Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: cluster analysis validation


From   "Dasinger, Lisa" <ldasinger@thezenith.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: cluster analysis validation
Date   Thu, 19 Apr 2012 13:30:13 -0700

Brendan, 

Thank you very much for the clarification on how to use the -ari- and
-permtab- commands.  I've been able to use them successfully (as well as
carry out your initial suggestion).  Thank you for all your help.

Lisa

Date: Tue, 17 Apr 2012 21:15:34 +0100
From: brendan.halpin@ul.ie (Brendan Halpin)
Subject: Re: st: cluster analysis validation

On Tue, Apr 17 2012, Dasinger, Lisa wrote:

> I've downloaded your programs, and I will try your suggested solution.
> What is the input for the ari command?  I don't see a help file for
it.
> Would it be the dataset variable (old vs. new) and cluster group
> variable (with values 1..6)?  Also, I see that the permtab command
seems
> to require a square matrix, but I would have a 6 x 2 matrix.

Both -ari- and -permtab- compare two classifications of the same size
(number of categories). If your two cluster solutions are o6 and n6

. ari o6 n6

will calculate the Adjusted Rand Index for the comparison, and

. permtab o6 n6 

will rearrange n6 so that it agrees as much as possible with o6, and
then cross-tabulate them. If you do -permtab o6 n6, newvar(p6)- it will
create in p6 a copy of the rearranged n6. 

Both of those commands serve to compare the old and the joint
classifications, if you wish to do that. My initial suggestion (where
the 6x2 table comes into it) is to compare the distribution of the two
waves across the joint classification. In the simplest sense, this
means examining the percentages within wave, but you could extend it to,
say, a multinomial logistic regression (with the cluster solution as the
dependent variable) and wave as one of the explanatory variables. 

If you want more info on the Adjusted Rand Index, there are some notes
in the form of comments in the ari.ado file -- its presence in the
package was something of an afterthought, so I never set up a help file.
ARI works in terms of pairs: if every possible pair of observations that
have the same value in one classification have the same value in the
other, and every pair that has different values in one has different
values in the other, the agreement is perfect and ARI is 1.0. Otherwise
the index is less than one.


Regards,

Brendan

PS: Note that since permtab permutes one of the classifications, its
runtime rises factorially, to the extent that it is useless from about
10 categories up. -permtabga- estimamtes an approximate solution for
larger classifications. 
- -- 
Brendan Halpin,   Department of Sociology,   University of Limerick,
Ireland
Tel: w +353-61-213147  f +353-61-202569  h +353-61-338562;  Room F1-009
x 3147
mailto:brendan.halpin@ul.ie    ULSociology on Facebook:
http://on.fb.me/fjIK9t
http://teaching.sociology.ul.ie/bhalpin/wordpress
twitter:@ULSociology








Date: Tue, 17 Apr 2012 11:36:07 -0700
From: "Dasinger, Lisa" <ldasinger@thezenith.com>
Subject: Re: st: cluster analysis validation

Thank you, Brendan,

I've downloaded your programs, and I will try your suggested solution.
What is the input for the ari command?  I don't see a help file for it.
Would it be the dataset variable (old vs. new) and cluster group
variable (with values 1..6)?  Also, I see that the permtab command seems
to require a square matrix, but I would have a 6 x 2 matrix.

Lisa


Date: Mon, 16 Apr 2012 15:57:30 -0700
From: "Dasinger, Lisa" <ldasinger@thezenith.com>
Subject: st: cluster analysis validation

I've run a cluster analysis in Stata 11.2 based on three continuous
variables using -cluster wardslinkage- with the default
similarity/dissimilarity measure to generate 6 groups.  I'd like to know
if there is a way to apply the same cluster analysis to a new set of
data.  In other words, is there a way to run a new dataset through the
old cluster analysis and see how new observations are classified, akin
to running a regression equation and then taking a new dataset to obtain
out of sample predictions?  
 
