Varieties of Sensitivity Analysisfor Mediation

• Activity Number: 484
• Type: Invited
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Tyler VanderWeele*+
• Companies: Harvard School of Public Health
• Keywords: Causal inference ; Direct and indirect effects ; Mediation ; Sensitivity analysis ; Confounding ; Measurement error

Abstract

Methods for direct and indirect effects have been popular within the social sciences for decades and have recently been receiving increasing attention from the statistics community. To interpret estimates of direct and indirect effects causally several strong identification assumptions are required. In most application settings these will not hold. Sensitivity analysis techniques can be useful in assessing the extent to which violations in the assumptions would changes substantive conclusions. Several sensitivity analysis techniques will be presented and illustrated including sensitivity analysis for unmeasured mediator-outcome confounding, for time-dependent confounding, and for measurement error of the mediator. Several applications will be presented in which such sensitivity analysis reasoning sheds considerable light on the applications in question. Examples will be presented from perinatal epidemiology, psychiatric epidemiology and genetic epidemiology.