GEE

[Kala-3]

I am using SPSS to replicate work some colleagues did in SAS.  It has
been a relearning experience--long time since I did much work with
claims data, and it was in SAS so I am learning SPSS as I go
along--although I suspect SAS today would be just as different.  They
used GEE since the data set consists f annualized claims summaries for
up to three years per person. I am trying to replicate their work using
our own claims in a very different geographic area (rural vs urban etc).
I can figure out most of what I need from the manuals but am not sure of
a  couple of things.

We are working with a fairly large insurance claims set--around 40,000
person/years over 3 years, each record is a member/year, with aggregate
expenditures for some specific disease categories, flags for these
categories, and number of encounters for the treatment of interest (we
are looking at the impact of periodontal care on other conditions).
Also some minimal demographic  data, and zip code which I've used to
link to census and other environmental information on a geographic level.

1) Insurance claims tend to follow a log distribution so they (and I)
converted the dependent variable (expenditure for x) to L(n) of the
value (also annual adjustment using CPI so they are constant $s) .  But
as I read the GEE it seems like it is possible to specify the log
transformation in the model so you don't have to scratch your head over
the interpretation (geometric mean) of the conversion back .  Can you
advise me how to do that, or, alternatively, how to convert back?  It
looks like the approach I took 20 years ago (Heckman?) is not considered
best practice any more.

2)  In specifying the correlation in the repeat measure they chose the
option "independent" but from my reading of the manual it seems like
since these are a series of sequential annual summaries of medical
claims data AR(1) might be more appropriate.  I am way over my head here
and would love to be pointed to a discussion of this.

3)  In the work we are replicating, they also looked at specific
conditions and compared expenditures for people with differing levels of
perio treatment.  When doing this they started with people who had each
condition, so it's a different denominator each time.  Should we instead
be either looking at the entire group or doing a two-stage analysis,
first of WHETHER they have a condition, and THEN comparing
expenditures?  Does SPSS have any handy tools for this, or do I have to
build up a 2-stage model step by step?  I am mostly using the drop-down
menus until I have a run where I didn't leave out something essential (I
am still climbing up my learning curve) then cut and paste into the
syntax tool and drop in alternative dependent variables and covariants
(eg condition-specific complications dummy) for the various conditions
we are looking at  so if you have a two-stage model that I can copy that
would be great.

4)  There are clearly strong geographic differences in the treatment
variable, (some counties  have NO patients with more than 2 perio
visits)  which may reflect lack of providers, different practitioner and
patient norms, and coding issues--likely all three.  To separate these
effects from individual differences  I created dummies representing 3
groups of counties with differing overall rates of perio visits.  I also
have geographic-level (33 dental districts--a couple per county)
information on # of dental providers  and included a provider-patient
ratio in some runs although that doesn't seem to make much difference.
How do I know if I am taking this one apart too many ways, or if I ought
to add an interaction term?  Including all of these doesn't seem to blow
up my results--adding the 3-level dummy actually seemed to calm things
down a bit.

Any help including pointing me to more detailed support would be greatly
appreciated.  I am using the SPSS online materials, and  some NCSU pages
that are at about my level of comprehension.

Thank you !

Kala Ladenheim, PhD, MSPH
Medical Care Development
Augusta, Maine

=====================
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