Stas Kolenikov wrote:
If SAS does it, it does not mean it is such a great idea. And
propensity score matching people rarelly care about any other
complications that may be arising from the complex data structure, in
my experience.
First, check out the FAQ:
http://www.stata.com/support/faqs/stat/xtweight.html which talks about
the conceptual foundations for use of weights. Propensity score
weights are neither frequency, variance, or sampling weights; they are
more like kernel weights in non-parametric regression.
At any rate, my understanding of GEE is that a contribution to the
objective function is from the whole panel: you compute the residuals,
then, for each panel, you compute the quadratic form with the
residuals using the working correlation matrix, and then the whole
result is multiplied by the weight and added to the total. How exactly
would the different weights go into that quadratic form? SAS might
have found some algorithmic implementation (e.g., multiply each
residual by the square root of the weight before wrapping the
residuals around the correlation matrix), but I would personally want
to see a Biometrika paper that would justify this before I apply any
such method.
On Fri, Jul 17, 2009 at 11:40 AM, Ariel Linden<[email protected]> wrote:
> This is a question more directed at the Stata folks than to the listserve
> per se.
>
> Is there a reason why xtgee does not allow different weights/person/wave? It
> gives an error message stating "weight must be constant within personnumber"
>
> While I hate to invoke the phrase, "but SAS does it", I am forced to. There
> is a growing body of literature in which the propensity score weighting
> method is applied to longitudinal data. Thus, by it's very nature, weights
> will differ within individuals over each wave.
>
> I recogize GLLAMM as an option, but it is not very user friendly and
> inordinately slower than other models within this family.
>
> Consider this a plea for improvement.:-)
--------------------------------------------------------------------------------
Ariel, what kinds of working correlation structure are used in the literature on
propensity score weighting in longitudinal data with GEE? If it's all PROC
GENMOD; . . . REPEATED SUBJECT= . . . / TYPE=IND; SCWGT . . . ;, then one
approach that might be worth considering would be something like -glm . . .
[iweight= . . . ], cluster( . . . ) . . .-.
According to the user's manual for PROC GENMOD, WEIGHT or SCWGT just divides the
dispersion parameter by the observations' SCWGT values. I didn't find anything
there about what more, if anything, happens when the REPEATED option is also
invoked. If the answer is "not much", then can you mimic PROC GENMOD for those
models where the dispersion parameter is one (binomial, Poisson, negative
binomial) with the average of the weights (or equivalently its inverse), as in
-summarize myweights, meanonly-, -xtgee . . ., corr( . . . ) family(binomial)
scale(`r(mean)') [robust]-? The analogous operations for Pearson chi-square or
other scales would be something like fitting an unweighted model, weight the
resulting scale value by the average of the weights, and then use this as the
fixed scale value for a subsequent fit. I suspect that this approach wouldn't
given anyone a sense of theoretical gratification--it wouldn't do an
observation-by-observation weighting of the dispersion parameter that PROC
GENMOD apparently does--but it might suffice in practice.
Neither of these would address any of the issues that Stas raises.
Joseph Coveney
*
* 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/
