Freight-Demand Modeling to Support Public-Sector Decision Making (2010)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

13 From the gaps identified in Chapter 2, the research team identified research topics that would represent an advance in the state of freight modeling in the short term, be consistent with the topics identified throughout the outreach process, and could utilize available data sources. At the research panel’s direction, the freight model research considered for this proj- ect concentrates on short-term model application improve- ments (rather than completely new methods). Other research efforts, such as the Strategic Highway Research Program (SHRP) 2 C20 Project on freight model improvement, will consider long-term model improvements to develop over the next 10 years or more. From the list of research topics developed by the research team, the research panel selected the following three topics that address critical gaps in existing freight demand models: • Chaining of freight activities, • Temporal and seasonal impacts, and • Modal diversion consistent with the geographies of public agencies. The selected topics would develop transferable parameters for models that can then be applied to various settings. They also are intended to identify techniques for quickly and in- expensively developing model parameters that could be employed by others who might be developing or applying freight models to support public decision-making. Although this section focuses on the three topics selected for further research, it also includes a discussion of why the remaining topics were not recommended for advancement. 3.1 Topics Selected for Further Research Chaining of Freight Activities Chaining of freight activities could be addressed by use of GPS data. Trip chaining of commercial trucks, including those moving freight, requires information on the nature of truck tours, particularly the number of stops, the average imped- ance between stops (e.g., time), and the nature of the land use at each stop on the tour. Many truck fleet operators subscribe to GPS services provided by vendors that collect and electron- ically distribute the GPS information provided by trucks equipped with units they sell. Although their business model is to provide GPS information to truck fleet operators, many vendors currently store the historical GPS information, and in all cases it would be easy to store these data. It should be pos- sible to process the historic GPS information to obtain better truck trip chaining data. Temporal and Seasonal Impacts The methods to collect GPS data discussed above include the ability to identify the time of day at which a truck stops and starts its trips. This information could be used to develop time- of-day allocation tables for a number of urban areas. These allocations may be borrowed for use in other urban models, or the GPS methods developed may be used to develop or update truck time-of-day allocation tables for models that consider time of day assignments. Additionally, the GPS information could be used to determine the behavior of specific types of trucks (e.g., long-haul trucks). It has been observed that long- haul trucks may make intermediate stops in urban areas to rest during congested time periods, or to wait until loading docks or port states are open to receive trucks. Documentation of this behavior and the intermediate stop locations could improve the time-of-day response of urban models. The seasonal movement of freight could be addressed by using information on the monthly flow of trucks from state weigh in motion stations, border crossings, and FHWA’s Vehicle Travel Information System (VTRIS). Although this information does not provide any information about the commodities carried by these trucks, the FHWA’s Freight Analysis Framework, Version 2 (FAF2), Highway Link and C H A P T E R 3 Research to Fill Critical Gaps

Truck Data1 provides data on freight and nonfreight truck volumes. The assignment of disaggregation of the FAF2 origin-destination (OD) database to the network can provide estimates of truck volumes by Standard Classification of Transportation Goods two-digit (SCTG2) commodity. By identifying the link flows at the locations with seasonal truck percentages as freight, nonfreight, and/or by SCTG2 com- modity, it may be possible to identify the appropriate sea- sonal factors that should be used for these commodities. Modal Diversion Consistent with the Geographies of Public Agencies Modal diversion requires data outside of the area covered by most DOT and MPO models supporting public decisions. The development of national multimodal databases of flows, behavioral characteristics, and networks may be needed to address this issue. Research needs to be undertaken into the variables that are considered in the mode choice of freight decisionmakers. Behavioral characteristics typically are deter- mined though a preference survey. There are national freight flow databases that could be adapted as revealed-preference databases. Other national databases exist that could provide values for some of the decision variables that might be part of the modal decision. An analysis of the modal choices, as revealed in existing freight databases together with available information on explanatory variables using standard mode- choice regression software, may provide useful insights into which variables are important in modal-choice decisions for freight, as well as the degree to which these explanatory vari- ables are considered. 3.2 Topics Not Selected for Further Research The topics presented in Section 3.1 were the focus of addi- tional research as documented in the remainder of this sec- tion. Topics that were not selected for additional research, but were identified as addressing critical gaps, are described in this section. Freight Data to Support Model Specification, Calibration, and Validation The focus of NCFRP Project 6 was intended to exclude data needs. Although the quality and availability of data was a principal issue raised by the interview respondents, there are other ongoing TRB and U.S.DOT projects and programs to address this issue, both as short-term incremental improve- ments and long-term improvements. Better Methods to Consider Nonfreight Trucks Expanding on a paradigm developed by Hunt and Stefan,2 trucking activity in a model area can be considered to have the following four components: 1. Interregional freight, typically trips with at least one trip end external to the region that is being modeled. Exam- ples include long-haul truckload, less than truckload, and private trucks. 2. Intraregional freight truck tours, those trucks that move individual shipments of goods locally within a model region. Examples include parcel pickups and delivery trucks dis- tributing goods to retail, office, and commercial establish- ments, as well as homes. 3. Intraregional service truck tours, those trucks that move in individual movements to offer services locally within a model region, to support construction, service, utility, and other service operations. Examples include trucks oper- ated by telephone and cable companies, contractors, and repair and service companies. 4. Fleet allocations and patrols, those trucks that are assigned to patrol or operate on fixed routes within a specific geo- graphic area or road links within the model region, rather than to move individual shipments of goods or services. Examples include garbage trucks, newspaper or mail deliv- ery, as well as roads and parks maintenance. Methods exist to address all of these components singly or in combination. For example, the FHWA’s Accounting for Commercial Vehicles in Urban Transportation Models already outlines procedures to better account for commercial truck activity, exclusive of intercity trucking, particularly Compo- nents 3 and 4 above. (It should be noted that Accounting for Commercial Vehicles identified many other commercial vehi- cle trips for business services—e.g., realtors, salesmen—that involve automobiles. Although this travel is a substantial por- tion of total commercial vehicle travel, it does not overlap with freight truck trips and these automobile commercial trips are not included in this proposed topic.) NCHRP 606 addresses methods to account for Component 1. The Quick Response Freight Manuals (1996 and 2007 editions)3–4 address 14 1 FAF2 Highway Link and Truck Data and Documentation: 2002 and 2035, Federal Highway Administration, Office of Freight Management and Operations, http://ops.fhwa.dot.gov/freight/freight_analysis/faf/index.htm. 2 Hunt, J. D., and K. J. Stefan, “Tour-Based Microsimulation of Urban Commercial Movements,” Transportation Research Part B 41, (2007): 981–1,013. 3 Cambridge Systematics, Inc., Comsis Corporation, and University of Wisconsin– Milwaukee, Quick Response Freight Manual; Federal Highway Administration, Office of Planning and Environment, September 1996. 4 Cambridge Systematics, Inc., Quick Response Freight Manual II; Federal High- way Administration, November 2007.

methods to account for Components 2 and 3. What is not available is some indication of the expected size of these truck components in model regions. It is possible to determine the base-year existing truck activity within a region from FHWA’s Highway Perfor- mance Monitoring System (HPMS). HPMS provides statis- tically based estimates of truck activity for the states and urban areas typically covered by models. The FAF network flow datasets, or improvements to those datasets, can be used to provide estimates of Component 1, and that portion of Component 2 covered by the FAF. FHWA’s Accounting for Commercial Vehicles in Urban Transportation Models5 outlines methods where vehicle registration data and/or air quality mobile source inventory data can be used to provide individual estimates for Component 4, and Components 2 and 3, respectively. For at least some urban areas, it may be possible to use all of these methods and data simultaneously to arrive at esti- mates of the total HPMS reported truck VMT in each region, which should be coming from the sum of the trucking com- ponents identified. This information on the allocation of freight and nonfreight truck VMT in the areas served by models should be of value in the calibration and validation of the truck components of those models. Better Incorporation of Labor and Equipment Productivity in Freight Models The economic activities that produce or consume freight can be expected to become more productive, such that the rela- tionship between economic activity and the freight required to support this activity may not match the relationships incorpo- rated into existing models. Similarly, the carriers who transport freight also continue to increase productivity such that the rela- tionships between the volume of freight and the number of vehicles required to move the freight, as included in existing models, would no longer be applicable. Although it might be possible to develop this information using time series of equipment and/or labor availability together with time series data of commodity flow, such datasets are not readily available. Where times series of com- modity flow databases do exist (e.g., Commodity Flow Sur- vey [CFS], FAF, TRANSEARCH), either there have been significant changes in the methods used to collect the data, or the commodity classification scheme has changed (e.g., Stan- dard Transportation Commodity Code [STCC] to SCTG). That same issue exists for the potential regression datasets (e.g., employment classified by the North American Industry Classification System [NAICS] versus Standard Industrial Classifications [SICs] industries). Local commodity, employ- ment databases, and surveys fare no better. Improved Methods for Nonhighway Freight Assignment In addition to truck freight activity on highways, domestic freight is carried by rail, barge, and other domestic water ves- sels, air, pipeline, and combinations of these same modes to travel between an origin and a destination. Air travel is not constrained to specific network links. Although water is con- strained to specific systems of public waterways, the options to divert from these waterways are limited and assignment by water is a trivial problem. Rail travel operates almost exclu- sively on privately owned tracks where the assignment serves the business needs of the railroad trains, not the freight being carried. Although improved rail routing methods would seem desirable, this topic is being pursued by other research, includ- ing research by the railroads. Additionally, rail assignments serve the multistate business interests of the railroads and those interests are not easily confined to the areas served by a state or MPO model. Simplified Methods for Considering the Economic Impact of Freight Improvements Although transportation projects advanced as freight projects may have significant impact on the economy, those projects also have benefits to passenger and other travel. Additionally, transportation projects intended to benefit gen- eral travel may have significant impacts on the transport of freight; those benefits to the transport of freight may have sig- nificant impacts on freight-dependent industries as well as other sectors of the economy. Given that the scope of this topic transcends freight, the subject may be pursued more appropriately in conjunction with more comprehensive meth- ods to better account for the economic impacts of transporta- tion projects and to properly attribute those benefits to all sectors of travel demand, including freight. Better Consideration and Forecasting of Trip by Empty and Repositioning Freight Vehicles Although understanding the movement of empty and repo- sitioning freight trucks is important, it is a subtopic to devel- oping a better understanding of the chaining of freight activities discussed previously. Thus, additional research on this topic should be deferred, unless subsequent research on truck chaining does not adequately advance the understand- ing of the movement of empty vehicles, including the reposi- tioning of trucks for the next trip. 15 5 Cambridge Systematics, Inc., Accounting for Commercial Vehicles in Urban Trans- portation Models, Federal Highway Administration, 2004.

