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.
48 A review of the literature and current practices in freight modeling (see Appendices D and E) revealed several informa- tion gaps. As discussed in Chapter 3, these gaps relate mainly to emerging land uses where there is a lack of modeling approaches that can effectively reproduce the characteristics of FTG. The objective of this chapter is to fill those information gaps by assessing the efficacy of different analysis techniques and land use classifications through a set of case studies. In addition, the procedure presented in this report can become the benchmark for future FTG studies. The case studies include the following: â¢ An establishment-based dataset with 76 furniture stores in Midwestern States, already with basic land use informa- tion and company characteristics (the company asked that its name not be divulged). â¢ An establishment dataset with about 400 completed ques- tionnaires of receivers of cargo in Manhattan and Brook- lyn containing information about deliveries and company characteristics. â¢ An establishment dataset with about 400 completed ques- tionnaires of private and common carriers of cargo in Northern New Jersey and New York containing freight trip information and company characteristics. â¢ Comparison of deliveries for a number of grocery stores in Manhattan (NY) and eight supermarkets in the Puget Sound region. These datasets include economic information about the busi- nesses and facilities, their locations, size, revenues, industries and lines of business, and trip data (e.g., number of truck trips per day/week, shipment sizes, frequencies, empty trucks, type of trucks, hours of operations, and in some cases, truck origins and destinations). In addition, the datasets are com- plemented with land use information. Description of the Datasets This section includes a description of the various datasets. Midwestern States Furniture Chain Dataset The data for furniture Chain A (the companyâs real name not divulged) contain the information of 76 stores in 18 states in the Midwestern and Eastern parts of the United States. Figure 9 shows the number of stores by State: Illinois has 17 stores; followed by Ohio (16 stores); Michigan (8 stores); and Indiana (7 stores). The team studied and analyzed data for Illinois, Ohio, Michigan, and Indiana as four individ- ual case studies; the North East-Mid Atlantic states (Con- necticut, Delaware, Maryland, Massachusetts, New York, Pennsylvania, Virginia, and West Virginia) and the rest of the Midwest states (Iowa, Kentucky, Minnesota, Missouri, Wisconsin, and Nebraska) were studied as separate case stud- ies. This produces a total of six individual case studies. Each dataset includes the number of deliveries to the stores; the number of pallets per delivery; store addresses; store loca- tion characteristics (i.e., off-mall-base and mall-base stores); and store types (i.e., combo, conventional, and outlet stores). Combo type refers to stores that sell both conventional and childrenâs furniture. On the other hand, outlet type refers to stores that handle returned and outdated furniture. As addi- tional information, the distribution center manager revealed that the company is shifting the stores from the malls to off- mall locations. These new stores are mostly the combo type. Stores receive one or two weekly deliveries throughout the year. For this chain, the most important information for tracking the performance of stores is the number of pallets delivered per week. A pallet is the basic unit of measure used for recording and planning the volume of shipments, the approximate dollar value of each shipment, and the num- ber of trucks required for a delivery. According to the inter- viewee, the average dollar value of furniture per pallet is $515. Up to 28 pallets can be shipped in a 53-foot container. Most trucks leave the distribution center when they are almost full. Thus, it is possible to approximate the number of truck trips originating from the distribution center by dividing the total number of pallets by 28. However, with this data, it is not C h a p t e r 6 Case Studies
49 possible to approximate the number of truck trips to each store, since one routing covers 2 to 3 stores. For example, the maximum number of pallets that a store receives is 18, which is below the maximum capacity of a 53-foot container. Instead, the weekly pallet information can be used to estimate its relationship with the store-related characteristics. Since the information on the number of employees, sales volume, and size of individual stores was difficult to obtain from the survey, it was purchased from InfoUSA. However, the correlation plot reveals that employees and sales volume are highly, though not perfectly, correlated. Because it is not certain whether the correlation is real (the most plausible case is that the company allocates the number of employees at each store based on sales) or artificial (e.g., InfoUSAÂ® estimates one of them based on the other), caution is essential when using this data. To avoid multicollinearity, only one of the variables is used to estimate the regression models. As sales is not a use- ful variable for planning purposes, the number of employ- ees was chosen as the independent variable for the regression models. However, employment data was available for 58 of the 76 stores of the dataset. The regression models considered only the observations that have available employment data. Figure 9 shows the number of stores by state. New York City (NYC) Carriers and Receivers Dataset As part of a project conducted for the New York State Department of Transportation (NYSDOT), disaggregated data was collected at the establishment level through two surveys targeting carriers and receivers. The questionnaire inquired about company attributes and operational and FTG patterns, in addition to how participants would react to differ- ent scenarios concerning off-hour deliveries. To develop the data collection plan, records were purchased from the Dun and Bradstreet (D&B) database for Manhattan and Brooklyn. Taking into consideration the area of study, and shipping and receiving patterns, companies were randomly selected from the purchased sample. Companies were selected for each of the SIC in the D&B database. In the random selection, more weight was placed on businesses prone to the transport of commodities (focus of this study) and less on the ones on service related industries. The receiver sample was selected from the list of receivers in Manhattan with more than five employees. For the Manhat- tan carriersâ case, companies were selected from two groups: for-hire carriers (those that provide services to the open mar- ket); and private carriers (those that provide transportation service to a parent/related company). The selected carriers had at least 25 employees and were based in some counties of New York and New Jersey. For the Brooklyn case, the sample plan considered Brook- lyn receivers/intermediaries and carriers from Brooklyn and New Jersey. Pure receivers only receive goods, while inter- mediaries both ship and receive goods. A filter was used con- sidering companies with more than five employees. After designing the data collection plan, the surveys were sent to the Eagleton Institute of Rutgers University to obtain the data, using computer-aided telephone interviews (CATI). A complete description of the data collection plan is found in the Project: Potential for Off-Peak Freight Deliveries to Com- mercial Areas website1 and final report2. The data collection process resulted in a sample for the Manhattan and Brooklyn receivers comprised of 362 com- plete observations. Table 27 shows the breakdown of their 1 1 17 7 1 3 1 1 8 4 3 1 2 16 5 1 1 3 0 2 4 6 8 10 12 14 16 18 N um be r of S to re s State Number of Stores 18 States 76 Stores Figure 9. Number of stores by states. 1http://www.rpi.edu/~holguj2/OPD/index.html 2HolguÃn-Veras, J. (2006). Potential for Off-Peak Freight Deliveries to Congested Urban Areas Rensselaer Polytechnic Institute. http://www.rpi.edu/~holguj2/OPD/ OPD_FINAL_REPORT_12-18-06.pdf
50 primary industry types. As highlighted, about a quarter of the companies are in the food related sector, and another third are in the wholesale durable and non-durable goods trade. Other sectors include: retail, construction, apparel and acces- sory stores, and furniture and building materials. Additional information available in the receivers sample consists of: number of deliveries received, type of facility, employment and commodities received, among others. The dataset was complemented with Dun and Bradstreet information. In terms of business size, most of the establishments in the sample were small- to medium-sized, with about 5 to 50 employees (80%) (see Table 87 in Appendix I). This is con- sistent with the overall breakdown for establishments located in Manhattan and Brooklyn, where more than 90% are in this same range of employment (U.S. Census Bureau 2010b). The data collection process resulted in a sample for New York and New Jersey carriers that consisted of 339 complete observations. Table 1 shows the breakdown of the sample by industry sector. As shown in Table 28, close to 45% of the sample was rep- resented by motor freight transportation and warehousing, and approximately another 40% was represented by whole- sale trade, both durable and non-durable goods. Round- ing out the top six sectors were food related establishments (4%); other transportations services (3%); and the construc- tion sector (3%). All other sectors represented in the sample each accounted for 1% or less. As with the receivers sample, data on number of trips, type of facility, employment and commodities transported, among other data were collected for the carriers sample. In terms of business size measured by number of employees, most of the sample is small- to medium-sized. Establishments with less than 50 employees account for almost 75% of the carriersâ sample (see Table 88 in Appendix I). The NAICS, adopted in 1997 to replace the SIC system, is the standard used by Federal statistical agencies in classify- ing businesses. Accordingly, the sample was reclassified by the NAICS. The NAICS is a more disaggregated system than SIC SIC description Number of establish- ments % of establish- ments 58, 54, 20 Eating and Drinking Places, Food Stores, Food and Kindred Products 88 24.31% 51 Wholesale Trade: Nondurable Goods 58 16.02% 50 Wholesale Trade: Durable Goods 58 16.02% 59 Miscellaneous Retail 46 12.71% 17 Construction-Special Trade Contractors 18 4.97% 56 Apparel and Accessory Stores 15 4.14% 57 Home Furniture, Furnishings, and Equipment Stores 13 3.59% 52 Building Materials, Hardware, Garden Supply, and Mobile Home Dealers 10 2.76% 23 Apparel and Other Finished Products Made From Fabrics and Similar Material 8 2.21% 15 Building Construction-General Contractors And Operative Builders 7 1.93% 34 Fabricated Metal Products, Except Machinery and Transportation 5 1.38% 25 Furniture and Fixtures 6 1.66% 39 Miscellaneous Manufacturing Industries 5 1.38% 22 Textile Mill Products 4 1.10% 24 Lumber and Wood Products, Except Furniture 5 1.38% 27 Printing, Publishing, and Allied Industries 2 0.55% 26 Paper and Allied Products 2 0.55% 55 Automotive Dealers and Gasoline Service Stations 2 0.55% 94 Administration of Human Resource Programs 2 0.55% 32 Stone, Clay, Glass, and Concrete Products 1 0.28% 16 Heavy Construction Other Than Building Construction-Contractors 1 0.28% 30 Rubber and Miscellaneous Plastics Products 1 0.28% 35 Industrial and Commercial Machinery and Computer Equipment 1 0.28% 36 Electronic and Other Electrical Equipment and Components, Except Computer 1 0.28% 38 Measuring, Analyzing And Controlling Instruments; Photographic, Medical 1 0.28% 74 Business Services 1 0.28% 96 Administration of Economic Programs 1 0.28% Grand Total 362 100.00% Table 27. Breakdown of receivers by SIC, NYC.
