Testing Simple Effects for Generalized Linear Regression

#1
Hello,

I have found significant interaction effects when analyzing a negative binomial generalized linear regression model. However, I am not sure how to further explore the interaction. I am currently using SPSS. How can I conduct follow-up simple effects to examine the interactions? Can I simply add something on to the syntax?

Thank you!
 

hlsmith

Omega Contributor
#2
How are your two variables formatted? I am guessing they are both binary?


I have not used negbin before, but I am guessing you can look at the odds ratios for one term stratified by the other term. I think it is usual to stratify on the term that may be easier to intervene on. So if I had employment status and alcohol consumption (y/n), I would look at the odds ratio of employ status on outcome for non-drinkers, then do it again but for the drinkers.
 
#3
I have one continuous variable (risky environment) and one binary variable (gene; carriers vs. non-carriers) predicting alcohol use (continuous). What are you saying makes sense, in my case I should look at the odds ratio of the risky environment among carriers and then again among non carriers. What I am unclear about is how exactly to do this. Should I run another regression (including all covariates?) without the interaction term among only carriers and then again with only non-carriers? Any help would be well appreciated!
 

hlsmith

Omega Contributor
#4
So your outcome is continuous and you have a binary*continuous interaction term, correct? If so you wouldn't use odds ratios, but plot the slopes for gene carriers and non-gene carriers on a graph, so you would have two incongruent slopes on the same graph. Since there is interaction you would assume the lines would be disordinal (cross each other). If needed you could slap confidence bands on them as well. See Figure 1 in the below article. That image is very crude, but conveys a vague depiction of an interaction.




Smith HL, Sidwell RA. Trauma patients over-triaged to helicopter transport in an established rural state trauma system. J Rural Health. 2013;29(2):132-139. (PMID: 23551643)

 
#5
Exactly, a continuous outcome with a binary*continuous interaction term. My advisor wants me to report (1) the effects of one predictor (that is, an environmental factor) on the DV when another predictor is set to be zero (that is, non-carriers) and (2) the effects when another predictor is set to be one (that is, carriers).

I see how creating the figure would be a visual aid, but is there any supplemental follow-up tests that will produce a p-value to help support my interaction finding?

Thank you!
 

hlsmith

Omega Contributor
#6
I am guessing you are running a multiple linear regression model, correct? If so, your model would be: Y = Bo + B1(X1) + B2(X2) + B3(X1*X2) + error term.


Where:
Y= continuous dependent variable
Bo = intercept (reference binary group slope)
B1 = slope for continuous variable
B2 = slope for binary non-reference group
B3 = slope for multiplicative interaction


If you model it this way, if the interaction term is significant then an interaction between the binary and continuous variable is present.
 
#7
Yes, a multiple linear regression. Should I simply calculate that by hand or there a way to do it in SPSS to obtain p-values?

Thank you!