3.3 Truck Trip Generation, Distribution, and Chaining Information Trip chaining of commercial trucks, including those mov- ing freight, requires specialized data not readily available to apply these methods in common usage. Information on the nature of truck tours, particularly the number of stops, the average impedance between stops (e.g., time), and the nature of the land use at each stop on the tour can only be established though expensive surveys. GPS data has been used in a number of passenger surveys to collect data on passenger tours to assist in developing pas- senger models.6,7,8 This information has included the deploy- ment of GPS devices in passenger vehicles, by passengers, and the processing of that GPS data to determine information concerning tours, including trip ends, the nature of the land use at trip ends, the time between trip ends or stops, and the organization of stops into tours. The use of GPS in these surveys improved the quality of the information collected, increased the response rate, eased the burden of data entry by the passengers, and added additional information that could not otherwise be collected. Although it was necessary to deploy GPS units in these pas- senger studies, it is possible to collect much of the same infor- mation for truck activities without the need to deploy new GPS units. Many truck fleet operators subscribe to GPS serv- ices provided by vendors. These vendors currently collect and electronically distribute GPS information provided by trucks equipped with units they sell. Although their business model is to provide GPS information to truck fleet operators, many vendors store the historical GPS information, and in all cases it would be easy to retrieve these data. It should be possible to process the historic GPS information to obtain data to use in better defining truck trip chaining. GPS data maintained by vendors typically will have an anonymous vehicle identifier, geographic latitude/longitude and coordinates, and time. In order to successfully process GPS information for this purpose, it will be necessary to identify stops and starts from truck GPS data, and to identify the land use at these stops and starts (e.g., in the Calgary truck chaining model, land uses are classified by five categories: low-density, residential, retail and commercial, industrial, and other high- density employment). Most metropolitan areas maintain land use maps in the form of shapefiles. Using standard geographic information system (GIS) processing tools, it is possible to match the GPS records of stops and starts by latitude and longitude with these land use records to determine the land use of the truck trip end. Sequential stops can then be grouped using the anonymous vehicle identifier and the sequence of stops; it may be possible to determine time between stops or stops per day. The information that could be available from providers of GPS data for truck operators is considerable, both in the number of firms involved and the geographies that are cov- ered. In order to prepare the data to be used in this topic, it was necessary to determine which vendors of GPS services to truck operators could provide historical information with sufficient detail to determine basic information concerning trucking behavior, at a reasonable cost. In order to determine if the information collected could be adapted to a number of geographic settings, it was necessary to choose metropolitan areas with various geographic locations, representing a vari- ety of sizes and densities. Selection of GPS Vendors For the purpose of this research topic, the selection of a GPS vendor had to satisfy several criteria. The GPS vendor must provide services nationally in order to have similar GPS data for the variety of metropolitan areas that would be con- sidered. Selection of different GPS vendors for each metropol- itan area was not practical because of the differing reporting formats and standards that might be used by those vendors, as well as the administrative effort in acquiring data from var- ious sources. The GPS vendor must not only offer GPS track- ing services to truck operators, it also must store and be able to provide this information from a historical database. The truck operators that are the customers of the GPS vendor must include firms that primarily offer services within the metropolitan area, in order to provide the desired informa- tion on trucking operations within metropolitan areas. The historical data provided by the GPS vendor must include, at a minimum, information on GPS events identified by vehicle, which includes the date, time, and location. Ideally, the sta- tus of the truck associated with the GPS event (e.g., moving or stopped) also will be available. Of the GPS data vendors assessed for this study, it was determined that one best met the research parameters, for the following reasons: • Meets the basics technical requirements (time stamped latitude/longitude data tracking truck trips and stop/start activities); • Serves mostly metropolitan area short-haul trucks, which would be those most likely to have frequent daily stops within the same metro area; • Provides substantial archived historical data for truck movements within U.S. metropolitan regions; and • Is available for a reasonable cost ($12,000 per month). 16 6 Bricka, S., and C. R. Bhat, A Comparative Analysis of GPS-Based and Travel Survey-Based Data, Technical Paper, Department of Civil Engineering, Univer- sity of Texas at Austin, July 2005. 7 NuStats, Kansas City Regional Household Travel Survey: GPS Study Final Report. Mid-America Regional Council, Kansas City, 2004. 8 Lawson, T., et al., GPS Pilot Project: Phase VI Final Report, New York Metropol- itan Transportation Council, May 2009.

Selection of Metropolitan Areas The research team selected four pilot cities for GPS data col- lection and processing. Although there were no technological limitations influencing the number of metropolitan areas that could be investigated, constrained project resources required a limited sampling. Metropolitan areas with active truck and/or freight studies were considered desirable because the data may have immediate application. To ensure that the data developed could be broadly applied, it was desirable to select metropoli- tan areas with various sizes, densities, and geographic loca- tions. Similarly, it was desirable to select metropolitan areas that have existing data for comparison against the GPS find- ings. Finally, the metropolitan area must have a shapefile of land uses, with attributes of land use coded in a conventional format, which could be used in processing the GPS data. Based on these criteria, GPS data were obtained for the following metropolitan areas: • Los Angeles, which currently is processing GPS data from October 2008 whose results could be compared to this study. This is a large metropolitan area, with low-density land use development. It has an active freight study that may be able to utilize any data developed. It has a land use shapefile available that could be used to process the data. • Chicago, which has active freight and truck studies. This is a large metropolitan area with high-density land use devel- opment. It has a land use shapefile available that could be used to process the data. • Phoenix, which has active freight and truck studies. It has recently completed a commercial vehicle survey that included the development of a land-use-to-land-use inter- change matrix. It is a mid-sized metropolitan area with low-density land use development. It has a land use shape- file available that could be used to process the data. • Baltimore, which has active freight and truck studies. It is a mid-sized metropolitan area with low-density land use development. It has a land use shapefile available that could be used to process the data. Although these four cities were selected, for preliminary research, duplication of the data collection and processing techniques described below could be completed by other metropolitan areas relatively easily with minimal resources. GPS File The GPS vendor provides data feeds of its subscribers for a user-specified period of time (typically in one-month incre- ments, within user-specified bounding “boxes” of latitude and longitude). These GPS products use wireless technologies to transmit GPS location and engine condition information to its central locations for transmittal to its subscribers, and pro- vides its subscribers with a number of product lines based on how often the data is transmitted. In addition to regular trans- mittals of data, the subscriber may query (i.e., “ping”) the GPS unit, which will generate and transmit information. All GPS and engine information received from these products is cen- trally stored and available for a historical period of 5 years. The information is provided in both XML and CSV file for- mats. The primary difference between the file formats is in the manner in which the data are stored and accessed. Some attrib- utes are meaningless when the status of the unit is “parked” or “stopped” (e.g., heading or speed would have no meaning for a stopped unit). Similarly, other information is meaningless when the vehicle is moving (e.g., stop duration is meaningless for a moving unit). The XML format includes only the data items appropriate for the specified status, defined as • Stop-Not Moving/Engine On; • Park-Not Moving/Engine Off; • Moving-Vehicle in Motion; and • Status-0 for moving, 1 for short stop, 2 for medium stop, 3 for long stop. For determining information about truck stops, records with a status of “Moving” or “Status-0” need not be processed. For this study, GPS records were acquired for the month of September 2009. Latitude and longitude boxes were defined to encompass the areas covered by the metropolitan area land use shapefiles. GPS records falling outside of this area were dropped. All remaining records were processed and sorted to provide the required information by metropolitan area (and land use). GPS records for Saturdays, Sundays, and the Labor Day holiday (September 7, 2009) were excluded in the calcu- lation of average weekday truck information. The remaining records were processed to produce the following information: • Number of trucks—Number of unique GPS IDs. • Number of GPS events—Transmittals that trigger a GPS event. • Number of stops—Number of GPS event records exclud- ing moving and maintenance records. In a chain of trips by the same vehicle, a stop is both the destination of one trip and the origin of the next trip. • Number of stops per truck per day—Number of stops divided by the number of trucks, adjusted by the operating days of the trucks. • Airline distance to next stop—From records in time sequence sorted by ID and by date. Records include fields indicating event time. Travel time is the difference between event times for a given stop and the next stop in sequence for that same GPS unit (truck). For information by land use stop, the next stop need not be of the same land use. • Airline distance to next stop—From records in time sequence sorted by ID and by date. Records include fields 17

indicating event time. Airline distance is the difference, expressed in miles as the great circle distance between the latitude and longitude of a given stop and the latitude and longitude of the next stop in the sequence for that vehicle on that day for the GPS unit (truck). For information by land use stop, the next stop need not be of the same land use. • Mileage to next stop—From records in time sequence sorted by ID and by date. Records include fields indicating cumulative vehicle mileage (odometer reading). Highway distance between stops is the difference between cumula- tive mileage for a given stop and the next stop in sequence for that same GPS unit (truck). For information by land use stop, the next stop need not be of the same land use. For the four pilot cities, processed data for GPS events and stops are included in Table 3.1, and for distances in Table 3.2 (number of event and number of stops). Although not pre- sented, these data could be processed to determine additional 18 Metro Area Land Use Number of Trucks Number of GPS Events Number of Origins Percent of Origins by Land Use Number of Origins per Truck per day Total 6,901 3,926,611 853,049 100% 9.05 Industrial 5,702 640,084 202,187 24% 3.32 Low density 4,631 230,703 44,876 5% 2.10 Other high-density employment 5,830 1,420,470 164,858 19% 3.73 Residential 4,919 773,228 176,728 21% 3.56 Los Angeles Retail and commercial 6,083 862,126 264,400 31% 3.92 Total 3,290 1,955,033 432,311 100% 10.59 Industrial 2,730 357,130 116,749 27% 4.38 Low density 2,441 241,271 39,584 9% 2.28 Other high-density employment 2,650 554,915 77,209 18% 4.14 Residential 2,298 348,463 76,076 18% 3.29 Chicago Retail and commercial 2,888 453,254 122,693 28% 3.97 Total 2,797 1,044,132 258,578 100% 8.61 Industrial 1,894 186,544 52,747 20% 3.24 Low density 2,058 273,476 36,396 14% 2.53 Other high-density employment 1,310 59,669 18,996 7% 2.75 Residential 1,917 287,941 78,102 30% 3.84 Baltimore Retail and commercial 2,343 236,502 78,937 28% 4.25 Total 2,851 1,491,659 436,758 100% 10.80 Industrial 2,446 179,345 54,718 13% 0.90 Low density 2,554 409,615 92,673 21% 4.33 Other high-density employment 2,163 146,677 43,062 10% 0.97 Residential 2,258 418,321 135,281 31% 2.72 Phoenix Retail and commercial 2,693 337,701 111,024 25% 3.49 Table 3.1. GPS-derived truck characteristics.

information, such as median, standard deviation, and distri- bution around the median. Similarly, in addition to calculat- ing information for average stops, the same information can be calculated by stop sequence (1st, 2nd, 3rd, etc.). The abil- ity to develop this information may assist in developing chaining and/or distribution models. As can be seen in Table 3.1, the GPS data provide a tremen- dous number of records on events. The number of records processed ranged from more than 1 million records in Balti- more to more than 3.9 million in Los Angeles. A large number of these records were GPS events in which the vehicle was mov- ing. For this research, only records in which the vehicle was stopped are of interest. The number of records that are stop events, where a trip originates, ranges from more than 258,000 in Baltimore to more than 853,000 in Los Angeles. The num- ber of origins per truck per day of operation ranged from a low of 8.6 origins per day in Baltimore to a high of 10.8 origins per day in Phoenix. The differences could indicate different levels of truck activities in that metropolitan area, but may indicate that the subscribers of GPS services in these metropolis areas represent a different mix of fleets with different truck activity. Although the number of events by land use do add to the metropolitan totals, the truck rates by land use are not additive. This is due to the fact that some truck trips may begin in one land use type and end in another land use type. The land use activity reported by the GPS data varies by metropolitan area. Stops by truck in industrial land uses are 24 percent of the total in Los Angeles, 27 percent in Chicago, 20 percent in Baltimore, and 13 percent in Phoenix. Stops by trucks in retail and com- mercial land uses are 31 percent of the total in Los Angeles, 28 percent in Chicago, 28 percent in Baltimore, and 25 percent in Phoenix. Stops by trucks in residential land uses are 21 per- cent of the total in Los Angeles, 18 percent in Chicago, 30 per- cent in Baltimore, and 31 percent in Phoenix. This could represent different patterns of truck usage in these areas, but more likely reflects the bias of different fleets of trucks that 19 Metro Area Land Use Travel Time to Next Stop (Minutes) Airline Distance to Next Stop (Miles) Mileage to Next Stop (Miles) Circuity Ratio Total 43.61 7.44 11.27 1.51 Industrial 54.59 8.71 13.19 1.51 Low density 43.36 8.26 12.62 1.53 Other high-density employment 41.11 8.56 12.75 1.49 Residential 37.96 4.88 7.47 1.53 Los Angeles Retail and com mercial 41.27 7.58 11.48 1.51 Total 40.90 5.73 10.82 1.89 Industrial 44.81 5.97 13.37 1.88 Low density 43.28 7.40 13.92 2.06 Other high-density employment 36.77 5.38 11.11 1.52 Residential 43.07 4.73 7.18 1.68 Chicago Retail and com mercial 37.80 5.85 9.86 2.24 Total 37.92 4.64 9.53 2.05 Industrial 44.16 4.97 10.27 2.07 Low density 37.84 5.13 9.22 1.80 Other high-density employment 36.43 3.87 6.57 1.70 Residential 37.43 4.14 11.03 2.67 Baltimore Retail and com mercial 34.73 4.93 8.32 1.69 Total 40.61 4.72 7.73 1.64 Industrial 54.46 6.09 10.95 1.80 Low density 40.08 5.26 8.45 1.61 Other high-density employment 44.80 4.66 7.01 1.50 Residential 34.14 3.24 4.99 1.54 Phoenix Retail and com mercial 40.10 5.30 8.92 1.68 Table 3.2. GPS-derived average trip characteristics.

