Tomas Lind <[email protected]>:
Yes, you are specifying each observation as a cluster with that
-svyset- call, which is what you want. What do you mean the data are
not "truly Poisson" exactly? Do you mean that you know that the
conditional mean E(y|X)=exp(Xb) for some b, but you don't know much
about the distribution of errors, in which case -poisson- is
consistent for b (see also help file for -ivpois- on SSC), and you
want SEs that are robust in various ways, i.e. give approximately
correct inference under various error distributions? [Personally, I
would make vce(robust) the default and have users step up to vce(cl
var) or step down to vce(oim), but I suppose there is no hope of that
change, even with a new change-oriented administration here in DC.]
On Tue, Jan 27, 2009 at 11:20 AM, Tomas Lind <[email protected]> wrote:
> My data is a stratified (on geographical areas) crossectional survey
> weighted for design and dropout. Within each strata is a simple random
> sample of a few hundred observations. I need robust standard error because
> data are not truly Poisson (but with robust standard error it works
> according to litterature).
>
> I intended to do analysis like this:
>
> svyset [pweight=wgt] , strata(str)
>
> svy: poisson y x1 x2 x3 , irr
>
> Will it work?
>
> /Tomas
>
> Re: st: Robust SE and svy
> Austin Nichols
>
> Tomas Lind <[email protected]>:
> I would say rather that -svy- is the same as vce(cluster psu) plus
> weighting plus FPC etc. if specified. A cluster-robust VCE is also
> heteroskedasticity-robust, so you can think of vce(robust) as
> "included in" or "a special case of" vce(cluster psu) in some sense,
> and the two are the same if each cluster contains exactly one
> observation (which is the case if you -svyset _n- or -svyset,srs- as
> Steve implies below). In most applications, clustering is more
> important than weighting which is more important than any FPC, but
> weighting can change point estimates whereas clustering only affects
> VCEs. I usually leave off any FPC because I am interested in
> model-based inference about an abstract data-generating process, not
> about describing the finite population from which the survey data was
> drawn, but note that this is not quite right for a stratified
> sample--google "korn graubard superpopulation inference" for some
> relevant literature.
>
> Do you in fact have a "simple random sampling cross-sectional study"
> or are you using a survey with a cluster design? Either way, you may
> want to specify the vce(cluster var) option to account for possible
> clustering by geographical area etc.; if there is no clustering, the
> vce option will have little effect (i.e. there is little harm is
> incorrectly clustering, as long as you have many balanced clusters, so
> that no cluster contains more than about 2% of the data).
>
> On Tue, Jan 27, 2009 at 9:04 AM, Steven Samuels
> <[email protected]> wrote:
>> --
>> Yes it does. The default for -svyset- is vce(linearized), which is the
> same
>> as vce(robust) in the non-survey setting.
>>
>> If you are doing a Poisson or other regression model, you should not use
>> the fpc anyway. That is appropriate only for descriptive statistics
> about
>> the sampled population. A good reference for all this is: Lohr, S. L.
>> (1999). Sampling: Design and Analysis. Pacific Grove, CA: Brooks Cole
>> Publishing Company. If you will be doing much survey analysis, I also
>> recommend that you purchase the Stata survey manual.
> \>
>> -Steve
>>
>> On Jan 27, 2009, at 3:30 AM, Tomas Lind wrote:
>>>
>>> Does svy imply that I have robust SE?
*
* 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/
