Help with ordinal regression model with multiple likert items

Interfaces brms
1

I am trying to predict the length oh hospital stay based on participants’ responses to a depression questionnaire. The questionnaire consists of 10 items to which participants respond on a likert scale from 1 to 5.

Normally I would add the items on a likert scale up and use a single score for each person. However, I recently read this article https://journals.sagepub.com/doi/10.1177/2515245918823199 and would like to do this properly taking into account all items. However, I have never used this kind of analysis before, and I struggle to adapt these recommendations to a situation where multiple likert items are used.

I defined the model as this:

brm(
  formula = 
    HospitalStay ~ 1 + ResponseToDepressionQuestionnaire 
                     + (1 | Participant) 
                     + (1 | ItemOftheQuestionnaire), 
  family = cumulative, 
  data = depression
)

Data is in long format with 10 rows for each participant.

Could anyone please help me and check that this is correct? Is there anything else that I need to add?

Thanks.

1 Like
2

Hi @statscretin welcome to the Stan forums.

I don’t work much with this kind of data but it looks like you’re trying to adapt something from @matti and @paul.buerkner’s paper on ordinal regression in psychology so hopefully one of them or someone else here more experienced with likert data than I am will be able to let you know.

3

Thank you

4

Hi,

HospitalStay is an integer or? Looks like you should use a Poisson or perhaps Negative-Binomial to model that in that case. ReponseToDepressonQuestionnaire on the other hand could be modeled as a monotonic variable mo() where you set a suitable Dirichlet() prior?

5

Thanks. Length of hospital stay is recorded as a continuous value

6

Aaah, so a real value which can only be positive, e.g., log-normal()?

7

Your code is for fitting a model an ordinal-scale response variable, not ordinal predictors. I recommend taking a look at monotonic effects for an approach that’s more suitable to your data.

3 Likes
8

I agree with @Christopher-Peterson. Here’s another example of that approach.

2 Likes