0
$\begingroup$

I know the approach using student t-test to derive p-value from spearman rho coefficient. I also saw permutation test for this operation, but I am somehow confused and prefer to stick to the z-score formula $\frac{x-\mu}\sigma$.

In other words, how can I convert Spearman rho coefficient to the normal approximation z statistic with formula z-score formula $\frac{x-\mu}\sigma$?

Link to related unanswered question: Spearman rho statistical significance value (z)

$\endgroup$
  • $\begingroup$ Please let me know in comment if I am missing something. $\endgroup$ – Woeitg Feb 12 '17 at 15:20
  • $\begingroup$ Link to related unanswered question: stats.stackexchange.com/questions/217422/… $\endgroup$ – Woeitg Feb 12 '17 at 15:23
  • $\begingroup$ I don't think you can use $\frac{x-\mu}{\sigma}$ to convert Spearman's $\rho$ to z-score here. You need Fisher transformation as the linked wikipedia page on the question in your comment shows. Check this question and Nick Cox's answer: stats.stackexchange.com/questions/59714/… . I don't think it is hard to use Fisher transformation to obtain z-score. $\endgroup$ – T.E.G. - Reinstate Monica Feb 12 '17 at 15:46
  • $\begingroup$ The problem is I get different p-value using fisher transformation and t-distribution. For example in the wikipedia xample, rho=-0.1757 and p-value=0.62. Using Fisher transformation. I get F(r)=Arctan(-0.1757)=-0.1775 and z-score=SQRT((10-3)/1.06)*-0.1775=-0.456. Using normal distribution table, pvalue==2*(1-NORMSDIST(ABS(-0.456)))=0.648 which is different from wikipedia example $\endgroup$ – Woeitg Feb 12 '17 at 16:19
  • 1
    $\begingroup$ I see and I am sorry if I misled you, but don't get confused with Glen_b's answer. That question is about comparing two separate correlations ("running separate Spearman's analyses for each group"). I sent it to show that using Fisher transformation for Spearman's $\rho$ is admissible. $\endgroup$ – T.E.G. - Reinstate Monica Feb 12 '17 at 16:54
4
$\begingroup$

Let $R^x_1, R^x_2,\ldots, R^x_{n_x}$ be the ranks of $x$ and $R^y_1, R^y_2,\ldots,R^y_{n_y}$ be the ranks of $y$. Spearman correlation is determined as \begin{equation} r_s = 1 - \frac{6D}{n^3-n} \end{equation} where \begin{equation} D = \sum_{i=1}^n d_i^2 = \sum_{i=1}^n (R^x_{(i)}-R^y_{(i)})^2 \end{equation}

Significance testing of $r_s$ is based on the standard normal distribution. Using the expected value of $D$ as \begin{equation} E(D) = \frac{n^3-n}{6} = \frac{n(n-1)(n+1)}{6} \end{equation} \begin{equation} V(D) = \frac{n^2(n+1)^2(n-1)}{36} \end{equation} \begin{equation} Z = \frac{D - E(D)}{\sqrt{V(D)}} \Longrightarrow \frac{x-\mu}{\sigma} \end{equation}

$\endgroup$
  • $\begingroup$ you're welcome. It is the closed form approximation that works, so testing is not really needed, unless you mean "debug your code." $\endgroup$ – JoleT Feb 12 '17 at 17:22

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.