Date: Fri, 9 Jun 2000 08:50:48 -0700
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Dale Glaser <glaser@PACIFIC-SCIENCE.COM>
Subject: Re: eta squared in UNIVARIATE GLM
Content-Type: text/plain; charset="iso-8859-1"
If I recall, technically it is a partial eta-squared, as opposed to eta
squared; if you use the MANOVA option (via syntax only), it provides both
eta squaredand partial eta squared, whereas with the GLM option only the
partial is reported even though it is mistakenly labeled as eta squared
(Dave Nichols please correct me if I'm off base here)..........dale glaser
From: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU] On Behalf Of
Sent: Thursday, June 08, 2000 9:01 PM
Subject: eta squared in UNIVARIATE GLM
I am trying to come to grips with the eta squared term in the UNIVARIATE GLM
procedure of SPSS (using version 9.0). I am using one way and 2-way
factorial ANOVAs to produce power and effect size estimates.
Several textbooks mention a term labelled eta squared or R-squared which
appears to be interpreted in the same way as the eta squared value in SPSS
(percent variance explained). Textbooks such as Cohen (Statistical power
analysis for the behavioral sciences) and Kirk (Experimental design) mention
that eta squared should be related to effect size in the following manner:
f-squared = eta squared / (1- eta squared)
However, when I attempt to use a text book derived formula for estimating
eta squared and then compare it to the SPSS value, they are not the same!
This occurs even in a 1-way ANOVA (using a simple example) when degrees of
freedom estimation is straight forward and not term dependent.
Can anyone explain why the SPSS eta-squared term does not appear to be
related to the effect size term in the way that Cohen and Kirk and others
say that it should be?
Nigel Perkins, BVSc, MS, Dip ACT, FACVSc
Senior Lecturer, Veterinary Epidemiology
Private Bag 11222