51 4The Tax Parcel Attributes Table also contained zoning designations. This designation was compared with the zoning polygons and found sixteen incon- sistencies; upon further investigation it was found that they do not influence the outcome of the models. SIC, but when comparing both systems at the 2-digit level, NAICS is more aggregated. Furthermore, a comparison of the aggregated industries (see Table 89 in Appendix I), reveals that some 2-digit aggregated SIC industries correspond to more than one 2-digit NAICS. In the receiversâ sample, for instance, some establishments in the food stores (SIC 54) industry match to manufacturing (NAICS 31) industry, and others to retail trade (NAICS 44). This has major implica- tions for modeling and planning efforts, as resulting FTG patterns for an industry identified using one industry clas- sification system cannot be generalized for the other. In addition, the dataset was geo-coded using addresses of the establishments to include the land use information. The geo-coded firm location was spatially joined to zoning poly- gons (the reference file can be found in NYC Department of City Planning website3) that contain the land use codes designation used by NYC. The team obtained the Tax-Lot Polygon Feature Class of the Department of Financeâs Digital Tax Map that was merged with the PLUTOâ¢ attribute data for 2006, defined as the attribute table of property informa- tion associated with each tax lot4. The spatially joined data was exported into Excel for analysis; this file was then con- verted to comma separated values (csv) format and exported to econometric software, where the complete dataset was used for FTG research. After studying the dataset, it was found that the establish- ments are located in 81 different land use categories, as defined in the City of New York Zoning Resolution (NYCZR). These classes were reorganized into the groups shown in Table 29. These groups are related to commercial, manufacture, and residential land uses. Table 29 shows the distribution of establishments by land use; more than a quarter of the establishments are in the M-1 Light Manufacturing district that typically includes wholesale trade. A quarter of the establishments are located in districts zoned Residential. These establishments provide retail and other services to neighborhoods, thus playing an important role for estimating FTG. (A detailed description of land uses can SIC SIC description Number of establish- ments % of establish- ments 42 Motor Freight Transportation and Warehousing 150 44.25% 51 Wholesale Trade: Nondurable Goods 65 19.17% 50 Wholesale Trade: Durable Goods 65 19.17% 58, 20 Eating and Drinking Places, Food Stores, Food and Kindred Products 12 3.54% 47 Transportation Services 9 2.65% 17 Construction-Special Trade Contractors 8 2.36% 59 Miscellaneous Retail 3 0.88% 30 Rubber and Miscellaneous Plastics Products 3 0.88% 34 Fabricated Metal Products, Except Machinery and Transportation 3 0.88% 26 Paper and Allied Products 3 0.88% 73 Business Services 3 0.88% 56 Apparel and Accessory Stores 2 0.59% 52 Building Materials, Hardware, Garden Supply, and Mobile Home Dealers 2 0.59% 23 Apparel and Other Finished Products Made From Fabrics and Similar 2 0.59% 57 Home Furniture, Furnishings, and Equipment Stores 1 0.29% 15 Building Construction-General Contractors And Operative Builders 1 0.29% 25 Furniture and Fixtures 1 0.29% 39 Miscellaneous Manufacturing Industries 1 0.29% 24 Lumber and Wood Products, Except Furniture 1 0.29% 55 Automotive Dealers and Gasoline Service Stations 1 0.29% 33 Primary Metal Industry 1 0.29% 35 Industrial and Commercial Machinery and Computer Equipment 1 0.29% 36 Electronic and Other Electrical Equipment and Components, Except Computer 1 0.29% Grand Total 339 100.00% Table 28. Breakdown of carriers by SIC, NYC. 3http://www.nyc.gov/html/dcp/html/bytes/dwnzdata.shtml
52 on their characteristics. In this system, land uses are classified by refining traditional categories into multiple dimensions, such as activities, functions, building types, site development char- acter, and ownership constraints. (A further description of the dimensions and the classification can be found in the supple- mental materials and appendices of the Task 11: Case Studies Report, available online6). Table 30 shows the breakdown of the receivers among the LBCS categories. As shown in Table 30, slightly more than 20% of the estab- lishments are related to retail activities. Wholesale trade of durable goods accounts for 17%, while non-durable goods represent 15% of the total. In addition, food service has a significant share, approximately 16%. In terms of LBCS activity, goods-oriented shopping takes place in 75% of the establishments, while a minor percentage uses land for plant, service, and restaurant activities. There is a significant degree be found in the supplemental materials and appendices of the Task 11: Case Studies Report, available online5.) As shown in Table 29, more than two thirds of establish- ments in the sample are located in commercial and manufac- turing districts. Furthermore, two thirds of establishments in commercial zones are in the Central business district; and about two thirds of establishments in the manufacturing zones are in light manufacturing districts. Overall, results show that the dataset covers the spectrum of land use catego- ries defined in the NYCZR. After completing the dataset with the required information, the team initiated the process of estimating FTG models. The process is discussed in the following section of this chapter. A similar analysis was made for the database using the LBCS developed by the APA (American Planning Association, 1994). LBCS provides a consistent model for classifying land uses based Land Use Land Use Description Number of establish- ments % of establish- ments Light manufacturing district (M1) Manufacturing district: light industries include woodworking service, auto storage and repair shops, and wholesale service and storage facilities 102 28.18% Residential district (R) Residential district: all residence districts permit most community facilities, such as schools, houses of worship and medical facilities. Certain facilities are not permitted or are restricted in size 100 27.62% Central business district (C5) Central business districts: offices, high-end retail establishments and continuous retail frontage 44 12.15% Central business district (C6) Central business districts: corporate headquarters, large hotels, entertainment facilities, retail stores and high-rise residences 40 11.05% Heavy manufacturing district (M3) Manufacturing district: heavy industries that generate traffic or pollutants. Typically include power plants, solid waste transfer facilities and recycling plants, as well as fuel supply depots 32 8.84% Retail district (C1) Small retail and service shops: grocery stores, restaurants and beauty parlors 18 4.97% General goods district (C4) Large stores with general goods: specialty and department stores, theaters and other commercial and office uses 13 3.59% General services district (C8) Heavy repair shops and automotive: automobile showrooms and repair shops, warehouses, gas stations and car washes (all commercial uses are permitted in C8 Districts) 5 1.38% Retail and services district (C8) Small retail and service shops: same as C1 but permits funeral homes and repair services 4 1.10% Middle manufacturing district (M2) Manufacturing district: middle ground between light and heavy industrial districts, more noise and vibration than in M1 are allowed, smoke is permitted and industrial activities need not be entirely enclosed 4 1.10% 362 100.00%Grand Total Table 29. Breakdown of receivers by land use according to NYCZR, New York City. 5Available online at http://transp.rpi.edu/~NCFRP25/downloads.shtml 6Available online at http://transp.rpi.edu/~NCFRP25/downloads.shtml
53 of correlation between function and activity, which prevents considering both as independent variables in econometric models of FTG. NYC Whole Foods Market Dataset Whole Foods Market is a chain of grocery stores offer- ing natural and organic foods with more than 270 stores in North America and the United Kingdom. The company has 54,000 team members, 9 distribution centers, 9 regional bake-houses, and more than 8 billion in sales during 2009. The team has the delivery information for five of their stores located in Manhattan, NY. Table 31 shows store names and LBCS Function Land Use Description Number of establish- ments % of establish- ments Retail Retail sales or service: automobile sale, heavy consumer goods sale, durable consumer goods sale (exclude grocery stores), consumer goods, retail food and beverage not included in the âGrocery LBCS functionâ 76 20.99% Durable goods Wholesale trade establishment: durable goods 62 17.13% Food service Food services: those that prepare meals, snacks and beverages for immediate consumption as primary economic function 57 15.75% Nondurable goods Wholesale trade establishment: nondurable goods 54 14.92% Miscellaneous 2 Include other economic use of the land such as communication and information, education and other institutions, construction related business 33 9.12% Miscellaneous 1 Manufacturing: include wood, paper, and printing products, chemicals and metals manufacturing, and miscellaneous manufacturing 26 7.18% Grocery Retail sales or service: include grocery stores, supermarkets, bakery, specialty food stores, fruit and vegetables stores, beer, wine and liquor store 25 6.91% Pharmacy Retail sales or service: include pharmacies, drug stores, cosmetic and beauty supplies, scientific and technical services 16 4.42% Textiles Manufacturing: include establishments that transform natural or synthetic fiber into products or manufacture textile products by knitting, cutting, and sewing fabric. 13 3.59% 362 100.00% LBCS Activity Land Use Description Number of establish- ments % of establish- ments Goods Shopping: Goods oriented shopping 271 74.86% Plant Plant, factory, or heavy goods storage or handling activities 39 10.77% Service Shopping: Service oriented shopping 35 9.67% Restaurant Restaurant type activity 16 4.42% Other Office activities 1 0.28% 362 100.00% Grand Total a) By LBCS function Grand Total b) By LBCS activity Table 30. Breakdown of receivers by LBCS function and activity, NYC. Store name Location (NYC) W. F. Union Square (USQ) 4 Union Square South W. F. Columbus Circle (CIR) 10 Columbus Circle W. F. Bowery (HOU) 95 East Houston St W. F. Tribeca (TRB) 270 Greenwich Street W. F. Chelsea (CHE) 250 7th Ave Table 31. Whole foods market Manhattan stores.
