-------- Original Message --------
Subject: RE: Discriminant analysis
Date: Mon, 22 Dec 2008 07:11:24 -0800 (PST)
From: andrea cardini <[email protected]>
To: [email protected]
Dear Sarah,
there's a bug in SPSS DA. I found that out too as, when I coded my groups
with large (>20 or 100) non consecutive numbers, I used to get an error
message saying that there was not enough memory for doing the analysis. If
I recoded groups as you did, everything was fine and results similar (or
identical) to those of other statistical software.
Cheers
Andrea
At 07:15 22/12/2008 -0500, you wrote:
-------- Original Message --------
Subject: RE: Discriminant analysis
Date: Thu, 18 Dec 2008 16:10:18 -0800 (PST)
From: Sarah Degroot <[email protected]>
To: <[email protected]>
References: <[email protected]>
I think I figured out the problem: SPSS (both 11.0 and 15.0) was having
trouble with the cross-validation when the group identifiers were not
consecutive numbers. For example, my groups were labeled 55, 181, 197,
273, and 274. Another look at the cross-validation case-wise statistics
showed that the second highest group in most cases was non-existent,
e.g. 1, 143, 127, 220. I re-numbered the groups (1, 2, 3, 4, 5) and
received a cross-validated with 84.0% correct classification, which I
have reason to believe is about right for these data.
FYI, those of you using SPSS.
________________________________
From: morphmet [mailto:[email protected]]
Sent: Thu 18-Dec-08 9:41 AM
To: morphmet
Subject: Discriminant analysis
-------- Original Message --------
Subject: Discriminant analysis
Date: Thu, 18 Dec 2008 09:38:39 -0800 (PST)
From: Sarah Degroot <[email protected]>
To: <[email protected]>
I'm having a problem with discriminant analysis. I am using SPSS 15.0.
With my ln-transformed or raw corolla truss data, I get correct original
classifications of 96.0 to 100%. Cross-validated classification is 0.0
to 2.7% correct. Random groups of the same data gives 54.7% correct
originally and 20.0% correct cross-validated (which makes sense given
that I am trying to discriminate 5 groups).
It seems that there is some signal in the data, otherwise I'd think the
original classification for the random groups would be higher. But I am
very puzzled why the cross-validated classification of the real groups
is so low, at or near 0% correct, i.e. worse than random. It's like
something is forcefully insisting on the wrong answer...
Any idea why the cross-validated classifications are so poor?
Thank you,
sarah.degroot at cgu.edu
--
Replies will be sent to the list.
For more information visit http://www.morphometrics.org
<http://www.morphometrics.org/>
--
Replies will be sent to the list.
For more information visit http://www.morphometrics.org
--
Replies will be sent to the list.
For more information visit http://www.morphometrics.org
