I'm trying to compute correlation between two genes across multiple samples from a microarray analysis (Affy). The pearson correlation changes depending on whether i compute the correlation in the original intensity domain or in log2-domain. I'm told that i should be doing this in log2-space, but i'm not sure what the reasoning is?

I'm not an expert on statistics. But usually you got better statistic properties after log transformation. For example, in some cases, you'll have a distribution closer to normal distribution after taking the log, which is more trackable statistically.