Clive,
A few points here: (1) stay away from iweights -- in general, they have no
statistical meaning in stata and exist as a way for programmers to modify an
estimator in a particular way, (2) pweights are the appropriate weight for
sampling issues -- as is often the case with more complicated regression
models, my guess is that the estimator for a random effects probit model
with sampling weights has never been derived, (3) if you control for all
variables that are used to determine your sample weights (and these are
exogenous to your outcome variables) and cluster on households (or PSUs, if
available), your results will be asymptotically equivalent to an estimator
that takes sampling weights directly into account.
Steve
-----Original Message-----
From: Clive Nicholas [mailto:[email protected]]
Sent: Saturday, 8 November 2003 8:48
To: [email protected]
Subject: st: Weighting in -xtprobit-
All,
OK, yet another question I wished to ask the list that I'll probably get
flamed for, but I think it's interesting (and essential for me to fix).
I've recently fitted some -xtprobit- models quite recently. It took time,
but bar one exception, all the models ran with decent results. As a new
Statabod, I was rather pleased with myself (not that that's ever too
difficult). Then I realised something: since this was panel survey data I
was using, I didn't weight the analysis!! :((
I had to start over trying to use Stata's weights. Unfortunately, it won't
run [pw=...], [aw=...] or [fw=...]. It ran [iw=...), but it returned a model
full of low coefficents and highly insignificant p-values. Given that the
data was sampled from households at random, [aw] and [fw] do not appear to
be appropriate (certainly going by their descriptions in [U], 23.16. Option
[pw] looks to be the most appropriate, but Stata returns the error: "pweight
not allowed in random-effects case r(101);".
It's been suggested to me that, instead, I may be able to reduce the effect
by entering the number of adults in household, age and marital status and
other such variables in the model. This is because Scottish households were
oversampled and that people in small households have a much larger chance of
selection than people in large households in this British survey (e.g. a
person in a one adult household has twice the selection chance of a person
in a two adult household). I've entered the last two of these in my models,
but not the first (yet).
So to bring this mini-epic to its cliffhanger, could adding such variables
get round the problem of not being able to formally use -[pw=(weight)]- in
-xtprobit-? Or are there other automatic fixes?
Thanks in advance. :)
Yours,
CLIVE NICHOLAS,
Politics Building,
School of Geography, Politics and Sociology,
University of Newcastle-upon-Tyne,
Newcastle-upon-Tyne,
NE1 7RU,
United Kingdom.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
The information contained in this document is intended only for the
addressee and is not necessarily the views nor the official
communication of the Department of Labour. All final/official papers
which are sent from the Department will be sent by non-electronic
means, on appropriate letterhead, signed by authorised personnel.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
