Testing for interaction effect btw two continous variables.

#1
Hi, I have 3 variables, A, B and C. All of them are continuous, assymetrical in distibution.

Using sperman correlations, I've found that A/B (A divided by B) is significantly
correlated with C. A by itself does not correlate significantly with C, neither does B correlate significantly with C by itself.
(Of note A*B does not correlate with C using Spearman's rho.)

Because of my finding of A/B correlating with C, I suspect that there is an interaction
effect going on. I wonder how I should investigate if this is the case in a meningful way.

My hypothesis - which would fit nicely with previous research and studies - is that higher values of A, combined with lower values of B (thereby creating a higher A/B fraction) interact so that A/B correlates positively with C.

I have tried visualizing and examining this using all sorts of plots and/or calculations, and it seems like my hypothesis holds - but I need help finding a way to calculate or visualize this that would be statistically sound, so that it actually could be published in an article.

All help appriciated!
 

hlsmith

Omega Contributor
#3
Based on the context, does your hypothesis make sense?


I may recommend creating scatter graphs for all of these relationships so you can better see what is going on. I slightly wonder if there is a non-linear shape that you are able to get rid of with the division.
 
#5
What kind of regression model should I use?
Does it matter if the dependant variable ("C") is continous or a discrete variable?
Sorry, foor my noobness :)
 
#6
hi,
why not build a multiple regression model C~A+B+A*B and see what you get?

regards
What kind of regression model should I use? Does it matter if variable C is discrete or not?

Linear Enter Regression with two blocks (block one A+B, block two A*B) ?

Edit: So I did the Linear Enter Regression with two blocks (block one A+B, block two A*B). Variable C is dependant variable and discrete in nature. (Does this matter? Does it have to be continuous for this to work?)

The R square change for the second model is 0.001 compared to 0.099 for model 1. Sig F. change is much above 0.05 for both models.

So I think it is safe to say that there is no significant interaction effect between a and b if I have made this correctly?
What I still wonder though is why A/B correlates with C significantly when neither A or B alone does. Shouldn't this
impy some sort of interaction going on? Why don't I find any? :\
 
Last edited: