2
$\begingroup$

My question is not about the definition of the two rank correlation methods, but it is a more practical question: I have two variables, X and Y, and I calculate the rank correlation coefficient with the two approaches. With the Kendall-tau-b (which accounts for ties) I get tau = 0 and p-value = 1; with Spearman I get rho = -0.13 and p-value = 0.44.

Why does Kendall tell me there is exactly no correlation, while Spearman does see a nonzero correlation?

Improve this question
$\endgroup$
5
  • $\begingroup$ Spearman's rho and Kendall's tau are calculated in different ways. Why do you believe that something is going wrong with the Kendall tau calculation (and not the Spearman one)? $\endgroup$
    – Stephan Kolassa
    Oct 25, 2017 at 17:09
  • $\begingroup$ Because I get a P-value of 1... Is not that suspicious? $\endgroup$
    – Roland
    Oct 25, 2017 at 18:42
  • $\begingroup$ $p=1$ is a direct consequence of $\tau_b=0$, so the question is why $\tau_b=0$ and $\rho= -.13$ is a problem for you. $\endgroup$
    – Stephan Kolassa
    Oct 25, 2017 at 18:46
  • $\begingroup$ Then, I re-formulate my question: why according the Kendall-tau there is an exact no-correlation, while Spearman is less strict on that? $\endgroup$
    – Roland
    Oct 25, 2017 at 19:52
  • $\begingroup$ Thank you. That clarification makes for a much better question. I took the liberty of editing it into your post (and upvoting) and answering. Hope my answer is helpful. $\endgroup$
    – Stephan Kolassa
    Oct 25, 2017 at 20:43

1 Answer 1

Reset to default
2
$\begingroup$

Spearman's $\rho$ and Kendall's $\tau$ are calculated differently. That is, they have different notions of "correlation". (As does Pearson's $r$.) Thus, they will output different correlation coefficients.

I don't see anything surprising about having one type of correlation zero and one nonzero.

Here is a fun little dataset with $\rho<0$, $r=0$ and $\tau>0$:

> xx <- 1:6
> yy <- c(-0.2,2.5,1,2,3.5,-1)
> plot(xx,yy,pch=19)
> cor(xx,yy,method="spearman")
[1] -0.02857143
> cor(xx,yy,method="kendall")
[1] 0.06666667
> cor(xx,yy,method="pearson")
[1] 1.748437e-18

correlation example

So different measures of correlation will output different results. Your next question probably is which coefficient you should use. Happily enough, we already have a question on this: Kendall Tau or Spearman's rho?

Improve this answer
$\endgroup$

Your Answer

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

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