Re: A question about Proc Survey

lee.kitty@YAHOO.COM wrote:
>I'm interested in knowing more about proc surveymeans/surveyreg etc. I'm
>reading David's recent sugi proceeding.
>
>http://www2.sas.com/proceedings/sugi31/193-31.pdf

You should have come to the SUGI presentation. That was more
entertaining than the paper.

>I know this topic came up a lot on sas-L. Just a quick question---what if I
>have data from some probability sampling but have NO idea what kind of
>sampling method the survey house used, i.e. I can't tell if it's a
>stratified sample or not, even if it is, I have no idea what variable(s)
>are
>stratified on. Can I still use proc surveymeans/surveyreg?

Not really.

If you cannot get enough information to decide on stratification or
clustering or sampling weights, then you're stuck. Even papers like
Dumouchel and Duncan indicate that you have to have some idea
what the samplnig weights are, and what are rational surrogates for
the strata.

If you cannot get info about the design effects and the weights from
the survey house, then play hardball. Make a business case to your boss
that the survey house is just wasting your company's money and he/she
needs a new survey company. Explain that the data from the survey
house can come up with drastically different answers depending on
the real weights and design effects, and the survey house won't tell
you, and that maybe they don't *know* this stuff and they're just
ripping you off. Once you get some sort of buy-off on this idea, get
your boss to pound on the survey house to squeeze this out of
them. Of course, they may really *be* ripping you off. They may
have totally inadequate samplnig designs with faulty fieldwork and
lax data collection, so that their data are garbage.

>David's examples suggest that results can be significantly diff. If I'm in
>the dark with no information about the sampling method of my data and use
>regular proc means/proc reg, is there a way to gauge how off are my
>estimates are?

Not really. You can make wild guesses about the sampling weights
and the design effects and plug those into the models, but that only
tells you a worst-case situation, not a meaningful one. But if you have
ten different scenarios to show your boss, and they yield a wide range
of results that are a lot wider than your company can endure, then
that's evidence that you desperately need the design information, and
the survey house is possibly in violation of their contract with your
company since the data are not usable.

I'm sure the survey house would *love* a nice big lawsuit that would
cost them a huge amount of business as well, and put them in a really
bad light. :-)

Sorry I couldn't be of more help,
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330

__________________________________________________ _______________
Express yourself instantly with MSN Messenger! Download today - it's FREE!
http://messenger.msn.click-url.com/g...ave/direct/01/