Discriminant analysis and prediction of membership

[andrea1977]

Hi,

I have conducted a cluster analysis to individuate cluster of individuals on the basis of their scores on the BIG FIVE questionnaire, then I computed the disriminant analysis to individuate the linear combination of such dimensions that better discriminate between clusters. However, now I am interested in using these results to catgorize another sample of individuals on the basis of my previous analysis.
How could I predict membership of the new sample's partcipants on the basis of the previous sample's participants catgorization ?

thnx!

[statisticsdoc]

Andrea,

This syntax should get you started.  This example assumes that you have
saved cluster membership in a variable called PersonalityCluster, and that
there are three clusters (you can change the number of clusters as needed).
The Big Five personality variables are represented by the variables var1
var2 var3 var4 var5  .

discriminant groups =  PersonalityCluster(1,3)
    variables = var1 var2 var3 var4 var5
     / save class prob
    /STATISTICS=BOXM COEFF RAW TABLE /  rotate coeff .

The output from this run will give you a table called Classification
Function Coefficients.  You can use these coefficients to compute a variable
that is the likeness of each new case to the cluster,  In this example, the
table will have five weights and a constant that can be used to compute the
likeness of the case to cluster 1, and the same for cluster 2 and 3:

Compute class1 = weight11*var1 + weight12*var2 + weight13*var3 +
weight14*var4 + weight15*var5 + constant1  .
Compute class2 = weight21*var1 + weight22*var2 + weight23*var3 +
weight24*var4 + weight25*var5 + constant2  .
Compute class3 = weight31*var1 + weight32*var2 + weight33*var3 +
weight34*var4 + weight35*var5 + constant3  .


And then assign the case to a predicted cluster (pred_clus) depending on
which likeliness score is highest

if ( ( class1 gt class2) and (class1 gt class3) and (class1  gt  class4 ) )
pred_clus =  1  .

if ( ( class2 gt class1) and (class2 gt class3) and (class2  gt  class4 ) )
pred_clus =  2  .

if ( ( class3 gt class1) and (class3 gt class2) and (class3  gt  class4 ) )
pred_clus =  3  .


Best Regards,

Stephen Brand


www.StatisticsDoc.com


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
andrea1977
Sent: Sunday, November 25, 2012 3:18 PM
To: [hidden email]
Subject: Discriminant analysis and prediction of membership

Hi,

I have conducted a cluster analysis to individuate cluster of individuals on
the basis of their scores on the BIG FIVE questionnaire, then I computed the
disriminant analysis to individuate the linear combination of such
dimensions that better discriminate between clusters. However, now I am
interested in using these results to catgorize another sample of individuals
on the basis of my previous analysis.
How could I predict membership of the new sample's partcipants on the basis
of the previous sample's participants catgorization ?

thnx!



--
View this message in context:
http://spssx-discussion.1045642.n5.nabble.com/Discriminant-analysis-and-pred
iction-of-membership-tp5716445.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command SIGNOFF SPSSX-L For a list of
commands to manage subscriptions, send the command INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Art Kendall]

The old -fashioned way is to add the new cases to your original file.  The new cases would have a system missing value on the cluster membership variable.  First recode the membership variables so sysmis values are replaced with a value that is outside
assuming you have 5 clusters and have added value labels that reflect your interpretation of the profile try something like this (untested)

recode membership (sysmis = 6) (else=copy) into membership.
add value labels membership 6 'from new sample'.

then rerun the same syntax you used before, still with 5 groups.  The new cases will be ungrouped.  The classification phase will show which cluster the old cases and the ungrouped cases would be assigned to.

Unless you have many thousands of cases, it should not be much effort to try a few clustering methods on all of the cases you now have.

<soapbox>
Clustering is a heuristic exploratory method.  I have been using it since 1971.  I never rely on a single combination of agglomeration method and similarity coefficient.  Also it is somewhat an art to decide on the number of cluster to retain.  TWOSTEP as one of the methods to use does give an AIC/BIC for a range of number of clusters.
</soapbox>
 

On 11/25/2012 3:17 PM, andrea1977 wrote:

Hi,

I have conducted a cluster analysis to individuate cluster of individuals on
the basis of their scores on the BIG FIVE questionnaire, then I computed the
disriminant analysis to individuate the linear combination of such
dimensions that better discriminate between clusters. However, now I am
interested in using these results to catgorize another sample of individuals
on the basis of my previous analysis.
How could I predict membership of the new sample's partcipants on the basis
of the previous sample's participants catgorization ?

thnx!



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Discriminant-analysis-and-prediction-of-membership-tp5716445.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD