Traditional migration modelling involves an estimation of a migration model, using population at origin, population at destination, distance and other social, economic and environmental variables to explain migration flows. Modelling performance is assessed for the migration model as a whole. However, it remains unclear if a migration model can fit the relative emissiveness and attractiveness of specific regions better than the spatial interaction effect between pairs of regions. Recent studies show that the parameters of the log‐linear model for a migration matrix can fully describe the relative emissiveness Pi, attractiveness Qj of specific regions and the spatial interaction effect between pairs of regions Fij. By calculating and comparing the contributions of the modelling errors of the relative emissiveness Pi, attractiveness Qj and the spatial interaction effect Fij to the overall modelling errors of migration flows, this paper reveals which factors of the migration process can be modelled more or less accurately using the case of regional migration in China for the period 2005–2010. According to the Poisson migration model for China, the modelling errors of the constant K, the relative attractiveness and emissiveness caused mean relative errors of 4.92%, 11.80% and 10.61% in migration flows, respectively. The spatial interaction caused a mean relative error of 17.74% in migration flows, which was much greater than the errors caused by the constant K, the relative attractiveness and emissiveness. Thus, the spatial interaction is the most important factor, which cannot be modelled very well in migration modelling. Copyright © 2016 John Wiley & Sons, Ltd.