Sebastian Juhl

Spatial autocorrelation is a ubiquitous phenomenon in cross-sectional data that poses notable challenges for statistical inference since unmodeled spatial dependence can cause common econometric methods to produce biased and inconsistent parameter estimates. While political scientists so far predominantly rely on parametric spatial regression models, semiparametric spatial filtering techniques constitute a valuable alternative. By using the eigenfunction decomposition of a transformed connectivity matrix, the filtering approach generates a synthetic proxy variable from a linear combination of judiciously selected eigenvectors. Eigenvector selection is performed in a supervised or unsupervised fashion and different selection criteria can be employed. This synthetic variable acts as a surrogate for omitted spatial effects and removes spatial autocorrelation from the model residuals. This study introduces eigenvector-based spatial filtering to political science and discusses its strengths and limitations in comparison to parametric spatial models. Analytical results and Monte Carlo simulations show that spatial filtering resolves spatial misspecification problems in regression models. Overall, this study concludes that this semiparametric approach constitutes a very flexible and intuitive strategy that requires only few assumptions. Given the widespread use of cross-sectional data, spatial filtering is applicable to a great variety of empirical analyses when space is just a nuisance. A new R package facilitates the application of this technique.