Date:         Wed, 19 Sep 2007 20:11:11 +0000
Reply-To:     morse@edoug.org
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Doug Morse <morse@EDOUG.ORG>
Organization: Vanderbilt University
Subject:      Re: ANCOVA - slopes are heterogeneous - what next?
Comments: To: sas-l@uga.edu

hi,

just to start i should say i'm knowledgeable to no advanced expert re:
statistics.  having said that, i've a good deal of experience with ANCOVAs and
MANCOVAs, and i don't believe that what Peter is suggesting is correct.

homogeneity of the regression lines is an important assumption of the
ANCOVA/MANCOVA models, and simply reporting the interaction does not fix the
problem.  if the slopes of the regression lines are different, the lines cross
each other somewhere, and one group has higher Y values for one interval and
lower Y values for another interval.  everything i've ever encountered re:
covariance models such as these has stated that one cannot proceed with the
analysis if this assumption is not met and some other analysis will need to be
conducted.  unfortunately, i've never had any luck finding out what these
"other" analyses should be.  like you, i've run across the johnson-neyman
technique and get the sense it's well-established, but as of yet have no
actual experience with it.

  Use of the Johnson-Neyman Technique as an Alternative to Analysis of
  Covariance.  Nursing Research. 45(6):363-366, November/December 1996.
  Dorsey, Susan G.; Soeken, Karen L.

just my $0.02.  as with the original poster, i too would be pleased to hear
more from this group about alternative models available when this assumption
isn't met (and certainly further elaboration on why this assumption is so
important).

doug


On Wed, 19 Sep 2007 12:24:11 -0400, Peter Flom
<peterflomconsulting@mindspring.com> wrote:
>  I would not rely on any test of the interaction, rather, I would figure out if there *ought* to be an interaction and I would also plot it and *see* if there was an interaction.  If either of these got a 'yes' answer, then I'd include the interaction term.
>
>  Once you suspect an interaction, you can proceed as usual.  But instead of analyzing only main effects, you have to analyze the interaction, too.  You don't say what x, y, and cov are, but it appears that x and y are class variables and cov  is a continuous one.
> ...