A0989
Title: Causal learning with invalid instruments for high-dimensional imaging responses
Authors: Yiting Wang - University of Virginia (United States)
Chunlin Li - University of Virginia and Iowa State University (United States)
Shan Yu - University of Virginia (United States) [presenting]
Abstract: Large-scale neuroimaging studies provide high-dimensional imaging phenotypes for understanding brain-related diseases and their associations with genetic, environmental, and clinical factors. However, causal inference remains difficult due to unmeasured confounding. Instrumental variables (IVs) are commonly used to address confounding, but identifying valid IVs is often difficult in practice. Existing IV methods face significant limitations when applied to high-dimensional imaging data, particularly in handling large-scale data and incorporating the intrinsic spatial structure of images. To address these challenges, a novel framework is proposed, causal image-on-scalar regression with invalid instrumental variables (CISR-IIV), which enables estimation of spatially varying causal effects in the presence of potentially invalid IVs. The approach integrates nonparametric spatial smoothing to capture the spatial structure of imaging data, combined with a lasso-based instrumental variable selection strategy to handle potentially invalid instruments. Rigorous theoretical guarantees are established for the CISR-IIV framework, including selection consistency and the asymptotic distribution of the estimated causal effects. Building on these theoretical results, asymptotic confidence intervals and data-driven simultaneous confidence regions are further constructed. CISR-IIV is applied to the Alzheimer's disease neuroimaging initiative to illustrate its effectiveness.
