Your math is wrong, rayiner didn't mention neighborhood variation at all, and Simpson paradox is relevant to changes over time or resampling, not simple rate comparisons