Subject: Re: Sudaan vs SAS, SPSS, etc
From: mcohen@cpcug.org (Michael Cohen)
Date: 17 Nov 1996 02:37:13 GMT
Thomas R. TenHave (ttenhave@BIOSTATS.HMC.PSU.EDU) wrote:
:
: >John Roden wrote:
: >
: > I understand that for datasets like NCES surveys which use a multistage
: >> sampling design, the cases are multiplied by a weight so they
: >> approximate the original population.
[snip]
:
: If the sampling fraction is small and there is no clustering
: in the sample design, then the standard errors produced
: by SAS with the correct weights (i.e., the weights sum
: to the sample size) are accurate.
I think I know what was meant by "correct weights" but I want to make sure
everyone understands why NCES weights are produced the way they are. They
sum to the size of the population, not the size of the sample. This way,
one can get an estimate of a total by computing a weighted sum. These
weights work fine with SUDAAN, PC-WesVar, PC-Carp, and other packages
produced with survey data in mind.
All NCES sample surveys have some clustering. This is because of the
natural hierarchical nature of the data, sampling efficiency, etc. There
may also be unequal probability sampling (e.g. with probabilities
proportional to enrollment), oversampling of certain groups, systematic
sampling, etc.
Having said all this, if one had to use SAS or SPSS to do, say, regression
on NCES sample survey data, one would need to renormalize the weights to
the sample size first. One should apply a design effect to the standard
errors afterwards.
I work for NCES. No official support by the U.S. Department of Education
is intended or should be inferred.
--
Michael P. Cohen home phone 202-232-4651
1615 Q Street NW #T-1 office phone 202-219-1917
Washington, DC 20009-6331 office fax 202-219-2061
mcohen@cpcug.org