54 dors serving each store. As shown in Table 32, the store at Columbus Circle (CIR) has the highest number of deliver- ies per week, while Bowery (HOU) receives about half of the deliveries received by Columbus Circle. From Monday to Friday, the average number of daily deliveries per store ranges between 20 and 39; these numbers drop significantly to between 7 and 11 deliveries per day during Saturday and Sunday. Data indicate that Tuesday is the busiest day of the week for most of the stores. The number of vendors ranges from 46 to 87; Columbus Circle (CIR) is the store served with the largest number of vendors. A detailed hourly breakdown for the number of daily deliveries is presented in Table 90 in Appendix I. Seattle Region Grocery Stores Dataset The data for the grocery stores in the Seattle Region includes information from eight stores. These are spread across the Puget Sound metropolitan area (see Figure 11). The stores are all adjacent to major arterials, and have similar square footage, from 23,000 to 53,500 square feet. All are part of national grocery chains except for the Puget Sound Con- sumer Cooperative (PCC), which is a regional grocery chain. It is part of a nine-store chain owned by approximately 45,000 members living within the Puget Sound region, and the largest consumer-owned natural food retail cooperative in the United States. Five of the surveyed stores are Qual- ity Food Centers (QFCs), which is one store banner of the Kroger Corporation. This is one of the nationâs largest gro- cery retailers, operating 2,468 stores in 31 states under nearly two dozen banners. Two stores are Safeway, another national chain that operates under eight store banners. One store is an Albertsons, which is also part of a national chain that recently became part of the SUPERVALUE family as one of 18 store banners. All of the stores except PCC have company trucks and operate through regional distribution centers. It is important locations. Figure 10 shows the location within the Manhattan street network. The information available includes the weekly delivery schedule and delivery times. Table 32 shows the number of deliveries per store per day, weekday deliveries per employee, weekly deliveries per employee, and the number of ven- Figure 10. Whole foods market Manhattan stores locations. M T W R F Sa Su W. F. Union Square (USQ) 173 26 28 27 26 30 15 7 159 0.92 0.16 22 46 W. F. Columbus (CIR) 193 35 48 40 34 36 9 9 211 1.09 0.20 30 87 W. F. Bowery (HOU) 167 25 25 23 13 13 13 3 115 0.69 0.12 16 58 W. F. Tribeca (TRB) 173 28 32 31 26 37 14 1 169 0.98 0.18 24 52 W. F. Chelsea (CHE) 140 32 27 36 33 30 11 4 173 1.24 0.23 24 68 Total 846 146 160 157 132 146 62 24 827 0.98 0.18 116 311 Vendors Week del/emp Weekday del/emp Store name Emp. Deliveries Sub- Total Week del/day Table 32. FTG information for whole foods market stores.
55 to note that the data have been made available to the team courtesy of the research team at the Puget Sound Region Council (PSRC), which conducted the work reported in Ta et al. (2010) and McCormack and Bassok (2011). Information data was gathered by interviewing individual grocery store managers. In addition, manual on-site truck counts were conducted. Data from manual counts tested the accuracy of the estimates of daily truck deliveries pro- vided by grocery store telephone interviews. Information includes: truck trips, average number of truck deliveries per day, empty trucks, type of trucks, location of facilities, store characteristics, typical hours of deliveries, among others. Information relevant for the present report is summarized in Table 33. Methodology This section describes the modeling approach used to characterize the relationship between FTG and land use. FTG refers to the number of vehicle trips required to trans- port a demanded quantity of goods. It is closely related to FG. However, while FG is associated with the amount of goods demanded by establishments, FTG is determined by the number of vehicles needed to transport them (HolguÃn- Veras et al. 2011). For the analysis, FTG is quantified using the number of deliveries per establishment. This approach enables the estimation of FTG as a function of land use or industrial sector and employment. Knowing establishment characteristics, FTG can be readily estimated. To estimate trip generation models, three different approaches were used: standard trip generation rates; linear regression; and Multiple Classification Analysis (MCA). (For a description of MCA, see OrtÃºzar and Willumsen 2001). The analyses were performed using the industry classification sys- tems SIC and NAICS, land use classification systems, and the NYCZR and LBCS at the disaggregate establishment level. For the case of linear regression models, the analyses used total employees per establishment as the independent variable, Interviews Manual Count 1 Manual Count 2 Average* (del/day) QFC Wallingford 23,000 80 7 am-12 pm Mon to Sat 10 25 22 19 QFC Kirkland 28,000 70 5 am-11 am Mon to Sat 9 15 19 14 QFC Mukilteo 37,000 70 5 am-11 am Mon to Sat 10 18 17 15 QFC Capitol Hill 46,984 100 5 am-11 am Mon to Sat 9 14 18 14 QFC Lynnwood 53,500 72 5 am-10 pm Mon to Sat n/a 13 n/a 13 Safeway Othello 26,092 n/a n/a Mon to Sat n/a 15 15 15 Albertsons Kent 46,000 60 5:30 am-10:30 am Mon to Sat 15 11 15 14 PCC Issaquah 23,000 95 6 am-2 pm All days 13 23 30 22 Deliveries Per Day Note: *The average of the 3 columns is used as the observed number of deliveries (per day) for model comparison. Store and Location Square Footage Emp. Delivery Hours Delivery Days Table 33. Seattle region grocery stores information. Figure 11. Grocery stores locations in the Seattle region.
