Error Analysis of Regional Migration Modeling

Much methodological advancement has been made in the modeling and analysis of regional migration. Previous migration modeling, however, has been done in a black box. The overall performance of a migration model is evaluated with the contribution of all explanatory variables, including regional attributes and spatial interaction effects. This research uses a new method to estimate migration modeling errors by their sources. Following the notion of migration spatial structure, the observed or estimated regional migration matrices of a migration system can be described by four factors: the overall effect, the relative emissiveness and the relative attractiveness of specific regions, and the effect of spatial interaction between pairs of regions. By calculating the contributions of migration factors to the modeling error, this article reveals which factors of the migration process can be modeled more or less accurately using the case of regional migration in China for the period between 2005 and 2010. A network spatially filtered Poisson migration model is estimated for China. Error analysis shows that the modeling errors of the constant K, the relative emissiveness, and attractiveness caused weighted absolute mean errors of 1.20 percent, 14.60 percent, and 15.57 percent in migration flows, respectively. The spatial interaction caused the greatest weighted absolute mean error of 31.55 percent in migration flows. Thus, the spatial interaction effect remains the most difficult to model. The findings of this research point to directions to improve migration modeling. More efforts should be made to improve the approach to model the effect of spatial interaction.