Jerry Dallal wrote:
>
> It wasn't ironically and has nothing to do with 5%. As Marvin Zelen
> has pointed out, one-tailed tests are unethical from a human
> subjects perspective because they state that the difference can go
> in only one direction (we can argue about tests that are similar on
> the boundary, but I'm talking about how they are used in practice).
> If the investigator is *certain* that the result can go in only one
> direction, then s/he is ethically bound not to give a subject a
> treatment that is inferior to another.
Basically correct, but I would argue that "inferior" can only
reasonably be interpreted in a braod context,in which it is not
synonymous to "more effective". A treatment may be slightly less
effective but superior because of lesser side effects (it is said that
castration, done early enough, will cure male pattern baldness - any
takers? Or how would you feel about a treatment involving major thoracic
surgery with a 100% success rate against the common cold?)
It may interfere less in quality of life (think of some dialysis
techniques that require somebody to spend several hours a day on a
machine). It may be ethically preferable for other reasons (consider the
debate about the use of fetal cells in the treatment of Parkinson's).
A slightly less effective treatment may be within somebody's means
whereas they simply cannot afford the absolutely most effective
treatment. It is said that the sort of intensive
one-physician-one-patient medical care enjoyed by many heads of state
has a significant health benefit. Should we all have our own private
doctors and who will care for _them_? It is nice to say that nobody
should be refused the best medical treatment because of cost, but there
seems no limit to the expensive medical procedures that can be developed
if there is a will to pay for them. Realistically, the human race is not
going to dedicate all its resources to health care; and even if we did
there would _still_ not be enough to go around.
Finally, a treatment may be more effective but unusable because its
effectiveness has not been demonstrated in a generally accepted way in
clinical trials, so that it is not permitted in some jurisdiction.
All this said, such "pidgin Bayesian" one-tailed tests are indeed
silly. A one-tailed test may be done where there is an _interest_ in one
tail, as in acceptance sampling, but not to artificially exaggerate a
p-value.
-Robert Dawson
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
