14
$\begingroup$

It is common to talk about the linear correlation, Pearson's $r$, between two random variables $\{(x_1,y_1),(x_2,y_2),\ldots,(x_n,y_n)\}$ as having two components: a) the copula and b) the marginal distributions of $x$ and $y$. In contrast, the rank correlation, Spearman's $\rho$, depends only on the copula.

As a reminder, the calculation of $r$ and $\rho$ are identical except that in the case of $\rho$, the variables are first transformed to ordered ranks $R(.) = 1,2,\ldots, n$.

$$r = \frac{\operatorname{Cov}(x_i,y_i)}{\sqrt{\operatorname{Var}(x_i)}\sqrt{\operatorname{Var}(y_i)}} \quad, \quad \rho = \frac{\operatorname{Cov}(R[x_i],R[y_i])}{\sqrt{\operatorname{Var}(R[x_i])}\sqrt{\operatorname{Var}(R[y_i])}} .$$

Ranks follow a uniform distribution by definition. Let's further assume a normal bivariate distribution for nontransformed $x$ and $y$.

Now my question: I want to express $\rho$ as a transformation of $r$. It seems that this should be a fairly simple function that somehow involves a mapping between the normal to a uniform.

My intuition of $\rho$ is that it is a weighted function of $r$ with large weights placed on small distances in the middle of the distribution; again, this would seem proportional to the normal distribution itself.

Is anyone able to work out the specific expression?

Improve this question
$\endgroup$
6
  • $\begingroup$ Unless you can quantify what you mean by "approximately" normal, there is no possible correct answer. The reason is that in many common, meaningful definition of "approximate," a mixture of a bivariate normal and any other distribution (with tiny weight) is "close" to bivariate normal, but by placing the center of the contaminating distribution far from the center of the normal distribution you can make $r$ attain any value in $(-1,1).$ This operation scarcely changes the distribution of Spearman's $\rho,$ proving there's no relationship between $r$ and $\rho$ whatsoever. $\endgroup$
    – whuber ♦
    Commented Aug 25, 2019 at 13:01
  • $\begingroup$ I appreciate this and am fine with an expression that holds in the normal case. (Hence my last sentence: Assume normality if that helps.) Thankful for any help. $\endgroup$
    – pengzell
    Commented Aug 25, 2019 at 13:04
  • $\begingroup$ Do you just want to compare expectations or do you want to derive the joint distribution of $(r,\rho)$ for every $n$? $\endgroup$
    – whuber ♦
    Commented Aug 25, 2019 at 13:12
  • 1
    $\begingroup$ I would be happy with expectations! I see that my original question was perhaps unnecessarily wordy... $\endgroup$
    – pengzell
    Commented Aug 25, 2019 at 13:16
  • $\begingroup$ To aid with intuition somewhat, it might be worth noting the close connection between (suitably standardized) ranks and a marginal transformation to standard uniform (in particular, with increasing sample size). $\endgroup$
    – Glen_b
    Commented Aug 25, 2019 at 23:18

1 Answer 1

Reset to default
13
$\begingroup$

I think I found the answer. In Pearson's "On further methods of determining correlation" (1907) he derives the expression: $$ r=2 \sin \Big(\frac{\pi}{6}\rho\Big), $$ which implies, $$ \rho= \frac{6}{\pi} \sin^{-1} \frac{r}{2}. $$ This made me realize that for the bivariate normal case, $r$ and $\rho$ are, as a rule of thumb, near identical. I would still be curious if anyone has an answer for more exotic distributions but that is a separate question.

Improve this answer
$\endgroup$
1

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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