make up the GPS vendor’s customer base in these metropoli- tan areas. Land Use Interchange Matrix Trip chaining recognizes that the probability of making a truck trip in a tour depends both on the type of activity the truck is serving at its current stop, as well as the type of activ- ity at the next stop. For example, a truck that is currently stopped at a manufacturing facility might be expected to make its next stop at another manufacturing facility. This informa- tion could be used to weight the attractiveness of truck trip distribution for individual trips, and to organize these trips into chains (tours). The data on the probability of making a truck trip from one activity, as determined by the land use in which the GPS stop event was located, to the activity serving the next stop for the same vehicle was determined by examin- ing and processing the GPS records. The results of the inter- changes of individual truck trips, based on the land use activity at the originating stop and the terminating stop, are shown in Table 3.3. The percentage of truck trips by the land use in interchange at the origin and the destination of trips in Table 3.3 total to 100 percent within each metropolitan area. Even on this basis, the tables show similar patterns of interchanges. If the cells are weighted based on the total trips to or from that land use pattern, the data appear even more consistent. In each case, as would be expected, the activity within a land use (e.g., trips with a manufacturing land use as the origin and a man- ufacturing land use as the destination) is the highest value 20 Los Angeles Destination Industrial Other High-Density Employment Retail and Commercial Residential Low Density 14.80% 2.20% 4.00% 1.70% 0.50% 64.00% 9.50% 17.20% 7.20% 2.20%Industrial 63.30% 12.40% 13.00% 7.10% 10.10% 2.40% 10.80% 2.90% 1.30% 0.50% 13.20% 60.60% 16.20% 7.10% 2.90% Other High-Density Employment 10.10% 61.30% 5.50% 5.50% 10.20% 4.10% 1.00% 18.40% 5.10% 1.00% 13.10% 3.00% 58.40% 16.20% 3.00%Retail and Commercial 17.60% 19.00% 60.20% 21.80% 19.00% 1.50% 1.20% 4.40% 14.50% 0.70% 6.90% 5.20% 19.60% 65.10% 3.30% Residential 6.60% 6.60% 14.30% 62.20% 14.80% 0.60% 0.50% 0.90% 0.80% 2.30% 11.30% 10.20% 17.70% 15.60% 45.20% O ri gi n Low-Density 2.50% 2.90% 2.90% 3.40% 45.90% Chicago Destination Industrial Other High-Density Employment Retail and Commercial Residential Low Density 15.30% 2.10% 3.50% 1.50% 1.20% 64.80% 8.90% 15.00% 6.40% 4.90%Industrial 64.30% 13.60% 12.60% 6.30% 13.70% 2.20% 9.10% 2.40% 1.10% 0.90% 13.90% 57.70% 15.40% 7.20% 5.90% Other High-Density Employment 9.20% 59.00% 4.70% 4.70% 10.90% 3.70% 2.40% 15.60% 5.40% 1.70% 12.80% 8.40% 54.10% 18.70% 5.90%Retail and Commercial 15.60% 15.70% 55.20% 22.40% 20.10% 1.50% 1.00% 4.90% 14.70% 1.20% 6.90% 4.40% 21.00% 63.50% 5.30% Residential 6.60% 6.60% 17.20% 61.00% 14.50% 1.30% 0.80% 1.80% 1.30% 3.50% 14.80% 9.30% 20.60% 15.50% 39.80% O ri gi n Low-Density 5.40% 5.30% 6.40% 5.60% 40.80% Table 3.3. GPS-derived land use interchange matrix.

within origins and destinations, and in all but a few cases (low-density interchanges in Los Angeles, Chicago, and Bal- timore, and other high-density employment in Baltimore and Phoenix) it is the majority of trips to or from that land use. Even when those intra land use exchanges are not the major- ity of truck trips, they are still the highest percentage. Trip Characteristics As discussed previously, trip chaining recognizes that the probability of making a truck trip in a tour depends both on the type of activity the truck is serving at its current stop and the type of activity at the next stop. It also depends on the characteristics of trips between these stops. In addition to being able to identify the land use at the destination for trips from a given land use origin, the GPS information can be used to estimate the travel time and distances between stops in the chain. The averages of these times and distances in total also can be used to develop friction factors for truck trip dis- tribution models. This same information by land use can be used to develop friction factors between specific types of land uses that might be used in trip chaining. The distance for a trip was calculated both as the airline distance between the latitudes and longitudes reported for the GPS records, as well as the difference in odometer readings reported by these GPS records. It is worth noting that the GPS odometer reading 21 Baltimore Destination Industrial Other High-Density Employment Retail and Commercial Residential Low Density 11.90% 0.80% 3.20% 2.30% 1.60% 60.40% 3.80% 16.10% 11.40% 8.20%Industrial 60.60% 11.00% 11.90% 6.80% 12.30% 0.80% 2.90% 1.40% 1.20% 0.60% 10.80% 42.10% 20.30% 17.70% 9.00% Other High-Density Employment 3.80% 42.70% 3.70% 3.70% 4.80% 3.20% 1.40% 14.30% 6.50% 2.60% 11.30% 5.10% 51.00% 23.20% 9.30% Retail and Commercial 16.10% 20.80% 53.00% 19.50% 19.80% 1.50% 1.10% 5.60% 20.50% 2.70% 6.90% 3.50% 17.50% 64.20% 8.40% Residential 6.60% 16.30% 20.70% 61.40% 20.30% 1.80% 0.60% 2.50% 2.90% 5.70% 13.10% 4.70% 18.50% 21.60% 42.10% O ri gi n Low Density 9.00% 9.20% 9.20% 8.70% 42.80% Phoenix Destination Industrial Other High-Density Employment Retail and Commercial Residential Low Density 6.30% 0.60% 2.70% 1.10% 1.40% 52.20% 5.30% 22.00% 9.40% 11.10%Industrial 52.60% 6.60% 10.80% 3.50% 6.60% 0.60% 4.00% 2.00% 1.70% 1.10% 6.70% 42.50% 21.50% 17.60% 11.60% Other High-Density Employment 5.20% 41.50% 5.10% 5.10% 5.40% 2.50% 2.10% 12.40% 5.40% 2.90% 10.10% 8.20% 48.80% 21.40% 11.50% Retail and Commercial 21.10% 21.40% 49.60% 16.50% 14.20% 1.50% 1.80% 5.10% 21.20% 3.20% 6.90% 5.50% 15.60% 65.50% 10.00% Residential 6.60% 18.40% 20.30% 64.50% 15.90% 1.40% 1.20% 2.80% 3.50% 11.80% 6.90% 5.70% 13.40% 16.70% 57.20% O ri gi n Low Density 11.90% 12.10% 11.20% 10.50% 57.90% Note: For each cell in the table: the first value, shown in bold, is the percent of the total table, the second value, shown in italic, is the percent of the origin, and the third value, shown in regular type, is the percent of the destination. Table 3.3. (Continued).

may be an actual odometer reading read directly from the truck’s equipment. It may also be a computed odometer reading based on the continuous GPS readings, in addition to the recorded and reported readings. The truck-generated odometer reading was used in this research to compute truck trip characteristics. This is because the GPS odometer read- ing will be incorrect when the device loses a satellite signal lock (e.g., if travelling in a tunnel or due to other sky block- ages). The difference between the average airline distance and the actual highway mileage shown is an indication of the amount of circuity of the trip on the highway network. The travel time was computed as the time difference between stop event times, adjusted to account for the dwell time at a stop. In order to calculate the dwell time at a stop, the GPS data were examined to identify the first moving GPS event record after a stop. The GPS vendor recorded this event when the vehicle speed after a stop reached 5 mph. This was considered to be a reasonable approximation of when the out- bound trip began. A flaw was discovered when it was deter- mined that the recorded stop events could include both business stops and stops captured due to congested traffic. For the purposes of this research, only business stops were of interest. Despite efforts to apply filters to exclude traffic stops, no consistently successful filter was identified. Based on this feedback, the GPS vendor is intending to change the record- ing logic to distinguish between stops with ignition on (a stop in traffic) and a stop with the ignition off (an activity stop). This may be a better way to address this issue in the future. The processed data for the four selected metropolitan areas are included in Table 3.2. Although not presented, these data could be processed to determine additional information (e.g., median, standard deviation, and distribution around the median). Similarly, in addition to calculating average times and distances between stops, the same information can be cal- culated by stop sequence (1st, 2nd, 3rd, etc.). The ability to develop this information may assist in developing chaining and/or distribution models. This research shows that subscription GPS data from trucks may be an inexpensive way to determine a variety of character- istics that could be used in truck trip distribution and chaining. The data developed appear reasonably consistent and credible. Before these data—or other data that would be developed in a similar manner—could be fully utilized, questions regarding data expansion need to be addressed. GPS subscription data are made anonymous before release in order to protect the identity of trucking clients. In order to develop meaningful dis- aggregations or expansions to types of trucks, more detailed information should be developed. Based on this research proj- ect, the GPS vendor is investigating methods to store informa- tion (e.g., the first eight characters of a Vehicle Identification Number [VIN], which could provide sufficient information to develop disaggregations and survey expansion factors while preserving anonymity). 3.4 Consideration of Temporal and Seasonal Impacts Freight flows are traditionally expressed as tons per year. This is true when freight flows are reported as multimodal flows (in CFS or FAF), as modal flows (e.g., in the STB Way- bill for railroads), or facility flows (as in the U.S. Army Corp of Engineers’ Waterborne Commerce Statistics for ports). Conversely, freight flows, particularly vehicle flows, are typi- cally expressed in vehicles per day for capacity expansion and design decisions, and as vehicles per hour to support opera- tional decisions. In order to use existing freight and truck modeling processes to support infrastructure decisions, it would be useful to develop factors that can be used to convert annual freight flows to daily and hourly flows. Databases of truck vehicle movements are for specific geo- graphic locations on highway networks. These databases of truck classification counts can not distinguish trucks by body type or by the contents or type of freight being carried. The annual modal and multimodal commodity freight flows reported by trucks are for origins and destinations of the freight flows, not for the highway locations along the routes between origins and destinations that would correspond to the truck counts. To address this difference, the research team first identified methods to assign commodity truck OD flows to the highway network. This would allow the identification of freight flows by commodity at highway locations corresponding to the truck counts. These truck counts could be used to develop monthly and hourly factors. These factors could then be applied to com- modity flows at each specific location. The flows by commodity will vary on the highway network, and the monthly and hourly factors from counts will vary by location. However, if the commodity flow is principally of a specific commodity, the variation in truck counts should also reflect variation for this commodity. For example, if the com- modity truck flows at a location hypothetically consisted of only a single commodity, then the monthly and hourly vehicle counts at this location could be expected to represent the monthly and hourly factors for that commodity. Where no sin- gle commodity dominates, it is proposed that the truck flow pattern at any single location will reflect the seasonal and tem- poral flow pattern of the underlying commodities. Although this method cannot be expected to develop factors that would apply to specific locations, the aggregation of the resulting pat- tern across all locations and commodities is expected to reflect an average national distribution of commodity freight flows by month and by hour. For this research topic, the first step will be to develop a method to assign the truck commodity flows to the highway network. The second step will be to develop monthly and hourly factors from national databases of truck counts. The third step will be to apply those factors to the commodity 22

