Print

Print


While PREFSCAL is indeed designed to do internal 
unfolding analysis appropriately, I would be very 
wary about using ALSCAL for internal 
unfolding.  The last I heard ALSCAL doesn't use 
the appropriate objective function for unfolding 
analysis, which SHOULD be SSTRESS-2, the analogue 
of STRESS-2, which Joe Kruskal and I concluded 
long ago should be used instead of STRESS-1 if 
KYST or later versions of MDSCAL are used for 
unfolding analysis.  (This was the reason Joe 
included STRESS-2 as an alternative objective 
function in later versions of MDSCAL and in 
KYST.)  Use of the incorrect objective function 
will generally lead to a nominally perfect 
solution (with SSTRESS = 0 in the case of ALSCAL) 
which in fact is degenerate;  e.g., with all 
subjects clustered into a single point and all 
stimuli clustered into a second one.  This 
solution can easily be shown to yield zero STRESS 
(or SSTRESS, in the case of ALSCAL) but to retain 
NONE of the information in the preference 
data.  Other more complicaed degeneracies can 
occur as well, but this simple one is sufficient 
to rule out using ALSCAL for unfolding.  Thus I 
would advise using PREFSCAL but not ALSCAL for 
unfolding of a subjects by stimuli preference 
data or for any other rectangular off-diagonal 
conditional proximity matrix, to use Coombs's term for this type of data.

