Multinomial Logistic Regression - Interaction Effect

#1
Hello there,

I would appreciate it very much if someone could help.

In my study, participants saw a picture of a man or woman either with or without a cigarette. So I have a 2(Male, Female) x 2(smoker, non-smoker) experimental design.

The question is whether participants would choose to (1) Date, (2)Not Date or (3)Become Friends (with the person on the picture).

I have put Gender and Smoking Status into the Co-Variate Box in SPSS and put the DV into the DV-Box with (2)Not Date as a reference category.

I used a cutomized model, that is, I put Gender, Smoking Status and Gender*Smoking Status into the model by using the forced-entry method.

After analyzing the data I found that the likelihood ratio test of the overall model was nearly significant (p=0.070). However, the analysis revealed that the likelihood ratio test of Gender*Smoking Status was significant, x2(2) = 4.32, p = 0.02.

Can I still interpret this interaction effect even if the overall model is not significant?


And suppose, the overall model was significant and I wanted to do follow-ups, how would I be able to compare, say, the smoking woman to the non-smoking woman with regard to the 3 outcomes?
 

hlsmith

Omega Contributor
#2
Add some more detail about this:


"After analyzing the data I found that the likelihood ratio test of the overall model was nearly significant (p=0.070). However, the analysis revealed that the likelihood ratio test of Gender*Smoking Status was significant, x2(2) = 4.32, p = 0.02. "


LR test = test of empty model versus saturated model?
 
#3
Add some more detail about this:


"After analyzing the data I found that the likelihood ratio test of the overall model was nearly significant (p=0.070). However, the analysis revealed that the likelihood ratio test of Gender*Smoking Status was significant, x2(2) = 4.32, p = 0.02. "


LR test = test of empty model versus saturated model?
Thanks for your prompt reply! :)

But does the overall model not need to be significant in order to continue with interpreting the interaction effect?

As for the follow-ups, would it not be sensible to split the data and then just look at each level of gender seperately, so as to look at: smoking woman vs. non-smoking woman and smoking man vs. non-smoking man?