Chris Ryan asked (rethorically?):
Are there *any* circumstances in which stepwise multiple regression
would be the preferred approach?
--------------------------------
I can think of one example:
In the thalidomide disaster an epidemic of rare malformations occurred.
There were no strong, but dozens of weak hypotheses about the etiology.
In such a situation I think an extensive case-control study followed by
a stepwise logistic regression analysis is the most sensible thing to
do.
If I was doing a data-mining exercise of this nature in an extensive
case-control study, then I think I would define a propensity score for each
candidate cause, calculate confidence limits and P-values for the
propensity-adjusted odds ratio for each candidate cause, and then use one
or more of the multiple test procedures available from the -smileplot-
package, downloadable from SSC (see Newson et al., 2003). This method would
have the advantage that I could then make meaningful confidence statements
about my set of "discoveries", which I could not do with a clear conscience
using stepwise regression.
