Behavioral specifications of network autocorrelation in migration modeling: An analysis of migration flows by spatial filtering

This research is concerned with the fact that migration flows between two regions are most likely related to other migration flows in a regional system. This phenomenon is called network autocorrelation. Because the presence of network autocorrelation violates the independence assumption that is frequently applied in migration estimation procedures, statistical results without accounting for network autocorrelation are likely to be biased and can potentially produce misleading conclusions. This research aims at [a] investigating the underlying behavioral and structural mechanisms leading to network autocorrelation and its operationalization in a network link matrix, [b] the statistical identification of network autocorrelation in empirical migration systems, [c] the explicit incorporation of network autocorrelation into a migration model by adopting the novel spatial eigenvector filtering approach, and [d] demonstrating the usefulness of the proposed methodology by applying it to interstate migration system of the U.S. during from 1995 to 2000.Network autocorrelation among migration flows can be explained by reflecting on how potential migrants may search for a destination within a regional system. Specifically, as migration is a spatial choice process, it is important to comprehend how migrants perceive space and choose a destination in the space. Network autocorrelation can be incorporated into modeling spatial search behavior of migrants by specifying a proper network dependency structure. In this research, concentrating on competing destination effects and intervening opportunities, two different criteria were proposed: which are the connectivity through a joint node and the spatial association between nodes.This research proposes spatial eigenvector filtering as a method to model network autocorrelation in Poisson regression. As the spatial eigenvector filtering method is conceptually easy to comprehend and produces robust spatial autocorrelation models, the method can be utilized to isolate network autocorrelation and to control for the effects of network autocorrelation onto a Poisson regression model. As a result, the potential biases in the estimated model parameters and their standard errors are adjusted in the spatially filtered Poisson regression model. The spatial eigenvector filtering approach for network autocorrelation in a migration model is demonstrated by an empirical analysis of the interstate migration in the U.S. during 1995--2000. It is shown that the network filtered model is able to adjust the potential biases in the regression parameters of exogenous variables and to control for network autocorrelation.