Regarding Simpson's paradox, you can see a good treatment here [0] describing whether splitting a group is correct or not (based on causality); see Section 2.2 specifically.

[0] http://ftp.cs.ucla.edu/pub/stat_ser/r414.pdf