Paired T-Test Debate

Hey guys! I was having a little debate with a doctor on her study about which statistical test to use. I was always under the assumption that a paired t-test was for "pre-post" tests until this whole idea of "matched controls" rocked my world. This was her study design...

Study Design
28 kids who had a disorder
28 matched controls
Outcome Variable: Stomach Fat Percent (continuous)

She wanted to see if there was a difference between the Disorder group and the Control group. I hadalways thought to run an Independent Samples T-Test until the doctor told me otherwise. Can anyone give me insight on what test to run and the theory behind it? Thank you!


Fortran must die
I agree with GretaGarbo:p and of course Jake and Dason....

If the elements are logically linked to each other you do a paired t-test. A problem I have with matching is that you can never be truly sure they share all values in common except the experiment - as you can for example when simply looking at the same person at two points in time. But that is a design not statistical issue.


Ambassador to the humans
But you can't be sure of that even in the case you mention. The point isn't to get *perfect* matching. It's to reduce variation by providing a pairing that makes sense. You do your best to eliminate the other factors that you want to control for by matching. You don't need to do a *perfect* job of this to see a reduction in the variance.


Fortran must die
Statistically I am sure that is true. But from a design perspective if the individuals are not identical except on the experiment you can never be certain that the experiment caused the result. The say you can with random assignment, but even ignoring this is rarely done in practice, it never seemed logical to be that if the groups end up being systematically different on some factor after random assignment that this has no effect simply because you assigned randomly.

Regardless it is a design issue not statistical.


TS Contributor
With matching, you have no independent observations (I suppose)
in each groups. So a dependent samples test should be appropriate.
On the other hand, I have often seen such designs analysed with
independent sample tests. Admittedly, in my first published study
I used an independent samples test for a matched pairs desgin, and
no reviewer objected. So, is it a matter of different viewpoints,
or is the independent samples analysis just wrong?

With kind regards




Fortran must die
The independent and dependent test calculate variation differently I thought since in the latter case the observations are related to each other, but in the former they are not. This makes one test invalid for the other.


TS Contributor
Try a scatter plot of the "matched" pairs. If you see a relationship a paired approach should be appropriate. However, if you see a shotgun pattern use a 2-sample t.