I read Bürkner & Vuorre (2018)'s tutorial paper on ordinal regression models in psychology. I was in particular interested in the “multiple Likert items” aspects. I found a few answers to my questions in the tutorial the paper, but some left with no response.
In fact, I am using multiple Likert scales as explanatory variables, all containing multiple items, but obviously these do not have all the same size. Similarly, I wish to correlate these Likert-responses to an index of visual exploration from eye-tracking data.
Entering the data as you suggest is most suitable for brms (1 item per row) makes a lot of blank dataset cells because of the different questionnaire size. It seems embarrassing for bmrs… Do you know how to deal with that?
The eye-tracker index is only 1 value for each participant. Should I replicate these values in this column for each cell related to each participant?
I would be very grateful if you had some insights for me on that part. Please find attached an example dummy dataset resembling the one I will analyze (I already followed some of the tutorial’s advice, e.g. one row for each item of each participant).
If you’ve measured mood and depression both with multiple items then you could think of first aggregating both into composite scores (via factor analysis or some other method, either within or -outside the model), then use the two scores to predict eye tracking. The ordinal nature of your predictors is not that important but that depends on what method you choose to use for aggregating them (if you choose to do so).
So you suggest that I first calculate some kind of “total score” for each participant and each questionnaire? And then running an ordinal regression using these scores as predictors? Won’t this make a substantial number of “intercepts”/“thresholds” (e.g. if total questionnaire scores range from 0 to 100, it will output 99 thresholds, right?).
The way I understood your data + question is that you’ve measured one eyetracking score per person, and 7 or so depression items and 5 mood items per person, and then you want to predict eyetracking from depression and mood. If you use mood and depression as predictors (however wrangled), there are no distributional assumptions about them because they are predictors. So choosing the data model depends on what the eyetracking score is. Maybe you can think of it as gaussian so you could do something like a vanilla lm(eyetracking ~ mood + depression, ...) model.
Or in other words:
So you suggest that I first calculate some kind of “total score” for each participant and each questionnaire?
Suggest is a strong word but that might make sense in many cases.
And then running an ordinal regression using these scores as predictors? Won’t this make a substantial number of “intercepts”/“thresholds” (e.g. if total questionnaire scores range from 0 to 100, it will output 99 thresholds, right?).
No, the ordinal model / intercepts issue relates to the outcome of your model, which seems to be eye tracking.
Glad to hear I was able to help. If you think one of the posts above was ‘the answer’, would you mind marking it so, so that other readers can quickly find it?