Selection Bias in Causal Mediation Analysis

• Activity Number: 183
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Linda Valeri*+ and Brent Coull
• Companies: Harvard and Harvard School of Public Health
• Keywords: Mediation analysis ; Non-ignorable missingness ; Selection bias ; Sensitivity analysis

Abstract

An important goal across the biomedical and social sciences is to identify and quantify the importance of pathways through which an exposure affects an outcome. Selection bias might severely undermine the validity of inferences on direct and indirect causal effects in observational as well as in randomized studies. The phenomenon of inappropriate selection may arise through several mechanisms. We focus on instances of non-ignorable missingness due to self-selection in the study, drop-out, and non-response. Our approach is non-parametric and allows the investigator to evaluate the impact of selection bias in causal mediation analyses. In particular, we study the sign and magnitude of selection bias in the estimates of causal direct and indirect effects. Under some simplifying assumptions, we show that the bias formulae are particularly easy to use in a sensitivity analysis. The sensitivity analysis can be applied to causal effects on the risk difference and risk-ratio scales irrespective of the estimation approach employed and extend to clustered and longitudinal data settings. The results are illustrated via an example in the context of environmental epidemiology.