Date:         Fri, 26 Apr 1996 16:07:16 -0700
Reply-To:     Feldesmanm@pdx.edu
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         Marc Feldesman <h1mf@ODIN.CC.PDX.EDU>
Subject:      Discriminant "Structure" Coefficients

Proc Candisc provides "structure coefficients" for the coefficients of
the canonical variates.  Proc Discrim does not.  Proc Discrim provides
much more useful information than does Proc Candisc, but I'd like to be
able to get the "structure coefficients" anyway.  I know that I can get
them by running both procs, but candisc doesn't give me the option of the
quadratic discriminant function, and the structure coefficients would be
useful in that circumstance too.
 
I can't find any reference to how these are computed other than a vague
reference to the fact that they are "correlation coefficients between the
original variables and the canonical variates (scores?)".
 
I've run proc corr on the original variables (scored for each case) and
the canonical scores for each case on each canonical axis.  The results
are similar to (but not identical with) those reported by proc candisc.
 
Does anyone know which set of variables are actually being correlated to
generate these coefficients, and which set of output data needs to be
associated with which set of input data to derive these "coefficients".
I've looked at 4 multivariate texts, all of which talk about "structure
coefficients", but never discuss how to compute them.
 
Thanks.
 
 
  ***********************************************************************
  * Dr. Marc R. Feldesman     | email: feldesmanm@pdx.edu| Beyond every *
  * Professor and Chairman    | fax:  (503)-725-3905     | credibility  *
  * Anthropology Department   | voice: (503)-725-3081    | gap lies a   *
  * Portland State University | pager: (503)-870-2515    | gullibility  *
  * Portland, Oregon 97207    | home:  NOT               | fill         *
  ***********************************************************************