Residual Spatial Correlation in Two-Way Error Panel Data Models

Sammanfattning: This thesis examines the spatial autocorrelation in residuals of two-way error panel data models. Three types of models are examined: the standard linear panel data model, the dynamic panel data model, and the spatial lag panel data model. A known theoretical result for the linear model, that the within estimator applied to independent observations results in a spatial correlation in the residuals which is proportional to the inverse of the number of observed individual units, is supported in a Monte Carlo study. Similar Monte Carlo results are shown for the dynamic and spatial models. The Monte Carlo study shows the effect of residual correlation on the maximum likelihood estimation of the spatial model and on residual tests for spatial correlation. Randomization tests for spatial correlation are formulated and their properties are evaluated. The results suggest that a randomization test for local spatial autocorrelation is the most suitable test for samples large in number of individual regions and small in number of time points. The model estimation methods and tests for residual spatial autocorrelation are applied in an empirical examination of regional unemployment in Southern Sweden. The study shows that the spatial lag model is sensitive to choice of spatial weight matrix and indicates the presence of a spatial structure which is not fully captured by the applied models and weight matrices.