Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: predicted values in svy glm l(log) f(poisson)
From
Steven Samuels <[email protected]>
To
[email protected]
Subject
Re: st: predicted values in svy glm l(log) f(poisson)
Date
Thu, 30 Dec 2010 14:53:18 -0500
Doug:
Another suggestion for estimating ATT: Abandon -glm-; do a propensity
score analysis which predicts "exposure"; match each "exposed" person
to one or more unexposed on the basis of the propensity score; get the
difference in outcome between exposed and unexposed (mean difference,
if there is >1 matched unexposed); and use -svy: mean- to compute the
mean difference and standard error weighted by the exposed person's
survey weight. Ordinarily you'd need a -bootstrap- to account for
uncertainty in the propensity score. However there is evidence you
don't have to account for the uncertainty, because the estimated
propensity score better adjusts for confounding than the unknown
"true" score.
You might find cases you have to exclude because you can't find
matches for them. In the language of causal modeling, they are not in
the area of "common support'.
A google search for "propensity score" weights "survey data" turned up
related ideas, such as:
http://www.demographic-research.org/volumes/vol23/25/23-25.pdf
http://www.jds-online.com/file_download/94/JDS-233.pdf stratifies on
propensity score, then does a further adjustment
Steve
Steven J. Samuels
[email protected]
18 Cantine's Island
Saugerties NY 12477
USA
Voice: 845-246-0774
Fax: 206-202-4783
--
Actually, I was wrong. -predictnl- will give you standard errors only
for individual contributions to ATT, not for the average. You'll need
write a program that incorporates the original -glm- and the creation
of the counterfactuals, and outputs the ATT; then -bootstrap- that
program. Unless you are lucky enough to be working with a replicate-
based data, you'll also have to create your own replicate weights with
Stas Kolenikov's -bsweights- program, ("findit bsweights"). However I
recommend against using -bsweights- unless you are very familiar with
sampling (and bootstrap) theory.
Steve
On Dec 28, 2010, at 2:49 PM, Steven Samuels wrote:
In Stata 10, the easiest way will be to use -predictnl- after -svy:
glm-. Please read the FAQ next time. Section 3.3 states:
"The current version of Stata is 11.1. Please specify if you are using
an earlier version; otherwise, the answer to your question is likely
to refer to commands or features unavailable to you."
Steve
On Dec 28, 2010, at 2:08 PM, Douglas Levy wrote:
Is there a way to do this in Stata 10?
On Thu, Dec 23, 2010 at 4:59 PM, Steven Samuels <[email protected]>
wrote:
Actually, the following code will work whether or not exposure was a
stratum
variable at any stage.
Steve
Steven J. Samuels
[email protected]
18 Cantine's Island
Saugerties NY 12477
USA
Voice: 845-246-0774
Fax: 206-202-4783
**************************CODE BEGINS**************************
sysuse auto, clear
svyset turn [pw= trunk]
replace foreign = foreign +1 //convenient for -margins-
// foreign =2 is the treated group
svy: glm rep78 mpg weight i.foreign, link(log) family(poisson)
margins, subpop(if foreign==2) at(foreign=(1,2)) post
vce(unconditional)
// _at2 is foreign as foreign _at1 is foreign as domestic
lincom _b[2._at]- _b[1._at] //ATT
margins, coeflegend //If you forget the coefficient names
lincom _b[2._at] - _b[1bn._at]
***************************CODE ENDS***************************
Use -margins-, but without knowing the survey design it's hard to
say more.
Were separate samples taken from the "exposed" and "unexposed" units
(whatever they were)? Were the PSUs stratified by exposure status?
Describe
the design and your -svyset- statement.
Steve
On Dec 23, 2010, at 2:03 PM, Douglas Levy wrote:
I am now revisiting this issue, having, with Steve's guidance, settled
on option #2 from my original post. I.e., estimate glm model; predict
daysmissed for exposed=1; predict daysmissed for the exposed group
when exposed is set to 0; take difference of the [weighted] means of
the predictions.
Now my question is, how can I put confidence bounds on the difference
in the mean predictions?
I thank the group for any help it can offer.
Best,
Doug
On Tue, Oct 26, 2010 at 1:34 PM, Steven Samuels
<[email protected]> wrote:
--
Your second suggestion would be an estimate of the average effect of
treatment (exposure, here) among the treated (ATT). For an
overview of
possibilities, see Austin Nichols's 2010 conference presentations;
his 2007
Stata Journal Causal Inference article; and the 2008 Erratum, all
linked at
http://ideas.repec.org/e/pni54.html.
Holding covariates at the means in non-linear models can be
dangerous.
For an example, see
http://www.stata.com/statalist/archive/2010-07/msg01596.html and
Michael N.
Mitchell's followup.
Steve
Steven J. Samuels
[email protected]
18 Cantine's Island
Saugerties NY 12477
USA
Voice: 845-246-0774
Fax: 206-202-4783
On Oct 26, 2010, at 11:24 AM, Douglas Levy wrote:
I have complex survey data on school days missed for an exposed and
unexposed group. I have modeled the effect of exposure on absenteeism
using svy: glm daysmissed exposure $covariates, l(log) f(poisson). I
would like to estimate adjusted mean days missed for the exposed and
control groups, but I'm not sure of the best way to deal with this in
a non-linear model. There are a couple of methods I've encountered,
and I would be grateful for some thoughts on the pros and cons of
each.
1. Estimate glm model. Reset all covariates to their [weighted]
sample
means. Predict daysmissed when exposed=0 and when exposed=1.
2. Estimate glm model. Predict daysmissed for exposed=1. Predict
daysmissed for the exposed group when exposed is set to 0. Take the
[weighted] means of the predictions.
3. Other suggestions?
Thanks.
-Doug
*
* 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/
*
* 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/
*
* 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/
*
* 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/
*
* 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/
*
* 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/