RE: st: logit with iweight

["Bontempo, Daniel E" <deb193@ku.edu>]

In a quick test, it seems to do the same [i.e., not put weighted N in e(N)] thing with pweights, and I also saw that some postestimation commands are not to be used with SVY prefix. However, it does use the weighted N with fweights. We can also consider the validity of iweights separate from the software's performance (undocumented, and changing across versions) when iw are used.

The issue in our analyses is to have cases oversampled in the study weighted in a manner that corresponds to their prevalence in the population. Our longitudinal sample had 80% children at risk of reading difficulty, and only 20% not at risk. However in the general population we would typically have 80% not at risk and 20% at risk.

We do not want to adjust for the purposes of inferring prevalence in the population, but only to have the predictive utility of IV in our analyses of this sample "speak to" the performance of these same IV when used by a school district on the entire population of students. This is why we down-weight the influence of "at risk" cases. I think this may be the exact kind of thing for which iweights were intended - after all they are there for a reason.

It is possible to select integer (frequency) weights that accomplish the same thing, and you get the exact same b-weights you get with iweights, so the weighting is doing what it is intended to do. However iweights gives a weighted N very close to the actual N, and fweights greatly inflate the sample size. In doing any power analysis, or using any stepwise selection, results do vary with sample size.

So the question is, "Is there a statistical reason not to report weighted N with iw and pw as it does for fw?"



-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Maarten Buis
Sent: Thursday, March 15, 2012 9:20 AM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: logit with iweight

On Thu, Mar 15, 2012 at 3:15 PM, Bontempo, Daniel E wrote:
> I found iewights does not change the reported N for logit regression, although the desired weighting is reflected in point estimates. As a result, when adding CoxSnell R2 to stored results, the wrong N is used for calculation.

To quote -help weight-:

 "iweights have no formal statistical definition; any command that supports iweights will define exactly how they are treated.  Usually, they are intended for use by programmers who want to produce a certain computation."

So you probably don't want to use iweights anyhow.

Hope this helps,
Maarten

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany


http://www.maartenbuis.nl
--------------------------
*
*   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/



*
*   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/