Date: Fri, 28 Mar 2003 12:49:48 -0800
Reply-To: BRUNDAGE Thomas W <Thomas.W.Brundage@STATE.OR.US>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: BRUNDAGE Thomas W <Thomas.W.Brundage@STATE.OR.US>
Subject: Re: Testing the global null hypothesis
That is the issue of experiment-wise error versus test-wise error.
If your whole experiment (the global test) is not significant, should you
bother to look for 'significant results' in any portion of the model dependent
on the whole?
A second point - if you are doing a series of statistical tests on the same
data, then you should know how many such tests you will be doing before you
start, and adjust the alpha to reflect that number. This is the Bonferroni
adjustment to the value of alpha, and is meant to provide an experiment-wise
level of protection at the test-wise level. For example, if you are doing a
study and want an alpha of 0.05 (global) and plan on doing 12 tests, then you
would divide the 0.05 by 12 to get 0.0042. If any of your individual tests
are significant at this level you can be comfortable that the results are not
spurious.
thomasb
thomas brundage
statistician
Department of Human Services
Oregon Public Health
800 NE Oregon St., #827
Portland OR 97232
voice: 503-731-3448
FAX: 503-872-5398
>>> rbyers@HSC.USF.EDU 3/28/2003 11:25:00 AM >>>
I am running logistic regression and there is a discussion here about
satisfying the global null test with a p>/= 0.05 before you can use the
variables that are significant i.e. have a p>/= 0.05.
The question is does the model need to have a significant global null
test before on can use the variables in the model that are significant
in asserting a relationship exists.
Robert W. Byers, Ph.D., M.S.W.
Coordinator of Research & Statistics
The Harrell Center for the Study of Family Violence
College of Public Health
University of South Florida
13301 Bruce B. Downs Blvd. MHC 1701
Tampa, FL 33612-3807
Tel: 813-974-7835
Fax: 813-974-7830
email: rbyers@hsc.usf.edu
"Let everyone become all they are capable of becoming."
