Using Commodity Flow Survey Microdata and Other Establishment Data to Estimate the Generation of Freight, Freight Trips, and Service Trips: Guidebook (2016)

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.

1 This summary encapsulates work done under NCFRP Project 25, “Freight Trip Generation and Land Use” (jointly funded as NCHRP Project 08-80), and NCFRP Project 25(01) “Estimating Freight Generation Using Commodity Flow Survey Microdata.” Throughout this guidebook, “Phase 1” refers to NCFRP Project 25 and “Phase 2” refers to NCFRP Project 25(01). Phase 1 was reported in the joint NCHRP Report 739/NCFRP Report 19: Freight Trip Generation and Land Use (Holguín-Veras et al. 2012). Since then, the work has expanded significantly. A comprehensive account of the entire project has been captured in this guidebook. The main objective of NCFRP Project 25 and NCFRP Project 25(01) was to study the relationship between freight trip generation (FTG) and land use “. . . to develop a handbook that provides improved freight trip generation rates, or equivalent metrics, for different land use characteristics related to freight facilities and commercial operations to better inform state and local decision-making.” To achieve this objective the research team: • used the Commodity Flow Survey (CFS) to estimate 1,409 (342 linear and 1,067 non-linear) freight production (FP) models for the entire United States, and separate models for the states of New York, California, Ohio, Texas, and Wyoming for 37 different industry sectors; • collected additional establishment-level data in the New York City metropolitan area and in New York State’s Capital Region—pooling funds with a grant from SHRP2 Project C-20, “Freight Demand Modeling and Data Improvement,” to the Capital District Transportation Committee (CDTC) serving the Albany, New York region—about FTG (both production and attraction); • added these data to the databases already in the team’s possession and used these data to estimate 62 freight trip production models (both linear and non-linear, 31 models each) and 70 freight trip attraction (FTA) models (both linear and non-linear, 35 models each) for 12 industry sectors; • collected data about service trip attraction (STA) using pooled funds with the CDTC SHRP2 Project C-20 grant, and used the data to estimate 118 STA models (both linear and non-linear, 59 models each)—the first reported in the literature—for 21 industry sectors (budget constraints prevented collection of data about service trip production [STP]); • collected additional establishment-level data, again pooling funds with the CDTC SHRP2 Project C-20 grant, about freight generation (FG), both production and attrac- tion, and used these data to estimate 49 FP (19 linear and 30 non-linear models) and S u m m a r y Using Commodity Flow Survey Microdata and Other Establishment Data to Estimate the Generation of Freight, Freight Trips, and Service Trips: Guidebook

2 using Commodity Flow Survey microdata and Other Establishment Data to Estimate the Generation of Freight, Freight Trips, and Service Trips 50 freight attraction (FA) models (19 linear and 31 non-linear models) for 11 industry sectors; • collected additional establishment-level data jointly with the CDTC SHRP2 Project C20 about the relationship between FG and FTG, and used these data to estimate 18 models exploring the relationship between FP and freight trip production (FTP) (8 linear and 10 non-linear models), and 20 models exploring the relationship between FA and FTA (9 linear and 11 non-linear models) for 12 industry sectors; • analyzed data and incorporated a set of employment rate models based on a national sample of FTG from 2008 that had been collected as part of a study to assess the impacts of congestion across the United States, with data about shipments received and sent out from 1,000 establishments (see Chapter 8); • consolidated the FTG models in the literature in a database to assist practitioners interested in using the models (http://transp.rpi.edu/~NCFRP25/FTG-Database.rar); and • developed the Freight Trip Generation Estimator Software at the ZIP code and 2-digit North American Industry Classification System (NAICS) code levels for public use (https://coe-sufs.org/wordpress/ncfrp33/appendix/ftg/). The research identified three principles that are central to the development of models that will be able to inform transportation planning and traffic impact analyses. The most important principle is the need to distinguish between FTG (i.e., the generation of vehicle trips), and FG (i.e., the generation of the cargo that is transported by the vehicle trips). FG is an expression of economic activity performed at a business establishment by which input materials are processed and transformed generating an output that, in most cases, is transported elsewhere for further processing, storage, distribution, or consumption. FTG, in contrast, is the result of the logistics decisions concerning how best to transport the FG in terms of shipment size, frequency of deliveries, and vehicle/mode used. The shipper is able to change shipment size to minimize total logistic costs by transporting more cargo (i.e., FG) without proportionally increasing the corresponding number of trips (i.e., FTG). Therefore, FTG cannot be generally assumed to be proportional to business size because large estab- lishments could receive larger amounts of cargo without concomitant increases in FTG. This has major implications for FTG modeling, as current modeling practices implicitly assume proportionality between FTG and such business size variables as square footage and employment. The second key principle is the need to account for service trips, which have been over- looked as a component of commercial vehicle activity. These service trips are generated by technicians, service providers, and the like, who visit an establishment to perform various services. Service trip generation (STG) is the number of service trips generated by a com- mercial establishment. The STG is made up of service trip production (STP), which is the number of vehicle trips leaving the establishment to perform services at other locations. The counterpart of STP is service trip attraction (STA), which is the number of vehicle trips arriving at the establishment to perform a service activity. Different metrics could be used to measure the transportation activity generated at a given establishment. To simplify the exposition, the term freight and service activity (FSA) refers to all activities related to freight and service. The term metric of freight and service (metric of FSA) is used to designate all potential ways to measure the transportation activity generated by the FSA (i.e., FG/FTG/STG).

