Thank you Joseph for your input and thoughts.
You will see from those references I provided earlier that the IPTW weights
are established for each person for each wave. The general model consists of
estimating the propensity score at a given wave (the probability of getting
treatment at that wave conditional on past characteristics and even past
treatment status). This is repeated at each wave for each person. After a
propensity score is estimated for each person/wave, the IPTW is then
calculated).
So there is a specific weight for each person/wave, and it indeed differs.
Robins, Hernan and Brumback (2000)* suggest using the SCWGT option in SAS
Proc Genmod together with the "repeated" option and specifying am
independence working correlation matrix.
In Stata, my work-around so far has been to use GLM with the pweight option
in which I refer to the IPTW variable I created (as described in the first
paragraph above). There is no need to rescale the weights.
I have not considered using survey modeling procedures in Stata to handle
these data, and I am not sufficiently familiar with the tools to know if
they will handle panel data with varying weights.
Again, I don't know how SAS GEE models handle varying weights per
person/period, but it is available, and researchers use it. Since I am a
Stata die-hard, I would hate to have to use SAS because I can't find a
solution in Stata.
Ariel
* Robins JM, Hernan MA, Brumback B. Marginal structural models and causal
inference in epidemiology. Epidemiology. 2000;11:550?560.
Date: Mon, 20 Jul 2009 09:35:33 +0900
From: "Joseph Coveney" <[email protected]>
Subject: Re: st: multiple weights per person in GEE?
Ariel Linden wrote (excerpted):
I have been using GLM with vce(cluster) with the IPTW weight, but the SE is
much larger than that produced in SAS using GEE. For example, with a beta
coeficient for a treatment variable of 2.47, GLM in stata gives me a SE of
0.484 (CI = 1.53, 3.43) while GEE in SAS gives me SE of 0.013 (CI = 2.45,
2.50).
This is a pretty meaningful difference, and in several models this can
change the treatment effect from being positive to one of non significance.
-
----------------------------------------------------------------------------
----
I take it that the difference you're seeing in SEs with identical point
estimates is between
PROC GENMOD;
. . .
REPEATED SUBJECT = . . . / TYPE = IND;
SCWGT . . .;
and
glm . . . [<weight>= . . .], cluster(. . .) . . .
If so, then these are indeed larger differences than would be expected if
the
two packages mean the same thing by "weight" in this context. You've
probably
already considered the following and more, but just in case:
1. What kind of weights are you declaring the IPTW to be in Stata? Fewell et
al. (2004) used Stata's -pweight-.
2. Related to that, does PROC GENMOD need scaling of the weights so that
they
sum to the number of observations?
3. Is it possible to cajole Stata into allowing the time-varying weights
that
you want by viewing the observation time points in the same manner as waves
of a
survey and setting the model up as a survey data analysis task?
Joseph Coveney
Z. Fewell, M. A. Hernán, F. Wolfe, K. Tilling, H. Choi, J. A. C. Sterne.
2004.
Controlling for time-dependent confounding using marginal structural models.
_The Stata Journal_ 4(4):402?420.
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
