Dear Fabricio,
No, Zelig does not automate this. However, if you follow Thomas'
suggestion in the survey design object input to Zelig, Zelig should be
able to accommodate this since it passes all the *.survey models to
Thomas' package.
Best,
Olivia
2009/10/10 Fabrício Mendes Fialho <fabriciofialho(a)gmail.com>om>:
Dear Olivia,
Thanks for your message.
Yes, if the sample is not stratified by religious groups (and this survey is
not), I will not have representative samples for these subsets.
However, recently I read a Thomas Lumley's paper ("Analysing Survey Data in
R", R News 3(1), 2003) in which he states that:
"Estimates on a subset of a survey require some care. Firstly, it is
important to ensure that the correct weight, cluster and stratum information
is kept matched to each observation. Secondly, it is not correct simply to
analyse a subset as if it were a designed survey of its own. The correct
analysis corresponds approximately to setting the weight to zero for
observations not in the subset. This is not how it is implemented, since
observations with zero weight still contribute to constraints in R functions
such as glm, and they still take up memory. Instead, the survey.design
object keeps track of how many clusters (PSUs) were originally present in
each stratum and svyCprod uses this to adjust the variance."
According to this statement, subset analysis (for instance, comparing blacks
and whites, men and women, catholic and protestant, and so on) would be
possible since we keep all cases and survey design information while running
our analysis, just attributing zero weight to cases outside the subset of
interest.
My question would be whether or not there is a Zelig command that automatize
this 'zero weighting process'. (but, of course, it is possible that my
question in fact does not make sense.)
Thanks for your attention.
Sincerely,
Fabricio
2009/10/9 Olivia Lau <olivia.lau(a)post.harvard.edu>
Dear Fabricio,
The answer depends on the original survey design. If the sample was
stratified by each of the religious groups, then you can subset to the
group or groups of interest, and analyze data from just those groups
using logit.survey. If the sample was stratified using other
background characteristics, then subsetting to the religious groups
*and* employing survey weights will not create a representative
sample. In this case, you have to refer to the survey design (which
should be in the documentation of the particular survey) for guidance.
If you know something about the alternative target populations (e.g.,
protestant, catholic, etc), then you could possibly reweight the data
after subsetting.
Yours,
Olivia
2009/10/7 Fabrício Mendes Fialho <fabriciofialho(a)gmail.com>om>:
> Hi there,
>
> I would like to know how may I obtain analysis for
> subgroups/subpopulations
> using Zelig. For instance, let suppose I am analyzing data on political
> participation and religion, and I would like to estimate the effects of
> political sofistication on electoral outcome. My data is from a complex
> sample survey. Let also suppose respondents were classified in
> "protestant",
> "catholic", and "others". Considering theoretical issues, it
would be
> interesting to estimate the effect of political sofistication and
> electoral
> outcome just for the "protestant" group.
>
> How should I proceed to obtain this coefficient only for a subgroup (for
> instance, "protestants")? How can I select this subgroup in my sample
> when I
> am running my analysis? For instance, I would like to run a
> "logit.survey"
> model, and I intend to include survey design information (PSU and
> weights)
> in my model. I would like to know how can I run analysis for this
> subgroup
> without just delete the other cases from my dataset since this procedure
> would distort the original sample design. Should I just set the weight
> to
> zero for observations outside the subset of interest? Or is there other
> way
> to proceed? And if I intend to run a "logit" model (not a
> "logit.survey"),
> what should I do?
>
> I looked for directions about this in Zelig manual, but I did not find
> instructions for this kind of analyses.
>
> Best regards,
>
> Fabricio
>
--
Fabrício Mendes Fialho
web:
sites.google.com/site/fabriciofialho/
e-mail: fabriciofialho(a)gmail.com
-
Zelig Mailing List, served by Harvard-MIT Data Center
Send messages: zelig(a)lists.gking.harvard.edu
[un]subscribe Options: