Bayesian Kruskal Wallis H Test

Comments

• EJ

  December 3

  I don't think there is, at the moment. There may be, but it will probably involve some effort to get it to work.

  EJ
• TarandeepKang

  December 3 edited December 3

  Hi EJ

  someone has been asking about this in the forum:

  https://github.com/jasp-stats/jasp-issues/issues/284

  
  

  And it looks as if part five of this paper might help at least a bit?

  Yuan, Y., & Johnson, V. E. (2008). Bayesian Hypothesis Tests Using Nonparametric Statistics. Statistica Sinica, 18(3), 1185–1200.

  
  

  Best,

  
  

  Tarandeep
• Michael_Jasper

  December 3

  On that link it suggested converting to ranks and then using ANOVA. When doing this using frequentist stats first as a sanity check before trying Bayesian I got a non-significant result using ranks data and ANOVA whereas got a very significant result using non-ranks data with Kruskall-Wallis H test. And so I'm not convinced of the interchangeability here. Any comment(s) welcomed. A shame because I was excited by this at first, thinking it could be a magic bullet.
• Michael_Jasper

  December 3

  @EJ  Is there any way to get a BF10 factor from a p-value from the Bayesian Kruskal Wallis H Test? Even if just some approximation. I saw you're working on p-value to Bayes factor approximations: have you anything suitable? Understanding that it would be an approximation. Perhaps any approximation made for one-way ANOVA would be the closet around presently?
• Michael_Jasper

  December 4

  @EJ Reference my last post above, just noticed that @TarandeepKang wrote this which seems to be a very good option (all be it using the test statistic rather than p-value). Any comment on this option if you have any is welcomed. This seems a very nice option. Not a too simple equation but an equation to use nonetheless.

  "And it looks as if part five of this paper might help at least a bit?

  Yuan, Y., & Johnson, V. E. (2008). Bayesian Hypothesis Tests Using Nonparametric Statistics. Statistica Sinica, 18(3), 1185–1200."
• EJ

  December 6

  I am still looking in to these sorts of problems, with an eye to application of the 3p*sqrt(n) rule (https://osf.io/preprints/psyarxiv/egydq).

  EJ