Summary 3 The third and final principle is that the accuracy of FG/FTG/STG analyses depends on the following factors: • the ability of the classification system used in the analyses to group commercial establish- ments in a set of internally homogeneous classes; • the ability of the measure of business size used to predict the FG/FTG/STG; • the ability of the statistical technique used to capture the underlying relations that shape FG/FTG/STG; and • the use of disaggregate (establishment-level) models in conjunction with the corresponding aggregation procedure to estimate aggregate values (if needed). To ensure proper understanding and use of the terms, brief descriptions are provided. A classification system is a systematic way to group establishments into pre-defined groupings or classes (e.g., residential, commercial, and industrial). A measure of business size is the inde- pendent variable used to predict FG/FTG/STG, such as square footage of the establishment or total number of employees. The statistical technique is the process used to compute the parameters of the models. Among the widely ranging approaches available, two techniques were found to be particularly useful: (1) ordinary least squares (OLS) regression analysis, and (2) multiple classification analysis. The aggregation procedure is the technique used to obtain aggregate values of FG/FTG/STG from the establishment-level estimates produced by a disaggregate model. This routinely overlooked aspect is at the core of many of the problems reported by practitioners when producing FG/FTG/STG forecasts. As these factors form the backbone of the modeling effort, it is important to discuss their implications: • Estimate/use FG/FTG/STG models with classification systems with homogenous classes. The estimation and use of FG/FTG/STG models works best if the commercial establishments are grouped in classes that are as internally homogeneous as possible. In this case, the variables that measure business size have a better chance of being good predictors of FG/FTG/STG. If the classes group together very different economic activi- ties (i.e., the data for a given class will be very heterogeneous), the ability of business size variables to be a good predictor of FG/FTG/STG will be compromised. Industrial clas- sification systems—such as the Standard Industrial Codes (SIC) and the North American Industry Classification System (NAICS)—are designed to group together similar economic activities, maximizing the internal homogeneity of each class. By construction, these classification systems are better able to support appropriate modeling of FG/FTG/STG and therefore offer the best alternative for FG/FTG/STG modeling. The concern with using land use classification systems in FG/FTG/STG modeling is that they tend to use very aggregate land use classes (e.g., commercial, industrial) that group together disparate sets of economic activities, which undermines the ability of business size to be a good predictor of FG/FTG/STG. An exception is the Land-Based Classification Standards (LBCS), which classifies land use using five dimensions: (1) the activity (taking place at the establishment), (2) the function (type of enterprise being served), (3) structure type (building characteristics), (4) site development character (physical description of the land), and (5) ownership (e.g., public or private). If the activity dimension contains classes that are defined using an industrial classification system, the resulting classes would be as good as using SIC or NAICS in FG/FTG/STG modeling. • Use variables that correctly measure the intensity of FSA as predictors of FG/FTG/ STG. Variables such as square footage and employment differ significantly in their ability to be good predictors of FG/FTG/STG. As an example, three establishments of exactly the same square footage may produce different amounts of FG/FTG/STG depending on