56 Following the methodology described previously, two dif- ferent approaches were applied: standard trip generation rates (per establishment and per employee); and Ordinary Least Squares (OLS). Using the available information, the analyses were performed by disaggregating data at the establishment level. These included average deliveries per establishment, average deliveries per employee, and linear regression mod- els. For the case of regression models per state, the analyses used total employees per establishment and the interaction between employment and store location as the independent variables. On the other hand, for the regression models per store location, the analyses used total employees per estab- lishment and the interaction between employment and store type as the independent variables. This resulted from consid- ering possible data collection and forecasting implications of different explanatory factors. The resulting linear regression models for freight attrac- tion for each industry/land use took different forms. As done for NYC case studies, three different types of models are considered: type S with constant FTG/FG per establishment; type E with a trip rate per employee; and type C represent- ing a linear model with intercept and rate per employee (see Section 4). Data for Illinois, Ohio, Michigan, and Indiana were con- sidered as individual case studies to provide a broad range of information. The states with less than seven stores were analyzed in two groups according to their geographic loca- tion, and each group was considered an individual case study. The first group was North East-Mid Atlantic (NEMA) states: Connecticut, Delaware, Maryland, Massachusetts, New York, Pennsylvania, Virginia, and West Virginia. The rest of the states were grouped into a Midwest (MW) states group: Iowa, Kentucky, Minnesota, Missouri, Wisconsin, and Nebraska. Table 34 shows the best models found to estimate FTG in weekly deliveries, and FG in weekly pallets for each state or group of states. As shown, only two out of six case studies have an FG dependent on business size. For Michigan, the best way to represent FG and FTG is using a trip rate per employee. In the case of Ohio, the best way to represent FG and FTG is to use an OLS model comprised of a con- stant generation and a term, depending on the number of employees when the store is located off-mall. For the remaining of the cases, the best way to calculate trip gen- eration is to apply a coefficient per establishment that var- ies between 0.950 and 1.636 trips. Similarly, the best way to estimate FG is to apply a coefficient per establishment that varies between 8.862 and 11.364. The state attracting the most freight trips and pallets per store is Illinois. On the other hand, Ohio is the state attract- ing the smallest amount of freight trips and pallets per store. These results are hard to extrapolate because they only repre- after considering the data collection and forecasting implica- tions of different explanatory factors. Efforts were made to collect data describing the area of the studied establishments using tax parcels, but the results showed that this method produced questionable estimates. The area variable was therefore discarded from the disaggre- gated analyses. The resulting linear regression models, for freight attrac- tion for each land use, fall in one of the following classes: â¢ Type S: Constant FTG per establishment; only the inter- cept was statistically significant and conceptually valid. FTG does not depend on business size. â¢ Type E: Trip rate per employee; only the coefficient of employment was statistically significant and conceptually valid. â¢ Type C: Linear model with intercept and rate per employee; both the intercept and the coefficient of employment were significant and conceptually valid. The Root Mean Square Error (RMSE) was the measure used to assess which type of model is more suitable to esti- mate freight trip attractions and productions. This metric provides a good indication of the absolute fit of the model to the data. Consequently, a lower value of the RMSE means a better fit to the data studied (Greene 1993). In the NYC carriers and receivers study cases, when the regression analy- ses found that FTG depends on business size, MCA models were applied to estimate the trip rates for each stratum of employment and for each category of land use. (It does not make sense to use MCA stratified by employment level if this variable is not statistically related to FTG.) The research explored different employee groupings to select the number and width of each interval class. The resulting models were grouped according to the type of disaggregated model (S, E or C, as previously described). MCA models were then esti- mated, where appropriate, for the different groupings and combinations of employee intervals. Case Study Illustrations This section includes illustrative applications of the FTG modeling approach. Midwestern States Furniture Chain: A Comparison of Different Location Structures (In-Mall and Off-Mall) The dataset used to estimate trip and FG models contains complete information on employment, store location, store type, and state for each observation. There were 58 establish- ments with complete information.
57 As well as for the statesâ classification, standard trip gener- ation rates and regression analyses were analyzed using store location (in-mall -M- or off-mall -OM) as a categorical fac- tor to determine FG and FTG. As shown in Table 35, the best way to represent attracted freight is by applying a constant trip generation per store; approximately 10 pallets attracted weekly in-mall-based locations and 12 in off-mall-based locations. In terms of FTG, Table 35 also shows that an off-mall locationsâ attrac- tion depends on the nature of the store. Each conventional store and each outlet store located off-mall attracts approx- imately one weekly trip, while each combo store located off-mall attracts two trips weekly. A surprising finding is that the number of trips attracted by mall-based stores depends on the store size, which was not the case for the number of pallets attracted. This may reflect the different sales levels for the in-mall stores. In addition to the models using each state and location type as a factor, the team estimated models for the pooled data. Specifically, the analyses focused on estimating models that sent a biased sample of furniture stores, and these models are valid only for this specific sample. An important characteris- tic of the furniture chain studied is that it is a franchise. Being a franchise, the stores only receive deliveries from the mother warehouse as decided by the franchisor. Moreover, as there is just one distribution center, the number of trips attracted and the cargo attracted depend not only on the characteris- tics of the store, but specially on the logistic decisions made by the franchisor. This is not the case when a store receives deliveries from various suppliers. In terms of land use, the only observable dimension for the sample collected was the location of the store. The LBCS devel- oped by the APA7 uses the structure dimension to differentiate in-mall locations and off-mall locations. The aim of the present research is to test the assumption that structure type (as defined by the LBCS) is a statistically significant factor in FG and FTG. The methodology adopted is based on OLS analyses. Const. Empl. Empl. off mall c b1 b2 ILLINOIS 11 1.636 1.636 0.148 S 1.1 OHIO 14 0.950 0.074 1.214 0.130 C 1.175 MICHIGAN 7 0.179 1.286 0.179 E 1.263 INDIANA 6 1.000 1.000 0.137 S 1.297 NEMA 6 1.000 1.000 0.111 S 1.302 MW 14 1.143 1.143 0.088 S 1.226 ILLINOIS 11 11.364 11.364 0.956 10.696 OHIO 14 8.862 0.619 11.071 1.143 C 10.433 MICHIGAN 7 1.690 12.143 1.690 E 11.262 INDIANA 6 9.833 9.833 1.343 S 11.598 NEMA 6 9.167 9.167 1.004 S 11.671 MW 14 10.857 10.857 0.862 S 10.899 S RMSE TRIP GENERATION BY STATE FREIGHT GENERATION BY STATE Best Model State Obs. Avg. Del/est Avg. Del/emp Table 34. Freight generation and freight trip generation by state for furniture store chain. Const. Empl. Combo store c b1 b2 IN-MALL 31 0.793 0.039 1.161 0.115 C 1.051 OFF-MALL 27 1.182 0.818 1.333 1.128 C 0.909 IN-MALL 31 10.160 10.160 0.996 S 9.648 OFF-MALL 27 11.794 11.704 1.099 S 8.417 Best Model Location Obs. Avg. Del/est Avg. Del/emp RMSE FREIGHT GENERATION BY LOCATION TRIP GENERATION BY LOCATION Table 35. Freight trip generation by location of furniture stores. 7More detailed information can be found in http://www.planning.org/lbcs/
58 size and industry sector, estimated models are discussed herein. Standard Industrial Classification (SIC) System The first step in the estimation process was to aggregate the different industries (2-digit) into broader groups accord- ing to their economic sectors. Using the SIC codes, 11 cat- egories or groups were created. Eight of them were selected as freight-related and used in the estimation process: agri- culture, forestry, and fisheries (Group 1); mineral industries (Group 2); construction industries (Group 3); manufactur- ing (Group 4); transportation, communication, and utilities (Group 5); wholesale trade (Group 6); retail trade (Group 7); and food (Group 8). The other categories: finance, insurance and real estate; service industries; and public administration were not considered in the model estimation process because they are not freight-related, and there was not enough freight trip information available. For estimation purposes, the team used SICs which con- tained five or more observations in the OLS analysis. Table 38 shows the final models estimated for freight attraction. It is important to note that for groups 1 and 2 not enough infor- mation was available. Considering their relation to the indus- tries in group 3, the delivery rate of group 3 is recommended for groups 1 and 2. For the case of group 5 (there were no observations) a constant rate of one delivery per establish- ment per day was assumed. As shown in Table 38, more than half (53%) of the mod- els are constant rates per establishment (Type S), 28.6% are linear models with an intercept and slope (Type C), and the other 19% depend on business size (Type E). Table 91 in Appendix I shows the type of resulting model for the differ- ent SICs and groups. More than half of the resulting models were constant per establishment (Type S) models; therefore, business size alone may not always be a good indicator of freight trip attractions. The MCA is another estimation technique; it computes coefficients for each category of the predictors (e.g., employ- expressed the number of deliveries received, and the num- ber of pallets received, per week, as a function of company attributes such as employment, geographic location and characteristics, and the interactions between these attri- butes. Table 36 shows the best models found to estimate trip generation using the pooled data, while Table 37 shows the results for the number of pallets attracted by store using the pooled data. As shown in Table 36, trip generation varies according to the type of store, i.e., combo stores have a statistically sig- nificant higher attraction. Each combo store attracts 0.903 more trips every week than conventional or outlet stores, ceteris paribus. Additionally, as expected from the results shown in Table 34, trip attraction in Michigan depends on business size. In contrast to the findings on trip generation, outlet stores were found to have different patterns of FG. As shown in Table 37, each conventional or combo store attracts 10.42 pallets weekly (constant), while each outlet store attracts 13.75 pallets weekly (constant + outlet stores coefficient). According to this model, the number of pallets attracted depends on the store type, but does not depend on employment. It is noteworthy that this model has a poor statistical fit. NYC Receivers and Carriers Case: A Comparison Between Production and Attraction Activities Using Industrial Classification Systems (SIC and NAICS) Freight Trip Attraction Freight trip attraction refers to the number of truck trips attracted by an establishment as a result of its economic activity. Freight trip attraction is captured in the receivers surveys with the average number of deliveries received by each establishment in a typical day. Considering business Variable Name Coefficient t-value Regression model Intercept CONSTANT 1.097 23.201 Store characteristic Combo store COMBO 0.903 7.031 State and employment Employment in Michigan EM_MICH 0.040 2.363 Observations 58 F 25.770 Adjusted R2 0.470 Table 36. Freight trip generation model for pooled data (employment, store characteristics, location and state) for furniture stores. Variable Name Coefficient t-value Regression model Intercept CONSTANT 10.420 19.329 Store characteristic Outlet O 3.330 2.294 Observations 58 F 5.260 Adjusted R2 0.070 Table 37. Freight generation model for pooled data (employment, store characteristics, location and state) for furniture stores.
