Spearman's correlation

Researcher

I notice that JASP doesn't haven an option to do Spearman's correlations. Has this functionality simply not been implemented yet? Or is there some sense in which Spearman's correlations are inappropriate in a Bayesian context?

EJ

Dear researcher,

JASP does compute Spearman but not yet in a Bayesian framework. We have completed a Bayesian version of Kendall's tau (paper accepted pending minor revision) and hope to add this to JASP soon. Spearman might follow suit. Nonparametric analyses are a little more tricky within the Bayesian framework, but we do believe they are important and so we aim to develop them and include them in JASP.

Cheers,
E.J.

Researcher

Thanks for the quick response. I look forward to using the Bayesian Kendaull's tau once it has been implemented.

HRH

Hi - I was just wondering if there was a timeframe for adding the Bayesian Kendall's tau? as I would be really interested in using it for my analysis.

Thank you

EJ

It will not be in the next release (scheduled for a week or so after today) but it will probably be in the one after that. I would say it is a few months away.
Cheers,
E.J.

HRH

EJ - does this mean that your paper is also imminent? I'm looking forward to reading it!

EJ

The paper has a "minor revision from The American Statistician. We've almost completed the revision and I'm pretty confident. But you never know.

Philippe

Meanwhile, you added Kendall's tau as an option for Baysian non-parametric correlation analysis. I'm not sure whether there are implemented statistics that adjust the tau coefficient for ties. Anyways, I'm interested in Bayesian Spearman correlation analysis. I assume that transforming my data into ranked data with ties and conducting a Bayesian Pearson correlation would not be equivalent to a Bayesian Spearman correlation!? Otherwise it would have been relatively simple to implement it into JASP. But please correct me if I'm wrong. Another minor remark: It would be nice to have an option to output the exact p values in a future Version of JASP.

Thank you very much!

EJ

Hi Philippe,

About the exact p-values: check out next week's blog post (https://jasp-stats.org/blog/)
About Bayesian Spearman: no, unfortunately it is not that simple! For instance, if you have N=3 and the ranks perfectly correspond, conducting a test on the rank numbers gives a BF of infinity :-)

Cheers,
E.J.

EJ

That said, we actually have developed a Bayesian version of Spearman. Paper and implementation in JASP will hopefully follow soon.
Cheers,
E.J.

JoannevanSlooten

Hi E.J,

I'm happy to read you developed a Bayesian version of the Spearman rank correlation! I wondered whether this option is already available in JASP3 (on mac OS)?

Kind regards,

Joanne

EJ

Not yet, but Johnny is working on it

We do have a Bayesian version of Kendall's tau in there (which I personally find more elegant than Spearman)

E.J.

elienbellon

Hi,

I was wondering what the progress on the Bayesian version of the Spearman rank correlation is?

Thanks!

elienbellon

EJ

Me too! I'll ping Johnny.

E.J.

JohnnyB

Yes! I was waiting till all the internal JAPS rewrites were done, which almost seems to be the case. Implementing the Bayesian Spearman's rho and one-sample Wilcoxon tests is at the top of my JASP todo list =)

cleag

Hello !

Any news on the bayesian module for spearman correlations?

Thanks a lot!

rohanp16

Just following up on this and wondering when/if there were any updates/timelines? Thank you for all the work you do! :)

JohnnyB

Sorry, there were all sorts of things that popped up instead.. it's still on our radar, just not with such high priority (since we do already have Bayesian inference for a rank-based correlation coefficient in Kendall's tau).

Kind regards

Johnny

rohanp16

Fair enough Johnny, thanks for the prompt response, appreciate it.

andersony3k

From a frequentist perspective, there's never any need to have an algorithm that specifically tests a Spearman in contrast to a Pearson's correlation coefficient. To compute and test Spearman's one need only rank-transform each of the two variables and then conduct an analysis of Pearson's R on the transformed data. Shouldn't the same hold true for a Bayesian analysis? -- just rank-transform each variable, then conduct a JASP Bayesian analysis of the Pearson Correlation between the two rank-transformed variables?

TarandeepKang

Hi All,

@andersony3k according to the paper,

van Doorn, J., Ly, A., Marsman, M., & Wagenmakers, E.-J. (2020). Bayesian rank-based hypothesis testing for the rank sum test, the signed rank test, and Spearman’s ρ. Journal of Applied Statistics, 47(16), 2984–3006. https://doi.org/10.1080/02664763.2019.1709053

the difficulty seems to relate not having a clear likelihood function available. But some of the maths looks a bit scary for my taste, so let's hope Johnny will explain it to us!

I do however another unrelated question.

I was just wondering what we make of the advantages (according to the authors) of the methods outlined in this, paper over previous Bayesian methods of computing correlations? You have developed Spearman/ Kendall coefficients and they have added alternative formulations and a polychoric correlation

Rodriguez, J. E., & Williams, D. R. (2022). Bayesian Bootstrapped Correlation Coefficients. The Quantitative Methods for Psychology, 18(1), 39–54. https://doi.org/10.20982/tqmp.18.1.p039

andersony3k

I think the spirit of the frequentist significance test is to embrace assumption-violation. The frequentist test of Pearson's r assumes bivariate normality, while the frequentist test of Spearman's rho deliberately violates the normality assumption (by applying the Pearson test to wildly non-normal--i.e., rank--data). One could take the same approach within a Bayesian framework: Do the Bayesian test of the Pearson correlation coefficient when that coefficient has be computed on ranks, despite knowing that the result won't be strictly correct!

Probably a better approach would be as follows.

Knowing that the frequentist test of Spearman's rho has never been strictly correct (and knowing that it does not do a good job at handling ties), use JASP's Bayesian Kendall's Tau-b instead. (Tau-b does appropriately account for ties.)