I've gotten a number of microarray datasets from GEO and arrayexpress, each of which looks at similar chemical treatments of mammalian cells. I downloaded the processed data, and from this I can easily generate log2 fold change values for each probe between the treated sample and the untreated sample. Are there any caveats to comparing these log2 FC values between datasets? Is this type of analysis OK, or do I need to go back to the CEL files for each dataset and normalize everything in one particular way?
Normalizing CEL files from different studies is fraught with statistical problems.
What you want to perform is a meta-analysis. Try searching this site for that term with respect to microarrays - there have been several previous questions on the topic. If you're using Bioconductor, take a look at the RankProd package for a good introduction to the topic.