Estimating the Effect of Spatial Interventions in the Presence of Interference: A Causal Inference Approach

In many settings, it is of primary interest to estimate the effect of a spatially varying intervention that has a nonlocal (i.e., spillover) effect on its surrounding environment. For example, what is the effect of a coal-fired power plant on downwind air pollution concentrations, or how might the addition of a rural health clinic improve health outcomes for individuals living within a certain distance? Unfortunately, estimating causal effects from observational data in such settings is challenging, due to (i) the risk of confounding bias, and (ii) the potential for treatment interference, namely, that multiple interventions affect the same outcome locations. To address this problem, we introduce a framework for causal inference with spatial data in which causal estimands are defined as functionals of the potential outcome distribution under a set of stochastic interventions. Corresponding nonparametric identifying assumptions are considered which allow the estimands to be estimated from observational data in the presence of distance-limited interference, and an augmented inverse probability of treatment-type estimator is proposed. Notably, the estimator is constructed from a log-Gaussian Cox process model for intervention locations and a semiparametric outcome model of the spillover structure that accounts for spatial autocorrelation.  We use the proposed method to estimate the effect of large concentrated animal feeding operations (CAFOs) on air pollution in Iowa in settings with increasing and decreasing numbers of CAFOs.

Nathan Wikle is an Assistant Professor in the Department of Statistics and Actuarial Science at the University of Iowa. His research interests include spatial and spatio-temporal statistics, causal inference, and statistical inference for dynamical systems, often centered on applications in environmental health, environmental science, and ecology.

Host: Ran Chen