Discriminant Analysis Question

[Marcos Sanches]

Hi There!

Out of curiosity I ran a quick test (in SPSS13) with Discriminant Analysis (DA).

First I ran a DA to predict a 5 category variable, using some independent variables. I asked for the predicted group membership to be appended to the file. Then I ran the DA again using the predicted groups as dependent variable and the same independent variables as in the first analysis. I was surprised to see that the DA could not predict its own predicted groups with 100% precision.

A Regression Analysis will predict with R2 = 100% the predicted dependent variable from a previous Regression Analysis. Why that is not so with the DA?

This results made me think the DA is neither a optimum nor a consistent procedure for classification.

Thanks a lot!

Marcos

[John F Hall]

I don't think it's supposed to be 100% accurate, just to estimate a likely nominal category from a set of interval scale predictors.  I did it once with an in-school survey where gender was missing (the question was at the end of a self-completion questionnaire and disruption meant it wasn't completed by some pupils).  I used DFA to predict sex from attitude items and it got about 95% right.  Magazine readership and favourite TV program helped it to get to 100%.

 

On national surveys (UK) it was pretty good at allocating political party support from attitudes to trade unions, welfare benefit recipients, and other divisive issues.

 

----- Original Message -----

From: [hidden email]

To: [hidden email]

Sent: Saturday, May 29, 2010 4:07 AM

Subject: Discriminant Analysis Question

Hi There!

Out of curiosity I ran a quick test (in SPSS13) with Discriminant Analysis (DA).

First I ran a DA to predict a 5 category variable, using some independent variables. I asked for the predicted group membership to be appended to the file. Then I ran the DA again using the predicted groups as dependent variable and the same independent variables as in the first analysis. I was surprised to see that the DA could not predict its own predicted groups with 100% precision.

A Regression Analysis will predict with R2 = 100% the predicted dependent variable from a previous Regression Analysis. Why that is not so with the DA?

This results made me think the DA is neither a optimum nor a consistent procedure for classification.

Thanks a lot!

Marcos

[Marcos Sanches]

Hi John,

Thanks for your answer!

Yes, in a real world application I would be very happy if I got 90%. But my point is another one, I am not trying to predict an ordinary nominal variable, I am trying to predict the predicted nominal variable, which is an output of the DA itself. In other words, you run a DA and get predicted categories that are 95% accurate. Ok, now you run the DA again trying to predict your 95% accurate predicted categories. I would like to know why this second time it is not 100% accurate. (at least not in SPSS13), given that it is possible to get 100% accuracy with a linear function.

I am sorry, I knwo this is not really a SPSS question but it is bugging me...

Thanks 

Marcos

2010/5/29 John F Hall <[hidden email]>

I don't think it's supposed to be 100% accurate, just to estimate a likely nominal category from a set of interval scale predictors.  I did it once with an in-school survey where gender was missing (the question was at the end of a self-completion questionnaire and disruption meant it wasn't completed by some pupils).  I used DFA to predict sex from attitude items and it got about 95% right.  Magazine readership and favourite TV program helped it to get to 100%.

 

On national surveys (UK) it was pretty good at allocating political party support from attitudes to trade unions, welfare benefit recipients, and other divisive issues.

 

----- Original Message -----

From: [hidden email]

To: [hidden email]

Sent: Saturday, May 29, 2010 4:07 AM

Subject: Discriminant Analysis Question

Hi There!

Out of curiosity I ran a quick test (in SPSS13) with Discriminant Analysis (DA).

First I ran a DA to predict a 5 category variable, using some independent variables. I asked for the predicted group membership to be appended to the file. Then I ran the DA again using the predicted groups as dependent variable and the same independent variables as in the first analysis. I was surprised to see that the DA could not predict its own predicted groups with 100% precision.

A Regression Analysis will predict with R2 = 100% the predicted dependent variable from a previous Regression Analysis. Why that is not so with the DA?

This results made me think the DA is neither a optimum nor a consistent procedure for classification.

Thanks a lot!

Marcos