interpreting interaction terms

#1
Hello,

I hope that someone may be able to provide some help in my research. I have a 3-way interaction gender (male 0, female 1) by black (black vs white) by (uninsured vs insured) significant effect on 10 year survival.

so the interaction term is INSxGENXBLK

Now that I have found it is significant, how do I break down this term to make meaningful interpretation of this interaction?

Note: I am using SPSS

Thank you very much,

Derek
 

hlsmith

Omega Contributor
#2
Well 3-way makes it harder, but what you traditionally do is plot the survival curves to find the difference(s), so 8 curves. You can also compare the hazard ratios by strata groups but that is tedious as well.
 
#3
Hello, I would be interested in comparing strata, as some other interactions I have, have more than 2 strata. I basically have the pub written up, have analysed that data (to this point) and am stuck with this.

As you say it is tedious, I am not hoping for a walk through. Is there some resource that provides such guidance? I have read many journal articles, and while i have published (meta-analysis; t-tests, anova, etc), and understand a number of statistical techniques, the explanations in the literature seem to be different as to ways to break down these interactions; as well, are written with the assumption people already know all of the steps.

Thank you hlsmith, if you have any other advice, I'd be very grateful.
 

hlsmith

Omega Contributor
#4
I haven't come across many 3-way interactions. I think the curve approach is a good start to try and get a feel for which groupings may differ. Do you really care about the interpretation of any of these variables, or are you just controlling for them and they is a treatment/exposure variable you are actually interested in?

Also you never reported your sample size or proportion of people in each outcome group - and sparsity of data in these subgroupings can be an issue sometimes. E.g., so you have a subgroup of black uninsured males that all die, but there are only 6 people in this group, which may trigger the interaction to be significant. Though, given your study design can you say that these 6 people represent all people in the population with these characteristics? Provide a little more information please!
 
#5
Hello, thank you for the valuable information hlsmith. I am working on those crosstabs now to get individual cell values. As I feel I've covered a lot of reading and need to further my learning by seeing it done and doing the work-through myself, I have an epidemiologist friend of mine who is going to sit down for several hours with me and work me through breaking the interaction into strata, calculating RRs, and describing the interaction using a "by hand" method, which I would like to understand working through.

Thanks again for your input, I really appreciate you taking the time to respond.

Best,

Derek
 

hlsmith

Omega Contributor
#6
I would be interested in seeing what you come up with in case I come across one of these beasts!

Side note, did you remember to keep all of the lesser main effects and interaction terms in the model?
 
#7
Yes I did, I built the model using the lesser terms, as well as, introduced several adjusted-for variables (age and stage of cancer) I actually found a significant 4-way, as well as a 3-way. I will let you know for sure when I break it down, how it comes out.




Derek