 |
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: RES: st: using pscore and pstest
I do not have a definitive answer for you, Henrique. I assume that
you restricted your control samples to households which were eligible
to be beneficiaries. If so, I would not run the logistic regression
with the survey weights-the weights are not apt to represent the
restricted population. More fundamentally, propensity score methods
are designed to control for confounding. But, confounding is a
property of the sample, not of the population.
A more important concern may be the clustering of housholds in
"projects" and "subregions". This clustering will violate the
assumption of independence required by the programs you are using.
-Steven
On Apr 25, 2008, at 9:57 PM, Henrique Neder wrote:
Steven
The primary sampling units (PSUs) are the projects and ultimate
sample units
are the households. In the first stage we selected a number of
projects. In
the second stage we selected 3 households in each project. We have
in the
sample 318 beneficiaries and 404 controls.
My concern is that by not considering the logit model with weights
representing the expansion of the sample to the population, making
it more
representative (because the sample is not self-weighted) I will
eventually
estimate the average treatment effects with bias. But when I use
weights in
logit models estimations the matched samples does not balance. It
is very
strange that in the command psmatch2 there is not an option
provided for
weighting. In the command pscore this option exists. But the
estimation
commands coupled with the latter command (atts, attr) does not have
this
option. My question is: is it necessary to use weights in the
estimation of
atts when I used this in the balancing tests?
Henrique
-----Mensagem original-----
De: [email protected]
[mailto:[email protected]] Em nome de Steven
Samuels
Enviada em: sexta-feira, 25 de abril de 2008 08:57
Para: [email protected]
Assunto: Re: st: using pscore and pstest
-
Henrique, how were the program and beneficiary sub-samples selected?
Are the beneficiary and control sampling units projects or
households? Was there control of the number of the beneficiary and
control units--for example to achieve approximately equal numbers, or
minimum numbers of each?
-Steven
On Apr 24, 2008, at 8:32 PM, Henrique Neder wrote:
Dear
I have a problem with the use of the pscore and pstest commands. My
sample
is a cluster sample in that the sample observations are selected by
the
followings steps:
1 - The survey is made in five States and in each State we selected
some
number of projects (the primary sample units - PSUs) in all sub-
regions of
this State.
2 - In each project we selected a fixed number of households.
The sample weights are of the fweights form, calculated as the
ratio -
number of households in each sub-region (universe) / the number of
households (observations) selected in the sample and inside that same
sub-region. Alternatively, we have a pweight, because the projects
were
selected with PPS (with probability proportionate to the number of
the
households in each project in the universe). This sample of
projects were
selected as a sub-sample of the projects in a previous survey (that
is, the
second survey was a subsample of the projects in the first survey).
We have in the survey two subsamples: a sub-sample of a program
beneficiaries and a control sub-sample.
I executed the commands pstest and pscore using a logit model with
y = dummy
of participation program and yi (covariates) = variables that
explain the
participation. In the first execution I executed the commands with
propensities scores obtained through non weighted logit regressions
and some
models are balanced in the pstest and pscore commands. But when I
use the
weighted logits, the balancing tests are not satisfied (the pscore
command
stops and in the pstest the covariates are unbalanced with small p-
values
and the conjoint chi-square test with low p-value, too).
My questions are:
1 - It is necessary to execute the commands with propensities scores
obtained through weighed logit regressions to estimate the atts?
2 - It is adequate to test balancing with pstest and pscore without
weights
and estimate the atts with propensities scores obtained through
weighted
logit regressions?
3 - I read that a sample proposal is to weight the treated
observations with
the weight = 1 / propensity and the untreated with the weight = 1/
(1 -
propensity). In my case, is this solution appropriate? Does it
correct the
estimates considering the representation of the program effects in
the
universe?
Best regards
Henrique Dantas Neder
Universidade Federal de Uberl�ndia - Minas Gerais - Brazil
No virus found in this outgoing message.
Checked by AVG.
Version: 7.5.524 / Virus Database: 269.23.2/1389 - Release Date:
21/4/2008
08:34
No virus found in this outgoing message.
Checked by AVG.
Version: 7.5.524 / Virus Database: 269.23.2/1389 - Release Date:
21/4/2008
08:34
*
* 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/
No virus found in this incoming message.
Checked by AVG.
Version: 7.5.524 / Virus Database: 269.23.2/1389 - Release Date:
21/4/2008
08:34
No virus found in this outgoing message.
Checked by AVG.
Version: 7.5.524 / Virus Database: 269.23.2/1389 - Release Date:
21/4/2008
08:34
*
* 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/