Suppose that $X$, $Y$, and $W$ are random variables. Also consider the random variable $Z=XY$. I am interested in comparing the Pearson correlation coefficients $\rho(Z,W)$ to the correlation coefficients $\rho(X,W)$ and $\rho(Y,W)$. I have a data set where the correlation coefficients seem to roughly obey

$$\rho(Z,W) =\rho(XY,W) \stackrel{?}{=} \rho(X,W) + \rho(Y,W)$$

I put the question mark over the equality because I am wondering if this is a consequence of the math or something interesting about my data.

Is there any relationship between the correlation coefficient of a product of two random variables with a third r.v. and the correlation coefficient of each r.v. with the third r.v.?

This may not be true in general (though if so that is interesting). If it is only true in a limited set of cases, what assumptions must be made about the r.v.s (independence, distribution, etc)? If it is not true, why (counterexamples, general reasoning why such a thing shouldn't be true)? In my particular case I think that $X$ and $Y$ are correlated and normal, and that $W$ is log-normal (and clearly correlated to $X$, $Y$, and $Z$).

  • $\begingroup$ If $\rho(X,W)+\rho(Y,W)\gt 1$ this relationship obviously cannot be true in general. Such a relationship has no natural connection with properties of correlations--the nonlinearity of the function $(X,Y)\to XY$ indicates there should be no general relationship between correlations of $X$, $Y$, and $XY$ with other variables--so it most likely is accidental. It would be interesting only if it persisted across several sets of random variables with varying correlations among them. $\endgroup$ – whuber Jun 28 '14 at 16:09

It is certainly not true in general. Covariance is a bi-linear operator, so the product term is unlikely to decompose into something constructive. Here is a counterexample:

Assume $X$ is not degenerate. Take $Y=1/X$ a.s., $W=X$ a.s.
$\rho(Y,W)=\rho(1/X,X)\ne -1$ as $X$ and $1/X$ are not linearly related (a.s.)
So clearly the equality you suggested does not hold.

Note that $cov(X+Y,W)=cov(X,W)+cov(Y,W)$. Assuming that all the variables have unit variance (so that correlations are equal to covariances), the following must hold:
$\rho(XY,W)=\rho(X+Y,W)$. This can probably provide some insight into the necessary relationship between $X$ and $Y$.

| cite | improve this answer | |
  • 1
    $\begingroup$ Your counterexample presumes $X$ has no probability of being zero. Simpler counterexamples are easy to come by: e.g., take $X=Y=W$. $\endgroup$ – whuber Jun 28 '14 at 16:11

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.