Dear Mike, Dean, Ian (and Andrea, who kindly reached out separately),

Thank you so much, I really appreciate all your information, which makes so 
much sense! I will move forward with this in the future, particularly 
remembering the limitation on p-values by the number of permutations and 
(perhaps most importantly) the reminder that “significant” doesn’t always mean 
“biologically meaningful”.

All the best and thanks again,

Vera

From: [email protected] <[email protected]> On Behalf Of Mike 
Collyer
Sent: Tuesday, 26 May 2026 12:02 AM
To: Ian Dworkin <[email protected]>
Cc: Vera Weisbecker <[email protected]>; Morphmet 
<[email protected]>
Subject: Re: [MORPHMET2] Should multiple comparisons with pairwise() have 
p-value adjustments?

CAUTION:  External message. Only open links and attachments you're expecting.
Dear Vera,

Actually, summary.pairwise provides an option for adjusting the confidence 
level, if one wishes to use a different pairwise alpha than the experiment-wise 
alpha.  For example, one can use confidence = 0.99, if the pairwise alpha is 
deemed to be 0.01.  The upper confidence limits will adjust accordingly.

Regarding whether one *should*, we leave the opinion to the user rather than 
impose it, but there is one thing I wish to point out, with respect to 
permutational software.  Some software will subset data to include only 
observations associated with two groups and perform multiple two-sample tests, 
perhaps with different random permutations in each test.  RRPP does not do 
that.  It fits, say 10,000, linear models with10,000 sets of coefficients 
(rather than 10,000 * g * (g - 1) / 2 sets of coefficients that would be 
calculated for the all pairwise two-sample tests of g groups).  From the 10,000 
sets of coefficients, 10,000 matrices of pairwise distances among group means 
can be estimated.  So, for example, the 1,427th permutation of the pairwise 
distance between the 3rd and 4th group means uses the exact same randomization 
of residuals as the 1,427th permutation of the pairwise distance between the 
1st and 10th means (because both are estimated form the 1,427th estimation of 
coefficients).

If there were some borderline cases that I worry could by type I errors, I 
might consider the following, in addition to pairwise alpha:

1. The seed used (try the test changing the seed a few times to see how stable 
the P-values are).
2. The number of permutations (more is better).
3. Sample size (I might be more worried about a type I error with smaller 
samples).
4. Most importantly, effect sizes. (If a pairwise distance is “significant” but 
the effect size is small, does its significance mean as much, biologically?)

As a final comment, the P-values in RRPP::pairwise are limited by the number of 
permutations.  If 1,000 permutations are used, 0.001 is a lower limit.  
Therefore, a pairwise alpha that is really small because there are many groups 
to compare might also require many permutations to get P-values that are closer 
to exact values.  Failing to reject the null hypothesis that a pairwise 
distance equals 0 because a P-value is 0.001 would be perhaps unfortunate and 
using 500,000 random permutations perhaps excessive, where common sense 
(evaluation of effect sizes) would probably be more useful.

Hope that helps!

Mike


On May 25, 2026, at 9:23 AM, Ian Dworkin 
<[email protected]<mailto:[email protected]>> wrote:

HI Vera,

 pairwise() in geomorph/RRPP does not do any adjustments. The last I spoke to 
Mike and Dean about (when I prepped an online seminar about it during the 
pandemic... so a while ago), it was in part because it provides ALL pairwise 
comparisons. One of the things I point out to people using it, is that you need 
to a priori plan which you are going to use, and do your adjustments based upon 
that.

Cheers
Ian

On Mon, 25 May 2026 at 06:33, 
[email protected]<mailto:[email protected]> 
<[email protected]<mailto:[email protected]>> wrote:

Hi everyone,

I was recently asked if I implement any multiple comparison adjustments (like 
Holm) when using the pairwise() function in geomorph on RRPP-based 
procD.lm/procD.pgls fits involving a factor with more than two levels. I admit 
that this never crossed my mind, but I can’t see why there should not be a 
higher risk of getting false H0 rejections with increasing comparisons, as it 
would in other multiple comparison contexts.

It is easy to do an adjustment manually, but I do not think I have seen it done 
anywhere – it is not part of the pairwise implementation as far as I can see, 
and I did not spot a comment on it in the Collyer & Adams 2018 paper on RRPP. I 
was wondering if there is a reason for this, perhaps related to the residual 
randomisation procedure or building of confidence intervals?
Thanks a lot for any suggestions!

Vera

--
You received this message because you are subscribed to the Google Groups 
"Morphmet" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to 
[email protected]<mailto:[email protected]>.
To view this discussion visit 
https://groups.google.com/d/msgid/morphmet2/cdfd532c-084e-4dbd-bc7a-190f6808ed2cn%40googlegroups.com<https://groups.google.com/d/msgid/morphmet2/cdfd532c-084e-4dbd-bc7a-190f6808ed2cn%40googlegroups.com?utm_medium=email&utm_source=footer>.


--
Ian Dworkin
Department of Biology
McMaster University
Office phone 905 525 9140 ext. 21775
Lab phone 905 525 9140 ext. 20076
[email protected]<mailto:[email protected]>
dworkinlab.github.io<http://dworkinlab.github.io/>


--
You received this message because you are subscribed to the Google Groups 
"Morphmet" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to 
[email protected]<mailto:[email protected]>.
To view this discussion visit 
https://groups.google.com/d/msgid/morphmet2/CAGudrj%3DZLnucD8G3k5qDFXzKjWBZrdWg_bX5z9Q90Pt90-hNCQ%40mail.gmail.com<https://groups.google.com/d/msgid/morphmet2/CAGudrj%3DZLnucD8G3k5qDFXzKjWBZrdWg_bX5z9Q90Pt90-hNCQ%40mail.gmail.com?utm_medium=email&utm_source=footer>.

--
You received this message because you are subscribed to the Google Groups 
"Morphmet" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to 
[email protected]<mailto:[email protected]>.
To view this discussion visit 
https://groups.google.com/d/msgid/morphmet2/4E1E1FC1-6AE3-4891-B870-CBE4B5C832A0%40gmail.com<https://groups.google.com/d/msgid/morphmet2/4E1E1FC1-6AE3-4891-B870-CBE4B5C832A0%40gmail.com?utm_medium=email&utm_source=footer>.

-- 
You received this message because you are subscribed to the Google Groups 
"Morphmet" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion visit 
https://groups.google.com/d/msgid/morphmet2/SY0PR01MB109054AB6907CA5D29D555C77A90A2%40SY0PR01MB10905.ausprd01.prod.outlook.com.

Reply via email to