I am performing ordinal regression, I have 5 response categories and several predictors both continuous and categorical. I would like to add a predictor which is categorical but ordered (1, 2, 3, 4). I don't think it would be appropriate to apply the usual dummy coding for unordered categorical predictors, but when I searched for how to code this I did not find much information. In Steyerberg (2009) "linear coding" or "assuming linearity of the predictor effect" is mentioned, but without further details. Does it mean I just use my ordered values as they are, i.e. use them as a continuous variable?

  • 1
    $\begingroup$ I think that is what is meant. I haven't seen a great solution to ordered categorical variables. It seems like it is either dummy coding, assume linearity, or use some overly complex non-linear function. $\endgroup$ – charles Nov 27 '13 at 3:14
  • 2
    $\begingroup$ @Charles You could choose any simple nonlinear function you liked. You could also assign any numeric scores that make substantive sense, or use a variety of more systematic methods to assign scores based on the relative frequencies of the ordinal variable. $\endgroup$ – Nick Cox Nov 27 '13 at 10:40
  • $\begingroup$ @NickCox Thank you. Great advice. Hopefully I'll remember next time I'm frustrated by an ordinal variable. $\endgroup$ – charles Nov 27 '13 at 15:36
  • $\begingroup$ @Nick: What are those systematic methods? $\endgroup$ – Scortchi - Reinstate Monica Nov 27 '13 at 19:58
  • 3
    $\begingroup$ Scores calculated from the data somehow, e.g. normal or logistic scores, scores from correspondence analysis. I mean any method not based on intuition, guesswork or convention. $\endgroup$ – Nick Cox Nov 27 '13 at 21:24

You could check out Gertheiss & Tutz, Penalized Regression with Ordinal Predictors, & their R package ordPens. They say:–

Rather than estimating the parameters by simple maximum likelihood methods we propose to penalize differences between coefficients of adjacent categories in the estimation procedure. The rationale behind is as follows: the response $y$ is assumed to change slowly between two adjacent categories of the independent variable. In other words, we try to avoid high jumps and prefer a smoother coefficient vector.

  • 2
    $\begingroup$ More active discussion is here. $\endgroup$ – Ilya V. Schurov Jan 14 '16 at 10:54

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.