Date: Wed, 20 May 2009 01:54:41 -0300
Reply-To: Hector Maletta <hmaletta@fibertel.com.ar>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Hector Maletta <hmaletta@fibertel.com.ar>
Subject: Re: Classification: LPA and cluster analysis
In-Reply-To: <411323.60247.qm@web81804.mail.mud.yahoo.com>
Content-Type: multipart/alternative;
With hierarchical cluster analyhsis (CLUSTER command in SPSS) there is no
single solution: the procedure starts with N clusters of 1 member each, and
finishes with one cluster including all N cases as members. Therefore it is
not surprising that you found hierarchical cluster to come up with a "one
cluster solution": that is just the last step in the procedure, not "the
solution". What you have to do next is examine the various "solutions", with
1 to N clusters including all the intermediate results (the penultimate one
was a solution with two clusters) to see whether any of them is of your
liking. Remember, in all this, that clustering is not a parametric but a
heuristic procedure. There is no "correct" solution. You can check,
externally, which clustering solution is better for your particular
purposes. For instance, if you are interested in some particular criterion,
and we seek forming clustering that are maximally homogeneous internally,
and maximally distinct between them, in some other variable, you can use
one-way ANOVA with different clustering solutions to see which is best for
that purpose. Likewise, if you want to have a moderate number of clusters,
from 2 to six say, you can restrict yourself to those "solutions" and try to
choose the one you judge the best.
As each procedure uses a different algorithm to include or exclude cases
in/from clusters, it is not surprising either that solutions are not
necessarily coincident case by case. Even within the same procedure, say
Hierarchical or quick cluster, using different criteria may end up with
different clustering decisions for specific cases. Such is the nature of
clustering.
Hector
_____
From: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU] On Behalf Of
Dale Glaser
Sent: 20 May 2009 01:09
To: SPSSX-L@LISTSERV.UGA.EDU
Subject: Classification: LPA and cluster analysis
Good evening all.......I would be interested in gathering your insights into
classification differences in hierarchal and nonhierarchical cluster
analysis vs. latent profile analysis. I obtained (n = 111) a 2-class model
using Mplus (8 continuous level predictors), and decided to compare the
classification with cluster analysis in SPSS. Using the k-means
QuickCluster option and constraining to 2-cluster solution, the
classification results were very similar to the latent profile analysis.
However, using the hierarchical approach (with Euclidian distance measure
and average linkage method) essentially a one-cluster solution results. I
was searching some texts/aritcles today trying to find out why there may be
congruity between the finite mixture modeling and nonhierarchial cluster
analysis methods but not necessarily so with the hierarchical approach, but
I couldn't find any sources. One of my multivariate texts did state that
based on the seed/type of partitioning as well as type of clustering
algorithm, it may not be atypical to have discordance between the two types
of clustering methods (hierarchical vs. nonhierarchical) so I wonder if this
extends to finite mixture modeling?
Any insights would be most appreicated. thank you...............
Dale
Dale Glaser, Ph.D.
Principal--Glaser Consulting
Lecturer/Adjunct Faculty--SDSU/USD/AIU
President, San Diego Chapter of
American Statistical Association
3115 4th Avenue
San Diego, CA 92103
phone: 619-220-0602
fax: 619-220-0412
email: glaserconsult@sbcglobal.net
website: www.glaserconsult.com
[text/html]
| 
