The Local Voter: A Geographically Weighted Approach to Ecological Inference

Drawing inferences about individual behavior from aggregate ecological data has been a persistent problem in electoral and behavioral studies, in spite of important methodological advances. In a recent article Anselin and Tam Cho (2002) provided Monte Carlo evidence that King's Ecological Inference (EI) solution will produce biased estimates in the presence of extreme spatial heterogeneity. In this article we provide further empirical evidence that supports their findings and shows that in the presence of spatial effects the residuals of Goodman's naïve model exhibit the same spatial structure that King's local BBi estimates. Solving for extreme spatial heterogeneity, it is argued here, requires controlling the omitted variable bias expressed in the spatial structure of much ecological data. In this article we propose a Geographically Weighted Regression approach (GWR) for solving problems of spatial aggregation bias and spatial autocorrelation that affect all known methods of ecological inference. The estimation process is theoretically intuitive and computationally simple, showing that a well‐specified GWR approach to Goodman and King's Ecological Inference methods may result in unbiased and consistent local estimates of ecological data that exhibit extreme spatial heterogeneity.