If so, is there a way to evaluate how well the "old" analysis fits the
new data, e.g., by determining how similar/dissimilar each new
observation is to the observations in the cluster in which each is
placed, and whether the new observation is placed in the "best" cluster,
meaning the one that minimizes the distance between the observation and
the "center" of the existing cluster?

I am new to cluster analysis and am looking for a way to validate the
"old" cluster analysis.  
 
Lisa


Lisa Dasinger, Ph.D.

Data Reporting Manager
Claims Analytics

 

Zenith Insurance Company
Pleasanton Regional Office
4309 Hacienda Drive, Suite 200
Pleasanton, CA 94588
 

ldasinger@thezenith.com

www.TheZenith.com
 
- ------------------------------

Date: Tue, 17 Apr 2012 00:54:03 +0100
From: brendan.halpin@ul.ie (Brendan Halpin)
Subject: Re: st: cluster analysis validation

On Mon, Apr 16 2012, Dasinger, Lisa wrote:

> I've run a cluster analysis in Stata 11.2 based on three continuous
> variables using -cluster wardslinkage- with the default
> similarity/dissimilarity measure to generate 6 groups.  I'd like to
know
> if there is a way to apply the same cluster analysis to a new set of
> data.  In other words, is there a way to run a new dataset through the
> old cluster analysis and see how new observations are classified, akin
> to running a regression equation and then taking a new dataset to
obtain
> out of sample predictions?  

I would suggest pooling the two data sets, running a new cluster
analysis, and analysing the resultant 6*2 table (cluster classification
by old/new). That would test the extent to which the two data sets are
similarly distributed across a joint classification. If that's
acceptable (and the combined data set is small enough) it is a clean and
easy solution.

If you are concerned that the joint classification is not compatible
with the old classification, you can compare the old cluster membership
with the joint cluster membership, for the old data. A good measure of
agreement is the Adjusted Rand Index. Comparing cluster solutions is
tricky because they don't have "labels" -- there is no way of saying
that a given group in classification A is the same as any particular
group in classification B, apart from having shared membership. The ARI
takes that into account.

In theory you can relate the new data to the old classification by
calculating the distance from each new observation to the old cluster
centroids, but I don't know an easy way of doing that with Stata.


Regards,

Brendan


PS: I have code to estimate the ARI, and to re-arrange cluster solutions
to maximise similarity. If you are interested, check out:

   net from http:teaching.sociology.ul.ie/sadi
   net install sadi

and look at the -ari- and -permtab- commands.
- - -- 
Brendan Halpin,   Department of Sociology,   University of Limerick,
Ireland
Tel: w +353-61-213147  f +353-61-202569  h +353-61-338562;  Room F1-009
x 3147
mailto:brendan.halpin@ul.ie    ULSociology on Facebook:
http://on.fb.me/fjIK9t
http://teaching.sociology.ul.ie/bhalpin/wordpress
twitter:@ULSociology



Lisa Dasinger, Ph.D.
Data Reporting Manager
Claims Analytics
 
Zenith Insurance Company
Pleasanton Regional Office
4309 Hacienda Drive, Suite 200
Pleasanton, CA 94588
 
Phone: 925.416.5235
RightFax: 925.460.1235
Branch: 925.460.0600
ldasinger@thezenith.com
 
www.TheZenith.com
 

***********************************************************
NOTICE:
This e-mail, including attachments, contains information
that may be confidential, protected by the attorney/client
or other privileges, or exempt from disclosure under
applicable law.  Further, this e-mail may contain
information that is proprietary and/or constitutes a trade
secret.  This e-mail, including attachments, constitutes 
non-public information intended to be conveyed only to the
designated recipient of this communication, please be
advised that any disclosure, dissemination, distribution,
copying, or other use of this communication or any attached
document is strictly prohibited.  If you have received this
communication in error, please notify the sender
immediately by reply e-mail and promptly destroy all
electronic and printed copies of this communication and
attached documents.

***********************************************************



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index