It is commonly done, but invalid. See these:
van Walraven C, Davis D, Forster A, Wells G. Time-dependent bias was
common in survival analyses published in leading clinical journals. J
Clin Epidemiol 2004;57:672-82.
Beyersmann J, Gastmeier P, Wolkewitz M, Schumacher M. An easy
mathematical proof showed that time-dependent bias inevitably leads
to biased effect estimation. J Clin Epidemiol, Available online 10 July 2008
Doug
At 08:37 15/07/2008, Colin Macleod wrote:
>Dear Allstat,
>
>While reviewing several publications, I've noticed multivariate models where
>a post-baseline efficacy variable has been added along with treatment to
>analyse the effect on a change from baseline parameter. This has been done
>in cases where one would want to check whether a measure taken early during
>the study is a predictor of response for another variable later on.
>
>I have reservations about the use of post baseline parameters in models
>along with treatment and would much prefer seeing correlation analysis being
>performed instead. I am however having problems explaining to
>non-statisticians why this approach is inappropriate. Can anyone suggest
>related papers on this topic or a suitable explanation on why this approach
>is incorrect?
>
>
>
>Thank you,
>
>Colin
_____________________________________________________
PLEASE NOTE NEW EMAIL ADDRESS BELOW
Doug Altman
Professor of Statistics in Medicine
Centre for Statistics in Medicine
University of Oxford
Wolfson College Annexe
Linton Road
Oxford OX2 6UD
email: [log in to unmask]
Tel: 01865 284400 (direct line 01865 284401)
Fax: 01865 284424
www: http://www.csm-oxford.org.uk/
EQUATOR Network - resources for reporting research
www: <http://www.equator-network.org/>http://www.equator-network.org/
|