"In observational studies, we often have to rely on assumptions that are not directly testable."