Hi,
Last week I posted questions on the Socnet list and got
the responses below. I have also appended the email addresses of those who
responded with their responses.
Regards.
MY QUERY BELOW:
________________________________________
Hello again
everyone,
A colleague of mine
has overwhelmed me with some 500 matrices.Each matrix is a 15X15 cognitive map.
The expected hypothesis use QAP and are not so strongly supported.
Therefore, we are
exploring how to cluster these 500 matrices into groups of matrices depending on
certain characteristic of each matrix such as centralities of nodes, etc. The
end result could be a few clusters where matrices in these clusters are
similar in certain respects.
With this
clustering, we plan to add dummy variables for these matrices in a QAP
regression and hope to get more significance.
1. Is this a
statistically valid thing to do?
2. Which software
package facilitates clustering of matrices (unlike clustering nodes in a single
matrix) ?
3. Running a QAP
correlation on a hundred pairs of matrices is troublesome if I use the user
interface in UCINET. Is there a way one can write a simple small program to
automate this once the matrices are stored in certain known files? (e.g., like a
SAS procedure)
Thank
you.
Regards.
Nilesh
Saraf
RESPONSES :
_________________________________
Kathleen Carley and I have recently done some work on
this problem. The paper in question is
currently under review, but you can find a draft at http://legba.hss.cmu.edu/~eagle/distribution/multiv.1.0.pdf This paper provides a general approach to distance
methods (e.g., cluster analysis, MDS) and covariance analysis (e.g.,
PCA, CCA, network regression) for graph
sets under arbitrary labeling conditions. I have S routines which will implement these analyses,
some of which are in my network S toolkit.
(The rest will be added eventually, but aren't
currently in a userfriendly form; OTOH, the structdist and
gscor/gscov routines are all you really
need, and these are in the current
release.)
Hope that helps,
 Carter T Butts
[[log in to unmask]]
______________________
Hi Nilesh

My advice would be to cluster a structural summary of the matrix, and
the best one that comes to mind would be the triad census. That is, you
could:
a)
Calculate the triad census for each matrix, which results in a 16 element
frequency distribution for each
matrix
b)
you would then have a dataset with 500 observations and 16
variables
c)
run that through a cluster analysis program to get your clusters.
If it
were me, I'd do it all in SAS. I have programs for getting the triad
census, and then use the SAS PROC CLUSTER to do the clustering. It is also
fairly easy to write a QAP routine in SAS (I've one for regression and logistic
regression).
Let me know if you need any of these programs.
Best
of luck,
Jim Moody
[[log in to unmask]]
_________________________
Nilesh,
1. In UCINET, you can JOIN all the matrices
together into a single file containing 500 2way matrices: in effect a 15x15x500
matrix. Then you can run SIMILARITIES on that file, specifying MATRICES as the
dimension to correlate (at least in fairly recent versions of ucinet). The
result is a 500 by 500 matrix of similarity coefficients. You can then submit
this correlation matrix to any clustering routine.
2. In SPSS or SAS, it is not difficult to cluster
your 500 matrices. Just string out each one as a row of 225 numbers (call these
variables). So you have a new matrix that is 500 rows by 225 columns. Then
submit to clustering routine (specifying cases as the dimension to cluster).
3. I'm not sure what you are trying to do so is
difficult to comment on the larger issues. But I think you flirt with the Dark
Side whenever you manipulate the data expressly to "get more significance".
_________________________
3. Running a QAP correlation on a hundred pairs
of matrices is troublesome if I use
the user interface in UCINET. Is there a way one can write
a simple small
program to automate this once the matrices are stored in certain known
files? (e.g., like a SAS
procedure)
If you use Stata, you might want to check out a small
progam called, conveniently, qap at: http://ideas.uqam.ca/ideas/data/bocasug01.html
Gindo Tampubolon
[[log in to unmask]]
University of Manchester
________________________________________
Hello Saraf,
please contact Stephen Borgatti about
QAPprocedure and the application of it.
The tests are somewhat questionable as he said at the last Sunbelt
Conference in Budapest.
Gerhard Wuehrer
[[log in to unmask]]
______________________________
it is not
statistically valid.
___________________