Hi,

kindly see the following link:

https://www.youtube.com/watch?v=3pzCa4Whv74

So far, I have two answers telling me not to do discriminant analysis.

Could someone answer my question?

Afraid I do not understand your problem. According to my understanding one has only one dependent variable in Discriminant Analysis (the grouping variable), and several independent variables. I have done discriminant analyses with many more than 12 independent variables. Do you mean the dependent variable is limited to 12 groups in SPSS? Another consideration is that your sample size (number of cases) must be (much) more than the number of independent variables, but in some applications I got around that constraint by doing stepwise discriminant analysis.

Dear Francois:

Thank you for the clarification. My situation is that the dependent variable has 16 levels and my old SPSS package will only handle 12. A colleague has just told me that newer SPSS versions will handle 16 levels so I will seek out a more modern version. Thanks for your help.

• I am afraid your analysis. There are 120 combination of 2-class discrimination using 16 dependent variables. The best combination of independent variables are quiet different, This was common knowledge of pattern recognition. 

Hi, Paul Yarnold

I think we had better compared multi-class discriminant analysis with the partition.

I think the result of partition may be more better than 16-class discrimination. 

John,

I think that you are concerned   de facto  with one variable which may appear in K=16 categories. You have a sample of n data vectors given as a data matrix X of size nxd, where n stands for number of rows (say, samples of wine), and d for number of columns (say, some chemical attributes characterizing the given wine samples). Generally, the wine samples were recorded in K locations.

You would like  to construct an algorithm which could predict -- for a given  sample of wine -- in which location  this wine sample  was produced.

This is a typical problem which may be solved easily by ANN, Artificial Neural Networks, in particular by Multilayer Perceptron (MLP) with one hidden layer, d inputs and K outputs. You will obtain some scores expressing probabilities that a given data sample was produced in each of the assumed K locations.

There is a splendid package NETLAB written by Ian Nabney in Matlab. The package is free and can be downloaded from the author's homepage. Of course, you need license for the basic version of Matlab (certainly, your University has such licence, it is not free, however - without all the toolboxes - Matlab is not so much expensive).  Working with NETLAB you are not limited to 12 or 20 categories/classes, however, as was mentioned above, with larger number of classes also a larger data sample is needed. There is also a book by Ian Nabney: NETLAB, Algorithms for Pattern Recognition  or another book by C.M. Bishop: Neural Networks for Pattern Recognition explaining the foundations and building strict mathematical models.

 You will have much fun working with MLP. 

Anna

Dear Corey. 

The problem is not with the predictors but with the dependent variable which has 16 different categories. My old SPSS package can only handle 12. 

Dear Anna:

Thanks for your enthusiasm but I definitely do not want to use neural networks. I would like to use discriminant analysis as my independent variables are ratio scaled and I would like to develop a perceptual map of the results. Also, I like the easy to interpret results of discriminant analysis.