To respond to Pramod Adhikari and Carol Mathews questions :
I don't known what exactly SAS consider as essential unidimensionality,
specially for continuous variables, but in Item Response Theory (quality of
life analysis, Intelligence, psychology...), the notion of essential
unidimensionality is defined by a set of variables who measured a dominent
concept (and minor concepts). The main reference is Stout, A non parametric
approach for assessing latent trait unidimensionality. Psychometrika, 52(4)
: 589-617 (1987).
Stout and its colleagues present two way to construct clusters of
dichotomous or polytomous items who verify the essential unidimensionality.
The DETECT method and the HCA/CCPROX method:
Zhang and Stout. The theorical DETECT index of dimensionality and its
application to approximate simple structure. Psychometrika 64(2): 213-249
(1999).
Roussos and Stout. Effectiveness of using new proximity measures with
hierarchical cluster analysis to detect dimensionality of simulated data.
Techniocal report, University of Illinois, Departement of Statistics (1995)
. (Submit in 1995 and perhaps accepted in a journal).
In the same time, Molenaar and its colleagues proposed another method to
construct unidimensional sets of dichotomous and polytomous items: The
Mokken Scale Procedure.
Hemker, Sitjsma and Molenaar. Selection of unidimensional scales from a
multidimensionalitem bank in the polytomous Mokken IRT model. Applied
Psychological Measurement.19(4): 337-352 (1995).
If I don't know procedures under SAS or Stata who realize DETECT or
HCA/CCPROX, I can propose programs to realize MSP under SAS or Stata. These
program can be downloaded under my (in construction) web site:
www.anaqol.fr.st who proposed too other SAS and Stata programs to analyse
Quality of life data.
However, these methods can be used only with ordinal or dichotomous
variables.
I hope this helps.
Sincerly
Jean-Benoit Hardouin,
Regional Health Observatory
Orléans - France
Email : [email protected]
On Wed, 17 Sep 2003 11:10:55 +0100, Nick Cox <[email protected]> wrote:
Pramod Adhikari
SAS has a VARCLUS procedure (Variable Cluster Analysis) 'to divide a
set of
variables into nonoverlapping clusters in such a way that each
cluster can
be interpreted as essentially unidimensional' (Source: SAS
documentation).
Is there a similar command in STATA that can perform similar
analysis. With
the existing clustering methods, I can group the variables (after
transposing observations into variables) using 'generate' command
but the
standard output from VARCLUS procedure seems easier to interpret.
I can't comment on SAS's procedure, except that the sentence you quote
is
a very strong claim, promising something that in many datasets may be
impossible.
There is a weasel word, "essentially".
If I had a set of continuous variables, then I would explore such
questions in Stata using -pca-. One crux, however, is judging
closeness of
components to variables, for which the correlations between
components and variables are useful information.
In a posting on 28 August, I showed how to do this more
easily using -makematrix-from SSC.
PCA is less appropriate given categorical variables.
Nick
[email protected]
*
* 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/
--
Jean-Benoit Hardouin
37, rue Arrachart
41000 Blois
tél : 02 54 45 39 75
email : [email protected]
http://www.jb-hardouin.fr.st
http://www.anaqol.fr.st
*
* 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/