Doug Carroll

At 07:01 AM 2/5/2008, Art Kendall wrote:
>OOPS!
>
>That should have been PREFSCAL which is in the 
>SPSS CATEGORIES option developed by the people at Leiden.
>
>If you only have access to the Base option take 
>a look at ALSCAL  an older procedure.
>I have attached a PDF file that shows some of the capabilities of ALSCAL.
>
>Documentation of SPSS Algorithms is on the 
>installation CD with all the other manuals even 
>for options you have not purchased.  Since I 
>have all the options I don't know whether the 
>online help includes only options you have purchased, or all options.
>
>If you do not have SPSS, you can also see the algorithms at
><http://support.spss.com/ProductsExt/SPSS/Documentation/Statistics/algorithms/index.html>http://support.spss.com/ProductsExt/SPSS/Documentation/Statistics/algorithms/index.html 
>
>
>
>Art Kendall
>Social Research Consultants
>
>
>Art Kendall wrote:
>>This could be handled as 33 cases and 78 
>>variables in a    TWOSTEP procedure defining the variables a nominal.
>>
>>Another approach would be to reduce to the data 
>>to 13 rank variables according to how often 
>>each sound was preferred, and then clustering cases on the 13 variables.
>>
>>
>>
>>another way you could construct your input data would be like this.
>>You have 33 lower triangular matrices with 13 rows and 13 columns.
>>You can recode your data , e.g.,  -1 meaning 
>>the row sound was preferred, zero meaning a 
>>tie, and +1 meaning the column variable was preferred.
>>
>>This sounds more like a multidimensional 
>>preference scaling  (unfolding) application,
>>Which can be done with PROXSCAL in SPSS.
>>
>>Art Kendall
>>Social Research Consultants
>>
>>
>>
>>Arnaud Trollé wrote:
>>>
>>>Thank you all for your help.
>>>
>>>Sorry, you're right Art, I've been too 
>>>evasive, I should have begun by defining
>>>the framework of my study :
>>>During a listening test, 33 subjects were 
>>>presented 78 pairs of sounds (i.e.
>>>number of possible combinations between 13 sounds).  For each pair, the
>>>subject is asked to indicate which sound he 
>>>prefers, three possibilities : ``first
>>>sound preferred", "second sound preferred", 
>>>and "no preference". Actually, my
>>>data set consits of 33 cases for 78 
>>>categorical variables (all with 3 modalities).
>>>Before any other analysis, my first objective 
>>>is to find out whether there exists
>>>any sub-groups of subjects with distinct preference logics.
>>>So, my approach is exploratory. However, if 
>>>there exists any subgroups (with,
>>>for each, a meaningful size), I'm expecting at most a weak number of
>>>subgroups. Thus, I first went in for 
>>>partitioning methods such as the k-modes
>>>of which I've heard. But, I've got to few experience to even judge whether
>>>this method is one of the most adapted or not to my study case ?
>>>
>>>I hope these elements will help to work out a 
>>>little more my initial questionning.
>>>
>>>Best Regards,
>>>
>>>Arnaud.
>>>
>>>
>>>______________________________________
>>>De : Classification, clustering, and phylogeny estimation [CLASS-
>>><mailto:[log in to unmask]>[log in to unmask]] 
>>>de la part de Art Kendall [<mailto:[log in to unmask]>[log in to unmask]]
>>>Date d'envoi : lundi 4 février 2008 18:30
>>>À : <mailto:[log in to unmask]>[log in to unmask]
>>>Objet : Re: About Partitioning Categorical Data ...
>>>
>>>Please tell us more about your application? Are the values ordered? Are
>>>you trying to find groups of variables or groups of cases (rows,
>>>subjects, entities)?
>>>How many cases (rows) do you have?  How many variables?  Do all of the
>>>variables have 3 values? Are you trying to see how an existing partition
>>>of cases or variables works with other cases or variables?
>>>Often it is helpful to us to know the substantive meaning of your
>>>variables, and what a case represents.
>>>
>>>
>>>SPSS is widely available but there are also many specific purpose
>>>programs around depending on what you are trying to do.
>>>If SPSS itself does not have a procedure, you can call any R procedures
>>>from within SPSS. So you might be able to use several procedures.
>>>If you are partitioning variables into sets, then you might look at
>>>Categorical Principal Components analysis (CATPCA).
>>>If you are partitioning cases into sets, then you might look TWOSTEP
>>>which clusters cases based on either/both categorical and continuous
>>>variables
>>>If you have an existing 3 value variable, that you want to see how the
>>>cases with each value differ on another, TREES, CATREG, and DISCRIMINANT
>>>might be what you could use.
>>>If you have three sets of variables, you can confirm how well a three
>>>factor solution fits in CATPCA by specifying the number of factors you want.
>>>
>>>
>>>Art Kendall
>>>Social Research Consultants
>>>
>>>
>>>Arnaud Trollé wrote:
>>>
>>>>
>>>>Hello,
>>>>
>>>>I'd like to cluster categorical data (3 
>>>>categories) by means of a partitioning
>>>>method; I'm quite a beginner in that field 
>>>>and I would need to be enlightened.
>>>> >From a bibliographic review I carried out 
>>>> about that topic, it appeared to me
>>>>that a method is often used :the k-modes method. From her/his experience,
>>>>could anyone confirm or deny that it is the case ? If denied, which method
>>>>could be more "powerful" ?
>>>>
>>>>Thanks in advance.
>>>>
>>>>Best Regards.
>>>>Arnaud.
>>>>PhD Student in Acoustics.
>>>>Lyon, France.
>>>>
>>>>----------------------------------------------
>>>>CLASS-L list.
>>>>Instructions: 
>>>><http://www.classification-society.org/csna/lists.html#class-l>http://www.classification-society.org/csna/lists.html#class-l
>>>>
>>>>
>>>>
>>>>
>>>
>>>
>>>----------------------------------------------
>>>CLASS-L list.
>>>Instructions: 
>>><http://www.classification-society.org/csna/lists.html#class-l>http://www.classification-society.org/csna/lists.html#class-l
>>>
>>>----------------------------------------------
>>>CLASS-L list.
>>>Instructions: 
>>><http://www.classification-society.org/csna/lists.html#class-l>http://www.classification-society.org/csna/lists.html#class-l
>>>
>>>
>>>
>>---------------------------------------------- 
>>CLASS-L list. Instructions: 
>><http://www.classification-society.org/csna/lists.html#class-l>http://www.classification-society.org/csna/lists.html#class-l 
>>
>---------------------------------------------- 
>CLASS-L list. Instructions: 
>http://www.classification-society.org/csna/lists.html#class-l



   ######################################################################
   # J. Douglas Carroll, Board of Governors Professor of Management and #
   #Psychology, Rutgers University, Graduate School of Management,      #
   #Marketing Dept., MEC125, 111 Washington Street, Newark, New Jersey  #
   #07102-3027.  Tel.: (973) 353-5814, Fax: (973) 353-5376.             #
   # Home: 14 Forest Drive, Warren, New Jersey 07059-5802.              #
   # Home Phone: (908) 753-6441 or 753-1620, Home Fax: (908) 757-1086.  #
   # E-mail: [log in to unmask]                                   #
   ###################################################################### 
----------------------------------------------
CLASS-L list.
Instructions: http://www.classification-society.org/csna/lists.html#class-l