Re: st: ANCOVA for pre post designs

[Constantine Daskalakis <C_Daskalakis@mail.jci.tju.edu>]

At 06:12 PM 12/23/2003, David Airey wrote:
> This is a question for the biostatisticians on the list.
>
> I'm thinking of formulating a commentary on accepted research procedures 
> in my area that I think could be improved by observing basic statistical 
> arguments presented to researchers by biostatisticians.
>
> It has been suggested that in a randomized clinical trial design with 
> baseline (B) and followup (F) test measures comparing a control and 
> treatment group (G), performing an ANOVA on the ratio pre/post is the 
> worst choice of the 4 ways to deal with baseline differences:
>
> (1) post: analyze F by G
> (2) difference: analyze F-B by G
> (3) ratio: analyze F/B by G
> (4) ancova: analyze F = constant + b1*B + b2*G, for G differences
>
> In light of biostatisticians' suggestion (e.g., Vickers, BMC Medical 
> Research Methodology (2001) 1:6, 
> http://www.biomedcentral.com/1471-2288/1/6) that method (4) above is 
> preferred most and method (3) is least preferred, does it apply to 
> "prepulse inhibition" literature?
In large trials, (1) should be fine (at least, in terms of no bias). But 
(2) or (4) may be more efficient.

(3) above is similar in flavor to (2) if you view it on the log scale, i.e.,

(logF-logB) by G (or, equivalently, log(F/B) by G).

A technical question is whether the original measurements (B and F), or 
their difference on the original scale, or their log-ratio (ie, difference 
of logs) more closely conforms to the assumptions of linear regression 
(normality of residuals, homoskedasticity).

Still, I wouldn't do it on (F/B) but rather on log(F/B) if that looks good.

There is a difference in the underlying scientific model and 
interpretation, of course.

Does the treatment work additively (ie, adds a fixed amount, no matter 
where you start)? If so, the difference (F-B) would be a good choice 
(constant additive treatment effect across all values of B). And you'll be 
talking about the (arithmetic) mean difference for treatment vs. control.

But if the treatment works multiplicatively (ie, increases/decreases your 
original B measurement by a certain percent), then log(F-B) would be 
better. And then, by exponentiating the regression coefficients etc, you'll 
be talking about geometric mean ratio for treatment vs. control.

Finally, the choice between (2) and (4) depends on the correlation between 
baseline and follow-up measurements. I think that when corr(B,F) < 0.5, 
then (4) turns out to be more efficient; otherwise, (2) is better. I 
believe there's a paper by Liang & Zeger on this.






The documents accompanying this transmission may contain confidential 
health or business information. This information is intended for the use of 
the individual or entity named above. If you have received this information 
in error, please notify the sender immediately and arrange for the return 
or destruction of these documents.
________________________________________________________________

Constantine Daskalakis, ScD
Assistant Professor,
Biostatistics Section, Thomas Jefferson University,
211 S. 9th St. #602, Philadelphia, PA 19107
   Tel: 215-955-5695
   Fax: 215-503-3804
   Email: c_daskalakis@mail.jci.tju.edu
   Webpage: http://www.jefferson.edu/medicine/pharmacology/bio/  
*
*   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/