Using directed acyclic graphs to understandconfounding in observational studies