Implementing artificial neural networks in microsimulation

Sammanfattning: This licentiate thesis presents a work focused on investigating the possibility of using "artificial methods" such as artificial neural networks (ANN) as an analytical alternative to the conventional technique of regression. The work has been concentrated towards two different modules in the SVERIGE microsimulation model, an earnings module used for predicting individual earnings and a property value module used for predicting real estate prices. The results from this work, shows that the use of artificial neural networks can improve the performance (degree of explanation), and reduce the time needed to develop the individual modules. But it also shows that replacing regression equations with artificial neural networks have drawbacks, e.g. that the modules become less transparent. Another example is that the difficulties in analysing each variable's importance increases since a neural net does not easily provide any coefficient estimates.