Partial Identification Bounds and Path-Specific Effects: Two (More) Options When Faced with Exposure-Induced Confounding

• Activity Number: 97 - Mediation in the Presence of Post-Treatment Common Causes of the Mediator and the Outcome
• Type: Invited
• Date/Time: Monday, July 31, 2017 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Caleb Hilliard Miles* and Ilya Shpitser and Phyllis Kanki and Seema Meloni and Eric Tchetgen Tchetgen
• Companies: University of California, Berkeley and Johns Hopkins University and Harvard T.H. Chan School of Public Health and Harvard T.H. Chan School of Public Health and Harvard University
• Keywords: Mediation ; Nonparametric identification ; Partial identification ; Path-specific effect ; Natural direct effect ; Semiparametric

Abstract

A unique and challenging feature of the natural direct and indirect effects (NDE and NIE) is that, even in point-treatment settings, one must consider the potential confounding impact of covariates affected by the exposure. This is in contrast with total effects, for which this concern only arises in longitudinal settings. If covariates affected by the exposure confound the effect of the potential mediator on the outcome, then the NDE and NIE are not nonparametrically identified. I will propose two options for the investigator facing this conundrum. The first is to respect the inability of the data to identify these effects, and instead estimate partial identification bounds. The other option is to consider an alternative causal estimand that also has a mechanistic mediation interpretation, and may also be of substantive interest. Specifically, I will discuss identification and estimation of a path-specific effect that captures the effect of the exposure transmitted along the path through the intermediate variable of interest, but not through the exposure-induced confounders. I will illustrate both options in an HIV data set from the Harvard PEPFAR program in Nigeria.