4 using Commodity Flow Survey microdata and Other Establishment Data to Estimate the Generation of Freight, Freight Trips, and Service Trips the intensity and type of the economic activity being performed; and whether or not the establishments are empty, lightly used, or very heavily used. In contrast, variables such as employment are likely to be better explanatory variables because they rise and fall in concert with the level of economic activity. As a result, employment is a better predictor of FG/FTG/STG. • Use statistical techniques that can capture the underlying relations that shape FG/FTG/STG. The ability of shippers and service providers to consolidate cargo and service activities using the same vehicle trip leads to a situation in which FTG cannot be generally assumed to depend on business size. Thus, it is important to statistically determine if a business size variable is a statistically acceptable predictor of FG/FTG/STG. For this reason, the team recommends the use of statistical procedures—such as ordi- nary least squares (OLS)—and spatial econometric techniques that test the significance of independent variables. These techniques provide a solid way to assess the role played by independent variables. However, they impose a functional form to the model. Multiple classification analysis (MCA) does not impose a functional form. This characteristic enables the MCA parameters to freely change, making MCA a very useful technique. This feature also helps explain why MCA was frequently found to produce models that had the best agreement with the data. That said, MCA should only be used in cases for which other techniques have already determined that the measure of business size plays a statistically significant role (MCA does not support hypothesis testing). • Use establishment-level models with the corresponding aggregation procedure. As in passenger-demand modeling, disaggregate (establishment-level) modeling is the recommended approach when producing FG/FTG/STG forecasts. Disaggregate models are better able to capture the interconnection between FG/FTG/STG and the independent variables. However, disaggregate models do require the use of a suitable aggregation procedure. The research conclusively showed that not using the correct aggregation proce- dure leads to significant errors in the estimation of FSA. Most notably, the research revealed that the widely used process of obtaining aggregate estimates of FG/FTG/STG by multiplying total employment by a FG/FTG/STG rate per employment is only valid in the minority of cases for which the underlying model is one in which FTG is directly proportional to employ- ment. Not following the recommended procedures will lead to estimation errors. The three central principles identified in the research were tested using statistical estimation techniques. To this effect, the research team used FG/FTG data from: • three surveys conducted by the team that collected data from about 1,100 receivers and from more than 300 carriers; • a furniture store chain in the Midwest and supermarkets in the Puget Sound region and New York City; and • the 2007 CFS data. In cases for which the data were most complete, the team had access to establishment- level variables such as employment, location, size, revenue, line of business, some trip data (e.g., number of truck trips per day/week, shipment sizes), and land use information. Using the data, the research team estimated and assessed the performance of FG/FTG/STG models based on: • classification systems; • statistical technique used; • aggregation procedure used to obtain aggregate values; and • business size variables used as predictors of FG/FTG.

Summary 5 Both industrial classification systems (i.e., SIC and NAICS) and land use classification systems (i.e., LBCS, and New York City Zoning Resolution [NYCZR]) were tested. OLS, spatial econometric techniques, and multiple classification analyses also were tested, as were the standard aggregation used in practice and the aggregation procedures developed by the team. Finally, square footage and employment (where available as a business size variable) were tested. The comprehensive analyses conducted by the team led to the following insights: • Industrial classification systems should be used for FG/FTG/STG modeling instead of standard land use classification systems. The research revealed that using indus- trial classification systems as the foundation for the estimation of FG/FTG models is significantly better than using standard land use classification systems such as the NYCZR, or land use classification systems that can be applied nationally, such as LBCS. The industrial classification systems tended to produce models that were statistically stronger than those obtained using any of the land use classification systems. The best results were found when an economic measure of business size (i.e., employment), was used in combination with an industrial classification system (i.e., 2-digit SIC codes or 3-digit NAICS codes). The team would expect that using LBCS will produce better models than using the standard land use classification systems (such as NYCZR), particularly if the activity codes in the LBCS use industrial classification systems (e.g., SIC, NAICS) (Holguín-Veras et al. 2012). • Proportionality between FTG/STG and business size happens only in a minority of industry segments. The research revealed that in 51% of industry segments, the FTG is constant and does not depend on business size, indicated by employment; in 31% of cases, the FTG model is a function of a constant and a rate that multiplies the establish- ment’s employment; and in the remaining 18% of cases, the FTG model is proportional to employment at a constant FTG rate (Holguín-Veras et al. 2011). Similar results were found for STA: in 66% of the industry sectors, the STA is constant; in approximately 28% of sectors, the STA is proportional to employment; and in only 6% of the models, the STA reflects a combination of a constant and an FTG rate per employee. The most commonly used approach (the constant trip rate per employee) is correct in only a minority of cases, which should be a concern. • The models estimated at the establishment level are transferable, though more testing is needed to reach solid conclusions. The models estimated with New York City data were applied to supermarkets in the Seattle region. The models produced very good estimates of FTG. This result is very encouraging, though larger-scale testing is needed to reach definitive conclusions. (Holguín-Veras et al. 2013a). • The NCFRP Project 25 models generally outperform the models previously reported in the literature. The NCFRP Project 25 models were compared to models in the ITE’s Trip Generation Handbook: An ITE Recommended Practice (ITE 2004) and Trip Generation: An ITE Informational Report (ITE 2008), and the FHWA’s Quick Response Freight Manual (QRFM) and QRFM II (Cambridge Systematics, Inc., 1996, updated 2007). The results show that the NCFRP Project 25 models produce more accurate FTG estimates than the corresponding ITE and QRFM models (Holguín-Veras et al. 2013a). • MCA performed better than OLS models. For those industries with FG/FTG/STG dependent on employment, the research found that MCA performed better than OLS. This was the case for both industrial and land use classification-based models. Because MCA does not impose a functional form, the parameters can take values that are not restricted by a function (as in OLS). This flexibility increases the ability of MCA to replicate the input data (Lawson et al. 2012).

