On Thu, 16 Jan 2003 10:28:02 +0100, Michael Specka <[EMAIL PROTECTED]>
wrote:

> On 15 Jan 2003 07:09:49 -0800, [EMAIL PROTECTED] (Dennis Roberts) wrote:
> 
> >just as an example ... say you had a null hypothesis that the population 
> >correlation was 0 ... (this is typical for the null) and ... you had a 
> >sample size of about 1000 ... it would only take an r of about .06 ... for 
> >you to reject the null ... however, do you think that accounting for .06^2 
> >... = ..0036 or less than 1% of the criterion variance ... is enough such 
> >that you would be willing to use X (which might be data that is costly to 
> >collect) as a predictor of Y?

> Agreed. But there might be circumstances where even small
> correlations provide important information and justify certain
> actions. If the 0.06 correlation comes from a death rate of 
> 250 out of 500 in one treatment and 220 out of 500 in the other
> treatment...
> 

As a clinician, you take note that a 15% advantage is 
minuscule and not statistically significant, and hardly any
reason, by itself, to select between the two treatments.
- if it is "just-barely-significant",   you can't show that it
is robust against contingencies; and if it is not-significant,
your confidence limit includes the opposite *conclusion*.

I have never seen anyone generate meta-statistics on this, 
but I suspect that there has been poor replication for medical
conclusions "at the 5% level"  where the Odds ratio 
was anything from 1.20  to 1.50:  Far too many people
have pursued far too many hypotheses, and so the 
literature is strewn with these findings that have capitalized
on chance.

By the way, having the results 'make sense'  is a bit helpful, 
but it is no guarantee.  It is better than *not*  making sense.


-- 
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/                    .
=================================================================

Reply via email to