59 the analyses; however, the results revealed what seem to be anomalous results for establishments within the 31â40 employee bracket. This reflects the low number of obser- vations in that range. Therefore, the trip rates generated for employment above 30 were excluded from the analysis. This exclusion does not impact the significance of the anal- ysis because over 90% of the businesses in the sample area (Manhattan and Brooklyn) have less than 30 employees (U.S. Census Bureau 2010b). Since the models are employment dependent, as anti- cipated, the number of deliveries increases with employ- ment. In general, establishments in the building mate- ment and SIC). The coefficients are estimated in such a way that they provide the best possible fit to the observed data (i.e., minimize sum of squared errors). Type of disaggregate model (i.e., S, E, or C) and number of observations were the two factors used to group the industry segments. Given that employment was considered as the independent variable, only industries that exhibited a dependence on business size, based on the results of the OLS analysis (i.e., E or C), were considered. Table 39 shows the MCA coefficients for freight attrac- tion using SIC as the industrial classification system. Five bins with an interval width of 10 employees were used in c b 15, 16, 17 Construction* 25 2.160 S 0.869 15 General contractors & operative builders 7 0.129 E 0.938 17 Special trade contractors 17 2.106 S 1.365 21-39 Manufacturing* 45 3.156 S 3.420 23 Apparel & other finished products 7 3.571 S 1.178 24 Lumber & wood products, except furniture 5 0.067 E 0.764 25 Furniture & fixtures 6 2.167 S 1.067 34 Fabricated metal products 4 1.500 S 0.500 39 Miscellaneous manufacturing industries 5 2.280 S 0.280 50, 51 Wholesale Trade* 117 2.272 0.069 C 3.655 50 Wholesale trade - durable goods 58 3.986 S 4.740 51 Wholesale trade - nondurable goods 59 1.713 0.071 C 2.147 52, 53, 55, 56, 57, 59 Retail Trade* 84 3.371 S 5.384 52 Building materials... & mobile home dealers 9 0.369 E 1.672 56 Apparel and accessory stores 13 0.187 E 4.598 57 Home furniture, furnishings, equipment stores 13 3.769 S 2.189 59 Miscellaneous retail 47 3.349 S 4.067 20, 54, 58 Food* 83 1.826 0.090 C 4.813 20 Food and kindred products 3 2.000 S 0.032 54 Food stores 23 0.288 E 4.851 58 Eating and drinking places 56 1.307 0.081 C 3.091 *Group models 6 7 8 Description Best Model RMSE 3 4 Gr. SIC Obs. Const. / Empl. Table 38. SICâfinal models selected for freight trip attraction (deliveries/day). Building construction Lumber Wood Building Material Food Stores Apparel & Accessories Wholesale - durable goods Eating and Drinking Places 1-10 1.405 n/a 3.648 3.768 1.176 1.984 1.875 11-20 1.608 n/a 3.852 3.972 1.379 3.076 2.966 21-30 6.550 4.274 8.794 8.914 6.321 4.466 4.356 1.130 2.832 2.290 4.875 4.382 2.060 3.300RMSE seeyolp m E SIC Type of Model Type E Models Type C Models Industry Sector Table 39. Multiple classification analysis results for sic for freight attraction (deliveries/day).
60 Freight Trip Production Freight trip production refers to the number of truck trips produced by the source of the commodities, i.e., the ship- per. This is captured in the carriersâ surveys with the average number of trips made by each establishment in a typical day. Considering the industry sector and the size of the business, estimated models are discussed in the following sections. Standard Industrial Classification (SIC) System The modeling process undertaken using the carriers dataset followed the same methodology as the receivers. The resulting OLS models using SIC coding are shown in Table 42. Approxi- mately 42% of the models are of Type S, 33% are Type E, and 25% are Type C. The largest percentage of models estimates deliveries per establishment (Type S), which indicates that, as rials and hardware industry (SIC 52) and food stores (SIC 54) receive, on average, approximately two more deliveries per day than those in building construction (SIC 15) and apparel and accessories (SIC 56) industries; both of which have similar freight attraction patterns. Food stores (SIC 54) also receive almost twice the amount of deliveries than establishments in the lumber wood (SIC 24); eating and drinking (SIC 58); and wholesale tradeâ nondurable goods (SIC 51) industries, except for establish- ments with 11â20 employees. North American Industry Classification System (NAICS) This section discusses the estimated FTG models considering NAICS as the industry classification system. Table 40 shows the final disaggregate models from the OLS analysis. As shown, 60% of the models are constant FTG rates per establishment (Type S), with the remaining 40% being combined models with intercept and slope (Type C). These results confirm the previous finding that business size alone may not be a good indicator of freight trip attraction. The MCA models generated using NAICS for the receivers sample can be found in Table 41. It is worth noting that no models were solely dependent on number of employees (Type E); all the industries were represented by combined models (Type C). The results indicate that retail trade (NAICS 44) establishments, on average, receive one more delivery than those in wholesale trade (NAICS 42) and accommodation and food service (NAICS 72) industries; with the two latter industries having a similar FTG pattern. c b 1 23 Construction* 25 2.160 S 1.364 31, 32, 33 Manufacturing* 51 2.831 S 2.791 31 Food, Beverage, Tobacco, Textile, Apparel, Leather & Allied Product Manufacturing 21 2.400 S 1.295 32 Wood, paper, printing, petroleum & coal products, chemical, plastics, nonmetallic & mineral product manufacturing 10 4.420 S 5.483 33 Metal, machinery, computer, electronic, electrical, transportation, furniture & misc. manufacturing 20 2.490 S 2.483 3 42 Wholesale Trade* 117 2.272 0.069 C 3.655 44, 45 Retail Trade* 98 3.070 0.063 C 4.054 44 Motor vehicle, furniture, electronics, building material, food & beverage, health, gasoline, & clothing stores 69 2.458 0.132 C 4.298 45 Sporting goods, hobby, book, & music stores 29 2.724 S 4.352 6 72 Accommodation and Food* 56 1.307 0.081 C 3.091 * Group models Best Model RMSE 2 Gr. Description Obs. Const. / Empl. NAICS 4 Table 40. NAICSâfinal models selected for freight trip attraction (deliveries/day). Table 41. Multiple classification analysis results for NAICS for freight trip attraction (deliveries/day). Wholesale trade Retail trade Accomodation & food 1-10 2.443 3.543 1.902 11-20 3.341 4.442 2.801 21-30 5.685 6.785 5.144 3.658 4.197 3.355RMSE seeyolp m E NAICS Type of Model C Industry Sector
61 c b 15, 16, 17 Construction* 9 0.068 E 1.586 17 Special trade contractors 8 0.065 E 1.576 4 21-39 Manufacturing 16 1.625 S 1.364 42, 47 Transportation, Communication and Utilities* 157 2.718 0.038 C 3.970 42 Motor freight transportation & warehousing 148 2.764 0.035 C 3.850 47 Transportation services 9 0.076 E 5.758 50, 51 Wholesale Trade* 126 1.944 0.036 C 5.408 50 Wholesale trade - durable goods 65 0.059 E 3.628 51 Wholesale trade - nondurable goods 61 4.328 S 6.939 7 52, 53, 55, 56, 57, 59 Retail Trade* 9 1.889 S 0.875 20, 54, 58 Food* 12 3.000 S 5.164 20 Food and kindred products 11 3.182 S 5.167 RMSE 3 5 6 8 Obs. Const. / Empl. Best Model Gr. SIC * Group models Description Table 42. SICâfinal models selected for freight trip production (trips/day). Type C Special Trade Contractors Transportation Services Wholesale - durable goods Motor freight transportation 1-20 1.448 2.863 1.641 3.346 21-40 2.081 3.497 2.274 4.064* 41-60 4.882 6.297 5.075 4.782* 61-80 5.888* 7.303* 6.081* 5.500 >80 6.894 8.309 7.086 9.750 1.964 4.897 3.536 5.101 * Coefficients have been modified for consistency using interpolation SIC Type of Model Type E Model seeyolp m E RMSE Industry Sector Table 43. Multiple classification analysis results for SIC for freight trip production (trips/day). with the receivers, business size may not be a consistent indica- tor of FTG. Table 92 in Appendix I shows the different SICs and groups, arranged by model type. Table 43 shows the MCA results for the carriers sample for SIC. As previously mentioned, MCA is performed on mod- els that depend on business size. The table shows the MCA results with an employment distribution of five bins, and with an interval width of 20 employees. Unlike with the receiversâ sample, which had a consistent dip in the 31â40 employees bin, the carriers sample had some variations that were cor- rected using interpolation. The decrease in the number of deliveries might be related to a change in logistic decisions. However, this could also be a consequence of biased data for these specific bins in the studied sample. Further research needs to be done to better understand the FTG patterns for establishments with these employment levels. MCA results for SICs show that establishments in the transportation services industry (SIC 47), on average, gen- erate one more delivery than establishments in the construc- tion (SIC 17) and wholesale trade durable goods (SIC 50) industries; while the two latter SICs have similar FG pat- terns. Establishments in the motor freight transportation and warehousing (SIC 42) industry had the highest trip rates among the SIC models in all but 21â40 and 41â60 bins, with the most notable being between 41â60 employees, where the trip rate was at least half the number of all the other employ- ment dependent SIC models. North American Industrial Classification System (NAICS) Table 44 shows the final disaggregate models for the esti- mated freight trip production using NAICS as the industry classification. The results for NAICS indicate that 30% are Type S, another 30% were Type E, and the remaining 40% were Type C models, as shown in Table 44.
