I had advocated teaching students to use planned comparisons and
to eschew the one-tailed test.
On Sat, 23 Dec 2000, Mike Scoles wrote:
> Stephen,
>
> I agree completely with your 3rd recommendation [eschew the
one-tailed]
> If we must use NHST, then
> at least we should accept that we are dealing with a conditional
> distribution based on the null hypothesis being true. Our opinions about
> the direction in which the null hypothesis could be false is irrelevant,
> except under the rare situations in which it could be false in only one
> direction.
>
> However, recognition of this seems inconsistent with your first
> recommendation [use planned comparisons].
> Is it possible that there are limited numbers of ways in
> which a possible set of null hypotheses could be wrong? Should we ignore
> those differences that do not fit our theoretical models?
I'm not sure I understand Mike's concern. The planned comparisons
method doesn't mean restricting yourself only to those
comparisons whose outcomes fit a preconceived theoretical model.
It means that you decide beforehand which questions you want
answered. These are questions that will produce meaningful
information, whether or not they fit your theoretical
expectations (if you have any).
An example might help. Suppose you were testing people's
performance on three drugs with two different stress levels.
Drug
A B C
--------------------------------
low stress
high stress
----------------------------------
You might decide that the set of comparisons you were interested
in was:
drug A low stress vs high stress
drug B low stress vs high stress
drug C low stress vs high stress
low stress: drug A vs drug B
low stress: drug B vs drug C
low stress: drug A vs drug C
high stress: drug A vs drug B
high stress: drug B vs drug C
high stress: drug A vs drug c
This gives you 9 planned comparisons, which you can test using
the Bonferroni. There are other possible comparisons, but you
have no interest in their outcome. e.g. low stress drug A vs high
stress drug C. Those nine comparisons represent the only
questions of interest for you.
So the general method with planned comparisons is to decide just
which questions you want answered beforehand, and then answer
them. Of course, there's nothing to stop you from combining a
planned approach with a post-hoc method, but if you've included
all questions of interest in your planned set, it wouldn't be
necessary.
-Stephen
------------------------------------------------------------------------
Stephen Black, Ph.D. tel: (819) 822-9600 ext 2470
Department of Psychology fax: (819) 822-9661
Bishop's University e-mail: [EMAIL PROTECTED]
Lennoxville, QC
J1M 1Z7
Canada Department web page at http://www.ubishops.ca/ccc/div/soc/psy
Check out TIPS listserv for teachers of psychology at:
http://www.frostburg.edu/dept/psyc/southerly/tips/
------------------------------------------------------------------------
