Predictive limitations of spatial interaction models: a non-Gaussian analysis

TL;DR

This paper evaluates the predictive performance of spatial interaction models, revealing that common models like the radiation model underperform compared to simpler gravity models in commuting data.

Contribution

It introduces a statistical comparison method for spatial interaction models and demonstrates its application using US Census commuting data.

Findings

Radiation model performs worse than simple gravity model

Spatial interaction models fit data poorly in absolute terms

Adding parameters can improve model predictions

Abstract

We present a method to compare spatial interaction models against data based on well known statistical measures that are appropriate for such models and data. We illustrate our approach using a widely used example: commuting data, specifically from the US Census 2000. We find that the radiation model performs significantly worse than an appropriately chosen simple gravity model. Various conclusions are made regarding the development and use of spatial interaction models, including: that spatial interaction models fit badly to data in an absolute sense, that therefore the risk of over-fitting is small and adding additional fitted parameters improves the predictive power of models, and that appropriate choices of input data can improve model fit.