6 using Commodity Flow Survey microdata and Other Establishment Data to Estimate the Generation of Freight, Freight Trips, and Service Trips • The CFS can be efficiently used to estimate FP models. The use of the CFS in combi- nation with complementary datasets provides an efficient way to estimate FP models for the entire nation at various levels of geography. The successful use of the CFS microdata has tremendous implications because the models released—which do not contain com- mercially sensitive information—can be used to infer FP patterns in between the years that CFS data are collected. The use of the FP models in conjunction with publicly avail- able data could enable state departments of transportation (DOTs) and metro politan planning organizations (MPOs) to estimate the amount of cargo being produced at their jurisdictions at the ZIP code level (using the U.S. Census Bureau’s ZIP Code Business Patterns database) and lower (if the data are available). In turn, these estimates will enable planners to monitor changes in cargo flows and implement policies and programs if needed. This approach is significantly better than waiting for the next iteration of the CFS. • Statistically significant differences in FP patterns occur across the states. Although the team could not conduct a comprehensive examination of the effects of geography on FP, the limited research used strongly suggests that FP patterns vary from region to region, and frequently from state to state. Thus, using national models, or models from a nearby state, may lead to errors in the estimation of FP. Future research could tackle a compre- hensive estimation of FP models to identify state-level differences and commonalities. This research could have important impacts on future data collection efforts. States with similar FP patterns could pool funds to collect similar data, which will enable them to estimate freight activity very cost effectively. • Non-linear models typically provide the best representation of FP patterns. This is likely a consequence of scale economies of production, in which the larger the establish- ment, the higher the productivity for a unit of labor. About 76% of the FP models are non-linear. This finding has important implications for modeling because it implies that the standard aggregation procedures—which simply assume that FP is directly proportional to employment—are incorrect. More sophisticated aggregation procedures must be used. Of all the non-linear models estimated (linear-logarithmic, logarithmic-logarithmic, and logarithmic-linear), logarithmic-logarithmic models were found to provide the best agreement to the data. • Service trips must be accounted for. The STA models estimated in the project indicate that most establishments receive between 1–5 trips per week. The largest attractors of service trips on a per-establishment basis, according to the admittedly limited amount of data collected, are the service industries, such as Information (NAICS 51), Finance and Insurance (NAICS 52), Education Services (NAICS 61), Health Care and Social Assistance (NAICS 62), Entertainment (NAICS 71), and Other Services except Public Administration (NAICS 81). These establishments represent approximately 55% of the establishments and 51% of the employment in metropolitan and micropolitan areas in the country, so they create a very large amount of service trips. Moreover, service vehicles occupy the curb for extended periods of times (sometimes in the range of hours). As a result, they tend to control a disproportionate portion of the number of spaces allocated to commercial vehicles (both freight and service), making it difficult for freight vehicles to find suitable parking. The work conducted as part of NCFRP Project 25 has set the empirical and theoreti- cal foundation for the modern study of FG/FTG/STG. On the empirical front, the work has brought to bear more FG/FTG/STG data than any previous effort. These data include establishment-level FTG (both production and attraction) data, the CFS FP data, and newly collected STA data. These data have led to the estimation of more than 1,700 FG/FTG/ STG models covering dozens of industry sectors. On the theoretical front, the research has

Summary 7 made a strong case for a redirection of FG/FTG/STG research and practice that emphasizes the economic roots of these activities and leads to more accurate models that can be seam- lessly used in combination with publicly available employment data. The incorporation of these recommendations on traffic impact analyses, traffic engineering, medium/long term demand forecasting exercises, and land use planning will lead to an enhanced understanding of the role of freight and service activity (FSA) in metropolitan areas, and the transportation needs of these sectors. Policy makers need the more-complete picture of FSA that these models provide to enact policy and programs that will help these sectors fulfill their eco- nomic roles, while producing the minimal amount of negative externalities that impact the economy, the environment, and local communities.