I found similar threads on the forum, but none seems to apply to my case, so I decided to post the question.
My response variable is an ordered 1-4 likert scale. I asked three subjects to rate the linguistic quality of sentences pre and post a linguistic revision process carried out on the sentences. I decided that the best way forward is analysing the post scores, having pre scores as a predictor in ordinal regression.
Since I have three versions of the scale (3 subjects), if I use mixed-effects analysis with a random effect of subject, I'd end up having the same sentence with different pre scores in the data (from different subjects), which does not seem coherent.
Would it be OK to use averages (both for the dependent variable and for the pre score)? I would still treat the dependent variable as ordinal (now with 9 levels), instead of being 1-4, it'd have intermediate categories, say 1, 1.3, 1.6, 2...