Predicting Internal Migration on Individual level in Sweden Using Micro Data : A Performance Comparison of Logistic Regression and Neural Networks

Abstract: To know in advance who, how many, and to what destination people will migrate is importantto provide robust population projections. As migration prediction model comparisons inthe literature are limited, this study compares the predictive performance of Logistic regressionand Neural networks. To evaluate the performance of these models, the task was to predict individualmigration from other parts of Sweden to the Local labour market region of Stockholmfrom the year 2011 to 2012 using microdata. The result shows that the models have relativelyhigh AUC (0.9056-0.9136). A t-test for the dierence in proportions of True positives andFalse positives at the models respective optimal operating point on the ROC curve, shows thatthere is no significant dierence (p > 0:05) in the performance of these models. The modelsare also compared to a baseline model (i.e., a model which randomly predict who migrates ornot), which shows that they are significantly better than random prediction. However, due to arelatively high imbalance in classes, between migrants and non-migrants (1:200), the absolutenumber of false positives versus true positives at the optimal operating point is large and makesthe predictions less useful in practice. Since the performance between Logistic regression and Neural networks is not significantlydierent, this result indicates that it is for a practitioner also relevant to compare the modelsbased on other factors such as interpretability and complexity when choosing a model for aprediction task.