flows on the network corresponding to the count locations. The fourth step will be to aggregate those monthly and hourly flows and develop average national adjustment factors. Development of a Commodity Assignment The 1998 version of the Freight Analysis Framework (FAF1) produced maps of truck flows on the FAF1 highway network and made available (via download) highway network files of daily freight truck volumes on the highway links, in two widely used platforms, ESRI and TransCAD. These flows were pro- duced by converting the county-to-county tonnage flows by truck to daily truck flows through the use of annual-to-daily factors as well as tons-to-truck-payload factors. Although these data could have been used to produce highway flows by com- modity, the proprietary nature of the FAF1 data prevented the disclosure of highway link flows with this information. It also prevented the disclosure of the origin/destination/commodity (O/D/C) table at a county-to-county level, which could have been used to assign the truck flows to the highway network. Only the reporting of state-to-state flows was publicly available. The FAF2 flow database, which is the 2002 update to the FAF1 flow database, can be considered as a basic O/D/C table for 114 very aggregate zones, called FAF2 regional zones. Those FAF2 zones consist of the state portion of the largest metropol- itan areas, and the remainder of states or whole states outside of these metropolitan zones. FAF2 separately developed a high- way network, which included updated information and addi- tional detail beyond the FAF1 highway network. However, the regional zone structure of the OD table by commodity is not consistent with the assignment scripts for the FAF1 or the detail of the FAF2 highway network. In order to be compatible with the assignment scripts developed for the FAF1, the O/D/C table must be disaggregated to smaller geographic zones (e.g., coun- ties, as used in the FAF1). The FAF2 documentation9 describes a procedure for disaggregating the FAF2 regional zones to counties and other freight activity centers. That procedure is based on the share of the number of establishments in the activity center as a ratio of the number of establishments in the zone and the share of the HPMS truck VMT in the activity cen- ter as a ratio of the HPMS truck VMT in the zone. There is no reason to think that truck trip ends in an activ- ity center should be related to truck VMT in that activity cen- ter. For example, a major truck route with considerable truck VMT passing through an otherwise empty activity center (e.g., county) does not indicate that there should be trip ends in that activity center. Similarly, a measure of the level of intensity of the establishments within an activity center (e.g., employ- ment), not the number of establishments should be used to disaggregate freight flows. For FHWA, Cambridge Systematics developed a procedure that disaggregates FAF2 regional flows to county flows using the employment data for the zones that produce and consume freight. This procedure relies on county business patterns’ employment, for each county, in the industries that produce and consume freight. The level of use by each industry for each SCTG2 commodity in the FAF2 was established by regression. Additionally, flows through ports, airports, and border cross- ings were disaggregated based on the county in which the facility was located and the share of reported activity at that facility from other modal databases. This procedure is docu- mented in an unpublished FHWA report.10 These disaggregation methods were used to convert the 144-zone by 144-zone FAF2 flows (where the regional zones beyond 114 represent ports, border crossings, and interna- tional zones) into flows among the 3,140 counties in the United States. The FAF2 database is available for download from the FHWA website as a Microsoft Access database. This database was converted to a set of TransCAD matrices, one for each SCTG2 commodity. The FAF2 technical documentation also describes a proce- dure for assigning truck flows to the FAF2 highway network.11 That procedure uses the stochastic user equilibrium (SUE) assignment routine in TransCAD to assign the freight flow tables. This assignment procedure uses the information avail- able in the FAF2 network (such as capacity, total vehicle vol- umes, and free-flow speed) to calculate a congested time on each link. That congested time is used as the basic link imped- ance. That basic impedance is modified by additional infor- mation for each link, such as the number of lanes, the location of the link in urban areas, truck restrictions, truck route des- ignations, tolls, and any interstate designation of the link. The SUE assignment is based on that modified impedance. The results of the assignment using the Battelle FAF2 disaggregated database and this procedure is shown in Figure 3.1. For this research topic, TransCAD scripts were developed to implement the documented assignment procedure and were used to assign the county-to-county flows for each of the 42 SCTG2 commodities disaggregated from the FAF2 data- base. The resulting assignment of total truck tonnage is shown in Figure 3.2. The flow pattern appears similar to that of Fig- ure 3.1 where most flows are concentrated on major interstate highways and the trip ends are concentrated in the counties 23 9 Battelle Memorial Institute, “Chapter 4: FAF2 Truck O-D Data Disaggregation,” in FAF2 Freight Traffic Analysis, FHWA, June 27, 2007, http://ops.fhwa.dot.gov/ freight/freight_analysis/faf/faf2_reports/reports7/c4_data.htm. 10 Cambridge Systematics, Inc., Development of a Computerized Method to Sub- divide the FAF2 Regional Commodity OD Data to County-Level OD Data, prepared for FHWA, January 2009, unpublished. 11 Battelle Memorial Institute, “Chapter 5: Freight Truck Assignment and Cali- bration,” in FAF2 Freight Traffic Analysis, Federal Highway Administration, June 27, 2007.

with large populations or production and/or consuming industries. This assignment procedure provides link volumes for each of the 42 SCTG2 commodity flows that are not publicly avail- able for the FAF2 network. Although the assignment routine can produce flows for each of the 42 disaggregated SCTG2 commodities, there are known errors in the creation of flows for certain import and export flows. For example, in translat- ing from the Performance Monitoring System (PMS) com- modity classifications used in waterborne commerce to the SCTG2 commodity classification system used in FAF2, all manufactured goods were reported to move in SCTG 34, machinery. As shown in Figure 3.3, this results in higher than expected flows to and from ports such as Savannah, Georgia, and Charleston, South Carolina. This error is more pronounced at the SCTG2 level of detail and is less of an issue at higher levels of commodity aggrega- tion. Additionally, the level of detail for 42 SCTG commodi- ties has additional processing and reporting issues. Rather than processing the flows at the SCTG2 level, the flows were grouped to the nine classes of commodities used in the 2002 CFS. For these CFS commodity groups, the ton-miles of flow were calculated from the assigned link volumes and the dis- tances of those links. The result of this calculation is shown in Table 3.4. As shown, the total ton-miles calculated for trucks are approximately twice the reported value in the CFS. This is to be expected since the CFS has several major commodity gaps, referred to by the FAF2 as out-of-scope commodities. In addition, the CFS undercounts some categories of trade and movements of freight (e.g., in-transit movements, petro- leum products, and exports). The FAF2 includes these addi- tional flows. This summary of ton-miles indicates that the flows on the FAF2 network are reasonable and can be used in processing the later steps. Development of Monthly and Hourly Truck Factors FHWA maintains the VTRIS database of traffic counts taken at stations by automatic traffic recorders (ATRs), vehi- cle classification counters, weight-in-motion equipment, and weight enforcement stations as submitted by state DOTs. The VTRIS database includes the station description, vehicle clas- sification, and the time of the counts in a consistent format for all 50 states. Although this does not provide any information 24 Source: FAF2 Freight Traffic Analysis, Figure 3.3. Figure 3.1. Base-year 2002 FAF2 truck flow on FAF2 highway network.

about the commodities carried by these trucks, it is possible to develop hourly and monthly factors from the stations in VTRIS. Weigh-in-motion and weight information stations do not provide the continuous readings that would be required to develop monthly and hourly factors. The ATR counts are of total vehicles and do not differentiate between trucks and other vehicles, including automobiles. The vehicle classifica- tion counts do provide the ability to distinguish trucks from other vehicles and include locations that are counted contin- uously. The information in VTRIS for classification counts is recorded by hour and by date. The complete VTRIS database for 2007 was obtained from FHWA’s Office of Highway Pol- icy Information. From that VTRIS database, the tables of vehi- cle classification counts were selected. VTRIS contains records for 13,862 classification stations for the United States. However, in order to develop hourly allo- cation factors, a station needs to be operated without hourly gaps for weekdays. There were only 798 stations without gaps, and these could be used to develop hourly factors. Only sta- tions that are on the FAF2 network links can be applied to the commodity flows and multiple stations on the same FAF2 link have to be combined before they can be used. As a result, hourly factors could be developed for 623 links on the FAF highway network. The location of these stations is shown in Figure 3.4. To develop monthly allocation factors, a station needs to be operated without daily gaps for the year. There were only 200 stations without gaps that could be used to develop monthly factors. Only stations that match the FAF2 network links can be applied to the commodity flows and multiple sta- tions on the same FAF2 link have to be combined before they can be used. Monthly factors can be developed for 177 links on the FAF2 highway network. The location of these stations is shown in Figure 3.5. For each station, factors were developed for combination trucks—Vehicle Classes 8 through 13—according to FHWA’s Scheme F classification. These vehicles are those that would most likely carry freight. Apply Factors to the Commodity Flows on the Network For each of the 623 links that have complete hourly factors developed from counts, those factors were applied to the 25 Figure 3.2. Disaggregated FAF2 truck ton flows (all commodities).

26 Figure 3.3. Disaggregated FAF2 truck ton flows (SCTG 34, machinery). CFS Commodity Group SCTG Code Description Annual Ton-Miles (Millions) CFS1 01 to 05 Agriculture Products and Fish 398,154 CFS2 06 to 09 Grains, Alcohol, and Tobacco Products 210,034 CFS3 10 to 14 Stones, Nonmetallic Minerals, and Metallic Ores 382,451 CFS4 15 to 19 Coal and Petroleum Products 222,452 CFS5 20 to 24 Pharmaceutical and Chemical Products 268,038 CFS6 25 to 30 Logs, Wood Products, and Textile and Leather 330,683 CFS7 31 to 34 Base Metal and Machinery 437,008 CFS8 35 to 38 Electronic, Motorized Vehicles, and Precision Instruments 87,758 CFS9 39 to 43 Furniture, Mixed Freight, and Miscellaneous Manufactured Products 245,299 Total 2,581,876 Truck Ton-Miles—2002 CFS Table 2a 1,261,813 Table 3.4. Annual ton-miles traveled by each CFS commodity group.

27 Figure 3.4. VTRIS stations with valid hourly factors. Figure 3.5. VTRIS stations with valid monthly factors.

annual tonnage flows to produce estimated hourly flows for each of the SCTG commodities. For each of the 177 links that have complete monthly factors developed from counts, those factors were applied to the annual tonnage flows to produce estimated monthly flows for each of the SCTG commodities. Develop Average National Adjustment Factors The hourly flows for each location were used to develop a national hourly summary of freight flows. The resulting hourly freight flows were aggregated to the nine CFS commodity groups. An hourly distribution of the summary of freight flows was developed and that distribution is shown in Figure 3.6. It is noted that averaged over all locations, the hourly distribu- tion of each of the nine CFS commodity groups appears to be virtually identical. The monthly flows for each location were used to develop a national monthly summary of freight flows. The resulting monthly freight flows were aggregated to the nine CFS com- modity groups. A monthly distribution of the summary of freight flows was developed and that distribution is shown in Figure 3.7. It is noted that averaged over all locations, the monthly distribution of each of the nine CFS commodity groups appears to be virtually identical. As noted, for both hourly and monthly factors, the resulting distribution showed little variation by commodity group. This finding was not expected. It seems to suggest that while indi- vidual locations could show distribution patterns that differ significantly from national averages, absent any other local information, it is reasonable to assume that commodity groups all follow the same distribution pattern for both hourly and monthly flow. Although this pattern could be true for large commodity groups, it might not hold true for the very differ- ent SCTG2 commodities with these groups. The pharmaceuti- cal and chemical commodity group includes two commodities (SCTG 21, pharmaceuticals, and SCTG 22, fertilizers) that could be expected to follow very different patterns. However, the development of hourly factors from the commodity flow data for each of these commodities produces very similar results, as shown in Figure 3.8. It is therefore assumed that the method used to estimate the average hourly distribution of commodities would yield the same results for any commodity. Monthly flow patterns should be related to monthly pat- terns of production by commodity. This information can be verified from separate sources. The Federal Reserve Board tracks this information in its industrial production and capac- ity utilization statistics. The U.S. Census Bureau tracks this information in the Manufacturers’ Shipments, Inventories, and Orders Survey (M3). The data were examined for 2007, which was prior to the current economic recession. Both data sources show similar results, but the Census Bureau data is easier to use since it tracks data for the entire year, not by fis- cal quarter. The Census Bureau survey reports shipments as dollar values not tonnage, but if there is stability in commod- ity prices, the flow patterns for tons and value should be sim- ilar. Additionally, the Census Bureau reports these shipments by NAICS industry, which is similar, but not identical, to the 28 0.0% 1.0% 2.0% 3.0% 4.0% 5.0% 6.0% 7.0% 12 AM 2 A M 4 A M 6 A M 8 A M 10 AM 12 PM 2 P M 4 P M 6 P M 8 P M 10 PM Hour Beginning % Ho ur ly CFS1 CFS2 CFS3 CFS4 CFS5 CFS6 CFS7 CFS8 CFS9 Figure 3.6. Hourly distribution of truck commodity flows.

