Posthoc on interaction with SPSS

[plierouge]

This question was asked several times on this group, but no real answer
was given.

Can we do a posthoc analysis on the interaction with SPSS? I'm
currently using version 11. All other statistical softwares
(Statistica, JMPIN, SAS, Systat) I have used in the past can do it,
either directly in the menu or through the command line. It would be
strange that SPSS can't do this simple analysis. Maybe a "new" feature
in SPSS 15? ;)

Thanks,

Sébastien Plante

[Bruce Weaver]

"Post hoc analysis on the interaction" is pretty vague. If you say
exactly what you want to do, someone can probably suggest a way to do
it. E.g., if you mean analysis of the simple main effects, something
like this will work:

UNIANOVA
y BY a b
/emmeans = table(a)
/emmeans = table(b)
/emmeans = table(a*b) COMPARE(b) /* <---- COMPARE option added here */
/DESIGN = a b a*b.

**** The Univariate Tests shows F-tests for the simple effects
of variable B, with one F-test for each level of A.

Alternatively, you can use MANOVA to get the same F-tests on the simple
main effects.

MANOVA
y BY a(1 2) b(1 3)
/NOPRINT PARAM(ESTIM)
/METHOD=UNIQUE
/ERROR within+residual
/DESIGN= a, b W a(1), b W a(2).

--
Bruce Weaver
bwe...@lakeheadu.ca
www.angelfire.com/wv/bwhomedir

[plierouge]

I want to run a Tukey posthoc test on the interaction in a 2-way ANOVA
design..

UNIANOVA
var3 BY var1 var 2
/METHOD = SSTYPE(3)
/INTERCEPT = INCLUDE
/POSTHOC = var1 var2 ( TUKEY)
/CRITERIA = ALPHA ( .05)
/DESIGN = var1 var2 var1*var2

Basically, I would like to include "var1*var2" in the /POSTHOC line. Is
this possible?

Thanks,

Sébastien Plante

[Bruce Weaver]

If you mean you want to do all pair-wise comparisons of the cells
created by the factorial combination var1*var2, I think you'll have to
recast the thing as a one-way ANOVA. I.e., you'll have to compute a new
GROUP variable that codes the various combinations of var1 and var2.
E.g., if the levels of both variables are numbered consecutively from 1
to the number of levels, something like this would work:

compute group = (var1-1)*k2 + var2.
exe.

where k2 = the number of levels of var2.

[Richard Ulrich]

On 21 Oct 2005 06:54:34 -0700, "plierouge" <plie...@hotmail.com>
wrote:

- all the packages, I imagine, can do contrasts.

I'm surprised if they *all* offer, for two-way ANOVA,
the Tukey-HSD or other specific "post-hocs" that are
designed for one-way ANOVA. The ones with statisticians
involved, it seems to me, would defer to the objections of
their statisticians. Am I out of date? Up to now, I
have never seen any argument saying that searching among
all means is a good way to frame and test hypotheses in
a two-way ANOVA.

What is easier to justify, by logic, is having a style
of Bonferroni testing, using individual Ns, which adjusts
for some stated number of contrasts, where the contrasts
are drawn (say) within levels of one factor. However, I don't
know if that can be implemented and documented without
creating confusion; and if you understand what you are
doing, it is simple enough to divide a stated p-level by k.

Using those inappropriate post-hocs (Tukey, etc) seems like
a bad idea, and one that ought to get you in trouble with
reviewers. It is least appropriate, I think, if there is an overall
main effect for either variable.

--
Rich Ulrich, wpi...@pitt.edu
http://www.pitt.edu/~wpilib/index.html

[jacc...@fiu.edu]

Interaction effects are usually decomposed using interaction contrasts.
Simple main effects really do not address the issue of interaction, as
numerous statisticans have pointed out.

You can conduct interaction contrasts using SPSS but you have to do so
using the /LMatrix line in the GLM program and it is a bit of a hassle
to do because the contrast weighting strategies are somewhat esoteric.
But it can definitely be done there.

Another option is to request the parameter estimates as an output
option from the GLM program. These estimates contain terms within them
that reflect single degree of freedom interaction contrasts. To make
use of this, you need to understand how dummy variables work in linear
regression models and how to interpret coefficients associated with
product terms between dummy variables.

There is an inexpensive add-on package to spss that calculates
interaction contrasts for a variety of factorial designs in a
straghtforward way and that is very simple to use. Or, it will generate
syntax for you for the /Lmatrix line and you can cut and paste it into
spss syntax. It is at www.zumastat.com