# 3 outcomes: one ordinal regression or two logistic regressions?

Imagine, for example, I am fitting a model with the following data:

Dependent variable: disease
0 = no disease (100 cases)
1 = mild disease (300 cases)
2 = advanced disease (100 cases)

Explanatory variables:
x1+x2+x3

Please correct me if I am wrong: the proportionals odds assumption maybe does not holds there. Anyway, I could do one ordinal regression or two logistic regressions (if I consider 0 = no disease, 1 = mild/advanced disease and 0 = no/mild disease, 1 = advanced disease). In a sense, you can put the cutoff somewhat arbitrarily: in this case, the advanced disease is "the more serious form" of the mild disease. Keep in mind that both approaches could be equally informative in the clinical sense of the term (and a logistic regression could be easier to understand).

Which one is the overall best approach? One ordinal regression or two logistic regressions?

(I use R)

• I don't understand why you think you can't use ordinal regression. The ordinal package has a function nominal_test to test for the proportional odds assumption. Commented Jun 17, 2018 at 15:37
• Thank you for your input, I read this: stackoverflow.com/questions/23395433/… and I will study the issue. Do you have any suggestion (maybe a personal opinion) about the main question? Commented Jun 17, 2018 at 15:55
• I would want to use one model. Ordinal regression makes sense to me, unless I'm missing something . Commented Jun 17, 2018 at 16:01
• How about using multinomial logistic regression? Commented Jun 17, 2018 at 19:22
• Thank you for your answers. Multinomial logistic regression is a good advice, do you think that it is appropriate (in the broadest sense of the term) even if my dependent variable is ordered? Commented Jun 17, 2018 at 22:11