[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: proper use of aweight
Austin Nichols <firstname.lastname@example.org>
Re: st: proper use of aweight
Fri, 20 Apr 2012 12:45:10 -0400
Carlianne Patrick <email@example.com> :
I think the [U] 20.18 text is aimed at people who use [aw] for survey
data, without robust variance estimation, and thereby dramatically
understate p-values and overreject null hypotheses. If you use [aw]
and appropriate variance estimation, you should be fine.
By the way:
My parenthetical attack on the use of plural "codes" to mean lines of
code is not directed at you, but at the increase in such incorrect
usage that will eventually make it correct, by the unfortunate
property of language that a plurality rules. Any longtime Statalister
who is sensitive to English usage will know what I mean about the
increasing use of "codes" to mean "code" (rather than distinct codes
for ID or ICD or what have you).
On Fri, Apr 20, 2012 at 12:22 PM, Carlianne Patrick
> Thank you for the help and apologize for incorrectly using "posted code". I was referring to the supplemental .do files available online for several (non-STATA) journal articles. After reading the STATA reference manual [U] 20.18, it seemed aweight should only be used with mean data. The section states that: "There is a history of misusing such weights. A researcher does not have cell mean data, but instead has . . ." I didn't know if using it in the situations described was part of the history of misuse. However, it sounds like aweight is more flexible.
> Thank you for the article on propensity score weighting. It helped clear up some other issues!
> Best, Carlianne
> Carlianne Patrick, CEcD
> PhD Candidate
> Department of Agricultural, Environmental, and Development Economics
> The Ohio State University
> From: firstname.lastname@example.org [email@example.com] on behalf of Austin Nichols [firstname.lastname@example.org]
> Sent: Friday, April 20, 2012 11:07 AM
> To: email@example.com
> Subject: Re: st: proper use of aweight
> Carlianne Patrick <firstname.lastname@example.org>:
> I cannot speak to whether unreferenced "posted codes" (presumably, a
> contraction of "posted code snippets" that sounds terrible to native
> English speakers who also program, since code is a mass noun here) are
> correct in their use, but aweights are of very general use. Their use
> gives the same point estimates as other weights, fweights and
> pweights, but does not inflate the overall sample size as does the use
> of fweights, nor impose robust variance estimation as does the use of
> pweights. If you are unsure about how to do correct inference, in
> many cases you may well be better off bootstrapping and using aweights
> in each bootstrap resample. Unless you are matching; the bootstrap is
> inappropriate for matching.
> See also http://www.stata-journal.com/sjpdf.html?articlenum=st0136_1
> On Thu, Apr 19, 2012 at 6:21 PM, Carlianne Patrick
> <email@example.com> wrote:
>> I am new to STATAlist, so please excuse any posting faux pas I may make in my initial posting(s). Also, thank you in advance for any help.
>> I am using STATA11.2; however, my question is regarding the posted code from some published papers (and thus I don't know which version they were using). I would like to use the techniques in these papers for my analysis, but am concerned that the code isn't doing what the papers say it is doing.
>> After reading the available STATA documentation on aweight, it appears that this particular weighting option is really only suitable when the observations are means of underlying data. However, in the posted codes it is used to accomplish one of two things:
>> 1) Weight the data with lagged values of the dependent variable.
>> 2) Weight "matches" by the inverse of their number or log odds ratio (specifically, in situations where multiple untreated observations are "matched" to one treated observation).
>> I may misunderstand what aweights is doing. If not, then it appears that it is not the appropriate weighting option command to accomplish (1) or (2) if you are weighting by the log odds ratio. It seems appropriate only if you have already taken the average of the matches. Then, the average of the matches could by weighted by their number and the treated observation weighted by 1.
>> Is this the correct interpretation?
>> If it isn't, is there a reference that might help me understand why aweight is appropriate in these situations?
* For searches and help try: