I have a dataset in which each individual ranks four attributes of a product. This task (with the same attributes) is repeated three times under different conditions. The product is still the same, so I have only individuals’ characteristics as variables. The rank-ordinal logit model seems to be the natural method. However, I was wondering if there are some statistical methods that allows to determine if the individuals ranking is consistent across the three different situations. Moreover, which method suits the most? I was thinking to: if exists a sort of panel data rank-ordered logit estimates three different rank-ordered logit and then test for equality of parameters across the three estimations use the three ranking to obtain one unique ranking for each individuals, and then applied the rank-ordered logit

Thank you for your attention


This is probably a place for a proportional odds mixed effects model (see e.g. the ordinal package in R). Per-subject random effects handle the non-independence of responses within a subject. I don't know how to model across-time agreement per se but it would be easy to include a linear time effect in the mixed model (time would be a fixed effect) to assess drift in the typical responses across time.

  • $\begingroup$ Thank you Frank for your reply. However I have a doubt: does the proportional odds mixed effects model work when the dependent variable is an ordinal one? Because I have a ranking as d.v. $\endgroup$ – Andrea Sep 9 '19 at 14:30
  • $\begingroup$ @Andrea yes it does. If you use R see the package Frank cited and its documentation. $\endgroup$ – mdewey Sep 9 '19 at 15:32
  • $\begingroup$ @mdewey thank you. The fact is that my d.v is not ordinal but is a ranking. Here (cran.r-project.org/web/packages/ordinal/ordinal.pdf) I don't find my case $\endgroup$ – Andrea Sep 9 '19 at 15:57
  • $\begingroup$ @Andrea as far as I can see a ranking is a subset of ordered categorical variables. $\endgroup$ – mdewey Sep 9 '19 at 15:59
  • $\begingroup$ I think that in an ordered regression, the respondent, for example, provides an answer to a question among different level (for example very likely, likely, unlikely, very unlikely). In the rank ordered the respondent faces different alternative and ranks them, so it's a kind of specification of a multinomial logit model, with the difference that respondents instead of choosing the best alternative, ranks them (providing so more information) $\endgroup$ – Andrea Sep 9 '19 at 16:24

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.