At 09:56 AM 5/12/2004 +0100, you wrote:
It seems to me that lining them up is
imparting structure which goes beyond
the structure of data production. Also,
if different assays are on the same footing,
taking one as response adds another arbitrary
decision.
Nick
[email protected]
Nick,
I am not so sure of your concerns here,
1. If the process that generates these pairs can be construed as having
been a random process, it doesn't seem it should be a problem. The 2nd
level variable (batch) will account for the lack of independence within
batch, the question becomes then is there a further source of non
randomness within batch.
2. If there are non random elements in the generation of the results or
pairings, for instance an assay time sequence within batch, or a location
in the assay system such as location of wells on a plate, or an order that
the aliquots etc. were obtained, couldn't it just be added as an ordinal
predictor variable?
3. Selecting one as an outcome and one as a predictor should not change the
significance of the level one relationship. It would require that the two
models be run in order to evaluate the "batch" at level two for the two
assays separately.
Alternatively, if the relationship of the within pairs is a concern, I
imagine one could come up with a random subsampling scheme that included
different pairings than the default original listing and bootstrapped the
results.
Buzz Burhans
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]]On Behalf Of
> Winfield Scott
> Burhans
> Sent: 12 May 2004 01:18
> To: [email protected]
> Subject: Re: st: RE: : unpaired regression
>
>
> John,
> One more possibility, last one from me. Assuming your
> interest is in the
> variance between batches, use either xtreg or gllamm.
>
> Line up the results as Scott suggested, then do either xtreg
> or gllamm.
> Rather than being interested in the significance of the
> coefficient on the
> predictor assay, the outcome of interest would be the significance of
> either sigma_u (xtreg) or the level two term "batch" in gllamm. In
> gllamm, you could use -gllapred- with the ustd option to identify
> specific outlier batches
>
> xtreg assay1 assay2, i(batch)
>
> or
>
> gllamm assay1 assay2, i(batch) adapt
> gllamm, allc
>
> Buzz Burhans
>
>
> >> I have two measures of batch performance on which I'd like to
> >> perform a
> >> regression. The measurements are taken on separate samples
> >> from the batch,
> >> and typically look something like:
> >> Assay1 Assay2
> >> Btch1 5400
> >> Btch1 5320
> >> Btch1 5670
> >> Btch1 0.900
> >> Btch1 0.905
> >> Btch1 0.898
> >> Btch2 8600
> >> Btch2 7840
> >> Btch2 7550
> >> Btch2 0.962
> >> Btch2 0.955
> >> Btch2 0.943
> >> ...etc (on for multiple batches which show correlated
> >> measures for the two
> >> assays)
> >> -collapse- ing them to batch averages and then performing the
> >> regression is
> >> one approach, but it doesn't take variance of the measures
> >> themselves into
> >> account in the regression. Is there a system for performing
> >> this type of
> >> analysis?
> >>
> >
> > *
> > * For searches and help try:
> > * http://www.stata.com/support/faqs/res/findit.html
> > * http://www.stata.com/support/statalist/faq
> > * http://www.ats.ucla.edu/stat/stata/
> > *
> > * For searches and help try:
> > * http://www.stata.com/support/faqs/res/findit.html
> > * http://www.stata.com/support/statalist/faq
> > * http://www.ats.ucla.edu/stat/stata/
> >
>
> *
> * For searches and help try:
> * http://www.stata.com/support/faqs/res/findit.html
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
Buzz Burhans
[email protected]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
