Multicollinearity in binary logistic regression

zxc

New Member
#1
I have two correlated independent variable and I need to use them both (one as a 'interesting' variable, other one as a control variable). Because of the multicollinearity, the odds ratio estimate for the interesting independent variable would be false, right? What shoud I do to get a one good estimate of the 'interesting' variable's impact on dependent variable when the impact of the control variable is removed?

Great if someone could help...
 

Masteras

TS Contributor
#2
Try to break the collinearity, standardize the values of the variables, one or both of them. How about using only one (worst case scenario)?
 
#3
zxc,

If you have moderate multicollinearity, you don't really have a problem. You just interpret the odds ratio of the interesting variable as the effect of that variable with the effect of the control removed (just what you're looking for).

A correlation between two IVs has to be really high (depending on sample size) before multicollinearity becomes severe enough to cause instability in the regression coefficients. If you're concerned about severe multicollinearity, there are better ways to check it than a correlation. I would start with running condition indices and running the model on split sampes (if your n is big enough).

Karen
 

zxc

New Member
#4
Many thanks for your help, Karen and Masteras. My IVs are dummy-coded, so I can't standardize them. Luckily multicollinearity isn't too big problem...
 

zxc

New Member
#6
I need an estimate of the interesting variable's impact on dependent variable when the impact of the control variable is removed. If I added interaction term, I would get estimates for IV's impact on DV in different classes of control variable. I want only one estimate for IV's impact on DV when the impact of CV is removed -> interaction term is not needed. IV's impact on DV may be different in different classes of CV, but that isn't the main point for me.

I hope I haven't mis-understood anything...
 
#8
ok, then this is your call if you want to do this, but if the interaction is significant i suggest you do not remove it.
I had to think about this a little bit.

Certainly by not including an interaction term, you are making an assumption that there is not an interaction. But that is no different with two categorical predictors than it is with any two predictors. Any multiple regression, for example, without any interactions is making that assumption. You could argue that all interactions should be tested, but you could end up with a mess.

If the interaction is not of interest, and there's no reason to suspect there might be one, I'm not so sure it needs to be tested.

I know SPSS puts in interactions between two categorical variables by default (SAS doesn't). But theoretically, this isn't a special situation.

Karen
 

Masteras

TS Contributor
#9
the thing is that he will need to see how ell tha model fits, if the deviance is small compared to the chisquare so he might have to put the interaction.
 

zxc

New Member
#10
Ok.. So I can't think that if I excluded significant interaction terms, I would get estimate for IV's impact on DV among whole target population when the impact of CV is removed. I have to add significant interaction, and then I have to look at odds ratios separately in different classes of control variable.
 
#11
Ok.. So I can't think that if I excluded significant interaction terms, I would get estimate for IV's impact on DV among whole target population when the impact of CV is removed. I have to add significant interaction, and then I have to look at odds ratios separately in different classes of control variable.
No. You can think it (and you should :)). If you left out the interaction term, you would indeed get the estimate for IV's impact on DV among whole target population when the impact of CV is removed.

You can check for an interaction. But if you are not hypothesizing one, you are not obligated to. An interaction will not test the research question you are asking.

Karen