Abstract: År 2013 började Trafikverket processen att utveckla en stokastisk logistisk modul för den nationella godsmodellen Samgods. Den nuvarande deterministiska modulen har flera svagheter. i) den låter bara kostnadskomponenter påverkaAbstract: In 2013 the Swedish Transport Administration (STA) started the process to develop a stochastic logistics module for the national freight transport model Samgods. The current deterministic module has several weaknesses: i) it only allows cost components to influence the choice of transport chain and shipment size, whereas transport service factors such as reliability and damage risk also matter in practice, ii) it has a weak behavioral foundation as it is not based on observed data but on assumptions about transport agents’ behavior and iii) it can lead to implausibly large changes in modal shares following changes in policy variables (i.e. “overshooting”). A stochastic model has several features that overcome these weaknesses. It includes a stochastic component that accounts for omitted factors, is estimated on observed logistics choices in the Commodity Flow Survey (CFS) and has a smooth response function which mitigates the problem of overshooting. The first step in developing a stochastic module was performed by VTI and Significance (Abate et al. 2014) and included the estimation of discrete choice models for two commodity groups based on the 2004/2005 CFS. The project outlined how to implement a stochastic module in Samgods and provided recommendations regarding what kind of choice model specifications should be used. The second step was performed by the same researchers (Abate et al. 2016). Discrete choice models of the shipment size and transport chain choice were estimated for sixteen commodity groups. For two of these groups, the estimation results were used to implement a stochastic logistics module and produce demand elasticities that were compared to those produced by the deterministic module. The third step was performed by VTI, Significance, TØI and WSP (Vierth et al. 2017). This project provided a recommendation for a new commodity classification in Samgods following the adoption of the NST 2007 nomenclature by Eurostat. The new Samgods classification will influence the specification of the discrete choice models and determine the structure of the stochastic logistic module. VTI and Significance suggest the fourth step of the development of a stochastic logistics module. The project is in line with the topic “Varugruppsstrukuren – nu klassificering av statistiken, samlastning och ny stokastik logmod (prio 1)” in the research- and development plan of the STA. This project entails the estimation of discrete choice models for all relevant Samgods commodity groups using updated cost and time variables and the new Samgods commodity classification. It also includes a detailed evaluation of discrete choice models and specifications, which can be used as a foundation for a stochastic logistics module. The work is organized within two work packages (WP). WP1: Data generation Task 1.1 Updating cost parameters (VTI) This task will update the cost parameters (transport cost, warehousing cost etc.) for the logistics module in Samgods to the same year as the latest CFS (2016). This will be done by means of indexation of the current cost parameters and linking them to the commodities in the new Samgods classification. We will update the costs by this procedure to both their 2016 and 2017 value. The 2016 costs will be used to calculate costs for the shipments in the CFS 2016, because these data sources should relate to the same year. A more comprehensive revision of the cost parameters is outside the scope of this project. Task 1.2 Generating cost functions (Significance) Using the updated cost parameters, a new cost function will be produced to estimate commodity-specific transport cost and time variables for chosen and non-chosen alternatives (i.e. combinations of shipment size and transport chain) in the CFS 2016. One issue related to this task is how the results are affected by discretizing the continuous variable shipment size. We will investigate this by using an alternative set of weight class intervals into the generation of cost functions. Another issue brought up in the previous projects is the extent to which accuracy is lost by allowing only one transshipment choice per type chain. To examine this question, we will add an analysis of restricting the maximum number of transshipment to two. WP 2: Estimation and evaluation of discrete choice models Task 2.1 Identification of commodities require a stochastic logistics module (VTI) This task will identify those commodities that require a stochastic logistics module and those that can be modelled by the deterministic cost minimization approach currently used in Samgods. The deterministic approach is sufficient for those commodities that exhibit virtually no variation in the logistics choices. The work will be based on the observed choices of transport chain and shipment size in the CFS 2016 and on the new commodity classification. Task 2.2 Estimation of discrete choice models for all relevant commodities (VTI) This task will estimate discrete choice models of the joint choice of transport chain and shipment size for the relevant commodities identified in task 2.1. The data will be based on the CFS 2016, the new Samgods commodity classification and the transport cost and time variables generated in WP1. There are a number of possible model specifications, i.e. combinations of variables included and variable definitions. Previous reports included transport cost, time, value density, access to rail and quay, seasonal dummies, international shipments and dummies for the transport modes. Model specifications also include comparing linear, loglinear and spline formulations for transport costs. Commentators of these reports also suggested that cost and time could enter separately for each mode. Variables that include transport distance could also be included, as could indicator variables for zone-combinations, which would capture distance as well as local infrastructure availability. We are however limited by the fact that all variables included in the discrete choice model must also be available in Samgods. The suitability of these specifications depends partly on how accurately they predict firms’ choice of transport chain and shipment size. To determine this, we will estimate a model on a sub-set of the shipments in the CFS 2016, use the model to make predictions for the rest of the sample and compare the predictions to the observed choices. We will alter the model specifications iteratively based on the models’ predictive ability. One issue brought up in the previous project is how the cost variables are influenced by the input variables. We therefore propose that we in this task also investigate how the discrete choice models estimated under one set of cost and time variables perform under a different set of cost and time variables. In other words, we examine how representative the estimated coefficients are for cost variables that are higher or lower than those variables used to produce the coefficients. The multinomial logit model that we aim to use comes with the independent of irrelevant alternatives assumption. One risk with this assumption is that the model generates implausible substitution patterns. Choice models that do not come with this assumption includes the mixed logit and the nested logit. Mixed logit models come with very high computational demands and may not work with this large set of alternatives. Nested logit may be an option, however, and we propose that we in this task evaluate the suitability of this model as well.