62 egory, the table shows the error of applying the rates obtained from the models previously discussed. When FTG is con- stant, the delivery/trip rate per establishment found for the category is used; instead when FTG is employment depen- dent, MCA rates are used. As shown, although SIC models perform slightly better when estimating freight trip attraction for the overall data- set with a RMSE of 3.49 compared to 3.56 for NAICS, the individual SIC industry models perform differently. Only manufacturing (NAICS 31) and retail trade (NAICS 44), two out of eight categories, show significantly better results for SIC models. In contrast, NAICS models perform better than SIC models when estimating freight trip production, overall, Table 45 shows the MCA results for the carriersâ sample, for both SIC and NAICS. For the NAICS analysis, establish- ments in the construction (NAICS 23) industry averaged approximately one more trip than those in the manufacturing (NAICS 32) industry. Establishments in the transportation and warehousing (NAIC 49 and 48) industry both averaged at least one more trip than establishments in the retail trade (NAICS 44) industry. Comparison Between Industrial Classifications Coding Systems Table 46 shows the error metrics for the models estimated using NAICS industries as categorical factors. For each cat- c b 1 23 Construction* 9 0.068 E 1.586 31, 32, 33 Manufacturing* 28 2.214 S 3.599 31 Food, beverage, tobacco, textile, apparel, leather & allied product manufacturing 13 2.846 S 4.990 32 Wood, paper, printing, petroleum & coal products, chemical, plastics, nonmetallic & mineral manufacturing 7 0.023 E 0.648 33 Metal, machinery, computer, electronic, electrical, transportation, furniture & misc. manufacturing 8 1.750 S 1.639 3 42 Wholesale Trade* 124 1.755 0.036 C 5.094 44, 45 Retail Trade* 9 0.161 E 6.485 44 Motor vehicle, furniture, electronics, building material, food & beverage, health, gasoline, & clothing stores 5 0.993 0.021 C 0.237 48, 49 Transportation and Warehousing* 157 2.718 0.038 C 4.811 48 Air, rail, water, truck, transit, pipeline, scenic & sightseeing, & support activities 153 2.725 0.038 C 4.005 * Group models 4 5 Best Model RMSE 2 Gr. Description Obs. Const. / Empl. NAICS Table 44. NAICSâfinal models selected for freight trip production (trips/day). Construction Wood, paper, petroleum,coal, chemical, plastics manufacturing Wholesale Trade Motor vehicle, furniture, electronics, food & beverage retail Transportation 1-20 2.424 1.303 2.946 1.685 3.381 21-40 1.727 0.606 2.564 1.303 2.998 41-60 2.061 0.939 3.283 2.023 3.718 61-80 4.061 2.939 2.764 1.504 3.199 >80 5.121 4.000 7.609 6.348 8.043 1.074 0.934 4.650 0.618 5.219 NAICS Type of Model Type E Models Type C Models RMSE seeyolp m E Industry Sector Table 45. Multiple classification analysis results for NAICS for freight trip production (trips/day).
63 The City of New York Zoning Resolution (NYCZR) Zoning ordinances adopted by municipalities regulate the size and use of land and buildings, including location and den- sity. This is a key tool for carrying out municipal planning policy, along with the powers to budget, tax, and condemn property. In this context, NYC has been a pioneer in land use zoning since it enacted the nationâs first comprehensive zoning ordinance. The NYCZR classifies basic land uses including the three considered in this study: residential (R), commercial (C), and manufacturing (M). Within these classifications are subcate- gories for low-, medium- and high-density uses and/or build- ings (e.g., âlightâ manufacturing, âsingle-familyâ residential). Although similar zoning designations may be aggregated into district classifications, e.g., residential, these districts may allow other uses, such as ground-floor commercial or âgrandfatheredâ pre-existing non-conforming uses, such as auto repair shops. Recent zoning amendments favor âmixed useâ classifications that include commercial, residential, and work space, to promote âwalkableâ communities. In essence, residential districts also generate freight (NYC Department of City Planning 2010). As described in the NYCZR, districts are classified in ascend- ing order of density or operations. For example, residential districts are classified R-1 through R-10, in ascending order of density, while manufacturing districts are classified M-1, M-2, and M-3 depending on characteristics and specific operations. Commercial districts are also classified numerically by allow- able activities. For example, Central business district C-5 allows offices, high-end retail establishments, and continuous retail frontage, while C-6 allows corporate headquarters and large hotels (NYC Department of City Planning 2010). For estimation purposes, the authors used NYCZR land use classifications with more than five establishments. Where with RMSE of 4.80 for SIC compared to 4.90 for NAICS. Nevertheless, this performance is confirmed for only three out of nine categories, where the difference on the errors is larger than 10%. The industries where NAICS models give better freight trip production estimates are construction, manufacturing, and retail trade (NAICS 23, 32, and 45). The analyses have highlighted the implications of the industry classification systems used when modeling FTG, though, results may be impacted by data limitations, indus- try comparability between the systems, and the resulting aggregations performed. As discussed, 2-digit industry aggregations were used for both SIC and NAICS, as there were not enough observations for a one-to-one mapping. An analysis of the different industriesâ definitions revealed that the closest match would have been to use a two-digit NAICS with the 2-digit SIC industries; however, this was not pos- sible. Furthermore, FTG modeling efforts should use 3-digit NAICS and 2-digit SIC for higher level of detail, as these would allow an adequate trade-off between level of detail for disaggregate analyses and the number of observations to be collected. NYC Receivers Case: A Comparison of Freight Attraction Between Local Land Use Codes and Universal Standard Land Use Codes This section describes the FTG models developed based on two land use classification systems that are applicable to NYC. The first is the City of New York Zoning Resolution (NYCZR), developed in 1916 and updated regularly. The sec- ond is the LBCS, developed by the FHWA in partnership with the APA. These systems provide the basis for analyzing the effects of land use on FTG. RMSE (NAICS) RMSE (SIC) RMSE (NAICS) RMSE (SIC) Construction 23 1.364 1.275 1.074 1.859 31 1.295 2.501 4.990 4.934 32 5.483 5.568 0.934 1.051 33 2.483 2.283 1.639 1.659 Wholesale Trade 42 3.658 3.655 4.650 4.895 44 4.197 3.819 0.618 0.569 45 4.352 4.396 8.539 10.551 48 * * 5.219 5.145 49 * * 2.018 2.009 Accommodation & Food 72 3.355 3.300 * * Total 3.563 3.492 4.796 4.903 Transportation & Warehousing Description * No data available on the sample NAICS Code Attraction Production Manufacturing Retail Trade Table 46. Model Estimation Errors (RMSE) for each Industry Segment (Freight Trip Production and Attraction).
