Re: Canadian doctors coming to the US

From: Herman Rubin (hrubin_at_odds.stat.purdue.edu)
Date: 09/04/04 

Date: 4 Sep 2004 15:25:50 -0500

In article <I2KZc.1615$Vl5.1292@newsread2.news.atl.earthlink.net>,
George Conklin <nilknoc@earthlink.net> wrote:

>"Herman Rubin" <hrubin@odds.stat.purdue.edu> wrote in message
>news:ch7bd2$4eqk@odds.stat.purdue.edu...
>> In article <mqEYc.4785$JT3.18@newsread3.news.atl.earthlink.net>,
>> George Conklin <nilknoc@earthlink.net> wrote:

>> >"Orac" <orac@mac.com> wrote in message
>> >news:orac-EF355F.16333529082004@news4-ge1.srv.hcvlny.cv.net...
>> >> In article <cgt8fn$22fi@odds.stat.purdue.edu>,
>> >> hrubin@odds.stat.purdue.edu (Herman Rubin) wrote:

>> >> > In article <orac-A727DC.21270928082004@news4-ge1.srv.hcvlny.cv.net>,
>> >> > Orac <orac@mac.com> wrote:
>> >> > >In article <cgra53$4a0m@odds.stat.purdue.edu>,
>> >> > > hrubin@odds.stat.purdue.edu (Herman Rubin) wrote:

>> >> > >> In article <41309b16$0$18253$79c14f64@nan-newsreader-05.noos.net>,
>> >> > >> Lictor <ghostmlNOSPAM-REMOVE@online.fr> wrote:

                         ....................

>> > Herman denies that the usual test of statistical significance is a vaid
>> >test of how effective medicine is. Thus anything the government posts
>would
>> >automatically be invalid in his mind. I have never really understood why
>he
>> >insists that only Bayesian analysis be used, since he has never shown
>that
>> >the results would be changed if you change the test. Anyway......

>> A test of statistical significance tests whether the
>> observations would be "sufficiently unlikely" if the
>> null hypothesis is EXACTLY true, and NOTHING else.

> Based on this comment, we could never do any medical research since all
>variables cannot be controlled.

This is NOT what I said. One can do research with NO
variables controlled; this is what happens in astronomy,
meteorology, and economics, for example. To a large
extent, it is what is done in ecology and geology. We
need to use these techniques in biology (it is, to a
small extent) and medicine.

> Having said that, self-selection, social class and a whole lot of other
>variables have fouled medical findings now for many years. They also ignore
>things like stress, since it is not something a BENCH scientist does.

I have seen studies taking into account stress.

>> It does not take into account how effective the
>> procedure is under the alternatives, nor the magnitude
>> of the effect under the null, and hence does not balance
>> the various components of risk. If 50% of the people
>> exposed to a disease without treatment died, and 51%
>> of 1,000,000 people given a treatment survived, this
>> would be HIGHLY significant.

> So? People would see from the figures what the effect was and the risk
>too.

One can see this from the FULL figures, but NOT from the
significance levels. It does take calculation; the
original problem from which the above extremization was
made was 3 out of 3 versus 7 out of 8. The results were
approximately that if treatments of this type, if they
worked, worked more than 90% of the time, take the 3 out of
3, else the 7 out of 8. The prior information, usually
involving lots of subjectivity, MUST be used. Even in
science, statistical theory shows that objectivity is not
even desirable, in that there are better non-objective
procedures than the best objective ones.

>If there was a new
>> treatment given to 4, and they all survived, this
>> would not meet the customary significance level.
>> Which treatment would YOU choose?

> False choice too Herman. Today we are wondering if common treatments
>might make things worse, like HRT.

And for complications like the ones occurring with HRT,
it can take years to accumulate enough data. What do
you do in the many years for this to happen?

In medicine, we are concerned with decision making under
uncertainty. We have to balance the components of risk;
this is prior Bayesian behavior. One way to do it is
to look at the posterior distributions, but it is not
always possible to reasonably approximate this, even
when one can approximate prior Bayes risk well. 

-- 
This address is for information only.  I do not claim that these views
are those of the Statistics Department or of Purdue University.
Herman Rubin, Department of Statistics, Purdue University
hrubin@stat.purdue.edu         Phone: (765)494-6054   FAX: (765)494-0558
Quantcast