SCTG commodities in the FAF. Table 3.5 shows the reported value of shipments by industry by months as well as the cal- culated standard deviation of those monthly flows as a per- centage of average monthly flows. For all but a few industries, which are shown on five shaded rows, monthly variation as a standard deviation is less than 10 percent of the average monthly flow. For the industries excluding those in Table 3.5’s shaded rows, the distribution of the U.S. Census Bureau reported monthly shipments is shown in Figure 3.9. Similar to the esti- mated monthly distribution of flows by commodity shown in Figure 3.7, there is little variation of flows across the months. The industries, whose monthly standard deviation of flows exceeds 10 percent of the average, are shown in Figure 3.10. Even for most of the industries with the most variation, that variation is minimal. For those industries with the most vari- ation, since the reported data is shipment value not tons, it is conceivable that the variation is due to fluctuations in com- 29 Figure 3.7. Monthly distribution of truck commodity flows. Fertilizers 12 A.M. Percent 0% 1% 2% 3% 4% 5% 6% 7% 2 A.M. 4 A.M. 6 A.M. 8 A.M. 10 A.M. 12 P.M. 2 P.M. 4 P.M. 6 P.M. 8 P.M. 10 P.M. Pharmaceuticals Figure 3.8. Hourly distribution of pharmaceutical and fertilizer flows.

Industry January February March April May June July August September October November December Standard Deviation as Percentage of Average Food Products 44.1 44.3 47.1 44.7 47.7 47.9 46.3 50.2 49.9 51.3 50.6 49.4 5% Beverage and Tobacco Products 9.4 9.6 10.7 10.6 12.1 11.8 11.2 12.3 10.9 11.7 11.4 10.6 8% Textiles 3.0 3.1 3.2 3.1 3.2 3.2 2.9 3.2 3.1 3.1 2.9 2.6 6% Textile Products 2.4 2.5 2.7 2.6 2.6 2.8 2.6 2.7 2.5 2.6 2.4 2.1 7% Apparel 2.2 2.5 2.5 2.3 2.4 2.3 2.4 2.8 2.6 2.8 2.9 2.3 9% Leather and Allied Products 0.4 0.6 0.6 0.5 0.5 0.5 0.5 0.6 0.6 0.5 0.5 0.5 10% Wood Products 7.3 7.5 8.4 8.6 9.3 9.6 8.8 9.4 9.0 8.8 7.8 7.4 10% Paper Products 14.0 13.2 14.1 13.6 14.4 14.5 14.0 14.5 14.0 14.5 14.0 13.8 3% Printing 8.0 7.8 8.7 8.2 8.3 8.4 8.0 8.7 8.7 9.4 9.1 8.5 6% Petroleum and Coal Products 36.6 36.4 43.1 45.2 50.4 49.1 49.9 47.8 47.7 50.2 54.7 53.0 12% Chemical Products 51.8 50.8 58.0 55.9 58.2 57.1 55.2 56.9 53.7 58.0 54.3 54.3 4% Plastics and Rubber Products 16.5 16.0 18.1 17.7 18.9 18.6 17.4 18.9 17.3 18.8 17.1 15.3 7% Nonmetallic Mineral Products 9.1 9.0 10.4 10.4 11.0 10.7 10.2 10.9 9.8 10.6 9.4 7.9 9% Primary Metals 19.5 18.8 21.0 20.8 21.9 21.3 19.5 20.9 19.9 21.3 19.1 17.5 6% Fabricated Metal Products 25.3 25.3 28.0 27.3 28.8 28.9 26.7 30.0 27.7 28.6 25.9 24.1 7% Machinery 23.2 24.6 30.1 29.0 28.9 30.1 26.8 27.7 28.5 28.2 25.9 28.5 8% Computer and Electronic Products 28.4 29.6 35.7 30.1 30.4 36.3 27.4 31.0 35.9 31.5 32.4 38.1 11% Electronic Equipment, Appliances, and Components 9.3 9.6 11.3 10.6 10.9 11.3 10.0 11.2 11.3 11.0 10.3 10.1 7% Transportation Equipment 50.3 55.9 65.9 55.2 61.9 63.7 45.3 64.5 58.5 60.6 57.2 54.0 11% Furniture and Related Products 6.5 6.7 7.1 6.7 6.8 7.0 6.8 7.4 6.9 7.1 6.7 6.5 4% Miscellaneous Products 11.7 11.7 13.6 11.9 12.6 13.6 11.6 12.9 12.9 13.4 13.2 13.9 7% Source: U.S. Census Bureau, Manufacturers’ Shipments, Inventories, and Orders Survey (M3). Note: Shaded rows indicate those industries with a standard deviation of 10 percent or greater. Table 3.5. Value of shipments by industry (2007, in billions of dollars).

modity price (e.g., the value per ton of petroleum) or in inter- national currencies. The U.S. Census Bureau data generally confirms the findings of the proposed method, that there is lit- tle variation in commodity flows, on average, throughout the year. In the absence of local data showing specific local varia- tions, any policy considerations of commodity truck flows that have been converted from annual to daily flows based on averages need not be concerned about seasonal variations in those commodities. Similarly, when policy decisions need to consider the hourly variation of commodity flows, absent any specific local infor- mation, the hourly distribution of commodity flows according 31 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Dollars (in Millions) 2007 $0 $50,000 $100,000 $150,000 $200,000 $250,000 $300,000 $350,000 Figure 3.9. Industries with minimal monthly variation. Jan Transportation Equipment Wood Products Leather and Allied Products Petroleum and Coal Products Computer and Electronic Products Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Dollars (in Millions) 2007 $0 $10,000 $20,000 $30,000 $40,000 $50,000 $60,000 $70,000 Figure 3.10. Industries with the most monthly variation.

to Figure 3.8, should be considered to be approximately con- stant at 6 percent each hour for all hours from 1:00 A.M. to 5:00 P.M. This drops to only 5.5 percent in each of the hours before and after that period. Although this may appear incon- sistent with conventional wisdom that trucks travel at night, it should be recognized that the national average truck trip length reported in the 2002 CFS is only 173 mi. At this distance, which can easily be completed during one business day, it should be expected that the majority of truck commodity activity would occur during normal business hours, if not prevented by local conditions. Annual-to-Daily Factor The assumption of this research was that different commodi- ties have different seasonal and temporal variations. Because the finding was that, on average, all commodities have similar sea- sonal and temporal variations, the VTRIS data were exam- ined to address an additional issue—the factor that should be applied to all annual ton flows to convert to daily flows. Different practitioners use different adjustments to con- vert annual commodity flows to daily flows. Some practition- ers merely divide the annual tons by 365, the number of days in a year, neglecting any lower flow on Saturdays and Sun- days. Some practitioners divide the annual flows by 250 as an estimate of the number of working weekdays, which makes the assumption that there is no flow on Saturdays and Sun- days. The proper factor is expected to fall within that range that would imply there is flow on Saturdays and Sundays but that the flow is less than the flow occurring on weekdays. The VTRIS data, for those stations that had continuous counts for the entire year, were examined to determine how those truck volumes vary during an average week. The 542 mil- lion trucks that were observed at the 177 stations with complete counts have the weekday distribution shown in Figure 3.11. Forecasting models most often deal with average weekday conditions. From Figure 3.11, the average daily flow is 85 per- cent of the average weekday flow. The annual conversion to average weekday would then be 85 percent of the days in the year, or 310. It also is observed that there is variation in truck volumes during the week, with similar volumes on Tuesday through Thursday and lower truck volumes on Mondays and Fridays. The average of Tuesday through Thursday truck vol- umes is 81 percent of the average daily volumes. The annual conversion to midweek daily flows would be 81 percent of 365 days or 295. It should be noted that other qualitative estimates of the conversion of annual to daily flows have relied on evidence from other sources in which total flow on both weekend days is almost equal to the average weekday flow which, over 52 weeks, would mean a flow equivalent to 312 days. This is often reduced by an estimate of the number of holidays on which little flow is expected, a number ranging from 6 to 12 days, which would reduce the annual-to-daily conversion to a number between 300 and 306. This practice would appear to be consistent with the values derived from Figure 3.11. 3.5 Developing Mode-Choice Models for Freight Forecasting Policy decisions for freight commonly consider alternatives that could change the mode-choice decision for domestic freight, most often those that would shift freight from high- way modes (truck) to nonhighway modes (rail, inland water, or air). These policy alternatives would benefit from a better understanding of the factors that affect the mode-choice deci- sion for freight, including the relative importance of these fac- tors and how they should be considered in freight forecasting. To address this need, this research topic investigated the vari- ables used in mode-choice decisions and attempted to find 32 0% 20% 40% 60% 80% 100% 120% 140% Sun Mon Tue Wed Thu Fri Sat Figure 3.11. Weekday trucks at VTRIS stations as percentage of average daily trucks.

how those variables can be used in estimating a freight mode- choice model. In choice modeling, equations representing the choice decision are most often in the form of logit choice equations. These logit models may be multinomial, that is several vari- ables are considered simultaneously in the choice decision, and where there is no correlation between those variables. In order to determine the variables that are important in the choice decision, as well as to determine the relative impor- tance of these variables, it is necessary to examine a survey of those choices together with the values of variables that are rel- evant to those choices. When the survey reflects observed choices as well as the observed relevant variables, this is called a revealed-preference (RP) survey. When the survey is made of decisionmakers to determine their stated choices given a hypothetical set of values for relevant variables, this is called a stated-preference (SP) survey. Stated-preference surveys in the mode-choice decision for freight would require the identification of a statistically rele- vant sample of decisionmakers. An SP survey requires provid- ing those decisionmakers with sufficient hypothetical choice experiments. From their responses, the relevant variables in the freight mode-choice decision would be determined. That determination would be made by estimating the coefficients and parameters associated with those variables in a logit choice model. The freight mode choice is most often national in scope, which would require that the geographic scope of such a survey also be national in scope. Identifying these decision- makers and conducting the choice experiments is an expen- sive undertaking that is beyond the resources of all but the largest freight studies. In order for RP surveys to support the development of a freight mode-choice model, it is necessary for the survey to report flows for all modes in a consistent manner, over a period of time that is long enough, and for a geography that is large enough to capture modal decisions. Because the values of the choice variables will differ between different origins and des- tinations, the RP survey must report information for both the origins and destinations of freight. Because freight mode- choice decisions are assumed to be similar for freight that shares the same characteristics, the freight flows also should be reported separately for freight (e.g., by commodities) that is expected to behave similarly. Additionally, the choice variable and the observed decisions can not be expected to be the same over very large geographies. The reported geographies in the RP database must be at a scale where modal availability and modal characteristics can be assumed to be similar within the reported geography. Finally, the choice in the RP database should be complete trips between and origin and destination— that is, linked trips—that may involve several modes as well as the transfers between modes at intermediate points. Mode- choice decisions should not be made using unlinked trips that are separately reported for each modal component of a trip between an origin and a destination. Given these requirements, only a few commodity flow databases should be considered for use as RP databases, as follows: • Databases that are limited to single modes (for example, the STB Carload Waybill Survey for rail) can not be used because they reveal no information about the decisions for competing nonrail freight modes. • The publicly available CFS provides flows for both origins and destinations only as state-to-state movements, and entire states are not a scale of geography over which modal availability and characteristics can be considered similar. • The privately available TRANSEARCH database does pro- vide flows for seven modes between zones chosen as part of the data purchase, which can be as small as counties or, syn- thetically, as zip codes. The flows are reported in unlinked form, which although more suitable for determining the proper assignment to modal networks, provides incom- plete information for trips that use multiple modes. Addi- tionally, the cost of obtaining the entire TRANSEARCH database as county-to-county flows for the nation would be prohibitive. A single state database with flows at the county level within that state typically costs from $50,000 to $100,000 for a single year. • The FAF2 database provides information for all modes in a consistent format, including linked multimodal trips. How- ever, the zones in the FAF2 database are very large and some modes—especially water and rail—cannot be expected to be uniformly available throughout these zones. Although some records in the FAF2 commodity flow data- base may not be suitable for use in an RP survey, it contains enough suitable records that it might be processed for use as an RP survey. Variables in Freight Mode Choice A literature review was conducted to determine variables that would be important in the mode-choice decision for freight. Although not intended to be exhaustive, the variables that were determined to be important in the mode-choice decision for freight are • Characteristics of the mode, including capacity, trip time, reliability, cost; • Characteristics of the goods, including shipment size, shelf life, density, value; • Characteristics of the shipper, including production pro- cesses and shipper size; • Characteristics of the receiver, including receiver size and other consumption processes such as operating hours; and 33

