Re: Babies Affected by Dads' X-Ray Exams

["H.Wade Patterson" <hwade@talltown.com>]

John Moulder wrote, in part:
 
> These people have a large data base of birth outcomes and a broad range of
> environmental factors, social factors, psychological factors and genetic
> factors that might be associated with "adverse" birth outcome.  They can and
> will mine "mine" this database for hundreds of different associations.
> 
> John Moulder (jmoulder@its.mcw.edu)
> Radiation Biology Group
> Medical College of Wisconsin

And "finding" something is esentially certain because of the phenomenon
of "multiple testing".

The multiple testing problem arises whenever many factors are being
considered as possible "causes" of a phenomenon. Each can be tested for
significance at a predetermined level – say p, where 0<p<1 (usually
p=0.05 or p=0.01). When n such tests are made, each at level p, the
probability that none of the factors will appear "significant", when
none of them are related to the phenomenon under study, is equal to
(0.95)^n. Thus, when 2 are tested the probability that neither is
significant is 0.95 x 0.95 = .9025. When n = 10 this probability is 0.60
and when n = 20 it is 0.36. Thus, chances are greater than 50/50 that
when 10 are tested one or more will appear significant when none really
are. If 20 are tested the 64% of the time one or more will appear
significant. The usual way to correct for this is to divide the level of
significance by the number of test so that instead of testing at 0.05,
if ten are being tested, each should be tested at 0.005. If 20 are to be
tested, each should be tested at 0.0025, etc. This rarely, if ever,
done.

Nature has decreed that most things will have a random distribution. The
unthinkable alternative is for such to be absolutely evenly distributed.

Epidemiologists know all this so a big data base is like money in the
bank, or a license to steal; whichever metaphor you prefer.
-- 
Wade

mailto:hwade@talltown.com

H.Wade Patterson
1116 Linda Lane
Lakeview OR 97630
ph 541 947-4974