I just published another article on BlueCanary, this one is about Simpson's Paradox.
Would you believe it's possible to look for correlations within sub-populations, and find them in agreement across the board, but when you look at them in aggregate, the correlation reverses? If you're having trouble believing it, check out the article.