Recognizing and Overcoming Obstacles to Causal Inference Arising from Spatial Structures

Policy interventions can affect not only the directly targeted units but also neighboring units in close geographical proximity, as well as substitute units that are unaffected by the policy. For example, the Philadelphia beverage tax was found to be associated with an increased sales of taxed beverages in neighboring counties of Philadelphia, as well as an increased sales of untaxed alternative beverages and high-calorie foods in Philadelphia. We will consider two types of interference commonly observed in evaluating policy effects: interference due to (1) spatial proximity and (2) substitutability. Understanding and evaluating these spillover and substitution effects is crucial for assessing the comprehensive impact of policy interventions. In this talk, I will introduce novel causal estimands under counterfactual neighborhood interventions in the presence of spillover. These causal quantities are policy-relevant for designing effective policies for populations subject to various contextual scenarios. We will then establish identification conditions for estimating these effects and apply our proposed methods to investigate the effect of the Philadelphia beverage tax. 

Spatial confounding poses a significant challenge in scientific studies involving spatial data, where unobserved spatial variables can influence both treatment and outcome, possibly leading to spurious associations. To address this problem, we introduce SpaCE: The Spatial Confounding Environment, the first toolkit to provide realistic benchmark datasets and tools for systematically evaluating causal inference methods designed to alleviate spatial confounding. Each dataset includes training data, true counterfactuals, a spatial graph with coordinates, and smoothness and confounding scores characterizing the effect of a missing spatial confounder. It also includes realistic semi-synthetic outcomes and counterfactuals, generated using state-of-the-art machine learning ensembles, following best practices for causal inference benchmarks. The datasets cover real treatment and covariates from diverse domains, including climate, health and social sciences. SpaCE facilitates an automated end-to-end pipeline, simplifying data loading, experimental setup, and evaluating machine learning and causal inference models. The SpaCE project provides several dozens of datasets of diverse sizes and spatial complexity. It is publicly available as a Python package, encouraging community feedback and contributions. 

Causal inference and spatial statistics methodology are often detached. We reconcile these threads of the literature within the realm of interference and unmeasured spatial confounding. We provide new insights on the proper analysis of spatial data sets for learning causal effects, and establish how tools from spatial statistics can be used to draw causal inferences. From a causal inference prism, we introduce spatial causal graphs to study the complications that arise when investigating causal relationships from spatial data, and we provide new insights for spatial data analysis: spatial confounding and interference can manifest as each other, and statistical dependencies in the exposure can render standard analyses invalid. We propose a parametric approach based on tools amenable to spatial statisticians that accounts for interference and mitigates bias from local and neighborhood unmeasured spatial confounding. We show that incorporating an exposure model is necessary from a Bayesian perspective. The proposed approach is based on modeling the exposure and the outcome simultaneously while accounting for the presence of common spatially-structured unmeasured predictors. 

Environmental data are often observational and spatially dependent, making casual treatment effects difficult to estimate. Unmeasured spatial confounders, i.e., spatial variables correlated with both the treatment and response, can induce bias and invalidate inference. Spatial data can also be subject to preferential sampling, where the locations of the samples are driven by unmeasured covariates or even the assumed value of the response of interest. We propose a method that simultaneously accounts for unmeasured confounders in both the sampling locations and treatment allocation. We prove that the key parameters in the model are identifiable and show via simulations that the causal effect of interest can be reliably estimated under the assumed model. The proposed method is applied to study the effect of marine protection areas on fish biodiversity.