R help - Aug 2007 - LDA and RDA: different training errors

Hello

I try to fit a LDA and RDA model to the same data, which has two classes.
The problem now is that the training errors of the LDA model and the
training error of the RDA model with alpha=0 are not the same. In my
understanding this should be the case. Am I wrong? Can someone explain what
the reason for this difference could be?

Here my code:

LDA model:
==========% x is a dataframe
tmp = lda(response ~ ., data=x)
tmp.hat = predict(tmp)
tab = table(x$response, tmp.hat$class)
lda.training.err = 1 - sum(tab[row(tab)==col(tab)])/sum(tab)

RDA model:
==========% x is converted into a matrix without the response
% variable. This matrix is then transposed
tmp = rda(x, y, alpha=0, delta=0)
rda.training.err =  tmp$error / dim(x)[2]                         

% The training error provided by rda.cv() is also different
% from the training errors provided by lda() or rda()
tmp.cv = rda.cv(tmp, x=x, y=y, nfold=10)
tmp.cv$err / dim(x)[2] / 10


Thanks a lot!

dominic senn wrote:
> Hello
> 
> I try to fit a LDA and RDA model to the same data, which has two classes.
> The problem now is that the training errors of the LDA model and the
> training error of the RDA model with alpha=0 are not the same. In my
> understanding this should be the case. Am I wrong? Can someone explain what
> the reason for this difference could be?

I assume lda from MASS?
If you are using rda() from package "rda", I do not know, since the
help
page is not very specific in telling which parameter means what (but I 
guess one of them should be 1).

If you choose rda() from package "klaR", the help page tells you that
gamma=0, lambda=1
should produce identical results to LDA. (lambda=1 means that the pooled 
covariance matrix is weighted with 1 while the specific covariance 
matrices are weigthed with 0.


Uwe Ligges

> Here my code:
> 
> LDA model:
> ==========> % x is a dataframe
> tmp = lda(response ~ ., data=x)
> tmp.hat = predict(tmp)
> tab = table(x$response, tmp.hat$class)
> lda.training.err = 1 - sum(tab[row(tab)==col(tab)])/sum(tab)
> 
> RDA model:
> ==========> % x is converted into a matrix without the response
> % variable. This matrix is then transposed
> tmp = rda(x, y, alpha=0, delta=0)
> rda.training.err =  tmp$error / dim(x)[2]                         
> 
> % The training error provided by rda.cv() is also different
> % from the training errors provided by lda() or rda()
> tmp.cv = rda.cv(tmp, x=x, y=y, nfold=10)
> tmp.cv$err / dim(x)[2] / 10
> 
> 
> Thanks a lot!
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.