Posted presented at ACIC 2022.
- Philippe Boileau*: Graduate Group in Biostatistics and Center for Computational Biology, University of California, Berkeley
- Nima S. Hejazi*: Division of Biostatistics, Department of Population Health Sciences, Weill Cornell Medicine
- Ivana Malenica*: Graduate Group in Biostatistics, University of California, Berkeley
- Peter B. Gilbert: Vaccine and Infectious Disease Division, and Public Health Sciences Division, Fred Hutchinson Cancer Research Center
- Sandrine Dudoit: Department of Statistics, Division of Biostatistics, and Center for Computational Biology, University of California, Berkeley
- Mark J. van der Laan: Division of Biostatistics, Department of Statistics, and Center for Computational Biology, University of California, Berkeley
*: These authors contributed equally.
Causal mediation analysis has provided numerous tools for defining and estimating effects that may be endowed with mechanistic interpretations. Owing to the close alignment of such causal direct and indirect effects with the goals of modern scientific investigations, the natural direct and indirect effects have risen to enormous popularity in applications of causal mediation analysis. Unfortunately, these canonical causal effects have stringent requirements for their identification, making their practical use limited. As a result, considerable effort has been invested in developing alternative direct and indirect effect decompositions with relaxed identification requirements, yet such efforts often yield effect definitions with nuanced and challenging interpretations. By contrast, relatively limited attention has been paid to relaxing the identification assumptions of the natural direct and indirect effects. We present a theoretical study of the identification of the natural direct effect under unobserved baseline confounding of exposure-outcome and exposure-mediator pathways. Our novel identification strategy reveals that this causal effect may be learned even when the total causal effect remains unidentified. Using an intervention strategy that relaxes the commonly used but restrictive cross-world counterfactual independence assumption, we discuss how the natural direct effect may be assessed in randomized controlled trials. Revisiting prior studies of non/semi-parametric efficiency theory required for the construction of flexible, multiply robust estimators of the natural direct effect, we discuss its efficient estimation without imposing restrictive modeling assumptions on nuisance parameters.