Date:         Thu, 7 Mar 2002 17:16:00 -0500
Reply-To:     hoffman@ria.buffalo.edu
Sender:       "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From:         Joseph Hoffman <hoffman@ria.buffalo.edu>
Subject:      Multiple testing procedures for interactions in regression
Content-Type: text/plain; charset="us-ascii"

I have a general question about statistical strategy in interpreting the
significance of many interactions in regression or logistic regression. If
I have run several regressions, and in each one, entered multiple main
effect predictors, followed hierarchically by multiple product terms for
testing interactions (both 2-way and 3-way), these analyses quickly
produce a large number of coefficients that are each tested for
significance. The number of tests increases quickly when multiple
interaction terms are included. For example, a typical equation with only
a modest number of main effect predictors can easily have 30 terms,
including both main effects and interaction terms. The question then is:
should multiple testing procedures be used to evaluate the significance of
all the coefficients, and for the interactions, in particular?  Some
statistics books seem to recommend these procedures in connection with
interpreting contrasts in anova, or other multiple means comparisons, but
do not uniformly recommend this for regressions, or illustrate this
approach with examples.  Also, common practice in the social and
behavioral sciences seems to be not to bother with such corrections to
control for Type I error, except perhaps for adopting a stricter (smaller)
overall alpha level, e.g., .001 instead of .05.

Any recommendations or citations on this practical question would be most
welcome.

Thanks for your consideration.
 Sincerely,
Joe Hoffman.

================================
Joe Hoffman
Data Analyst
Research Institute on Addictions
State University at Buffalo
1021 Main Street
Buffalo, NY 14203
716-887-2219
FAX 716-887-2510
email:  hoffman@ria.buffalo.edu
=================================