I have two subgroups of structure (bone structure) and I want to test if there is any difference of size (area) between them, and if there is, how important this difference is. The first set is a group non-heated, and the second was heated. I wanted to do a Bayesian T-test (Mann-Whitney because my data isn't normal and the variance of the two groups isn't equal).

I am just not sure about the prior distribution to set for the test. I understand what is a prior distribution but I am not sure if I have any ... My take is that if I had some information about my priors, I think it would be more clear than that and I would know. Is it better to keep the default prior in JASP (cauchy = 0,707) if I am not sure about the priors distribution (I've seen this in some publications)? Is it often that we can't add some priors knowledge to those tests?

  • $\begingroup$ ". I wanted to do a Bayesian T-test (Mann-Whitney because my data isn't normal and the variance of the two groups isn't equal)." - what does this mean? $\endgroup$ – AdamO Apr 8 at 16:23

You have mixed a couple of ideas. A $t$-test is a parametric test (although the Bayesian $t$-test can relax the assumptions) while the Wilcoxon test treats Y as ordinal and uses only its rank ordering. To get a Wilcoxon test, use its generalization the proportional odds ordinal logistic regression model. The blrm function in the R rmsb package makes this easy and detailed case studies may be found here.

For the Bayesian $t$-test read this and see examples in BBR. BBR also has more background information about using the proportional odds model to get a Wilcoxon test.

  • $\begingroup$ Thank you @FrankHarrell. I read the section about assumption being more flexible with bayesian model in the link BBR. I thought they were as important as in the ''classic'' t-test, and particularly after reading some publication. I have a LOT of tests to do,~600 groups to compare (300 non-heated and 300 of the heated counterpart) and for 10 variables. Not all the samples are normal, not all the variance are equal, but the majority are. I wanted to apply the same test to all for comparison purpose. I am not sure how a bayesian model can relaxe those assumptions and if it can apply to my data? $\endgroup$ – user317954 Apr 8 at 14:23
  • $\begingroup$ Like this publication (researchgate.net/publication/…) $\endgroup$ – user317954 Apr 8 at 14:24
  • 1
    $\begingroup$ The most important Bayesian contribution besides making a $t$-test much more robust is in handling the 600 groups and 10 variables by having a unified model, perhaps a hierarchical one that borrows information. Any time you rely on significance tests, and especially any time do you many of them, expect bad things to happen. $\endgroup$ – Frank Harrell Apr 8 at 14:58

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.