Handling zero-response categories in in Adjacent-Category Models with brms


I’m analysing ordinal survey data using the brms package in R, specifically employing Adjacent-Category models with category-specific effects. I am investigating whether practitioners choose to teach statistics from a predominantly mathematical or conceptual perspective, or equal for both. Occasionally, I encounter convergence issues when a response category has zero responses for a certain group. Here is an example of my data:

Discipline Mathematical equalMathConc Conceptual
Bio 1 3 12
Biomed 1 3 6
Med 0 1 7
Psych 2 3 14

For the “Med” discipline, there are no responses in the “Mathematical” category. This zero-response seems to affect the model’s convergence and parameter estimation.

I have a concern - Does the zero-response category for one group affect the credible intervals for that specific category (i.e., “Mathematical” for “Med”), or does it also impact the credible intervals for other categories in the model?

If this zero-response is likely to substantially detrimentally affect the model, could a reasonable solution be to set the zero-response category to NA (i.e., treat it as missing) and run the model on the remaining data? I could transparently report this in my Methods, noting that no respondents from “Med” selected the “Mathematical” approach, hence it was excluded from the model for this discipline. I could also run the model with and without this modification to ensure that the exclusion does not significantly alter the results for other categories.

Here is the code I used for fitting the model:

brm_teaching_approach_var_3cat <- brm(teaching_approach_3 ~ cs(discipline),
                    data = survey_filt,
                    family = acat(link = "probit"),
                    file = here("models", "brm_teaching_approach_var3")

Any advice on how to handle this situation or alternative recommendations for dealing with zero-response categories in this modeling context (e.g., a hurdle model?) would be greatly appreciated. I had anticipated that zero-inflated wouldn’t be appropriate as I have only one zero.

Thank you!

  • Operating System: Windows 11
  • brms Version: 2.21.0

@willpkay I could be misunderstanding your use case, but are those zero values counts of responses? This does not seem to be a Likert scale that we’d usually apply ordinal regression to, but a contingency table of the number of practitioners who use each approach per discipline. This would motivate a different type of model.

Zero values on ordinal scales usually refer to a non-response pattern (e.g. missing data, ‘I don’t know’), and these can be handled separately in a variety of ways of differing complexity.