R help - Dec 2005 - Logistic regression to select genes and estimate cutoff point?

Hi, all, 
I am new to R or even to statistics. Not sure if the question has a answer. But
I couldn't find a straight forward answer in the help mailing list.
I need use MicroArray data to select several diagnostic genes between Normal
samples and Tumor samples and use these genes to predict unknow samples.
Since the sample size is so small and data doesn't follow normal
distribution, I am thinking to use logistic regression instead of Student T test
to select genes. To make the problem simpler, I assume each gene is independent
to each other without interactions.
My questions is how I should build up the model: one model for each gene or a
multiple variable model to include all genes? Which is the test to compare the
discrimination power of each gene? I am thinking it is Wald statistic for the
multiple variable model and Maximum likelihood for the single gene models? Am I 
correct?
To estimate the cutoff point, I guess the answer is the gene expression when
p=0.5 in the model. Am I on the right direction?
Any suggestion is appreciated!
Thanks a lot.
Lingsheng 

	[[alternative HTML version deleted]]

You could take a look at www.bioconductor.org
limma would be a good starting point.

hth
ido

Lingsheng Dong wrote:
> Hi, all, 
> I am new to R or even to statistics. Not sure if the question has a answer.
But I couldn't find a straight forward answer in the help mailing list.
> I need use MicroArray data to select several diagnostic genes between
Normal samples and Tumor samples and use these genes to predict unknow samples.
> Since the sample size is so small and data doesn't follow normal
distribution, I am thinking to use logistic regression instead of Student T test
to select genes. To make the problem simpler, I assume each gene is independent
to each other without interactions.
> My questions is how I should build up the model: one model for each gene or
a multiple variable model to include all genes? Which is the test to compare the
discrimination power of each gene? I am thinking it is Wald statistic for the
multiple variable model and Maximum likelihood for the single gene models? Am I 
correct?
> To estimate the cutoff point, I guess the answer is the gene expression
when p=0.5 in the model. Am I on the right direction?
> Any suggestion is appreciated!
> Thanks a lot.
> Lingsheng 
Just a comment: Do you not have a statistician to work with at your 
institution?  You are new to statistics and are asking a question that 
would be very difficult to deal with for someone with a PhD in 
statistics and 20 years of experience.  Some of the issues involved are 
multiple comparisons, false discovery rate, shrinkage, array geometry 
effects, nonparametric vs. parametric statistics, stability of selected 
genes, discovery validation, ...

-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University