Replies to This Discussion

Permalink Reply by Vibha on April 12, 2009 at 2:11am

Discriminanat analysis is the best method to validate clusters. A good hit ratio implies good clustering. Also check if the F-values for all the variables are significant, else remove the insignificant variables, and rerun the procedure.

• ▶ Reply

Permalink Reply by arup guha on April 12, 2009 at 4:56am

thanks for this. can you tell me a little of the theory behind this? as to what makes dcm the best method to validate clusters?

• ▶ Reply

Permalink Reply by arup guha on April 13, 2009 at 2:53am

i think logistic regression would be better ffor the same purpose. requires lesser assumptions, is more robust...

• ▶ Reply

Permalink Reply by Ralph Winters on April 26, 2009 at 6:03pm

Use logistic regression instead, since the coefficients are easier to interpret; especially if you will be presenting the results to senior management.

-Ralph Winters

• ▶ Reply

Permalink Reply by Tim Brennan on April 26, 2009 at 9:26pm

As I understand it, If the same variables are used in a Discriminant Function analysis to test a cluster solution then all the F ratios and significance tests are invalid since clustering (especially K-means or Ward) is optimizing essentially same criterion that DFA is based on. Thus, no independence between original solution and it's test.
So, the only significance tests are appropriate are those using variables NOT used in the original cluster analysis. Thus, DFA does not give an appropriate significance test of the quality of a cluster analysis solution. One would need a totally independent DFA that uses variables not used in the initial cluster solution.

This may be too rigid, and I would be glad to be corrected.

• ▶ Reply