The first issue for the student is to determine a sufficient sample
size. I'd suggest N > 100, the more the better.
We then will have a series of triples (Ai,Bi,Ci).
Take a subsample C* of the {Ci} with replacement. Compute rho(A,C*) and
rho(B,C*).
Use permutation methods to compare the two samples of rhos.
Phillip Good
who first used such a combined test in 1975.
-------- Original Message --------
Subject: Re: Randomisation test to compare dependent correlations
From: "Grant, Robert L" <[log in to unmask]>
Date: Tue, August 19, 2014 7:01 am
To: [log in to unmask]
I should have made it clear these three measures are performed on the
same study participant at the same time. Thanks Stas Kolenikov for
asking the crucial question there.
--
Robert Grant
Senior Lecturer in Health & Social Care Statistics
Kingston University & St George’s, University of London
[log in to unmask]
http://www.robertgrantstats.co.uk
@robertstats
From: Grant, Robert L
Sent: 19 August 2014 13:47
To: [log in to unmask]
Subject: Randomisation test to compare dependent correlations
Hello Allstaters
I’m supervising a student who will soon have data comparing two
potential measures A & B against a “gold standard” C. Null
hypothesis is that corr(A,C)=corr(B,C). I know of a t-test
implementation in the R package ‘psych’ which uses Fisher’s
z-transformation, but I am thinking that a randomisation test could be
more robust (could work with Spearman’s, for example) and simpler to
understand. Is anyone aware of any precedent for this? And would
randomly swapping A & B values be a sensible way to do it?
--
Robert Grant
Senior Lecturer in Health & Social Care Statistics
Kingston University & St George’s, University of London
[log in to unmask]
http://www.robertgrantstats.co.uk
@robertstats
This email has been scanned for all viruses by the MessageLabs Email
Security System.
You may leave the list at any time by sending the command SIGNOFF
allstat
to [log in to unmask], leaving the subject line blank.
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|