Controlling for spatial confounding and spatial interference in causal inference: modelling insights from a computational experiment

Causal inference is a rapidly growing field of statistics that applies logical reasoning to statistical inference to estimate causal relationships. Spatial data poses several problems in causal inference – namely, spatial confounding and interference – that require different strategies when designing causal models. In order to obtain valid inferences, existing nonspatial causal models must adjust for such spatial problems. Given the blossoming literature on spatial causal inference, this research analyzes the usage of spatial causal models under a priori knowledge and a priori ignorance of the spatial structure of data. We synthesize existing research directions in noncausal spatial modelling and causal nonspatial modelling by assessing the performance of 28 spatial causal models across 16 spatial data scenarios. We used ordinary least squares (OLS) models, conditional autoregressive (CAR) models, and jointly CAR models for outcome and treatment variables as the basis for the tested models, equipping them with a variety of spatial causal adjustments. We compare our results to principles of noncausal spatial modelling and investigate their implications for spatial causal modelling. Specifically, we show that noncausal spatial modelling guidance holds in causal spatial modelling workflows and demonstrate how researchers can leverage noncausal theory to great effect. In parallel, we introduce the spycause Python package of spatial causal models and data simulators to facilitate the widespread use of these models and to enable reproduction and extension of our work.