Statalist


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: RE: regression assumption question


From   Steven Samuels <sjhsamuels@earthlink.net>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: RE: regression assumption question
Date   Tue, 10 Mar 2009 15:22:33 -0400



I suggest that Moleps get get a copy of

DR Bock, Multivariate Methods in Behavioral Research, 1975, McGraw Hill, Section 7.3 (p 489+) or check the URL below. Bock's description applies to the comparison of two groups. The following is a sumary

If subjects were randomized to comparison groups (unlikely from Molep's description), then both analyses estimate the same quantity, but ANCOVA is more powerful.

If subjects were not randomized to groups, but can be considered as randomly sampled, then ANCOVA and analysis of change scores estimate different quantities. This is "Lord's Paradox. (Lord FM: A paradox in the interpretation of group comparisons. Psychol Bull 1967, 68:304-5. There is a nice discussion at: http://www.ete-online.com/content/ 5/1/2#B21) .

With the proviso that subjects are drawn randomly from defined populations (maybe hypothetical), the choice of analysis method depends on the purpose of the study:

If the purpose is, in Bock's words, to "determine whether a difference between the group mean final scores can be attributed to the difference in baseline scores," use analysis of covariance.

If the purpose is to assess effects on changes, then analyze the change scores.

As Austin states, problems arise because the baseline variable is measured with error. There are other issues if, in addition, assignment to comparison group or selection to follow-up was made on the basis of a factor correlated with the baseline score.

-Steve

On Mar 10, 2009, at 1:01 PM, Austin Nichols wrote:

moleps islon:
I tend to agree with Tony <Peter.Lachenbruch@oregonstate.edu>, but
note that if you believe that X affects the change in Y (dY = Ypost -
Ypre) but Y is measured with error pre and post, you may prefer not to
regress Ypost on Ypre and X, but rather to regress dY on X.  Since
Ypre is measured with error, its coef may be biased toward zero and
you may be able to reject the null that its coef is one even when that
is the true model, and you may also have bias in other coefs when you
include Ypre, esp if treatment levels are correlated with true
baseline Y levels.

mat c=(1,0,0,0\ 0,1,0,0.5\ 0,0,1,0\ 0,0.5,0,1)
drawnorm e1 y0 e0 x, n(1000) corr(c) seed(1) clear
g ypost=y0+x/2+e1
g ypre=y0+e0
g dy=ypost-ypre
reg ypost ypre x, nohe r
reg dy x, nohe r

But of course if Ypre does not have a coef of one in the true model,
you will introduce bias by imposing that constraint in a regression of
dY on X.   Nearest-neighbor matching (findit nnmatch) on pre-treatment
observables is another way forward here...

On Mon, Mar 9, 2009 at 3:58 PM, Lachenbruch, Peter
<Peter.Lachenbruch@oregonstate.edu> wrote:
My preference is to use the preop measure as a covariate.  If you use
the change, you are essentially forcing the preop to have a coefficient
of 1.  Sometimes people use the preop as a covariate for the change
score - this automatically induces a fairly high correlation with preop
- if you're not careful, you can believe it.

Tony

Peter A. Lachenbruch
Department of Public Health
Oregon State University
Corvallis, OR 97330
Phone: 541-737-3832
FAX: 541-737-4001


-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of moleps islon
Sent: Monday, March 09, 2009 11:49 AM
To: statalist@hsphsun2.harvard.edu
Subject: st: regression assumption question

Dear listers,
I've got data on 300 patients preop and postop using the VAS scale
(ordinal scale). I'm trying to locate factors predicting improvement
postop. However there are several questions pertaining to this that
I'm unsure of. 1)Do I violate the assumption of independence? I assume
there would be some correlation between preop and postop within the
patients. 2)Would you recommend using delta (preop-postop) as the
dependent variable or postop alone? The analyses so far show some
heteroscedasticity-in case i violate the independence assumption- is
it possible to do add both vce(robust) and vce(cluster id) ?

regards
Moleps
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index