I have 4 likert scale questions (coded 1= never through 5=always) that are to be dependent variables. I have 125 observations and 3 predictors. My initial thought was to just use ordinal regression on each question/dependent variable. However, given that there are 5 outcome levels, this is a lot to report. I am curious if there could be a meaningful way to combine the likert scale questions into one dependent variable. Thanks for any advice.

  • $\begingroup$ If the questions are related, I've seen people create an aggregated variable and then treat it like a single continuous predictor. I'm not clear on the theoretical motivation for that, so I'm not comfortable giving this an answer. $\endgroup$ – dankernler May 8 '18 at 16:56
  • $\begingroup$ Duplicate of stats.stackexchange.com/questions/10382/… $\endgroup$ – rolando2 May 8 '18 at 20:14

This will depend on the questions themselves. One approach would be to use a factor analysis approach to determine how well the items "load" on to the construct of interest. Are they all questions intended to measure the same underlying construct?

If you find that the four items are reasonably similar in their capture of the target construct, you might use a dimensionality reduction technique (PCA, median score across questions) to simplify. Note that this approach makes some assumptions about your items/data so you should think carefully about those assumptions.

The approach with the fewest assumptions is to proceed as you have described---one ordinal regression (e.g., cumulative link model) per item.

One tweak to simplify though (again assuming your questions target the same underlying construct) is that you could use a cumulative link mixed model (CLMM) where you predict the ratings for each of your four items based on your IVs and a random effect of participant (or whatever source generated your ratings).

model <- clmm(rating ~ IVs + (1|participant_id), data = your_data)

The value of creating an aggregated variable will vary depending on the number of items you have and 4 is on the low end of where that approach would be acceptable. I would not suggest this route but see this question for resources on this approach.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.