It's clear for me to count correlation and regression between two variables, e.g. X and Y. Assume, that X is a fuel consumption and Y is engine displacement. I've got a data set with 100 pairs X Y, and based on those I can calculate a correlation rate between X and Y which in this case is obvious. So I'm trying to find some dependency between X and Y and check how Y affects X.

But how about the situation when there are three variables (X, Y, Z) and I want to calculate correlation between X and (Y Z) as a group. Assume once again, that Z is a mass of vehicle and I want to check if mass of vehicle and engine displacement (Z and Y) influence on the fuel consumption X.

It's easy to everyone to notice a dependency in my ordinary example:

X may be called normal (is inside some range, expressed as number) when Y is normal and Z is normal, X is high when Y is normal and Z is high, X is very high when Y is low and Z is high, X is low when Y is low and Z is normal

and so on...

So, I've got a data set again, this time with 3 columns: X, Y and Z and I want to find (or not) the relationship between X and YZ.

I know that I can express Y and Z as a one number - some average, but then it makes no sense because of the fact that the extreme values will give the same average.

Look that there are only 3 variables here and the dependency is noticeable for everyone, but what about when there are more variables with some relationship between them which is hard to notice without computing?

  • 3
    $\begingroup$ I think looking into Linear Regression and Interaction terms might be of use for what you are trying to accomplish. $\endgroup$
    – DanTheMan
    Dec 6 '12 at 20:58

There is an extension of the concept of correlation to groups of variables called canonical correlation. In the case where one group is a single variable (X) and the other group has multiple variables (Y and Z), the canonical correlation is the same as the multiple correlation ($R^2$) that you get from a multiple regression.

So the easiest way to do this is to fit a multiple regression with X as response and Y and Z as predictors, and look at the $R^2$. This measures the strength of the linear relationship between X and the best possible linear function of Y and Z.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.