Message-ID: <alpine.LRH.1.10.0809241133070.21470@yoknapatawpha.pols.columbia.edu>
Date: 2008-09-24T16:34:06Z
From: Gregory Wawro
Subject: Weights for polr
In-Reply-To: <Pine.LNX.4.64.0809221616350.32341@homer21.u.washington.edu>
Thanks for directing my attention to the survey package and svyolr().
Best,
Greg
.............................................................
Gregory Wawro gjw10 at columbia.edu
Associate Professor phone: 212-854-8540
Dept. of Political Science fax: 212-222-0598
741 International Affairs http://www.columbia.edu/~gjw10/
Columbia University
New York, NY 10027
.............................................................
On Mon, 22 Sep 2008, Thomas Lumley wrote:
> On Mon, 22 Sep 2008, Gregory Wawro wrote:
>
>> Hello,
>>
>> I'm estimating an ordered logit model on a probability weighted survey
>> sample.
>
> You could use svyolr() in the survey package.
>
>> polr permits case weights with the "weights" option, but I cannot figure
>> out from existing documentation what it actually does with these weights.
>
> They are frequency weights.
>
>> I'm concerned about this because I get somewhat different results using
>> Stata's ologit command with the pweights option
>
> You should get the same point estimates, but different standard errors.
>
>> and very different results using proc logistic in SAS with its weight
>> option.
>
> Again, it should be the same point estimates but different standard errors.
>
>> So my basic question is whether or not it is appropriate to use the weight
>> option for polr with my data.
>
> No.
>
> -thomas
>
> Thomas Lumley Assoc. Professor, Biostatistics
> tlumley at u.washington.edu University of Washington, Seattle
>