Euclidean distance versus travel time in business location : a probabilistic mixture of hurdle-Poisson models : 55th ERSA Congress 2015, Lisbon

While the question of the specification of spatial weight matrix is now largely discussed in the spatial econometrics literature, the definition of distance has attracted less attention. The choice of the distance measure is often glossed over, with the ultimate use of the Euclidean distance. This paper investigates this issue in the case of establishments locating in the Paris region. Indeed, numerous works highlight the importance of transport infrastructure in the location model, which challenges the choice of the Euclidean distance in representing spatial effects. To compare the various distance measures, we develop a probabilistic mixture of hurdle-Poisson models for several activity sectors. Each model class uses a different definition of distance to capture spatial spillovers. The following distance measures are considered: Euclidean distance, two road distances (with and without congestion), public transit distance, and the corresponding travel times. Data were drawn primarily from the Census survey of establishments carried by the French National Institute of Statistics and Economic Studies. Data on the stock of establishments are given for the 1st of January 2007. In our sample, 763 131 pre-existing establishments were registered on the market. The number of newly created establishments in 2007 equals to 87 974. Based on the performed analyses we drew four main conclusions. 1) Overall, the obtained results are in line with the literature regarding the main determinants of establishment location. 2) Based on the Bayesian Information Criteria, we found that the proposed mixture of hurdle-Poisson models that uses two latent classes performs significantly better than the ?pure? hurdle-Poisson models based on a single distance measure, emphasizing the usefulness of our approach. By using the mixture hurdle-Poisson model we considerably decreased the level of BIC up to 42%. 3) From the overall level of estimated probabilities, we observed that for some transport-oriented sectors, such as construction, the peak road travel time is the most likely to correctly capture spatial spillovers. For other sectors, which do not rely so heavily on the transport infrastructure and which search the proximity to the potential client or user, such as real estate, the Euclidean distance tends to perform well to account for the linkage between neighboring areas. This tends to show that spatial spillovers are channeled by different means depending on the activity sector. 4) In addition, by allowing different distance measures to coexist within a hurdle-Poisson mixture model, the hurdle part of the model that uses the appropriate distance matrix significantly improves.