An empirical study of aggregation of alternatives and its influence on prediction: case study of car type choice in Sweden

In the car type choice models, alternatives are usually grouped into categories by some of their main characteristics such as make, model, vintage, body type and/or fuel type. Each of these categories contains different versions of the cars that are usually not recognized in the applied literature. In this study we empirically investigate whether including the heterogeneity of these versions in the modeling do matter in estimation and prediction or not. We have detailed data on alternatives available on the market down to the versions level of each model which enables us to account for heterogeneity in the model. We also have Swedish car registry data as demand. We estimate different discrete choice models with different methods of correction for alternative aggregation including nesting structure. We estimate these models on based on year 2006 Swedish registry data for new cars, predict for 2007 and compare the results. The results show that including heterogeneity of cars' versions in the model improves model fitness but it does not necessarily improve prediction results.