An empirical study on aggregation of alternatives and its influence on prediction in car type choice models

Assessing and predicting car type choices are important for policy analysis. Car type choice models are often based on aggregate alternatives. This is due to the fact that analysts typically do not observe choices at the detailed level that they are made. In this paper, we use registry data of all new car purchases in Sweden for two years where cars are observed by their brand, model and fuel type. However, the choices are made at a more detailed level. Hence, an aggregate (observed) alternative can correspond to several disaggregate (detailed) alternatives. We present an extensive empirical study analyzing estimation results, in-sample and out-of-sample fit as well as prediction performance of five model specifications. These models use different aggregation methods from the literature. We propose a specification of a two-level nested logit model that captures correlation between aggregate and disaggregate alternatives. The nest specific scale parameters are defined as parameterized exponential functions to keep the number of parameters reasonable. The results show that the in-sample and out-of-sample fit as well as the prediction performance differ. The best model accounts for the heterogeneity over disaggregate alternatives as well as the correlation between both disaggregate and aggregate alternatives. It outperforms the commonly used aggregation method of simply including a size measure.