1
$\begingroup$

Suppose we have three random variables X, Y, T.

X is positively correlated with T: we know $\rho_{XT}$ and it is greater than 0, let's say for the sake of argument 0.5.

Y is positively correlated with T: we know $\rho_{YT}$ and it is greater than 0; similarly about 0.5.

If we find that the product of X and Y, XY is positively correlated with T, does $\rho_{(XY)T}$ tell us anything new (like an interaction term in multivariate regression)? Or is high $\rho_{(XY)T}$ inevitable given $\rho_{XT}$ and $\rho_{YT}$?

Update: I identified a possibly duplicate question sum of correlation coefficients and correlation coefficient of product . Not quite a duplicate as the other question asks whether in the general case $\rho_{(XY)T} = \rho_{XT} + \rho_{YT}$ which it isn't. I am interested in the less specific question of whether high, positive $\rho_{(XY)T}$ is inevitable given high, positive $\rho_{XT}$ and $\rho_{YT}$, and whether $\rho_{(XY)T}$ tells us anything useful we didn't already know.

$\endgroup$
3
  • $\begingroup$ It may be inevitable if X and Y are restricted to be positive. Should I explore that case in my answer? $\endgroup$ – eric_kernfeld Dec 13 '17 at 16:48
  • $\begingroup$ Yes, please explore that - I didn't think it made a difference but your initial answer shows me it does. And in the case I'm dealing with all variables are positive. $\endgroup$ – Sideshow Bob Dec 13 '17 at 16:52
  • $\begingroup$ OK, I'll work on it. It's tricky. $\endgroup$ – eric_kernfeld Dec 13 '17 at 17:53
2
$\begingroup$

It is not inevitable. Suppose T is a shifted Bernoulli r.v. that takes the values -1 and 0 with probability 0.5. Suppose X=Y=T. Then XY is negatively correlated with T: as X, Y, T decrease to -1, XY is increasing.

EDIT: Suppose all the R.V.'s are positive. As a preliminary question, is it possible to have a negative correlation between X and Y? The answer is yes. For example, suppose X, Y, T each live in {1, 2}. Suppose T is again a 50-50 coin flip. Suppose that:

  • $X=Y=T$ with 20% probability,
  • $X\neq T$ but $Y=T$ with 40% probability
  • $X=T$ but $Y\neq T$ with 40% probability

This setup completely determines a joint distribution over $X,Y,T$ such that $Y\neq X$ with 80% probability, even though X and Y are each correlated with T (equaling it with 60% probability).

This possibility makes it hard for me to reason abstractly about XY and guarantee its behavior. It also opens up avenues for a counterexample: if X and Y are negatively correlated, then at least the motion of XY relative to Z could be dampened rather than amplified as one expects intuitively (or reversed, as when negative numbers are allowed).

In the example above, E[XY] is above 2, so XY is usually below its mean, even when T is above its mean. Unfortunately, there is one case where XYT is really big.

  • T, X, Y, (XY): % chance
  • 1, 1, 1 (1): 10
  • 1, 2, 1 (2): 20
  • 1, 1, 2 (2): 20
  • 2, 2, 2 (4): 10
  • 2, 1, 2 (2): 20
  • 2, 2, 1 (2): 20

My instinct is that this basic structure could be tweaked until cov[XY, T] is negative. Maybe I'll try it later today.

$\endgroup$
2
  • $\begingroup$ This has already clarified my thoughts a lot, thank you. The point you raise is that the actual values of x and y matter. In the general case you would suspect therefore that $\rho_{(XY)T}$ provides new information, even if we already know the other factors, and even if we know $\rho_{XY}$ as well? $\endgroup$ – Sideshow Bob Dec 13 '17 at 22:50
  • $\begingroup$ Yes, that is the idea. $\endgroup$ – eric_kernfeld Dec 14 '17 at 2:08

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.