Multinomial Logistic Regression - when to exclude predictor


Active Member

Multinomial Logistic Regression With 3 predictors X1, X2, X3 and 3 possible values for the dependent variable Y: A, B, C
When you build the model before getting the results you aren't sure if X3 should be in the model.

X3 is significant for ln(Pb/Pa) , but is insignificant for ln(Pc/Pa)

Will you include X3 in the model?


TS Contributor
My example before, if I recall, was for an independent variable with 3 levels and k-1= 2 dummy variables (and not removing a dummy because one was non-significant on an individual test).

In any case, let's clarify: you're saying Y (a,b, or c) is an unordered variate that you suspect is a function of X1, X2, and maybe of X3?

What are the X1, X2, and X3 defined as?

Can you share a copy of your output on the thread?
Last edited:


Active Member
Hi Ondansetron,

Yes, it was a different model but it remains me my current question :)

In your example, you can't use only one dummy if one is significant and the other is not significant.
In the Multinomial logistic model, I think that you can't use predictor to only part of the Y categorical values?
(You may separate it into several binary logistic couples, but in this case, I think the result will not be as good as one multinomial optimization.)

Y is a categorical varaible, there in no mean to the order.
Correct, you suspect that Y is a function of X1, X2, and maybe of X3.

X1, X2, X3 are continuous variables.

Thanks, O

Multinomial logistic.png


TS Contributor
My mistake, I accidentally left off "un-" in "unordered" DV. I corrected my post. If X3 goes against theory, there isn't strong evidence it's a predictor, but I would otherwise leave X3 in the model. I see the connection that you were making with my other comment on a dummy variable. Since the dependent variable has levels A, B, and C, this is the entire variable. It wouldn't make sense to me to remove X3 from a model for part of the DV. I would leave it in because there is some weak-moderate evidence it may be useful in predicting the DV and you can reduce the risk of coefficient bias if you include it in the model.

How many observations do you have and how do they split up into the levels of the DV?


Active Member
Thanks Ondansetron.

It makes sense :)

I assume the same conclusion is also for your example of the dummy variables?

The sample size is a little small total of 74 (A:25 B:30 C:19)

Thanks O