Ordinal predictor variable in regression


Fortran must die
When you have a nominal categorical predictor you just create dummies. With say a five level ordinal predictor (the classical case would be a likert scale predictor) do you just use the raw variable as the predictor? I always thought so, because the distribution or number of unique levels of the IV did not enter into the regression assumptions.

But recently a proff of mine said that doing this, say using a 5 point likert predictor, could generate nonsensical results.

Any suggestions?


New Member

From my understanding, ordinal predictors are different from those "truly" categorical in that the "distance" between, say, 1 and 2, 2 and 3, 3 and 4, etc. is the same. Under this assumption, using a likert predictor as a continuous scale is plausible. On top of that, I came across a paper, which argues that with more than 8 (as far as I remember) categories, there is no difference at all. I can provide you with more references on this assumption, if interested.
Lastly, it is possible to compute the predicted values for the dichotomized and "scale" predictors and examine the correlations between them.