Date: Tue, 28 Feb 2006 04:11:46 -0800
Reply-To: Dennis Fisher <dfisher@CSULB.EDU>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Dennis Fisher <dfisher@CSULB.EDU>
Subject: Re: QR:cluster analysis with binary variables
In-Reply-To: <20060228100458.54918.qmail@web26411.mail.ukl.yahoo.com>
Content-Type: text/plain; charset="ISO-8859-1"; format="flowed"
Adel,
Part of the issue is how many binary variables do you have
and how many axies do you think you will wind up with? My
experience with correspondence analysis is that it is used
for large two-way tables. I wind up with two axies. Two
variables would not make a very interesting cluster
analysis. An ideal number of variables for a cluster
analysis would be in the neighborhood of about half a
dozen up to about two dozen. This would be both
interesting and manageable. If you do a correspondence
analysis, I do not understand why you do not stop there
and interpret your results and present them. If you want
to do a cluster analysis, then just do a cluster analysis
with binary variables. I do not understand why you want
the additional comnplication of doing both.
Dennis
On Tue, 28 Feb 2006 11:04:58 +0100 (CET)
"adel F." <adel_tangi@yahoo.fr> wrote:
> Hi,
> Do you think, that a correspondence analysis approach
>for binary variables, followed by a cluster analysis
>using axies in the correspondence analysis step, is not
>appropriate to obtain the clusters?
> Adel
>
> "Dennis G. Fisher" <dfisher@CSULB.EDU> a écrit :
> I am glad you said this. In their book Aldenderfer and
>Blashfield attack
> this idea a great deal. Their example is that people
>will do a cluster
> analysis and then do a discriminant analysis on the same
>variables that
> made up the cluster and discriminate the clusters. They
>then try to
> claim that this is a measure of the validity of the
>cluster analysis.
> A&B come out strongly against doing this.
> Dennis Fisher
>
> David L Cassell wrote:
>
>> adel_tangi@YAHOO.FR wrote back:
>>
>>> Hi David,
>>> Thank you very much for your interesting reply; this
>>>encourages me to go
>>> further in my analysis. I will try to interpret the axes
>>>at the
>>> correspondence analysis step, as you suggest, in order
>>>to interpret the
>>> results of the cluster analysis easily.
>>>
>>> I am also thinking that after obtaining the clusters, I
>>>can do a
>>> multinomial analysis (proc catmod) with the clusters as
>>>dependent
>>> variable
>>> and using the original binary variables as independent
>>>variables,
>>> this can
>>> also explain the association between the original binary
>>>variables
>>> and the
>>> clusters.
>>
>>
>> I don't recommend this last part. You built the cluster
>>specifically as a
>> clustering
>> algorithm on linear combinations of (some of) the binary
>>variables. This
>> addition
>> only makes your reasoning particularly circular. You
>>will be forcing the
>> results to
>> show you the axes from the correspondence analysis. Any
>>variables which
>> were
>> important in those axes will come out as important here,
>>and any
>> variables
>> which
>> were overlooked in those axes (the ones you used) will
>>be overlooked
>> here.
>> I
>> do not think that this extra analysis will help you any.
>>
>> HTH,
>> David
>> --
>> David L. Cassell
>> mathematical statistician
>> Design Pathways
>> 3115 NW Norwood Pl.
>> Corvallis OR 97330
>>
>> _________________________________________________________________
>> FREE pop-up blocking with the new MSN Toolbar – get it
>>now!
>> http://toolbar.msn.click-url.com/go/onm00200415ave/direct/01/
>>
>
> --
> Dennis G. Fisher, Ph.D.
> Director
> Center for Behavioral Research and Services
> 1090 Atlantic Avenue
> Long Beach, CA 90813
> 562-495-2330
> 562-983-1421 fax
>
>
>
> ---------------------------------
> Nouveau : téléphonez moins cher avec Yahoo! Messenger !
>Découvez les tarifs exceptionnels pour appeler la France
>et l'international.Téléchargez la version beta.
| 