• Other logistic characteristics, including shipment frequency, inventory costs, loss and damage costs, etc. Although these variables might be important in the mode- choice decision for freight, determining values for many of these variables requires detailed information about ship- ments, which might be obtained from an SP survey but can not be expected to be available for all shipment records in an RP survey. Publicly available information was identified that could be used to develop data for the variables to support the use of an RP survey, including • Modal distances and impedances between U.S. counties from the Center for Transportation Analysis (CTA) at Oak Ridge National Laboratory (ORNL); • Detail by commodity, including shipment size and value from the FAF2 commodity flow database; • Employment by industry for the shipper and receiver regions from the U.S. Census county business patterns; and • Population by destination region from the U.S. Census. The utility equations developed for use in logit mode- choice equations include a constant for each mode, expressed as a difference from a base mode. The base mode, for which no modal constant will be estimated, is trucking. Separate equations were developed for similar commodities. The generally important mode-choice variables, as well as how those variables will correspond to the publicly available data and parameters in the RP estimation, are shown in Table 3.6. 34 Category Utility Variable Corresponding Variable to be Used in Revealed-Preference Utility Estimation Modal Characteristics Capacity Modal Constant Trip Time Modal Distance/Impedance Reliability Modal Constant Equipment Availability Modal Constant Customer Service and Handling Quality Modal Constant Modal Cost Modal Distance/Impedance Goods Characteristics Shipment Size Commodity Total Tons Package Characteristics Commodity Modal Constants Shipment Shelf Life Commodity Modal Constants Shipment Value Commodity Value per Ton Shipment Density Commodity Modal Constants Shipper Characteristics Production Processes Industry Employment at Origin Shipper Size Industry Employment at Origin Receiver Characteristics Consumption Requirements Industry Employment/Population at Destination Receiver Size Industry Employment/Population at Destination Other Logistic Characteristics Inventory Costs Commodity Modal Constants Loss and Damage Costs Commodity Modal Constants Service Reliability Costs Commodity Modal Constants Length of Haul Truck Distance Shipment Frequency Commodity Total Tons Table 3.6. Freight mode-choice variables.

Preparing the FAF2 for Use as an RP Survey The FAF2 commodity flow database provides separate tables for domestic freight flows, seaborne international freight flows, land border crossing flows, and air and other international modal flows. The designation of the international tables pro- vides information about the mode used internationally (e.g., sea, land border, or air) while the attributes within the table provide information about the mode used for domestic flows between a U.S. FAF zone and the U.S. port of entry/exit. The seven zones outside of the United States are very large and include the entire countries of Canada and Mexico, which leaves only five zones for all of the rest of the world. Addition- ally, these files are prepared from other commodity flow files and it has been confirmed by the FHWA Office of Freight Management that the correspondence between the commod- ity codes used in some of the international flow data files and the SCTG commodities codes used in the FAF2 is not correct (e.g., SCTG 34, machinery, actually includes flows for all man- ufactured products for international water flows and there are no international water flows assigned to other SCTG codes for manufactured goods). Because of the errors in commodity assignments, because the flows only include the domestic mode used, and because the international geographies are too large to ensure consistent modal characteristics and availabil- ity for the entire international zone, it was determined that only the records in the FAF2 domestic tables would be suitable for use in an RP survey. The FAF2 domestic table reports flows between 114 FAF2 regions. These regions include the state portions of the largest metropolitan areas, as well as whole states, or remainders of states outside of those metropolitan areas. The FAF2 regions representing whole states, or remainders of states outside of the metropolitan regions, are too large to ensure consistent modal characteristics and availability throughout the region. Therefore, all records that contain a whole state or a remain- der of a state zone as an origin or destination were not included for use in the RP survey. Finally, the separation of metro- politan areas into their state portions was intended to aid in developing summaries of freight flows at the state level. The reported FAF2 flows between FAF2 regions in the same met- ropolitan area but in different states will involve short dis- tances over which modal choice decisions most likely reflect production or logistic processes unique to the commodity and not decisions that should be considered in an RP survey. Therefore, records that are of freight flows within the same metropolitan area were not included for use in an RP survey. Finally, while the FAF2 includes records for Hawaii and Alaska, the mode-choice decision for shipment to or from these regions includes unique considerations and modes, and they were dropped from use in an RP survey. The FAF2 records that were included for use in an RP survey include those domestic flows between metropolitan regions, excluding flows that are reported as shipments within the same metropolitan area. It is assumed that the flows reported for these records are consistently available with the same char- acteristics for all modes. In order to use as an RP database, total flows between and origin and destination, as well as the flows by each mode, must be determined. The FAF2 database was reformatted to include the flow by each mode for each origin, destination, and SCTG2 commodity. Appended to these were the vari- ables for that origin and destination that were to be tested as explanatory variables in the choices represented by the observed modal flows. Modal distances were obtained from the CTA at ORNL. The CTA provides skim tables of distances and impedances between U.S. counties for highway (truck), rail, and water, as well great circle distances for air travel. These skim times are based on the paths identified using the ORNL modal freight networks. In addition to distances, the CTA skims include estimated impedances for each of the modes, as well as rail highway rail (RHR) impedances that represent the impedance using the respective rail and highway networks connecting through the intermodal terminals that are expected to serve that origin and destination pair. These CTA distances and impedances are for U.S. county- to-county movements. In order to use these distances with the FAF-region-to-FAF-region records in the RP data, a rep- resentative county had to be associated with each FAF region included in the RP survey, which includes only FAF2 metro- politan regions. The county with the largest employment in a FAF metropolitan region was chosen as the representative county for use in selecting distance and impedances from the CTA skim files. County employment, the surrogate for shipper charac- teristics, was obtained from county business patterns. The employment total for all of the counties in the FAF region was selected to test as an explanatory variable. In the same manner, the Census of Population was used to calculate the population of the region, the surrogate for receiver charac- teristics. From the FAF2 database itself, the total flow for the O/D/C record was added for use as a surrogate for shipment size. From that same FAF2 database, the value of the ship- ment by all modes also was added to the RP data for use as a surrogate for shipment value. An investigation of the RP database indicated that the SCTG two-digit level for commodities had insufficient records to use in estimating models for some commodities. The records were aggregated to the commodity groups used in the 2002 CFS as shown in Table 3.4 in order to provide sufficient records to estimate the mode-choice equations by commodity group. 35

Estimation of Mode-Choice Utility Equation Utility equations were estimated from the RP flows and the variables associated with those flows. The software used was an object-oriented software package designed for the maxi- mum likelihood estimation of generalized extreme value (GEV) models including multinomial logit models. Each of the RP variables listed in Table 3.4 were tested singly, as simple functions, as simple cross products (e.g., tons multi- plied by distance, or as cross products functions with other variables such as distance multiplied by the natural logarithm of value per ton). The initial estimation runs were used to determine which variables did not contribute significantly to the utility equations, as indicated by uniformly poor t-statistics. It was found that the surrogate for producer characteristics, employment at the origin and for receiver characteristics, and population at the destination, were not significant explanatory variables at the geographies tested. This does not necessarily indicate that these variables are unimportant, only that they are unimportant at the actual geographic scales as used in the RP survey. It is possible that at smaller geographic scales and for specific shippers (which, of course, would not be correlated with the total of all employment over an entire FAF region), these variables might be significant. The CTA impedances were found to be highly correlated with distance and, in fact, the CTA describes how they are com- puted from distances. Because impedances were so highly cor- related with distance, only modal distances were retained as utility variables. Modal constants by commodity were estimated and found to be significant and large. However, because these modal con- stants are associated with a number of general variables, it is not possible to determine which of the general variables are the most significant. Additionally, the estimation method only provides an indication that these modal constants are signifi- cant relative to an assumed zero value for the base mode, which was chosen to be “truck.” It provides no indication of what the absolute modal constant is for that mode because there is no ability to estimate the modal constant for the base truck mode. For example, the estimation that the rail modal constant is sig- nificant might indicate that any rail capacity, rail reliability, rail equipment availability, or rail customer service and handling quality are important considerations in mode choice but that does not indicate the relative importance of each, nor does it indicate the absolute utility for any of these, only the total rel- ative utility compared to that of the truck mode. The results show that modal distance, whether singly or in combination with other variables, is the most important vari- able as estimated from the RP survey. Since distance in this estimation serves as a surrogate for both modal cost and modal time, this is an expected finding. What the estimation also shows is that the size of the shipment, as indicated by the surrogate of annual tons moving between markets, is only sig- nificant as a cross product with the natural logarithm of dis- tance. This indicates that the impact of shipment size increases as distance increases, but it does so at a logarithmically decreas- ing rate. The estimation also shows that value per ton is only significant as a cross product with the natural logarithm of distance. This indicates that the impact of value increases as distance increases but it does so at a logarithmically increas- ing rate. The results of the estimation are shown in Table 3.7. The variables were chosen to provide, where possible, uniform consistency across commodity groups. Thus, a chosen vari- able might have lower than desirable significance, as indicated by its t-statistic where absolute values greater than 2.0 are gen- erally considered to be significant. However, that variable was retained to allow for comparison with other modes and com- modities. For those commodities where the variable clearly degrades the estimation, they were excluded. Although the estimation model was used primarily to show which variables are significant in freight mode choice, the esti- mated coefficients themselves can be used to gain insights as to how mode-choice decisions might change as these variables change. The sign of the variable coefficients in Table 3.7 indicates whether the modal utility increases (has a positive sign) or decreases (has a negative sign) as the variable increases. Thus, for SCTG 01-05, agricultural products commodity group, the truck utility for distance decreases (coeff. = –0.00423/mi) as distance increases. As expected and shown in Table 3.7, the modal utility decreases as distance (serving as a surrogate for modal cost and time) increases. The value of the modal coefficient for all modes within a commodity group relative to other modes within that same commodity group is an estimate of the relative utility of that mode to other modes. Thus for the agricultural products com- modity group, the rail distance coefficient of –0.00397 (which is a smaller negative number compared to the truck distance coefficient of –0.00423) estimates that rail as a mode has a higher utility compared to truck as distance increases—that is, its utility increases by 6 percent per mile, –0.00432/–0.00397, compared to truck mode, as distance increases. The size of a modal coefficient, compared to all modes in a commodity group, indicates the preference for that mode. Thus, for truck mode, the estimation from the RP data is that as distance increases, truck utility decreases the most for fur- niture and miscellaneous products (–0.01110) and decreases the least for pharmaceutical and chemical products (–0.00309). A review of the modal constants for each nontruck mode within each commodity group shows that for all commodity groups, the utility of nontruck modes compared to truck mode is estimated to have a lower utility, which means that it is less likely to be chosen than the truck mode. The size of the modal 36

