t think the stronger prior will solve your problem here (it will make the sd less, but there is something weird going on). The plot for Model A is how I would expect the distribution to look, but not like the one for B, even with a higher sd. @Solomon @simonb…

Hmm, what does the output look like for pp_check(group, type = "bars_grouped", group = "group"), where group is the name of your model with the group predictor variable?

I also am not sure why this is happening. One of the challenges of hierarchical ordinal models is that if someone always reports the highest or lowest value, their intercept keeps improving as it moves towards positive or negative infinity. Shrinkage from the hierarchical prior should keep this in c…

So contrary to all expectations (mine included), it turns out that setting a stronger (though not super-restrictive) prior really does help. By replacing the default student_t(3, 0, 2.5) with a normal(0, 1) prior in the model without the effect of group, suddenly the estimates for all parameters bec…

Interesting. Obviously the default is pretty wide, but I don’t see why the default prior works fine for Model A but not for Model B…that’s the part I would like to know. I would have expected similar behavior regarding the prior for both models.

Understanding odd estimates from cumulative probit model

Interfaces brms

ps2 November 7, 2022, 3:53pm 10

@Solomon: Like this:

Topic		Replies	Views
Choice of regularizing prior for the cumulative probit brms ordinal-response , priors , logistic-regression	4	176	May 28, 2025
Partial pooling on ordinal response (cumulative probit) brms ordinal-response	16	476	May 23, 2025
Priors for a novice General	27	4431	October 25, 2019
Results from monotonic effect versus dummy variables in brms Modeling	27	1788	March 14, 2020
Inappropriate parameter estimates in ordinal models brms	10	911	January 25, 2021

Understanding odd estimates from cumulative probit model

Related topics