Thread Subject: multiple Regression; reducing variables;

I would like do multiple linear regression (MLR) of my data.
Problem is, Ive more predictor variables (matrix X) than
dependent variables (Vector Y). To use MLR, ive to reduce
the number of my predictor variables and analyse the
regression coefficients.

[Matrix Y = 1*38]= [Matrix X = 38*48]

Matrix X contains binary data for nucleotide sequence
(encoded from A, T, G, and C). The binary data is
represented by 4 possibilities/position. So 48 columns in
Matrix X represent 12 positions (12*4 = 48). I cant run MLR
with this dataset. one of the idea to reduce the number of
the predictors is by setting value of one of the nucleotide
to be zero. But i am not sure how to do it? Whether to get
the regression coefficients before i set value of one
nucleotide value to zero. or something like robust
regression to used so that i can do iterative regression
calculations?
I hope im clear in my question? Any suggestions?

On 26 Nov, 08:09, "Vivek " <vivek_muta...@yahoo.com> wrote:
> I would like do multiple linear regression (MLR) of my data.
> Problem is, Ive more predictor variables (matrix X) than
> dependent variables (Vector Y). To use MLR, ive to reduce
> the number of my predictor variables and analyse the
> regression coefficients.

One way is to use the Singular Value Decomposition.
Use as many predictor variables as there are "significant"
singular values. That leaves two problems:

- How to discriminate between "significant" and "insignificant"
  singular values
- How to map the original predictor vatriables to the reduced
  set of predictors.

Rune