Sensitivity analysis for unmeasured confounding with high-dimensional outcomes via outcome reduction

Music City Center 

When multiple outcomes are of interest in an observational study, each outcome often exhibits different sensitivities to unmeasured confounding, where some outcomes are more sensitive to biases from an unmeasured confounder while others are less sensitive. Also, if the outcomes share a common low-dimensional structure the individual biases from unmeasured confounding also have (up to rotation) the same low-dimensional structure. Leveraging both features, we propose a novel procedure, LaunchODR, which conducts a sensitivity-aware dimension reduction for testing causal effects in observational studies with multivariate outcomes. Specifically, LaunchODR modifies standard dimension reduction methods to identify the shared low-dimensional structure and conducts the "least sensitive test" to assess average treatment effects across outcomes. Under some assumptions, we show LaunchODR asymptotically identifies the correct low-dimensional embedding, and the resulting test has Type I error control and consistency. We demonstrate LaunchODR on a population-scale epigenomic dataset to study the causal effect of Amyloid-β on multiple DNA methylation regions.

Sensitivity analysis

Unmeasured confounding

Dimension reduction

epigenomics

DNA methylation

Alzheimer's disease 

Biometrics Section