Estimation and implementation of joint econometric models of freight transport chain and shipment size choice

As part of the further development of the Swedish national freight model system (SAMGODS), we developed a stochastic logistics model in the form of a disaggregate random utility-based model of transport chain and shipment size choice, estimated on the Swedish Commodity Flow Survey (CFS) 2004-2005. Moving from the current deterministic logistics model within the SAMGODS model to a stochastic one, is important because it bases the model on a stronger empirical foundation. The deterministic model was not estimated on observed choice outcomes, but just postulates that the least cost solution will be chosen. We estimated logit models which explain the joint choice of shipment size (in discrete categories) and transport chain separately for sixteen different commodity types. A transport chain (e.g. truck-vessel-truck) is a sequence of modes used to transport a shipment between the locations of production and consumption. Transport cost, travel time and value density are some of the main determinants included in the models. It is important to note that by their very nature these probabilistic models account for the influence of omitted factors. A deterministic model effectively assumes that the stochastic component can be ignored – in other words, that the researcher has full knowledge of all the drivers of behaviour and that there is no randomness in actual behaviour. As a result of adding the stochastic component in the random utility model, the response functions (now expressed in the form of probabilities) become smooth instead of lumped at 0 and 1 as in a deterministic model. This in turn will address the problem of “overshooting” that is prevalent in a deterministic model when testing different scenarios or policies. For two of the commodity types (metal products and chemical products) for which we estimated a transport chain and shipment size choice model, we also implemented the model in the SAMGODS framework. The implementation takes place at the level of the annual firm-to-firm flows by commodity type between producing and consuming firms that are generated by the first steps of the SAMGODS model (PC flows between zones that have been allocated to individual firms at both ends). For every firm-to-firm flow, shipment size and transport chain choice probabilities are calculated and added over the firm-to-firm flows of the PC relation (sample enumeration, as used in several disaggregate transport models). From this, the aggregate OD matrices by mode can be derived straightforwardly, as well as results in terms of tonne-kilometres by mode. It was not possible to empirically model transshipment location choices, because they are not stated in the CFS. Therefore, the determination of the optimal transshipment points for each available chain type from the set of available locations is still done deterministically. The implemented models were applied to produce elasticities of demand expressed in tonne-kilometres for various changes in cost and time for road, rail and sea transport. These elasticities are compared to those for the same commodity types in the deterministic model and to the available literature. The elasticities clearly differ between the two models, they are usually smaller (in absolute values) in the stochastic model, as expected. In the paper, we report the basic differences between a stochastic and a deterministic logistics model, the estimation results for the sixteen commodities, the way the stochastic model was implemented within the SAMGODS model, the elasticities that we obtained for the implemented stochastic model and the comparison with elasticities from the deterministic model and the literature.