Ranked Data Analysis

[Femi Oladunni]

Hello Everyone,

I hope this meets you all well.

Please, which technique do you recommend is best to analyze a dataset were 500 individuals rated 5 elements of a profile in rank order (e.g. 1st, 2nd, 3rd, 4th, and 5th). No two elements they rate can have the same number, so they are actually ranking which element they like best, and the 4 they like less in order (Preferential Scaling). I want to know which element was rated more highly, doing a mean summary could explain this I assume, but I want to do something statistically professional and diligent for a publication. I also have controls like estimated income, gender, etc. And I would like to know how this ranking varies across controls as well.

Thanks

[Bruce Weaver]

Interesting question, Femi. It reminded me of this blog post by textbook author Thom Baguley:

• https://seriousstats.wordpress.com/2012/02/14/friedman/

As you can see, Baguley argues that repeated measures ANOVA on the ranks is often preferable to the Friedman test. Given that many authors nowadays recommend a multilevel modeling approach to analysis of repeated measures, I wonder if you could get a decent model using -mixed- with elements clustered within participants. The other possibility that comes to mind is -meologit-, again with the 5 elements clustered within participants. (Both ideas are still only half-baked at best, so apologies if they are completely daft.)

[Hong Il Yoo]

As a starting point, you can consider the rank-ordered logit model (-rologit- in Stata). You may need to -expand- your data depending on how your data set is currently structured. Once you -help rologit- and browse the example data file, you'll be able to get a good sense of the assumed data structure. Note that the rank-ordered logit model is not the same as the ordered logit model.

[Femi Oladunni]

Thank you so much, Bruce and Hong, I will checkout on your recommendations and see what comes up from there. I so much appreciate it.