interaction in logistic regression

#1
Hi,
I have run a moltinomial logistic regression and wanted to test for moderation effects of some variables so I tested the model with main effects only against a model where interaction effects where added.
Is not normal that the likelihood ratio test table does not show significance for the main effect of one variable (the only continuous variable in my model)? The table only shows a significant interaction effect. Am I doing something wrong?
Thank you very much
 

Dason

Ambassador to the humans
#2
Are you're saying that the interaction effect is significant but the main effects are not significant? Or are you saying that the only output you get is for the interaction effect?
 
#3
The second is correct. I have no output in the likelihood ratio test table for my continuous variable. I mean, the variable is there, it is in the "list" with the others but chi sq is 0 and df are 0 and so there's no significance value.
Hope it is clearer.
 

Dason

Ambassador to the humans
#4
Well the likelihood ratio test is comparing your full and reduced model. It can only say something about the differences between those models. Your main effects are in both models so it can't say anything about the main effect. What you're doing will only give you the ability to make a statement about the interaction effect.
 
#5
Good, this is what I wanted to get in the end! Thank you very much, that was very helpful! I am not sure whether I can pose onother question here or I need to open a new thread...the two questions related. Well, I'll post it here, if this is not the proper place, just tell me, I'll open a new thread.
Is there a way to choose the variables to be entered in the model? I want to test the full model against a nested one but in SPSS all the methods (forward entry, stepwise, etc..) do not allow you to choose which variables to drop. At least, I have no idea how to do that.
Thanks again!!
 

Dason

Ambassador to the humans
#6
I've never really used SPSS so I can't help you there. It sounds more like an SPSS question so you should probably start a thread in the SPSS section.
 

spunky

Can't make spagetti
#7
if you go analyze---->regression---->binary logistic you see how you get your usual logistic regression prompt screen? right under where you enter the dependent variable it says in blue letters "Block 1 of 1" there you can enter as many covariates as you want and then click on the "Next" button on the upper-right corner of that sub-section of the logistic regression window and you'll see how "Block 1 of 1" changes to "Block 2 of 2". you can enter more covariates and, if you run the regression, it will give you the chi-square-change statistic of the regression equation you entered originally in Block 1 VS the regression equation in Block 2 (Block 2 will include all the covariates you specified in Block 1 PLUS the ones you uniquely entered in Block 2 so the equations are nested and you can perfom the likelihood ratio test). if you want to drop variables then you just do it the other way around... enter as many covariates as you want in "Block 1 of 1" and in "Block 2 of 2" enter the exact same ones MINUS the ones you wish to drop, so the models keep being nested...

hope that helps.
 
#8
Hi, thanks for the hints but unfortunately I am running a multinomial logistic regression, not a binomial one.

Moreover, I checked again my likelihood ration test table and I am still having troubles understanding...Dason, as you said the likelihood ration compares my full and reduced model and main effects are in both models, than why I do have main effects for one of the variables but not for the other?
An example will help here:
I have two variables:
1. model (having versus not having a specific model in psychotherapy)
2. years of experience in psychotherapy
Plus I have the interaction between the two.
The likelihood ratio test shows sig values for the variable model and for the variable model*years of experience but not for years of experince alone.
Sorry for bothering you but I really need to understand this and it is driving me crazy!

Thanks a lot
 

Dason

Ambassador to the humans
#9
I guess I can't help you. It sounds like an SPSS issue. If you were using R I might actually be able to be of more assistance.
 

spunky

Can't make spagetti
#10
darn... totally missed that logstic VS multinomial, but i guess i've gotta agree with Dason. i'm learning SPSS as i go myself, i'm more of an R user... if CowboyBear reads this i'm sure he can help. he's i think the most proficient among us with SPSS...