LDA vs Cluster Analysis

[WeeG]

Hi Guys,

I have a basic question...

If I have a dataset, containing many variables, and based on that, I want to be able to divide (classification) the subjects into two groups (let's say, healthy and not healthy), which method is the best, Discriminant Analysis or Cluster Analysis ? What is the difference between them ?

One more thing, if I choose LDA, how do I check a multivariate normality assumption ?

cheers

[gianmarco]

Hi,
Leaving aside LDA, that I do not know too much, I was wondering if you are interested in knowing how variables possibly affect the group divisions you will find. In this case, you could consider to use use Correspondence Analysis (along or in alternative to Cluster Analysis).

Regards
Gm
Gm

[bugman]

Cluster analysis is more of an exploratory technique, while DA is a formal tool for testing statistical hypotheses about group membership (it is like a reverse MANOVA in a sense).

[mattc]

In general LDA is for supervised learning which means you know the class membership and want to use the IVs to predict membership.

Cluster analysis is an unsupervised learning method which means the class membership is unknown to you and you want to group subjects according to their IVs.

Hope it helps.

[terzi]

Hi everyone,

Mattc has clearly stated the main differences between Discriminant and Cluster Analysis: discriminant techniques will create models that can be used to predict future observations. I just wanted to intrude to talk about the Multiavriate normality assumption which can be verified with some formal tests, such as the Doornik-Hansen omnibus test or Mardia's measure of multivariate kurtosis/skewness. It can also be checked with some graphical procedures, plotting quantiles against Mahalanobis' distance.

Hope this adds something valuable