64 for establishments with less than 30 employees, but exhib- ited anomalies in the group of 31â40 employees. This likely results from the lack of a sufficient number of observations in the 31â40 employees group to support the MCA estimation. (The number of observations in this range was significantly smaller than for the other groups). For that reason the 31â40 employees group was omitted from the analysis. This omis- sion does not affect the relevance of the research because (as previously discussed) more than 90% of the establishments in Manhattan and Brooklyn have fewer than 30 employees. More research is needed to explain FTG patterns on estab- lishments with more than 30 employees. Table 48 shows the FTG rates for NYCZR land uses for different employment levels. some land uses had five or fewer, the individual classes were aggregated. For a more detailed description about grouping procedures, see (HolguÃn-Veras et al. 2011). For example, commercial establishments were divided into two groups: retail and service oriented establishments. Trip generation rates and linear regression models were estimated for various land use categories. Table 47 shows the final models estimated for freight trip attraction in the disaggregated land uses, and in the land use groups (i.e., C1C4C5C6, C2C8, Manufacturing, and Residential). All of the variables presented on the tables were found to be signifi- cant at the 95% confidence level. The best models were selected based on t-statistics and RMSE. The results indicate that for 73% of the FTG models, a constant coefficient produces the best results; for 18% of the models, FTG depends on employment; and for 9% FTG is a combined model, with a constant coefficient plus an employ- ment term. The results also indicate that employment depen- dent models (type E) are found in residential classifications with moderate- and high-density districts (R6, R6A, R7-2, R7A), while FTG in light manufacturing land use (M1-6) is better represented by a combined model (type C). For the remaining land uses, especially commercial land use, FTG is better represented by a constant coefficient. The next step was to conduct MCA analyses for those land use classifications with FTG that depend on employment (types E and C). The rates found were conceptually valid c b C1-9 8 4.900 4.900 0.090 S 6.007 C5-2 9 2.670 2.670 0.076 S 1.414 C5-3 22 3.509 3.509 0.089 S 3.448 C5-5 7 2.343 2.343 0.163 S 1.401 C6-6 6 2.200 2.200 0.072 S 1.633 C1C4C5C6* 115 2.760 0.063 4.179 0.127 C 0.050 5.417 C2C8* 7 4.286 4.286 0.137 S 3.692 M1-1 24 3.700 3.700 0.094 S 3.948 M1-2 11 1.909 1.909 0.754 S 1.240 M1-2/R6A 7 7.229 7.229 0.154 S 9.416 M1-2D 7 3.057 3.057 0.135 S 1.841 M1-6 31 1.287 0.069 2.271 0.121 C 0.135 1.926 M3-1 32 3.381 3.381 0.122 S 2.935 Manufact.* 138 3.216 3.216 0.115 S 4.000 R6 20 0.338 4.740 0.338 E 0.498 3.496 R6A 15 0.243 3.000 0.243 E 0.075 1.720 R6B 13 5.415 5.415 0.248 S 8.624 R7-1 5 1.960 1.960 0.073 S 1.795 R7-2 5 0.206 4.200 0.206 E 0.248 2.596 R7A 14 0.140 3.786 0.140 E 0.101 4.647 R8 10 2.660 2.660 0.125 S 1.470 Resid.* 10 2.660 2.660 0.125 S 4.427 Total 4.460 gnirutcafuna M tcirtsi D RMSE tcirtsi D laitnedise R R2 Adj. laicre m mo C tcirtsi D Best Model Gr. Land Use Obs. Const. / Empl. Avg. Del/est Avg. Del/emp Table 47. Freight trip attraction by NYCZR land use and type of models. C6* M1-6 R6 R6A R7-2 R7A R* 0-10 2.97* 1.49 3.28 2.37 0.95 1.14 2.73* 11-20 3.53* 2.25 4.38 3.47 2.05 2.24 3.28* 21-30 6.10* 5.67 8.33 7.42 6.00 6.19 5.86* 5.380 1.670 3.240 1.990 2.450 4.560 4.390 NYCZR seeyolp m E RMSE Notes: (1) Overall RMSE: 4.430 (2) *Based on group models Table 48. MCA results for daily freight trip attraction by the city of New York Zoning Resolution (NYCZR) land uses.
65 Comparison Between LBCS and NYCZR Upon analysis, the authors found that for both NYCZR and LBCS, most of the best models were produced using a constant coefficient only, and do not depend on business size (quan- tified as number of employees). When using NYCZR, land use classifications as the categorical factor, 73% of the land uses were best estimated when a constant coefficient is used. Similarly, when using LBCS Function as the categorical fac- tor, six out of the nine land use categories were best estimated using a constant coefficient. When LBCS Activity is used as the categorical factor, four out of the five land use categories were best estimated using a constant coefficient. These results are consistent with the findings when using the industry seg- ment as the categorical factor as described in HolguÃn-Veras et al. (2011). The evidence of non-employment dependent FTG for most land uses and establishments in NYC is indeed convincing. The regression model results indicate the RMSE is 4.46 for NYCZR and 4.62 for LBCS (see Tables 47 and 49); while the MCA models leads to RMSE of 4.43 for NYCZR and 4.60 for LBCS (see Table 48 and Table 50). This indi- cates that MCA performs slightly better than regression analysis for almost every land use classification where FTG is employment dependent, even though it is a very small difference with respect to the overall error. This is not surprising, because MCA has more degrees of free- dom than regression analysis, at the expense of higher data requirements. In essence, where FTG is constant, employment plays no role. Table 51 summarizes the best FTG coefficients for the Land-Based Classification Standards (LBCS) This section discusses the models estimated for NYC using the LBCS, and specifically, the factors most relevant for the purposes of FG: LBCS Function and LBCS Activity. Table 49 shows the final models estimated for freight attraction. The results indicate that for 67% of the FTG LBCS Function models, a constant coefficient produces the best results; for 11% of the models, FTG depends on employ- ment; and for 22% of the models, FTG is a combined model, with a constant coefficient plus an employment term (for LBCS Activity, this breakdown is 80%, 0%, and 20%, respectively). MCA models were estimated for the LBCS Function classifications in which FTG depends on employment. As in the previous section, there is an abnormal coefficient for establishments with 31â40 employees; therefore, these establishments were omitted from the results shown in Table 50. c b Retail 73 3.682 3.682 0.140 S 5.111 Grocery 24 0.217 5.225 0.217 E 0.039 5.280 Pharmacy 16 3.988 3.988 0.203 S 3.469 Food 55 1.307 0.081 3.100 0.111 C 3.147 3.092 Textiles 12 2.867 2.867 0.153 S 1.459 Miscellaneous 1 26 3.254 3.254 0.094 S 4.081 Durable Goods 60 4.387 4.387 0.173 S 5.736 Nondurable Goods 54 1.681 0.072 2.948 0.121 C 2.256 2.214 Miscellaneous 2 32 3.919 3.919 0.118 S 6.871 Total 4.622 Goods 265 2.588 0.067 3.811 0.141 C 4.583 4.565 Services 34 3.865 3.865 0.117 S 6.670 Restaurant 16 2.488 2.488 0.149 S 1.939 Other 1 0.400 0.400 0.080 S N/A Plant 38 3.132 3.132 0.100 S 3.483 Total 4.620 noitcnu F ytivitc A R2 Adj. RMSE Best Model Gr. LCBS Obs. Const. / Emp. Avg. Del/est Avg. Del/emp Table 49. Freight Trip Attraction by LBCS Function. Grocery Food Non Durables 0-10 4.13 1.86 1.92 11-20 5.44* 2.96 3.02 21-30 6.75 4.38 4.44 5.032 3.301 2.120 seeyolp m E RMSE Notes: (1)*Coefficient has been modified for consistency (2) Overall RMSE: 4.600 Table 50. MCA for freight trip attraction by LBCS function land uses.
