[Richard Williams]

Here is my paper on Ordinal Independent Variables:

https://methods.sagepub.com/foundati...dent-variables

https://methods.sagepub.com/foundati...dent-variables

I think anyone can access this article toll-free with the first link but if not you can email me for the pdf. Also, here is the link to the earlier online handout:

https://www3.nd.edu/~rwilliam/xsoc73...ndependent.pdf

[Mike Lacy]

I'd recommend a small addition to the mention of "scoring systems" in the article that Richard cites, namely the use of so-called "ridit" scores, which score a category using an approximation of the cumulative relative frequency for that category. For example, with p(1) = 0.3, p(2) = 0.5, and p(3) = 0.2, the ridit scores would be
(1/2) * 0.3 for category 1, 0.3 + (1/2) * 0.5 for category 2, and (0.3 + 0.5) + (1/2) * 0.2. for category 3. The -ssc describe egenmore- package incudes it as an option. I think of ridit score effects as relatively straightforward to interpret, since the unit of measurement is fractional rank (or percentile point if you like.) A slope of 0.15 (say) on a ridit-scored variable scaled as a percentile would indicate that an increase of 1 percentile point in rank is associated with a predicted increase in y of 0.15.

Of course, there are various assumptions underlying the plausibility of this approximate scoring, and it has the disadvantage of scoring based on the distribution of the sample, but it could tried as one among several possible ways to use an ordinal response in that model, and perhaps it would work well. Percentile rank might be attractive as a metric for an income variable.