On Thu, 10 Oct 2002 11:37:01 -0400, Bruce Weaver
<[EMAIL PROTECTED]> wrote:

> On Thu, 10 Oct 2002, Rich Ulrich wrote:
> 
> > On Thu, 10 Oct 2002 14:05:08 +0200, Jan Malte Wiener
> > <[EMAIL PROTECTED]> wrote:
> >
> > > hi,
> > > hopefully a simple one ->
> > >
> > > let's assume i have data like this:
> > > subject A: 1,0,1,1,0,1,1,1 -> mean=6/8=0.75
> > > subject B: 1,1,1     -> mean=3/3=1
> > >
> > > weighted arithmetic mean of A+B-> (8*0.75 + 3*1)/11 = 0.82
> > > -> well that was easy, but what if i do have 20 subjects like this and i
> > > want to compare their weighted arithmetic mean to the weighted mean of
> > > another group of 20 subjects ?? i guess i need to weight every single
> > > subject-mean before running any stat-test. and here is my problem: how
> > > do i weight the individual subject means ??
> >
> > Statistically speaking, that looks bad.  You don't have
> > this problem when comparing 20 vs 20 subjects,
> > if you are doing an Analysis of variance in the usual, legal,
> > legitimate way; you want to count each Subject once.
> > Weighting creates a problem of logic in computing 'error';
> > it is a problem of whether the analysis is by  ANOVA.
> >
> > If you still want to get the weighted average that you describe,
> > use "WEIGHT" .
> 
> 
> I just want to make sure I follow you here, Rich.  You're talking about
> WEIGHT as in a weighted regression (or GLM), right?  (As opposed to WEIGHT
> CASES BY some frequency count.)  I presume one would weight each subject's
> mean by the inverse of its variance, right?


The question (and example) was about weighting by the 
frequency count.  My answer was about the frequency
count, and counting a subject several times is the sort of 
weighting that screws up any simple t-test,  or correlation, 
etc.  
 -- When a sample in survey research is 're-weighted'
to match a population, there is a simple and crude approach
which is sometimes adequate.  Samples are weighted by 
proportion, and Ns are further  tweaked so the original N results.
The tests produced that way are approximately correct so long
as the adjustments are not far from 1.0,  and don't change 
(much) the effective standard errors. 
 -- The example here has counts that (a) count one person 
more than once, so that the DF  is grossly overstated, and 
(b) are notably unbalanced.

You raise a different question.  I have worried much less
about what you mention, weighting by variance.  But that, too, 
is more common when *testing*  is not at issue (I think).  


Does my argument, now or earlier, overlook something?
 - I faced survey questions once or twice, long ago, and
right now, I am working from memory  and from extrapolation 
from research principles.

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
.
.
=================================================================
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