Hi,

I have a dataset consisting of 2 groups of samples on a set of
variables, I would like to fish out a subset of variables or
combination of subset of variables to discriminate the 2 groups, so I
chose to use LDA from MASS library in R to do the analysis.

I foud out that wehther I normalize the data or not (mean 0, variance
1), I get the final prediction exactly the same, justified by plotting
the discriminating scores of the 2 groups of samples and obtaining the
exactly same shape of plot (certainly the scale of the scores are
different). And of course, the linear discriminant coefficients are
different under the conditions of normalized or not, thus picking out
different set variables based on the scale of the coefficients.

My question are:

is it correct that you would get the same final prediction whether you
normalize the data or not?

Then is it correct that the purpose of normalization is solely to
allow you to find out the discriminant power of the original power
because all original variables are on the same scale now so that no
variable will dominate others simply because its variance is too big.
But normalization will not change the prediction of the samples.

I also find out that if I did t tests for each of the original
variables, the t-statistics is not linearly related to the
discriminant power of the original variables whether normalize or not.
That is, a variable that has a large t statsitic may have very low
discrimant power justified by its discrimiant coefficient. how to
explain this?

The last question is my understanding of linear discriminant analysis
with multiple variables is that if 2 variables have high correlation (
like close to 1), then LDA will not give both variables high
discriminant coefficients, but will give only one of them high
coefficient since the 2 variables are redundant. However, I found out
that my results with the above data showed many variables with large
discriminant coefficients are highly correlated. is this normal?

Thank you very much
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to