A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect