0
$\begingroup$

I was asked for an alternative procedure to multiple comparisons which does not require independence of the multiple tests being applied. I´ve found some readings where Bonferroni´s correction is said to not require independence, I was presented Bonferroni´s correction before but not informed of this neither able to check.

So, could anyone give me an example of such procedure and/or clarify on the independence assumptions for multiple comparisson corrections.

Thanks in advance.

PS. Sorry, in case this is a duplicated question.

EDIT: The answer to the possible duplicate isn't what I needed. In fact, the question there is "...what is positive regression dependency and what is non-positive regression dependency?" and the OP is, basically, concerned about the Benjamini–Hochberg procedure while I'm looking for a more general answer.

$\endgroup$
  • $\begingroup$ Possible duplicate of FDR correction when tests are correlated $\endgroup$ – shadowtalker Apr 7 '16 at 19:36
  • $\begingroup$ My impression is the linked question isn't an ideal answer for the OP here. It isn't obvious that it provides "an example of such procedure" or clarifies "the independence assumptions for multiple comparisson corrections". The answer there is decent, but very spase, IMHO. $\endgroup$ – gung - Reinstate Monica Apr 7 '16 at 20:07
1
$\begingroup$

The Bonferroni, Bonferroni-Holm and more generally closed testing procedures constructed using Bonferroni tests do not require an independence of the test statistics. Bonferroni-Holm and suitably constructured closed testing procedures will be uniformly more powerful than a simple Bonferroni procedure so that I would normally prefer them. The Simes test / Hochberg procedure require non-negative correlation between the test statistics in order to preserve one-sided familwise type I error rates. Various other procedures e.g. Dunnett or Dunnett-Tamahane require specific known correlations, which would be the case e.g. in case multiple groups are compared to a single reference group. Depending on what kind of application we are talking about, it may also be possible to estimate the correlation from the data or to do some kind of bootstrapping procedure that accounts for correlations.

$\endgroup$
  • $\begingroup$ sorry for my novice doubt and wikipedia link, is it link what you mean by "closed testing procedure"? Is this case, how does it effect an approach for correcting p-values? $\endgroup$ – Raimundonetto1 Apr 8 '16 at 1:56

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.