Using directed acyclic graphs to understand confounding in observational studies.