Simon

[Sion]

I am an undergrad doing a project with MZ & DZ twin-pairs.  I'm
interested in using two dichotomous categorical variable to examine
differences on several IVs (most of them also dichotomous categorical).  My
analytic strategy  is to use SPSS GEE to account for the non-independence of
twin pairs.  I have several questions about this process -- I realize this
is long, so I've listed the questions in order of importance to me.  :)

Q1:  I am not confident that I've set up the data correctly.  I have
one line for each individual.  I have a DYAD variable that identifies which
twin pair each individual belongs to.  I am using this DYAD  variable as the
"Subject" variable on the "repeated" tab.  I have another numerical variable
that arbitrarily designates one twin as  twin A and one as twin B (so half
the individuals in the sample have a  2 for this and half have a 1).  I have
added this variable to "Within-Subjects."  Specifying the working
correlation matrix is extremely confusing to me.  Right now, I've got
"robust estimator" for covariance matrix. Someone (who I respect and does
this kind of stuff a lot) told me in passing that "independent working
correlation matrix" paired with robust estimator is the combination I want,
"even though it's completely counterintuitive."  I don't understand this
stuff very well, and I thought that the exchangeable working correlation
matrix would indicate that there is no reason that "twin Bs" would be
correlated with each other across the sample (and also no reason that
"Twin As" would be correlated with each other across the sample).  How does
the robust estimator relate to all this?  A very clear, "dummy" explanation
for this would be helpful.

Q2: When I run my analyses and use the two dichotomous categorical vars as
IVs, my "test of model effects" matches my parameter estimates.  However,
when I add an interaction effect, this changes, and my parameter estimates
go out of wack.  I know this has something to do with dummy coding -- right
now, I've just got one variable for each of the two categories, coded 0-1.
What can I do to make my parameter estimates match the test of model
effects?

Q3: Would I need to consider changing the nature of the working  correlation
matrix depending on the type and distribution of my IV?

Q4:  What assumptions do I need to meet to use the Poisson  distribution,
and how do I show that they're met?  How would using  this compare to log-
transforming my IV (somewhat unsuccessfully) and  using a linear
distribution?

ANY ASSISTANCE would be EXTREMELY HELPFUL.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD