A spectral confounder adjustment for spatial regression with multiple exposures and outcomes
On this page:
Spatial confounding, sometimes defined as missing confounders having spatial patterns, is hard to accurately identify and adjust, and even more difficult in multivariate settings. To remedy this problem, we propose a spectral method to adjust for spatial confounding for data with multiple exposures and responses. Specifically, we project spatial data onto the spectral domain, in which measurements for different scales are uncorrelated, and allow the coefficient estimates to vary by scale. We assume no confounding exists in the local scales but allow for global confounding, a more relaxed assumption than the no unmeasured confounding assumption required for giving coefficient estimates causal interpretations. To deal with the number of parameters needed for multiple exposures, responses, and scales, we use canonical polyadic (CP) decomposition to reduce dimensions in the three-way tensor. We demonstrate the effectiveness of the method on an extensive simulation study, use the method to analyze health burdens of per- and polyfluoroalkyl substances (PFAS), and discuss limitations of the method.
Keywords: spatial confounding, causal inference, tensor decomposition