Date:         Wed, 12 May 2004 11:33:18 -0400
Reply-To:     Colin_Valdiserri@marketstrategies.com
Sender:       "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From:         Colin Valdiserri <Colin_Valdiserri@marketstrategies.com>
Subject:      Re: Runnnig Means Procedure in K-Means Cluster
Comments: To: Art@DrKendall.org
In-Reply-To:  <40A23979.4040904@verizon.net>
Content-Type: text/plain; charset="US-ASCII"

Thank you for the feedback , and I do agree with never relying on just one
cluster run for the final solution.  In response to your question, the 20
items were not taken from scales, but more so each question was thought up
separately...  I did run principal components anlaysis and derived 6
factors from these 20 items, though only the first factor has a high
cronbach's alpha, the rest are around .30 - .50.  I am investigating these
items further through basic correlation matrices.

We will definitely run the discrim on the back end to validate the
segments.  I am also thinking about splitting the sample randomly in two
and comparing clustering solutions to look at stability.

Thanks again, I will look into the Two-Step prodecure further.  If you
don't mind and have the time could you expand on your resaons for
prefering the newer Two-Step procedure?

Colin M. Valdiserri, M.A.
Marketing Sciences Analyst
Market Strategies, Inc.
colin_valdiserri@marketstrategies.com
Phone: (734) 779-6826



Art Kendall <Arthur.Kendall@verizon.net>
Sent by: "SPSSX(r) Discussion" <SPSSX-L@listserv.uga.edu>
05/12/2004 10:49 AM
Please respond to
Art@DrKendall.org


To
SPSSX-L@listserv.uga.edu
cc

Subject
Re: Runnnig Means Procedure in K-Means Cluster






If you are each doing a few randomizations of the cases with different
runs it shouldn't make much difference.  I have been doing clustering
since 1971 and would never rely on a single run.

That being said, these days I would use TWOSTEP rather than k-means.  I
would follow the clustering with DFA's to to assess the quality of
different solutions.

Also are the 20 items grouped into scales based on constructs?

Art
Art@DrKendall.org
Social Research Consultants
University Park, MD  USA
(301) 864-5570



Colin Valdiserri wrote:

>Hi again everyone,
>
>I am currently working on a consumer segmentation project with another
>colleague.  We have about 1400 cases and 20 attitudinal items that are
>being used as our basis variables.  I usually select the "Running Means"
>option when deriving a clustering solution, if I remember correctly this
>option forces new centers each time a case is added to a cluster as
>compared to re-doing the centers after all of the cases have been added
to
>clusters.  To me it sound like the running means produces a more
>stable/reproducible solution.
>
>My colleague does not use the running means option and before suggesting
>it to them (being  a VP and all), I want to make sure that I am
>understanding it correctly and am offering a possible better solution.
>
>Thanks in advance.
>
>Colin M. Valdiserri, M.A.
>Marketing Sciences Analyst
>Market Strategies, Inc.
>colin_valdiserri@marketstrategies.com
>Phone: (734) 779-6826
>
>
>