Next: Subjective or objective Bayesian
Up: Frequentists and Bayesian `sects'
Previous: Bayesian versus frequentistic methods
Contents
As a last comment about frequentistic ideas I would like
to add here a nice dialogue, which was circulated via internet on
19th February 1999,
with an introduction and comment by the author,
the statistician George
Gabor [77]
of Dalhousie University (Halifax, N.S., Canada).
It was meant as a contribution to a discussion triggered by
D.A. Berry (that of Refs. [10] and [63])
a few days before.
``Perhaps a Socratic exchange between an ideally sharp, i.e not easily
bamboozled student (S.) of a typical introductory statistics course and his
prof (P.) is the best way to illustrate what I think of the issue. The class
is at the point where confidence interval (CI) for the normal mean is
introduced and illustrated with a concrete example for the first time.
- P.
- ...and so a 95% CI for the unknown mean is (1.2, 2.3).
- S.
- Excuse me sir, just a few minutes ago you emphasized that a CI is
some kind of random interval with certain coverage properties in REPEATED
trials.
- P.
- Correct.
- S.
- What, then, is the meaning of the interval above?
- P.
- Well, it is one of the many possible realizations from a collection
of intervals of a certain kind.
- S.
- And can we say that the 95collective, is somehow carried over to this particular realization?
- P.
- No, we can't. It would be worse than incorrect; it would be
meaningless for the probability claim is tied to the collective.
- S.
- Your claim is then meaningless?
- P.
- No, it isn't. There is actually a way, called Bayesian statistics,
to attribute a single-trial meaning to it, but that is beyond the scope of
this course. However, I can assure you that there is no numerical
difference between the two approaches.
- S.
- Do you mean they always agree?
- P.
- No, but in this case they do provided that you have no reason, prior
to obtaining the data, to believe that the unknown mean is in any
particularly narrow area.
- S.
- Fair enough. I also noticed sir that you called it `a' CI, instead
of `the' CI. Are there others then?
- P.
- Yes, there are actually infinitely many ways to obtain CI's which
all have the same coverage properties. But only the one above is a
Bayesian interval (with the proviso above added, of course).
- S.
- Is Bayesian-ness the only way to justify the use of this particular
one?
- P.
- No, there are other ways too,
but they are complicated and they operate
with concepts that draw their meaning from the collective (except the so
called likelihood interval, but then this strange guy does not operate
with probability at all).
- ...
-
It could be continued ad infinitum. Assuming sufficiently more advanced
students one could come up with similar exchanges concerning practically
every frequentist concept orthodoxy operates with (sampling distribution
of estimates, measures of performance, the very concept of independence,
etc.). The point is that orthodoxy would fail at the first opportunity had
students been sufficiently sharp, open minded, and inquisitive. That we
are not humiliated repeatedly by such exchanges (in my long experience not
a single one has ever taken place) says more about... well, I don't quite
know about what -- the way the mind plays tricks with the concept of
probability? The background of our students? Both?
Ultimately then we teach the orthodoxy not only because of intellectual
inertia, tradition, and the rest; but also because, like good con artists,
we can get away with it. And that I find very disturbing.
I must agree with Basu's dictum that nothing in orthodox statistics makes
sense unless it has a Bayesian interpretation. If, as is the case, the
only thing one can say about frequentist methods is that they work only in
so far as they don't violate the likelihood principle; and if they don't
(and they frequently do), they numerically agree with a Bayesian procedure
with some flat prior - then we should go ahead and teach the real thing,
not the substitute. (The latter, incidentally, can live only parasitically
on an illicit Bayesian usage of its terms. Just ask an unsuspecting
biologist how he thinks about a CI or a P-value.)
One can understand, or perhaps follow is a better word, the historical
reasons orthodoxy has become the prevailing view. Now, however, we know
better.''
Next: Subjective or objective Bayesian
Up: Frequentists and Bayesian `sects'
Previous: Bayesian versus frequentistic methods
Contents
Giulio D'Agostini
2003-05-15