Rich --

Well-stated comment!

Yup, my motto is "IF YOU CAN'T CONTROL IT, MEASURE IT AND PUT IT IN THE MODEL".

IMHO matched pairs on a measureble attribute, was a method used before the availability of computers. It is almost impossible to "match" on several attributes. So if the information is available anyway, then the "controlling" (matching) can be accomplished by including the information in the model. Furthermore, it is possible to investigate interactions, non-linearity and other questions that might be interesting.

--Joe

*********************************************************************** * Joe Ward Health Careers High School * * 167 East Arrowhead Dr. 4646 Hamilton Wolfe * * San Antonio, TX 78228-2402 San Antonio, TX 78229 * * Phone: 210-433-6575 Phone: 210-617-5400 * * Fax : 210-433-2828 * * joeward@tenet.edu Fax : 210-617-5423 * ***********************************************************************

On 10 Oct 1997, Richard F Ulrich wrote:

> Date: 10 Oct 1997 20:23:29 GMT > From: Richard F Ulrich <wpilib+@pitt.edu> > To: SAS-L@VTVM1.CC.VT.EDU, STAT-L@VM1.MCGILL.CA, > EDSTAT-L@jse.stat.ncsu.edu > Subject: Re: Matched-paired case-control analysis > > Eric Hardy (erich@algene.com) wrote: > : I have to conduct a 1:1 matched-paired case-control analysis. The cases and > : the controls are age-matched (+/- 2 years). I have 150 pairs of > : observations. > > : The problem is that my controls are a little bit older than the cases. The > : mean of the difference of age is 0.5 year, and it is significative. So when > : I'll run my conditional logistic regression, should I include the this > : variable in the model (since it will be significative) or should I exclude > : it from the model (since the observations are age-matched) ? > > Matched-pairs is a useful way to draw upon - to select a tiny subsample > from - a vast sea of untapped controls, if you have that luxury. But it > is not the powerful or wise way to do analyses unless the Matches are > something pretty special (like, family?). > > When the match is "age", analyze GROUPS while controlling for > age. And, analyze for the effect of age, as well, to see if it does > make any difference. I don't see any justification to throw away > degrees of freedom with the idiosyncratic, personal matches; and > the error term should be smaller, too, with a smoother function > for subtracting out an age-effect. (Take age, for instance, as > linear and quadratic effects, and/or control for age-groups, however > narrow that you require.) > > > > Rich Ulrich, biostatistician wpilib+@pitt.edu > http://www.pitt.edu/~wpilib/index.html Univ. of Pittsburgh >