The positivity assumption and marginal structural models: the example of warfarin use and risk of bleeding