Correcting experiment results We have performed a microarray screening of about 200 samples. In each sample we measure about 100 different variables. For technical reasons the screening of these 200 samples was divided into two batches with a couple of weeks interval between them. When all the data has been collected, I have performed principle component analysis (PCA) on the 200 x 100 table. 
When we look on linear projection of first 4 components (responsible for ~70% of the variability), we see a clear division between the two experiment batches. An illustration to what I have can be seen here: http://img594.imageshack.us/img594/3687/pca.png
What are the accepted techniques to address this issue?
 A: How did you normalise your microarray data? Standard ways are:


*

*Robust Multichip Average (RMA)

*Genechip RMA - this can be a bit slow for lots of samples.


This presentation gives a good overview of the two techniques.

There are two R microarray tutorial papers which may also help:


*

*A microarray analysis for differential gene expression in the soybean genome using Bioconductor and R (link to paper)

*Analysing yeast time course microarray data using Bioconductor (link to paper).


Both these papers provide the data and R commands.
Competing interest: I'm a author on the second paper.
A: I would think that the first step would be to examine the component loadings and the actual variables to see if you can identify why the two batches yielded discernible differences. Depending on the reasons for the differences you may or may not be able to use a statistical control to "correct" the results.  
However, if you have every reason to believe samples were randomly assigned to batches, that the your testing methods have not reached floors or ceilings in their ability to assess the variables of interest, and the actual scores themselves are not of interest but only their relative position, perhaps you could standardize scores along each variable by batch.
Edit:  It looks like csgillespie understands your research area and has provided you some good links.  All I was suggesting was that for each batch you could calculate a Z score for each observed variable.  This would have the effect of eliminating batch differences since for each variable in each batch the mean would be the same (0) and the standard deviation would be the same (1).
A: Using PCA only to analyze your gene expression data can lead to (possibly) wrong conclusions. The fundamental assumption behind PCA is equal genewise variation and this may NOT be justified almost always. The links to the paper to analyze time course data by csgillespie point to better ways.
