NetNews Usenet Archive 1992 #26/NN_1992_26.iso/spool/bit/listserv/statl/1898

Text File | 1992-11-05 | 2.5 KB | 52 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!gatech!paladin.american.edu!auvm!SEQ.HULL.AC.UK!R.A.REESE
Via:         UK.AC.HULL.SEQ;  5 NOV 92 14:39:29 GMT
X-Mailer:    ELM [version 2.3 PL11]
Original-Sender: [e-mail]
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.stat-l
Date:         Thu, 5 Nov 1992 14:33:38 WET
Sender:       "STATISTICAL CONSULTING" <[e-mail]>
From:         "R.A.Reese" <[e-mail]>
Subject:      Re: differences between correlation coeficients
In-Reply-To:  <no.id>; from "John Wendell" at Nov 4, 92 3:41 pm
Lines: 37

> measurements. Several questions: (1) Can these T-values be derived
> from the correlation coefficients and the sample size alone (and if so
> how, and if not how would they be derived from the raw data)? (2) If
> there was also a Y3 in the above table how would a test of an over-all
> difference in correlation coefficients be done (an F-test perhaps)?

Aloha!

First question: dunno, but I don't like to make any inferences based
only on the quoted correlations.  People that quote correlations and
p-values in isolation in my experience haven't looked at plots and
don't appreciate the limitations of testing a strictly linear effect.

How would I tackle it? Given the raw data, I'd use GLIM or similar
software to fit a regression to the combined data then add in a parameter
for the group. This gives an overall test for the improvement in fit
by allowing a different slope for each group (or a different intercept)
and each coefficient can be tested by a t-test. In the case of two groups,
these tests are equivalent; if more than two groups you have the standard
problem of interpreting multiple t's in the context of the overall test.
If you normalize the data and suppress the intercept, this is equivalent
to comparing the two correlations.

This is a good example of how the Generalized approach to data modelling
and hypothesis testing is a better m.o. than the cookbook approach of
looking for a specific test to meet each situation.  A good example paper
is "Statistical Modelling of a pharmacological experiment" by M.A.Adena,
The Statistician Vol 31 No 4 Dec 1982 pp279-292.

Sayonara ... to get me home via the pretty route.

--
(R.) Allan Reese        Janet:          [e-mail]
Head of Applications    Direct voice:   [phone]
Computer Centre         Voice messages: [phone]
Hull University         Fax:            [phone]
Hull HU6 7RX, U.K.