Unpacking Spatial Dependence: A New Experimental Design for Spatial Autoregressive Simulation

Simulating spatial patterns with varying levels of spatial autocorrelation is essential for evaluating models, estimators, and diagnostics. The spatial autoregressive (SAR) process is widely used for this purpose, but its main parameter affects both spatial autocorrelation and spatial heterogeneity. Only the former is usually acknowledged. Unfortunately, these unintended variance effects can bias Monte Carlo interpretations when ignored. This paper introduces a variance-stabilized SAR (VSSAR) process that decouples spatial autocorrelation from heteroskedasticity. The proposed approach preserves the spatial patterning implied by the SAR model while enforcing constant marginal variance, making it possible to more accurately assess the behavior of models and estimators under spatial autocorrelation. We validate the VSSAR method by replicating three canonical simulation experiments and demonstrate that it should serve as the new default data-generating process for simulation studies in spatial analysis when spatial autocorrelation is the primary object of interest. JEL codes: C1, C15, C21, C2