http://spssx-discussion.165.s1.nabble.com/How-data-entry-for-TURF-analysis-tp4635246p4660764.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
command. To leave the list, send the command