Skip to content

Bonferroni correction for multiple correlation tests

7 messages · Louise Cowpertwait, R. Michael Weylandt, David Winsemius +4 more

#
Please can someone advise me how I can adjust correlations using bonferroni's correction? I am doing manny correlation tests as part of an investigation of the validity/reliability of a psychometric measure.
Help would be so appreciated!
Cheers,
Louise
#
On Wed, Aug 29, 2012 at 6:23 PM, Louise Cowpertwait
<louisecowpertwait at gmail.com> wrote:
The observed correlation is an immutable property of the observed data
and the Bonferroni correction does not change it. Rather, it should be
applied to the p-values of the observed correlations, much as it would
be for any test. Those more statistically savy than I might jump in,
but I don't see why the p-values of, e.g., cor.test() would be
adjusted in a different way than those of t.test().

Consider a similar case for a set of t-tests: you see some data and do
the tests based on the sample means. It doesn't make any sense to
"adjust the mean" of your data, rather you might wish to adjust your
_interpretation_ of calculated p-values to account for multiple
comparisons.

Cheers,
Michael
#
On Aug 29, 2012, at 4:23 PM, Louise Cowpertwait wrote:

            
?p.adjust
David Winsemius, MD
Alameda, CA, USA
#
Indeed...

But it seems clear that,

1. This is off topic for this list.
2. Ms. Cowpertwait would be wise to seek the help of a local statistician.

does it not?

Cheers,
Bert

On Wed, Aug 29, 2012 at 6:48 PM, R. Michael Weylandt
<michael.weylandt at gmail.com> wrote:

  
    
#
On Wed, Aug 29, 2012 at 6:48 PM, R. Michael Weylandt
<michael.weylandt at gmail.com> wrote:
I am happy to be corrected, but under specific situations, I can see
an alternative correction method being appropriate.  For p variables,
the p x p correlation matrix has p * (p - 1) / 2 unique correlations,
however, once you know about some of the correlations, you actually
have some information about the other correlations.

Imagine the situation where p = 3 and cor(p1, p2) = .9, cor(p2, p3) =
0.  Is cor(p1, p3) free to be any possible correlation?  The answer of
course is no.  I am not sure what the exact rule would be, but this
would hold and increase for larger matrices.

  
    
#
At 00:23 30/08/2012, Louise Cowpertwait wrote:
Louise, apart from the excellent advice you have already received 
from others on the list I would suggest that if you really, really do 
not have a scientific theory and need go data fishing the correction 
named after Bonferroni is unlikely to be the best approach. There are 
other methods available in p.adjust to which your attention has 
already been drawn.
Michael Dewey
info at aghmed.fsnet.co.uk
http://www.aghmed.fsnet.co.uk/home.html