I concur.  I should have said "if TI used that approach it might explain 
the small differences in the results."



Art


Rich Ulrich wrote:
> posted and e-mailed.
> 
> On Mon, 06 Jan 2003 22:29:56 GMT, "Arthur J. Kendall"
> <[EMAIL PROTECTED]> wrote:
> 
> [ ... ]
> 
>>It looks as if the TI is using about 8.1 df. (or the harmonic mean of 
>>[10,7] ) which gives fractional df.
>>
> 
> I think you nailed that one.
> 
>>Since one SD is twice the other, the Levene F for non homogeneous 
>>variances is about 4.  So the "separate variance" interval estimate 
>>would be more appropriate and would result in small differences.
>>
> 
> 
>  - of course, you are not supposed to make the choice of
> test on this ad-hoc basis.  
> 
> To be more explicit: according to a couple of
> articles that I read (and I agreed on this point), it is a BAD
> practice to  'condition'  your choice.  You are not
> supposed to let the test of variances  determine 
> which t-test to believe.
> 
> For equal Ns, the difference in tests is *very* slight.  
> For unequal Ns, both the versions of t-tests mis-behave 
> rather badly in their  one-tail rejection rates.  The
> separate-variance tests (there are a couple of versions) 
> are  minimally more robust than the Students t, but not 
> enough (in my opinion) to justify favoring them.
>  - Insist on both tests showing the same, or consider
> transformations.  Your  "best test" will come if you find
> the transformation that is both logical and normalizing.
> A rank order transformation (non parametric test) is 
> usually available and convenient, and almost as good.
> 

.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to