Re: [Gretl-users] Behrens-Fisher problem
On Thu, 8 Jun 2006, Sven Schreiber wrote:
Thanks Allin; ok, let's see if I understood your replies
correctly:
1. The stand. dev. formula that is used in the 2-means test depends on
whether both samples are large (n>30) or not.
Yes, that's right.
2. If one of the samples is "small", then a formula is used
that appears
to be valid only in the case of equal variances.
True.
3. there is no known/accepted way to test an arbitrary difference of
means in small samples with unequal (and unknown) variances
So far as I understand it, yes.
Provided those assessments are correct, I would suggest the
following:
ad 1) I would like to see some mention of which assumption/formula is
used in the test output, or make it a user option.
Yes, that should be done.
ad 2) This is unfortunate given that the user explicitly choose the
option "do not assume equal variances". The easiest workaround is to use
the asymptotic formula instead and print a warning in the output that
the approximation may be poor due to the sample size.
OK, we'll do that.
Allin.