Hi,
I've received several replies to my previous post, but none of them
claimed to give an authoritative view on which statistic is best for use
in general contingency tables. But I suddenly remembered that one crucial
difference between Pearson's Chi-square and all the others is that only
Pearson's does not assume fixed marginals.
Now it also happened that in answer to one of my previous questions, Prof
David Cox gave me an email saying that in fact there is a living debate as
to whether or not margins should be fixed. The details of this debate I
don't know, but seems to be largely philosophical. I haven't heard the
case for fixing the margins, but I have heard one reason for not fixing
it. Howell (Statistical Methods for Psychology) said that in most cases
when we collect data, we don't fix the margins, so why should we when we
analyse the data? This seems convincing to me, but then the problem is
there is no statistic available, as far as I know, to deal with the case
when there are small counts.
Anyone would like to continue the discussion?
Tim
|