On Wed, 10 Apr 2002 19:57:25 +0200, "K.J.A. Postulart" <[EMAIL PROTECTED]> wrote:
... " Pairing of data is only usefull in the case, that samples can be taken pairwise and undergo some different treatment ( which may be very diverse ! ). It is absolutely necessary that the pair originates from the same material or source. If this cannot be guaranteed, paired data testing is useless." Useless? I have to disagree. If that were so, then (almost) every case-control analysis ever done would be wrong. It offends me less, if I read that to say, "The pairing in paired testing is often useless" -- instead of what seems to be the emphasis, "[such] testing is useless." I would be willing to agree that paired-*analyses* have been done too often in the past. You may form pairs, as many studies have done, based similar characteristics. It is safe to do this in a prospective study, before randomizing (randomize in pairs). After the study, the analysis - "testing" - might or might not use the pairs. These days, the computer programs are more available for using "statistical control of covariates" and there is more recognition of the wisdom and flexibility of the covariate-approach, compared to the paired-approach. Retrospective studies have used pairing, too, although it is more questionable. In addition to logical problems, paired-testing in retrospect usually results in discarding a large share of the collected data, which is definitely a *bad thing* and something to be avoided. The statement quoted above is distorted, in terms of what-implies-what. If you have data arising from some unique "material or source" that (likely) matters, then you *do* have to use pair, or use groups, or take the source into account -- until you show it does *not* matter. You don't need to use groups or pairs if there is a better way to control. For instance: you might match on age for the sake of random assignment; but it is usually better to *test*, in the end, using statistical control for age instead of the original pairs. Yes, there are occasions when pairing is strictly wrong -- Here is the only example I think of: pairing based on outcome, such as, sorting the outcome scores, and matching highest with highest etc. I went on this long in hope that addressed the original question to some extent. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
