http://spssx-discussion.165.s1.nabble.com/How-data-entry-for-TURF-analysis-tp4635246p4657823.html
For a computation formula, see
http://en.wikipedia.org/wiki/Brown%E2%80%93Forsythe_test
Having zero cases is a design problem. Does SPSS not drop the
missing category for you? - Sure, drop the zero category.
The denominator is a double summation, so there is no reason for
any division by zero - that can prevent some statistics, some times.
Is the program *saying* there is a zero cell, preventing the statistic?
Saying that "F statistics could not be used" is a mis-statement, as a
generality. The simple ANOVA tests are rather robust.
I think it was Frederick Lord who exaggerated, saying something like,
"Using a variance test to check the validity of a t-test is like using a
canoe to check the water conditions for the safety of a liner."
That's especially true when Ns are equal, as in designed experiments.
(He wrote that in the 1950s or so, back when most data was of that kind.)
Generally, homogeneity testing is useful for moderate size samples -
ANOVA is especially robust when the Ns are large, and the
variance test at the usual nominal level will reject far too often.
And the tests have almost no power when the Ns are small.
- More about the testing -
Sources that I read a dozen years ago, concerning the Levene test
and Student's t-test, seemed persuasive in arguing that one should
never "condition" your choice of the two t-tests on the outcome of
the variance test. (What to do instead was less consistent.) I believe
that applies elsewhere.
Unequal variance may bias the result in either direction, depending on
whether the large-variance group has the large N or small N.
As a practical matter, you should scan your data, and learn about
how it is generated; surprisingly often, dirty data needs correcting.
After that, I've avoided variance problems by using some natural
transformation, most often log or square root.
Otherwise, if you have reason to expect unequal variances, you
should expect to correct for it unless it turns out to be small enough
to ignore.
Hope this helps.
--
Rich Ulrich
Date: Tue, 2 Aug 2011 03:45:47 +0000
From:
[hidden email]Subject: Brown-Forsythe Issue in ANOVA
To:
[hidden email]
Hi,
I have some queries on Brown-Forsythe.
Query 1
I have an issue with ANOVA. One of my categories is having a 0 variance. Hence SPSS could not calculate the Brown-Forsythe's Statistics. The Levene Statistics showed that the homogeneity of variance was not met and hence F Statistics could not be use. I have checked and found out that the reason why one of my categories is having a 0 variance is because no respondents fell under this category. My instinct is to remove this category and perform an ANOVA again with K-1 categories. Is this a correct approach?
Query 2
My case happened because one category has 0 respondent in it and this caused the 0 variance issue. However I would like to also check if anyone had this similar issue before but under the following condition. (I kind of doubt the possibility of a 0 variance with all the categories !
having at least one respondents though.)
- All categories have respondents in them.
- At least one of the categories have 0 variance and caused SPSS unable to calculate Brown-Forsythe's Statistics.
- Lastly, the work around for this situation.
Thanks.
Dorraj Oet