On 29 Aug 2003 10:04:52 -0700, [EMAIL PROTECTED] (Brian Cuffel)
wrote:

> I am interested in understanding alternatives to the matched-pairs
> design such as "partial matching" and "marginal matching".  I have
> been unable to locate the write literature and I am not an
> epidemiologist.  Can anyone point me to references or help clarify the
> differences in these designs.

You might google for  < Rosenbaum   "propensity scores" > .

Matched pairs are, unfortunately, often weak matches --
The 9 year old with childhood leukemia is paired with the
11 y.o.  who has not....  "Two years"  is pretty close for adults,
but not (IMHO)  for children.

Bad matches like that don't help the statistical power.  

Dropping out all those possible, useful Normals is not
good, and does not help the statistical power.

If you use the matching, then you are *not*  investigating
the factor that you are matching on.  In practice, the real
data is apt to give you more *power*  when you use narrow
groups of age and sex (say)  than if you use matching;
plus, you get means and tests for the factors.

References for  "Propensity scores" should offer both 
discussion and advice.  Those are concerned with matching
on factors that are especially likely to be risk-factors.  

I think you can think of carefully devised scores done early,
for randomized studies,  which might seem to contrast
to scores that might be done in multiple ways that are 
experimental and exploratory, after the fact, as control
for covariates, for observational studies.
(If there's a big difference that I don't know about in 
how you have to do those, 
maybe someone will point the way....)

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
"Taxes are the price we pay for civilization." 
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to