Thanks for your response, very helpful.
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Steven
Samuels
Sent: Monday, April 07, 2008 10:31 AM
To: [email protected]
Subject: Re: st: IPW with OLOGIT
Michael,
I am not expert in this area. That said, your post is confusing to me.
* I don't see why you want to use -ologit-. The important part of a
treatment selection model is to model the probability of selection
into a treatment group. The outcomes of treatments might be ordered
by treatment, but there is no reason to assume a priori that the
selection probabilities should be ordered by treatment number.
* You state you have a count outcome, but your specification for the -
glm- command is for a binary outcome.
* The -glm- command does not take a -svy- prefix. Try "help
svy_estimation" for a list of commands that will take survey weights
and design factors.
*To use a survey weight, you would need to -svyset- your data after
creating your new weights; then you would need to use the -svy-
prefix for your command.
* Wooldridge's example of treatment selection is for the purpose of
estimating individual population means; then taking the difference
between those means. For that purpose you would need to define three
weights ipw1, ipw2, ipw3 based on p1, p2, and p3. (See help for -
generate-) Run your second regression model three times, each with a
different weight, you would average the predicted values over the
entire sample. The post-estimation command -predictnl- might compute
the difference between means.
* Wooldridge's method would estimate a difference between
populations means unadjusted for covariates. Is this what you want?
* Apparently Wooldridge's doubly-robust variance-matrices take into
account variability due to computing the propensity scores, although
I don't quite follow the argument . You could also bootstrap or
jackknife the entire process. As sums are over the entire sample,
then an estimated contrast in means is an average of the contrast in
the predicted values of individual predictions.
* As you have the same variables ("$var") on the right hand side of
your treatment and outcome equations, I don't see a need for the IPW
model at all.
-Steven
> I am trying to adjust for selection on observables using propensity
> scores as inverse probability weights (ipw), following Wooldridge
> ("IPW
> Estimation for General Missing Data Problems"). My dataset has a
> complex survey design with survey weights (svywt), and I want to
> adjust
> for selection bias of an ordered treatment (t1,t2,t3) on a count
> outcome. Can someone help me with the Stata code to compute the IPW
> using the predicted probabilities as propensity scores? I want to
> compute IPW by multiplying the survey weight*(1/propensity score),
> following Zanutto et al. (2005).
>
> ologit t_cat $var
> predict p1 p2 p3
>
> /*Here's where I need help*/
> ipw=svywt*[1/p1 ...]
>
> glm depvar $var t2 t3 [pweight=ipw], fam(bin) link(logit) irls robust
>
>
> Thanks in advance,
> Mike
>
>
> Michael F. Furukawa, PhD
> Assistant Professor
> Health Management and Policy
> W. P. Carey School of Business
> Arizona State University
> (480) 965-2363
>
> *
> * For searches and help try:
> * http://www.stata.com/support/faqs/res/findit.html
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
