Generalized Propensity Score Approach to Causal Inference with Spatial Interference

Abstract

Many spatial phenomena exhibit interference, where exposures at one location may affect the response at other locations. Because interference violates the stable unit treatment value assumption, standard methods for causal inference do not apply. We propose a new causal framework to recover direct and spill‐over effects in the presence of spatial interference, taking into account that exposures at nearby locations are more influential than exposures at locations further apart. Under the no unmeasured confounding assumption, we show that a generalized propensity score is sufficient to remove all measured confounding. To reduce dimensionality issues, we propose a Bayesian spline‐based regression model accounting for a sufficient set of variables for the generalized propensity score. A simulation study demonstrates the accuracy and coverage properties. We apply the method to estimate the causal effect of wildland fires on air pollution in the Western United States over 2005–2018.

Spatial causal inference in the presence of unmeasured confounding and interference

“…with unmeasured confounders [Schnell and Papadogeorgou, 2020, Christiansen et al, 2022, Papadogeorgou, 2022, interference [Giffin et al, 2022], or both , though model-free definitions using potential outcomes directly have also been employed [e.g. Verbitsky-Savitz andRaudenbush, 2012, Gilbert et al, 2021].…”

“…Interference has attracted a lot of attention in the last couple of decades [e.g. Sobel, 2006, Hudgens and Halloran, 2008, Manski, 2013, Aronow and Samii, 2017, among many others] with some studies that focus explicitly on how interference manifests in spatial settings [Verbitsky-Savitz and Raudenbush, 2012, Wang et al, 2020, Zigler et al, 2020, Antonelli and Beck, 2020, Giffin et al, 2022.…”

Nonparametric doubly robust estimation of causal effect on networks in observational studies

“…Specifically, when studying partial interference, the neighbourhood treatment may be defined at the cluster level, for example, it can be the proportion of treated nodes in a cluster. It's worth noting that and are widely used in previous work to integrate the effect of neighbours' covariates and the neighbours' treatment on the outcome (Aronow & Samii, 2017; Forastiere et al, 2020; Giffin et al, 2022; Ogburn et al, 2022; Sofrygin & van der Laan, 2017). Although Park and Kang (2022) avoid the assumption about and by using the covariate vector and neighbourhood treatment vector directly, they assume partial interference, and thus, the covariate vector and the neighbourhood treatment vector within the same cluster are of the same length.…”

Discussion on “Spatial+: A novel approach to spatial confounding” by Dupont, Wood, and Augustin

“…Another common setting that gives spatially smooth exposure is the study of neighborhood effects, for example, Giffin et al. (2020) regressed air pollution concentration onto kernel‐smoothed measures of wildland fire indicators. There are many other examples such as extreme temperature, some forms of air pollution, distance to a point source, and so forth.…”