Art,
Thank you so much for taking the time to provide these helpful, detailed
suggestions. I will try them out, and see how it goes.
Liza
From: Classification, clustering, and phylogeny estimation
[mailto:[log in to unmask]] On Behalf Of Art Kendall
Sent: Thursday, September 04, 2008 5:43 AM
To: [log in to unmask]
Subject: Re: cluster analysis validation technique
If you have SPSS here are some ways to do this.
the squared Euclidean distance is the sum of the squared distances on
each dimension.
If you have 10 z variables try something like this untested syntax.
which will find the distance of each case from each centroid.
create 60 variables for the centroids in a file with 1 "case" with a
variable called constant set to 1, and 6 sets of 10
cen1z1 to cen1z10 cen2z1 to cen2z10 ...cen6z1 to cen6z10
in your main file
compute constant=1.
match files file=main /table= centroids by constant.
do repeat
vector
distance= distance1 to distance6
/ z = z1 to z10
/ center1 = cen1z1 to cen1z10
/ center2 = cen2z1 to cen2z10
. . .
/ center6 = cen6z1 to cen6z10.
loop #i =1 to 6
compute distance(#i)=0.
loop #j = 1 to 10.
distance (#i) = distance(#i) + ((center(#i) - z(#j)**2).
end loop.
end loop.
If you do not have a huge number of cases and have a fairly powerful
machine a solution with less effort on your part but a lot of
computation for the machine might be this.
Just add 6 cases to the main each representing a centroid at the top of
the files and do PROXIMITIES on the large matrix and then delete the
columns you do not want.
Another way to look at the agreement between two solutions is to do the
clusterings with filtered cases saving the memberships.
Then do two DISCRIMINANTs, each time treating the other set of cases as
unclustered in the classification phase saving the assignments and
probabilities of membership on each pass.
Then CROSSTAB the assignments on the DFA with those from the original
clustering.
Art Kendall
Social Research Consultants
Liza Rovniak wrote:
Hi,
I am hoping someone here can help me with a "how to" question on running
McIntyre and Blashfield's (1980) nearest-centroid evaluation procedure
to validate the stability of my cluster analysis solution. I am a newbie
to cluster analysis, so this is my first time running this procedure.
I have a sample of about 900 observations and have randomly split the
sample in two (Sample A and Sample B). I conducted hierarchical cluster
analysis and then calculated the centroid vectors for a 3-cluster
solution on each of these two subsamples (i.e., steps 1 through 4 of
McIntrye and Blashfield's evaluation technique).
Step 5 of McIntrye and Blashfield's technique is to calculate "the
squared Euclidean distance for each of Sample B's objects from each of
the centroids of Sample A," and Step 6 is to assign "each object in
Sample B to the closest centroid vector." At this point, I am not sure
what buttons to press in SPSS to complete the analysis. One possibility
I tried is to use K-means cluster analysis to achieve these two steps,
but K-means uses simple Euclidean distance (not squared Euclidean
distance as recommended by McIntyre and Blashfield) to assign the
observations to clusters. Is this okay? (someone told me it was, but I
just want to double-check). I would greatly appreciate any guidance on
what buttons to press in SPSS/appropriate syntax to complete steps 5 and
6 of this analysis.
Thank you.
Liza Rovniak
Liza S. Rovniak, PhD, MPH
Adjunct Assistant Professor
Center for Behavioral Epidemiology & Community Health
Graduate School of Public Health, San Diego State University
San Diego, CA 92123
Phone: 858-505-4770, ext. 152; Fax: 858-505-8614
Email: [log in to unmask]
---------------------------------------------- CLASS-L list.
Instructions:
http://www.classification-society.org/csna/lists.html#class-l
---------------------------------------------- CLASS-L list.
Instructions:
http://www.classification-society.org/csna/lists.html#class-l
----------------------------------------------
CLASS-L list.
Instructions: http://www.classification-society.org/csna/lists.html#class-l
|