Commodity Truck Truck and Rail Water Rail Water and Rail Air Statistics Group Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Constant 0 0 -10.4 -7.72 -4.91 -14.21 -4.37 -25.83 -4.17 -23.41 -5.82 -18.15 Distance -0.00423 -0.85 -0.00188 -0.41 -0.00123 -0.44 -0.00397 -0.85 -0.00127 -0.27 -0.00418 -0.74 dist * log(kton) -0.00099 -1.75 -0.00062 -1.14 -0.00131 -2.53 -0.00020 -0.37 0.00024 0.78 -0.00269 -3.37 SCTG 01–05, agriculture products and fish dist * log($/ton) 0.00069 1.11 0.00058 1.03 0.00048 0.82 0.00050 0.87 -0.00024 -0.63 0.00098 1.38 Number of records: 3,280 Rho- square: 0.866 Constant 0 0 -5.72 -30.44 #N/A #N/A -3.78 -34.79 -3.15 -21.27 -6.68 -14.72 Distance -0.00821 -1.68 -0.00602 -1.33 #N/A #N/A -0.00567 -1.25 -0.00606 -1.36 -0.00707 -1.2 dist * log(kton) 0.00033 0.55 0.00041 0.73 #N/A #N/A 0.00061 1.09 -0.00177 -2.63 -0.00065 -0.85 SCTG 06–09, grains, alcohol, and tobacco products dist * log($/ton) 0.00058 1.09 0.00046 0.94 #N/A #N/A 0.00021 0.42 0.00054 1.11 0.00064 1.01 Number of records: 4,790 Rho- square: 0.820 Constant 0 0 #N/A #N/A #N/A #N/A -4.57 -27.88 -4.47 -13.98 #N/A #N/A Distance #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A dist * log(kton) 0.00067 1.28 #N/A #N/A #N/A #N/A 0.00105 2.26 0.00041 0.71 #N/A #N/A SCTG 10–14, stones, nonmetallic minerals, and metallic ores dist * log($/ton) #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A Number of records: 2,242 Rho- square: 0.879 Constant 0.00000 0 #N/A #N/A -3.17 -9.83 -3.66 -24.56 -4.45 -17.45 #N/A #N/A Distance -0.00615 -1.33 #N/A #N/A -0.00781 -2.88 -0.00653 -1.62 -0.00701 -1.63 #N/A #N/A dist * log(kton) -0.00118 -3.20 #N/A #N/A 0.00059 2.66 -0.00020 -0.66 -0.00060 -1.53 #N/A #N/A SCTG 15–19, coal and petroleum products dist * log($/ton) 0.00109 1.83 #N/A #N/A 0.00058 2.08 0.00110 2.11 0.00132 2.4 #N/A #N/A Number of records: 1,594, Rho- square: 0.706 Constant 0 0 -1.80 -12.18 #N/A #N/A -2.73 -47.49 -1.21 -24.57 -2.76 -33.54 Distance -0.00309 -1.73 -0.00685 -3.62 #N/A #N/A -0.00266 -1.59 -0.00351 -2.17 -0.00642 -2.97 dist * log(kton) -0.00105 -3.50 -0.00010 -0.35 #N/A #N/A -0.00020 -0.70 -0.00269 -9.62 -0.00250 -6.39 SCTG 20–24, pharmaceutical and chemical products dist * log($/ton) 0.00032 1.92 0.00025 1.5 #N/A #N/A 0.00001 0.09 0.00047 3.12 0.00070 3.55 Number of records: 10,302 Rho- square: 0.682 Table 3.7. Results of revealed-preference mode-choice estimation. (continued on next page)

Commodity Truck Truck and Rail Water Rail Water and Rail Air Statistics Group Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Constant 0 0 -6.40 -37.05 #N/A #N/A -4.91 -47.38 -1.51 -37.81 -6.23 -36.67 Distance -0.00886 -3.89 -0.00601 -2.81 #N/A #N/A -0.00608 -2.91 -0.01050 -5.14 -0.00931 -3.25 dist * log(kton) -0.00109 -2.47 -0.00069 -1.70 #N/A #N/A -0.00046 -1.13 -0.00171 -4.22 -0.00168 -3.12 SCTG 25–30, logs, wood products, and textile and leather dist * log($/ton) 0.00103 4.43 0.00074 3.39 #N/A #N/A 0.00065 3.07 0.00125 6.03 0.00122 4.18 Number of records: 13,689 Rho- square: 0.778 Constant 0 0 -6.95 -35.31 #N/A #N/A -4.13 -47.08 -1.36 -20.96 -4.56 -27.94 Distance -0.00822 -3.05 -0.00746 -2.99 #N/A #N/A -0.00885 -3.58 -0.00685 -2.76 -0.01070 -3.31 dist * log(kton) 0.00026 0.61 0.00060 1.50 #N/A #N/A 0.00101 2.50 -0.00196 -4.83 -0.00121 -27.94 SCTG 31–34, base metal and machinery dist * log($/ton) 0.00083 3.25 0.00080 3.37 #N/A #N/A 0.00075 3.19 0.00082 3.47 0.00122 3.99 Number of records: 10,949 Rho- square: 0.785 Constant 0 0 -5.26 -41.36 #N/A #N/A -3.95 -52.78 -0.76 -22.98 -3.09 -53.35 Distance -0.00725 -4.14 -0.00929 -5.58 #N/A #N/A -0.00917 -5.44 -0.00635 -3.98 -0.01070 -5.21 dist * log(kton) 0.00005 0.15 0.00087 2.62 #N/A #N/A 0.00081 2.41 -0.00124 -3.89 -0.00072 -1.74 SCTG 35–38, electronic, motorized vehicles, and precision instruments dist * log($/ton) 0.00063 4.33 0.00078 5.68 #N/A #N/A -0.00917 5.51 0.00064 4.86 0.00103 6.08 Number of records: 10,546 Rho- square: 0.575 Constant 0 0 -5.77 -34.26 #N/A #N/A -4.59 -34.32 -1.90 -40.64 -5.74 -29.15 Distance -0.01110 -5.56 -0.00982 -5.26 #N/A #N/A -0.01090 -5.85 -0.01370 -7.6 -0.01690 -7.08 dist * log(kton) -0.00217 -6.14 -0.00131 -3.95 #N/A #N/A -0.00116 -3.42 -0.00265 -8.22 -0.00250 -5.95 SCTG 39–43, furniture, mixed freight and misc. manufactured products dist * log($/ton) 0.00131 5.78 0.00110 5.14 #N/A #N/A 0.00110 5.16 0.00165 8.05 0.00201 7.49 Number of records: 12,940 Rho- square: 0.836 Note: #N/A means that no value is given. Table 3.7. (Continued).

constant estimates how much less useful that mode is than the truck mode. Thus for the truck/rail mode, the worst utility compared to truck mode is for agricultural products (–10.4) and the best comparison to truck mode is for pharmaceutical and chemical products (–1.8). When no value is given for a mode in a commodity group (the value is shown as #N/A), there were an insufficient num- ber of records in the RP data for values to be estimated for that mode. Thus, for the water mode, data were only sufficient to estimate coefficients for the agricultural products and the coal and petroleum products commodity groups. When no value was given for a variable within a commodity group, the RP data did not show that this variable was a significant explanatory variable for that commodity group. Thus, for the stone com- modity group, only the cross product of distance times the nat- ural logarithm of tons in thousands (ktons) was found to be a significant explanatory variable. As mentioned previously, the variables for thousands of annual tons shipped (ktons) and the value in dollars per tons were found to be significant explanatory variables for the mode-choice decision for any commodity group, but only as the natural logarithm of that variable taken as a cross product with distance. Thus, both the impact on the utility from ship- ment sizes (as shown by annual ktons) and value vary with dis- tance, but that effect decreases, varies logarithmically, as the variable increases. As before when the coefficient of the vari- able is negative, the utility increases as the variable increases, and when the value is negative, modal utility decreases as the variable increases. Although the statistical ability of the estimated model to explain the variation in mode choice was generally good, rang- ing from a Rho-square of 0.575 to 0.879, those estimates must be compared against observed mode shares. An examination of the model estimates was made where only distance varies, by setting the value in the model for ktons and dollars/ton equal to the average value for that commodity group. This allows the variation of distance, which was the most significant explana- tory variable singly and in combination with the other vari- ables, to be plotted and examined against observed mode shares in the RP data. For the agricultural products commod- ity group, Figure 3.12 shows the results of varying distance on the mode-choice estimates (shown as curves) against the observed mode shares (shown as bars). The model not only has a good statistical fit, it also appears to generally match observed mode shares. This is not the case for all commodity groups. For the stones and ores commodity group, Figure 3.13 shows the results of varying distance on the mode-choice estimates (shown as curves) against the observed mode shares (shown as bars). The model has a good statistical fit, but it does not appear to match observed mode shares. It generally also overestimates the truck mode share at large distances. As shown in Figure 3.14, the observed flows for this commodity group are most heavily represented by flows of less than 500 mi. An investigation was undertaken to see if the introduction of variables of distance by class would approve the ability of the model to estimate mode choice. For the re-estimation, the distance variable was estimated according to the following formula: if the distance is less than 500 mi, then distance1 would be equal to the distance and distance2 would be equal to 0; if the distance is greater than 500 mi, then distance1 would be equal to 500 mi and dis- tance2 would be equal to the distance minus 500 mi. Thus, for a distance of 400 mi, distance1 would take on a value of 400, and distance2 would take on a value of 0, while for a dis- tance of 1,600 mi, distance1 would take on a value of 500 mi, and distance2 would take on a value of 1,100 mi. As shown in Table 3.8, this did not significantly improve the model estimation. For two commodity groups, agricultural prod- ucts and wood products, shown as shaded rows in the table, 39 0.00 0.20 0.40 0.60 0.80 1.00 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1 0 0 0 1 1 0 0 1 2 0 0 1 3 0 0 1 4 0 0 1 5 0 0 1 6 0 0 1 7 0 0 1 8 0 0 1 9 0 0 2 0 0 0 2 2 5 0 2 5 0 0 2 7 5 0 3 0 0 0 3 2 5 0 3 5 0 0 Distance (mile) Pr ob ab ili ty truck truck/rail rail water/rail Water/Rail air Truck AirTruck/Rail Rail Figure 3.12. Mode share by distance for CFS Commodity Group 2 (estimated and observed).

the ability of the model to explain the variation in the data, as shown by the value of Rho-squared, decreased relative to those estimations with a single-distance variable. For all other commodities, the statistical fit improved only slightly, with increases in Rho-squared, between those shown in Table 3.8 and the single-distance variable estimates shown in Table 3.7, of at most 0.039. That increase was in the stone and ores commodity group. Figure 3.15 shows the results for the stones and ores com- modity group using the two distance variable estimations. The results of varying distances on the mode-choice estimates are shown as curves and the observed mode shares are shown as stacked column bars. Again, the model has a good statisti- cal fit, but it does not appear to match observed mode shares. Generally, it also overestimates the truck mode share at mid- range distances. It appears to better explain mode share below 500 mi, but as distance increases, it still does not show the expected decrease in truck mode share and increase in non- truck mode share. Because the two-distance class estimation did not produce the desired results, an additional investigation was undertaken where SCTG 14, sand and gravel, which represents most of the 40 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1 0 0 0 1 1 0 0 1 2 0 0 1 3 0 0 1 4 0 0 1 5 0 0 1 6 0 0 1 7 0 0 1 8 0 0 1 9 0 0 2 0 0 0 2 2 5 0 2 5 0 0 2 7 5 0 3 0 0 0 Distance (mile) Pr ob ab ili ty truck rail water/rail Truck Rail Water/Rail Figure 3.13. Mode share by distance for CFS Commodity Group 3 (estimated and observed). 0 5000 10000 15000 20000 25000 30000 35000 40000 1 0 0 3 0 0 5 0 0 7 0 0 9 0 0 1 1 0 0 1 3 0 0 1 5 0 0 1 7 0 0 1 9 0 0 2 2 5 0 2 7 5 0 Distance (mile) To nn ag e (kt on ) Truck Rail Water/Rail Figure 3.14. Volume by mode share by distance for CFS Commodity Group 3.

