You can test the interaction in a log-linear model with either the pearson
chi-square or the likelihood ratio chi-square.
Cheers,
David Cross
On Tue, 23 Nov 1999, Rich Ulrich wrote:
> On 23 Nov 1999 11:05:51 -0800, [EMAIL PROTECTED] (David
> Cross/Psych Dept/TCU) wrote:
>
> > Jin:
> >
> > Pearson's test is "conditional" only on the sample size (N) being fixed.
> > This is the classic test of independence, and also the test of interaction
> > in a log-linear model.
>
> - isn't the test in the log-linear model done with a Likelihood
> chisquare?
>
> > On Fri, 19 Nov 1999, Jin Kim wrote:
> >
> > > Dear list members
> > >
> > > I am a student studying contingency table analysis these days.
> > > I have a question about the underlying statistical concept of Pearson
> > > chi^2 test.
> > > My question is:
> > >
> > > Does Pearson chi^2 test assume that both row and column margins are
> > > fixed?
> > >
> > > In other words, I wish to know whether Pearson chi^2 test is
> > > 'conditional' like Fisher's exact test.
>
> I don't usually worry about definitions like this one --
> Does 'conditional' mean exactly the same as 'fixed margins'? The
> original theoretical development of Pearson's assumed margins were
> fixed.
>
> See F. Yates, 1984, "Tests of significance for 2x2 contingency
> tables', Journal Royal Statis. Soc., Ser A, 147:426-449. This is the
> same Yates who devised Yates's correction factor in 1934. (See
> references in Zar's textbook.) In JRSS, Yates chuckled at the fact
> that the assumptions for all three tests were formally the same. What
> I don't remember is whether they *have* to be the same, which does
> not seem right.
>
> The Pearson test with the Yates correction does a better job of
> reproducing Fisher's Exact test; the Pearson test without the
> correction does a better job of reproducing what you actually get by
> randomization of the fractions on either or both margins -- so,
> pragmatically, Pearson's behaves as if it were not on fixed margins,
> even though it was created with fixed margins. (I saw at least three
> ways to 'develop' the test when I took the relevant course; I don't
> think that *all* the ways assumed fixed margins. But, did they?)
>
> --
> Rich Ulrich, [EMAIL PROTECTED]
> http://www.pitt.edu/~wpilib/index.html
>