Date:         Fri, 17 Jul 1998 17:58:33 -0500
Reply-To:     "Nichols, David" <nichols@SPSS.COM>
Sender:       "SPSSX(r) Discussion" <SPSSX-L@UGA.CC.UGA.EDU>
From:         "Nichols, David" <nichols@SPSS.COM>
Subject:      Re: Ques: Confidence Levels in SPSS

SamplePower does not deal with complex samples. The new Wesvar Complex
Samples 3.0 program that SPSS distributes deals with complex samples.
 
I didn't read the original question as having anything to do with weights
or complex samples, but as a request for a confidence interval that related
to confidence in the sample data rather than to a population parameter, and
as I responded to the original question just now, I don't know that such a
thing exists.
 
David Nichols
Principal Support Statistician and
Manager of Statistical Support
SPSS Inc.
 
----------
From:  Hector E. Maletta [SMTP:hmaletta@overnet.com.ar]
Sent:  Monday, July 06, 1998 1:03 PM
To:  SPSSX-L@UGA.CC.UGA.EDU
Subject:  Re: Ques: Confidence Levels in SPSS
 
Cindy Wong wrote:
>
> Question:
>
> Is it possible to obtain a confidence level (%) from sample data (i.e. a
> statistic such as the mean) that would describe the confidence we have
in
> that sample as compared with the population (assuming normal
distribution)
> in SPSS?  If so how would this be done (SPSS 6.1)?  Any insight or
comments
> would be much appreciated.
 
 
Many SPSS procedures produce confidence levels (for instance T-TEST,
CROSSTABS and others). The values are based on the assumption that the
cases represent a simple random sample out of a population of infinite
size.
 
If your sample is not a simple random sample, the results will not be
correct: clustered samples effects tend to enlarge errors,
stratification tends to reduce them.
SPSS considers that the WEIGHTED number of cases is the size of the
sample. Thus, if you're using the WEIGHT command to 'expand' your
results to the size of the estimated population, you'd get an
exaggerated confidence level (SPSS will assume your 'sample' has the
size of your population). To correct this you should use unweighted
data, or (if sampling probabilities are different among cases, and thus
the weighting factor does not only expand the total but also give cases
different weight, you should use a trick recently explained in this same
list: create a new weighting factor = old factor x n/N where n=sample
size and N=estimated population size. This leaves clustering and
stratification problems aside, but at least the resulting confidence
levels will refer to a simple random sample with the same size as your
actual sample.
A recent addition to the SPSS family of products (sample power)
apparently deals with complex sampling designs. Classic SPSS knows only
simple random samples.
 
 
Hector Maletta
Universidad del Salvador
Buenos Aires, Argentina