A1248
Title: Partial identification for spatiotemporal causal inference under unmeasured confounding
Authors: Jiaxi Wu - University of California, Santa Barbara (United States) [presenting]
Alexander Franks - UC Santa Barbara (United States)
Abstract: Unmeasured confounding that varies over space and time can severely bias causal effect estimates from observational data. We introduce a unified framework that allows partial and, in some cases, point identification of causal effects in the presence of spatiotemporal confounding. The approach decomposes residual covariance into three latent components: unit-specific, time-invariant, and spatially invariant factors, treating the remaining variation as noise. Closed-form bias bounds for the unit-specific component are indexed by interpretable R-squared-style sensitivity parameters and become tighter under spatial exclusion restrictions or multi-resolution designs. Shared confounding across units is captured by low-rank factor models, and partial interference assumptions further sharpen identification by anchoring the latent structure. Simulation studies demonstrate that the proposed approach substantially reduces bias compared to spatial smoothing and double machine learning baselines, particularly when confounders exhibit non-spatiotemporal variation. We illustrate the method in a case study of the effect of prenatal PM2.5 exposure on birthweight in California. The estimated effect remains stable across a range of covariate adjustment sets, underscoring the method's robustness to unmeasured confounding.
