Discriminant Analysis Question re-post..

[Ranjana Dutta]

Hello all, 

I had posted this a couple days ago.  I haven't received any messages back, not even out of office replies :) So, I am thinking maybe something went wrong with the post. So here is another attempt. If there is an issue with the question, please draw it to my attention.

***

How can one test if Centroids of groups in Discriminant Analysis are significantly different from each other? I would think it should be possible..but am not sure how to go about it. The graphs shows the Means on the Discriminant Functions and variation around the means but no tabular output on variance or SD of the discriminant functions.

***

[statisticsdoc]

Ranjana,

 

Perhaps it would be informative to start with a MANOVA using the groups as the independent variable and the function scores as the dependent variables.  The follow-up tests will at least tell you which groups differ from one another in the univariate sense.   Please note that in principle it is possible for groups to differ significantly on a multivariate profile that is not reflected in the univariate post hoc tests.

 

Best Regards,

 

Stephen Brand, Ph.D.

 

www.StatisticsDoc.com

 

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Ranjana Dutta
Sent: Sunday, November 18, 2012 7:50 PM
To: [hidden email]
Subject: Discriminant Analysis Question re-post..

 

Hello all, 

I had posted this a couple days ago.  I haven't received any messages back, not even out of office replies :) So, I am thinking maybe something went wrong with the post. So here is another attempt. If there is an issue with the question, please draw it to my attention.

***

How can one test if Centroids of groups in Discriminant Analysis are significantly different from each other? I would think it should be possible..but am not sure how to go about it. The graphs shows the Means on the Discriminant Functions and variation around the means but no tabular output on variance or SD of the discriminant functions.

***

[Rich Ulrich]

Well, the DF's do not work pairwise, so they won't tell you directly
about the separate groups in any case -- if you are asking about
tests on pairs of groups.  IIRC, you can request the univariate tests
between groups (but not between pairs of groups).

The tests on Wilks's lambda are the usual overall tests.  There are
k-1  "canonical roots" if there are k groups.  You can use the tests
on separate roots, in decreasing order of importance, to see how
many roots are relevant. 

If you hypothesized that there was only one "dimension" that all the
groups fall on, then you might use Roy's Greatest Root test for the
overall. 

I never tried to make any use of the actual, k-1 discriminant functions
or of the k  "classification functions", one for each group. 

--
Rich Ulrich

---

Date: Sun, 18 Nov 2012 16:50:27 -0800
From: [hidden email]
Subject: Discriminant Analysis Question re-post..
To: [hidden email]

Hello all, 

I had posted this a couple days ago.  I haven't received any messages back, not even out of office replies :) So, I am thinking maybe something went wrong with the post. So here is another attempt. If there is an issue with the question, please draw it to my attention.

***

How can one test if Centroids of groups in Discriminant Analysis are significantly different from each other? I would think it should be possible..but am not sure how to go about it. The graphs shows the Means on the Discriminant Functions and variation around the means but no tabular output on variance or SD of the discriminant functions.

***