Commodity Truck Truck and Rail Water Rail Water and Rail Air Statistics Group Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Constant 0 0 -18.6 -5.92 -12.10 -3.47 -9.8 -8.60 -6.06 -9.27 -9.64 Distance <500 -0.00765 -3.74 0.011900 1.26 0.0107 1.45 0.00641 1.96 -0.00105 -0.36 0.00215 Distance >500 -0.00120 -2.13 0.00077 1.00 -0.00115 -5.37 -0.00141 2.62 -0.00060 -1.12 -0.00006 dist * log(kton) #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA SCTG 01–05, agriculture products and fish dist * log($/ton) #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA Number of records: 3,280 Rho-square: 0.848 Constant 0 0 -34.10 -4.71 #NA #NA -6.59 -13.25 -7.64 -7.32 -10.40 Distance <500 -0.01010 -7.06 0.04940 3.35 #NA #NA -0.00104 -0.58 0.00289 1.21 0.00100 Distance >500 -0.00119 -1.95 0.00004 0.06 #NA #NA 0.00024 0.33 -0.00002 -0.22 0.00093 dist * log(kton) -0.00011 -0.85 -0.00017 -1.27 #NA #NA 0.00004 0.10 -0.00012 -1.11 -0.00146 SCTG 06–09, grains, alcohol, and tobacco products dist * log($/ton) -0.00032 -0.78 -0.00019 -0.50 #NA #NA -0.00044 -3.18 -0.00266 -4.97 -0.00019 -8.11 0.73 -0.09 -5.83 0.24 0.93 -2.43 -1.13 Number of records: 4,790 Rho-square: 0.829 Constant 0 0 #NA #NA #NA #NA -7.31 -11.15 -8.26 -6.61 #N/A #N/A Distance <500 -0.01470 -4.79 #NA #NA #NA #NA -0.00550 -1.59 -0.00385 -1.08 #N/A #N/A Distance >500 -0.00200 -1.90 #NA #NA #NA #NA -0.00530 -4.28 -0.00007 -0.19 dist * log(kton) -0.00200 -1.90 #NA #NA #NA #NA -0.00530 -4.28 -0.00007 -0.19 #N/A #N/A SCTG 10–14, stones, nonmetallic minerals, and metallic ores dist * log($/ton) 0.00139 2.99 #NA #NA #NA #NA 0.00205 4.73 0.00020 1.50 #N/A #N/A Number of records: 2,242 Rho-square: 0.908 Constant 0 0 #NA #NA -4.19 -7.31 -5.87 -11.06 -7.48 -5.58 #N/A #N/A Distance <500 -0.00867 -4.44 #NA #NA -0.00744 -3.26 -0.00173 -0.74 -0.00031 -0.09 #N/A #N/A Distance >500 0.00091 1.08 #NA #NA -0.00537 -3.05 -0.00033 -0.33 -0.00015 -0.09 dist * log(kton) -0.00122 -6.01 #NA #NA 0.00046 2.81 -0.00049 -2.96 -0.00095 -3.54 #N/A #N/A SCTG 15–19, coal and petroleum products dist * log($/ton) 0.00027 1.14 #NA #NA 0.00015 0.73 0.00027 1.21 0.00047 2.37 #N/A #N/A Number of records: 1,594 Rho-square: 0.725 Constant 0 0 -8.01 -9.48 #NA #NA -5.34 -21.75 -4.12 -23.16 -6.97000 -19.90 Distance <500 -0.00750 -4.21 -0.00188 -0.78 #NA #NA -0.00161 -0.93 -0.00082 -0.49 -0.00274 -1.23 Distance >500 -0.00074 -0.39 0.00019 0.11 #NA #NA -0.00051 -0.29 -0.00172 -1.02 -0.00269 -1.2 dist * log(kton) -0.00105 -3.25 -0.00057 -1.90 #NA #NA -0.00022 -0.72 -0.00277 -9.27 -0.00212 -5.36 SCTG 20–24, pharmaceutical and chemical products dist * log($/ton) 0.00008 0.5 -0.00003 -0.2 #NA #NA -0.00020 -1.25 0.00025 1.59 0.00038 1.89 Number of records: 10,302 Rho- square: 0.717 Table 3.8. Results of revealed-preference mode-choice estimation with two distance classes. (continued on next page)

Commodity Truck Truck and Rail Water Rail Water and Rail Air Statistics Group Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Constant 0 0 -7.78 -11.97 #NA #NA -7.19 -15.13 -4.71 -23.48 -8.66 -8.58 Distance <500 -0.00659 -18.56 -0.00287 -2.00 #NA #NA -0.00158 -1.48 0.00179 3.3 -0.00005 -0.02 Distance >500 -0.00057 -7.94 -0.00024 -1.71 #NA #NA -0.00107 -8.71 -0.00006 -2.52 0.00028 1.61 dist * log(kton) -0.00057 -7.94 -0.00024 -1.71 #NA #NA -0.00107 -8.71 -0.00006 -2.52 0.00028 1.61 SCTG 25–30, logs, wood products, and textile and leather dist * log($/ton) #N/A #N/A #N/A #N/A #NA #NA #N/A #N/A #N/A #N/A #N/A #N/A Number of records: 13,689 Rho-square: 0.772 Constant 0 0 -12.90 -12.04 #NA #NA -5.48 -21.73 -5.78 -20.42 -7.42 -14.42 Distance <500 -0.01320 -4.92 0.00027 0.08 #NA #NA -0.01060 -4.22 -0.00204 -0.8 -0.00911 -2.73 Distance >500 -0.00771 -2.84 -0.00726 -2.89 #NA #NA -0.00876 -3.52 -0.00685 -2.74 -0.01060 -3.23 dist * log(kton) 0.00046 1.06 0.00078 1.96 #NA #NA 0.00120 2.99 -0.00165 -4.09 -0.00094 -1.66 SCTG 31–34, base metal and machinery dist * log($/ton) 0.00075 2.93 0.00072 3.08 #NA #NA 0.00068 2.92 0.00075 3.18 0.00113 3.69 Number of records: 10,949 Rho- square: 0.792 Constant 0 0 -13.70 -13.89 #NA #NA -7.86 -17.37 -3.72 -32.45 -6.29 -32.60 Distance <500 -0.01260 -7.51 0.00348 1.34 #NA #NA -0.00559 -3.00 -0.00439 -2.83 -0.00811 -4.04 Distance >500 -0.00561 -3.22 -0.00835 -5.01 #NA #NA -0.00825 -4.88 -0.00541 -3.4 -0.00938 -4.57 dist * log(kton) 0.00013 0.42 0.00093 3.13 #NA #NA 0.00087 2.91 -0.00114 -3.98 -0.00063 -1.69 SCTG 35–38, electronic , motorized vehicles, and precision instruments dist * log($/ton) 0.00049 3.42 0.00066 4.86 #NA #NA 0.00064 4.69 0.00051 3.95 0.00087 5.17 Number of records: 10,546 Rho- square: 0.597 Constant 0 0 -8.86 -8.16 #NA #NA -5.11 -14.94 -5.31 -26.91 -6.86 -15.24 Distance <500 -0.01390 -7.19 -0.00549 -1.88 #NA #NA -0.01220 -6.08 -0.00865 -4.73 -0.01700 -6.84 Distance >500 -0.01090 -5.62 -0.00991 -5.49 #NA #NA -0.01080 -6.06 -0.01390 -8.07 -0.01700 -7.34 dist * log(kton) -0.00185 -5.31 -0.00100 -3.04 #NA #NA -0.00085 -2.50 -0.00226 -7.08 -0.00209 -5.04 SCTG 39–43, furniture, mixed freight and misc. manufactured products dist * log($/ton) -0.00185 -5.31 -0.00100 -3.04 #NA #NA -0.00085 -2.50 -0.00226 -7.08 -0.00209 -5.04 Number of records: 12,940 Rho-square: 0.841 Note: #N/A means that no value is given. Table 3.8. (Continued).

short-distance flows within that commodity group, was sepa- rated for estimation and new estimates were made for the remaining commodities in that group. Those results are shown in Table 3.9. Although the model’s ability to explain the variance in mode choice is very high for SCTG 14, sand and gravel, when treated separately, as can be seen in Figure 3.16, that is due almost entirely to the fact that the share by modes other than truck is extremely limited for this commodity at all distance ranges. Additionally, with fewer records available for the estimation, the model cannot successfully estimate coeffi- cients for the distance across products with shipment size and value, used as logarithms of total tons and total value per ton. Those few distance ranges where other modes are observed to be used do not fall into a pattern. The estimated mode share for distances greater than 500 mi is relatively constant above 1,500 mi. Mode choice for this commodity must be assumed to be largely related to modal availability or pro- cesses unique to the production and/or consumption of this commodity. For the remaining flows in this commodity group, the exclu- sion of the records for SCTG 14 results in a poorer model esti- mation, as indicated by a decline in the Rho-square compared to that in Table 3.8. Additionally with fewer records available, this estimation for the remaining commodities in this group cannot successfully develop coefficients for the distance across products with shipment size and value, as logarithms of total tons and total value per ton. As shown in Figure 3.17, for the distances below 500 miles, the observed flows do appear to cor- respond to the estimated mode share, but again this may be largely due to the dominance of truck mode share over this dis- tance range. At distances greater than 500 miles, the share of other modes used do not fall into a discernible pattern with distance and the estimated mode share for distances greater than 500 miles is relatively constant above 1,500 miles. Mode choice for the remainder of this commodity group must be assumed to be largely related to modal availability, or pro- cesses unique to the production and/or consumption of this commodity. The research has shown that it is possible to develop RP databases from existing, publicly available sources. It has shown that modal distance, which is expected to be highly cor- related with modal time and costs, is the single most important consideration in mode choice. When local policy decisions can only impact the local component of modal costs and distances, and those local costs and times are only a small fraction of total modal costs and distances, the difficulty of influencing mode- choice decisions by local policies can be seen. The policy deci- sions that might be more subject to local control, such as shipper, and receiver characteristics, were not found to be significant variables in freight mode-choice decisions, at least as estimated by this RP survey. 43 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1 0 0 0 1 1 0 0 1 2 0 0 1 3 0 0 1 4 0 0 1 5 0 0 1 6 0 0 1 7 0 0 1 8 0 0 1 9 0 0 2 0 0 0 2 2 5 0 2 5 0 0 2 7 5 0 3 0 0 0 Distance (mile) 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 Pr ob ab ili ty truck rail water/rail Truck Rail Water/Rail Figure 3.15. Mode share by distance for CFS Commodity Group 3 (estimated with two distance classes and observed).

Commodity Group Truck Truck and Rail Water Rail Water and Rail Air Statistics Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat Coeff t-stat t-stat Constant 0 0 #NA #NA #NA #NA -151.00 -4.36 -9.76 -4.04 #NA #NA Distance <500 -0.00945 -1.26 #NA #NA #NA #NA 0.28700 3.82 0.00122 0.20 #NA #NA Distance >500 -0.00174 -1.31 #NA #NA #NA #NA -0.00224 -1.42 0.00023 0.53 #NA #NA dist * log(kton) #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA SCTG 14, sand and gravel dist * log($/ton) #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA Number of Records: 3,280 Rho-square: 0.954 Constant 0 0 #NA #NA #NA #NA -7.21 -10.66 -8.33 -6.01 #NA #NA Distance <500 -0.01210 -3.81 #NA #NA #NA #NA -0.00152 -0.42 -0.00194 -0.49 #NA #NA Distance >500 -0.00110 -0.95 #NA #NA #NA #NA -0.00132 -1.26 -0.00005 -0.13 #NA #NA dist * log(kton) #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA SCTG 10–13, stone, nonmetallic and metallic ore dist * log ($/ton) #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA #NA Number of Records: 4,790 Rho-square: 0.884 Table 3.9. Results of revealed-preference mode-choice estimation with two distance classes (STCG 14 and remainder of stone and ore commodity group).

45 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1 0 0 0 1 1 0 0 1 2 0 0 1 3 0 0 1 4 0 0 1 5 0 0 1 6 0 0 1 7 0 0 1 8 0 0 1 9 0 0 2 0 0 0 2 2 5 0 2 5 0 0 Distance (mile) 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 Pr ob ab ili ty truck rail water/rail Truck Rail Water/Rail 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1 0 0 0 1 1 0 0 1 2 0 0 1 3 0 0 1 4 0 0 1 5 0 0 1 6 0 0 1 7 0 0 1 8 0 0 1 9 0 0 2 0 0 0 2 2 5 0 2 5 0 0 2 7 5 0 3 0 0 0 Distance (mile) 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 Pr ob ab ili ty truck rail water/rail Truck Rail Water/Rail Figure 3.16. Mode share by distance for CFS Commodity Group 3, SCTG 14 (estimated with two distance classes and observed). Figure 3.17. Mode share by distance for CFS Commodity Group 3, SCTG 10-13 (estimated with two distance classes and observed).