66 tial exists for universal or transferable FTG models. To assess the potential of LBCS will require applying and comparing results to other local land use classifications (e.g., Seattle, WA, or Portland, OR). Comparison Between Institute of Transportation Engineers and Land Use Models This section compares the performance of the models based on NYCZR and LBCS with the current benchmark in this field of study, The Institute of Transportation Engi- neers (ITE) Trip Generation Manual. The ITE has produced a series of manuals to estimate the number of vehicle trips generated by a facility or establishment located in a par- ticular âland use.â The latest version of this manual is the 8th Edition (Institute of Transportation Engineers, 2008). In this manual, trip generation rates are provided based on trip generation studies submitted to ITE by public agen- cies, consulting firms, universities, developers, and oth- ers. These include average freight truck trips for several categories of land use, including truck terminals, indus- trial parks, warehouses, mini warehouses, high-cube ware- NYC study dataset. Cities with characteristics analogous to NYC, especially Manhattan and Brooklyn, or with similarly situated neighborhoods, may have similar FTG patterns, in which case the findings of this report could be extrapolated. However, it is important to emphasize that the results found are valid only in the context of this dataset. The next step of the analyses compared the performance of models using local zoning classifications (NYCZR) with models using a standardized classification system such as the LBCS. The RMSE analysis, performed on the complete dataset using the best models for each land use classifica- tion (including models for pooled data and grouped land uses) found that the overall error for NYCZR was 4.21, as compared to 4.53 for LBCS. According to this RMSE analy- sis, classifying land uses using NYCZR gives more accurate results for a local dataset, which is not surprising. The reason is related to the data limitation (i.e., the high level of correla- tion between LBCS dimensions) that prevents taking advan- tage of the multiple dimensions of the LBCS. Although LBCS was not found to be superior to the NYCZR, it is not possible to make definitive conclusions about LBCS merits for FTG modeling. In fact, it is only through the use of LBCS that intercity comparisons can be made and the poten- Land Use Trips/ Establish- ment Land Use Trips/ Establish- ment Land Use Trips/ Establish- ment Land Use Trips/ Establishment C6-6 2.20 M1-2 1.91 R7-1 1.96 Textiles 2.87 C5-5 2.34 M1-2D 3.06 R8 2.66 Misc. 1 3.25 C5-2 2.67 M1* 3.14 R6B 5.42 Retail 3.68 C5* 3.17 M2* 3.22 Misc. 2 3.92 C5-3 3.51 M3-1 3.38 Pharmacy 3.99 C2* 4.29 M3* 3.38 Durables 4.39 C8* 4.29 M1-1 3.70 C4* 4.86 M1-2/R6A 7.23 C1-9 4.90 C1* 5.03 *Use this group coefficient when no detailed land use coefficient is found a) Constant trip attraction per establishment NYCZR LBCS C6* M1-6 R6 R6A R7-2 R7A R* Grocery Food Non Durable 0-10 2.97 1.49 3.28 2.37 0.95 1.14 2.73 4.13 1.86 1.92 11-20 3.53 2.25 4.38 3.47 2.05 2.24 3.28 5.44 2.96 3.02 21-30 6.10 5.67 8.33 7.42 6.00 6.19 5.86 6.75 4.38 4.44 *Use this group trip rate when no detailed land use coefficient is found b) Trip rates per establishment by employment size seeyolp m E NYCZR LBCS Table 51. Summary of daily freight trip attraction by land use, New York City.
67 Area-Based Models As found in the review of the freight systems and land use, there are variables that play a significant role in the estimation of FTG. For example, the previous sections have shown the estimated disaggregate models for FTG based on employment for different industry segments and land use categories. Given that the estimated models are based solely on employment (due to limited data availability, and despite team efforts to include area as independent vari- ables for the case study datasets), it was also important to understand the relationship between FTG and the estab- lishmentsâ areas. To estimate the relationship between employment and area, two different datasets were used. On one hand, merg- ing the Tax-Lot Polygon Feature Class of the Department of Financeâs Digital Tax Map with the 2006 PLUTOâ¢ attribute data for NYCâdefined as the attribute table of property information associated with each tax lot8âprovided aggre- gated estimates of the establishmentsâ areas. In general, the PLUTOâ¢ data files contain three basic types of data: tax lot characteristics; building characteristics; and geographic/ political/administrative districts. The dataset also contains information about different types of tax lot areas (see Table 93 in Appendix I). On the other hand, the County Business Patterns provided data on the total number of establishments, employment, first quarter and annual payroll, and number of establish- ments by nine employment-size classes by detailed industry for all counties in the United States, the District of Columbia, Puerto Rico, and the Island Areas (American Samoa, Guam, Northern Mariana Islands, and Virgin Islands)9. Information for only NYC was used for the analyses. After the two datasets were assembled, aggregates at the ZIP code level of employment, number of establishments, and the different areas were estimated. Linear regression houses, assisted living facilities, state departments of motor vehicles, United States post offices, research and develop- ment centers, free standing discount stores, hardware/ paint stores, wholesale markets, furniture stores, and qual- ity restaurants. The research, to ensure comparability, focused on those establishments present in both the ITE Manual and the NYC dataset: hardware/paint stores, wholesale markets, and furniture stores. In some cases, ITE models do not consider a specific directionality for freight trips. As there is not enough information in the ITE Manual, the authors assume that truck trips are a fraction of total vehicle trips, in which case freight-related trips entering and leaving the establishments would be the same proportion as pas- senger trips. This is not the case for the models devel- oped in this study, as the NYCZR and LBCS only estimate freight trip attraction (in terms of deliveries). As any deliv- ery attracted by an establishment produces two trips (one loaded, entering the establishment, and one empty, leav- ing), each delivery is multiplied by two to obtain the total number of trips attracted. Future research will focus on freight trip production. Table 52 compares the RMSE for the establishments located in the specific land uses that can be classified in all three: NYCZR, LBCS and ITE. If high employment establishments are included in the analysis, RMSE increase substantially. In general, NYCZR and LBCS models perform better than the ITE rates. Table 52 shows that the overall RMSE is usually larger for ITE models than for the land use models developed in this study. For furniture stores and hardware/ paint stores, the models have a similar performance. For wholesale stores, which account for 82% of the sample, NYCZR and LBCS models have a superior performance (i.e., RMSE is 18% lower). When considering the complete sample, the total RMSE of NYCZR and LBCS models is about 30% lower than ITEâs. In essence, the models devel- oped in this research give more accurate estimates for trip attraction. Trip Attraction Establish- ments Trip Attraction RMSE* Trip Attraction RMSE* Trip Attraction RMSE* Hardware/Paint Stores 60 8 43 4.1 36 4.4 35 4.8 Wholesale Markets 823 102 1527 15.9 659 12.9 791 12.9 Furniture Stores 51 14 42 6.4 99 5.9 63 5.7 Total 934 124 1612 14.6 794 11.9 889 11.9 *RMSE is computed at the establishment level ITE land use ITE employment models NYCZR LBCSNYC Sample Table 52. Estimation errors (RMSE) for freight trip attraction models by land use classifications. 8http://www.nyc.gov/html/dcp/html/bytes/applbyte.shtml 9http://censtats.census.gov/cbpnaic/cbpnaic.shtml
68 models using employment or number of establishments per ZIP code as dependent variables, and the different areas as independent variables, were estimated. Table 53 shows the resulting models. Users can choose the best type of model depending on the available area information. The adjusted R2 value generally exceeds 0.60. As expected, there is no sig- nificant relation between residential area and employment or number of establishments as shown by the low R2. With these models, the number of employees can be estimated from area aggregates, which allows the user to implement the FTG models discussed in previous sections. Independent Variable Area Coefficient* Adjusted R2 1.41E-03 0.77 (12.51) 6.70E-05 0.82 (25.19) 2.47E-03 0.92 (25.19) 1.10E-04 0.85 (16.44) 1.90E-03 0.33 (4.70) 1.04E-04 0.48 (6.11) 3.86E-03 0.88 (19.51) 1.63E-04 0.73 (11.02) 2.21E-02 0.79 (13.13) 1.06E-03 0.85 (16.59) employment number of establishments employment number of establishments Building Area Commercial Area Residential Area Retail Area Office Area employment number of establishments employment number of establishments employment number of establishments Table 53. Employment and number of establishments vs. area models. Application of Synthetic Correction This section discusses the theoretical background of syn- thetic correction and explains the need for this procedure to adjust some of the models found in the literature. The objective of this procedure is to correct existing models to account for the differences in FTG patterns for both small and large establishments. In fact, the empirical evidence from the FTG models estimated with establishment-based data indicates that FTG rates depend on business size. As discussed, small establishments tend to generate propor- tionally more trips than large establishments. This leads to a situation in which a constant trip rate underestimates the FTG of small establishments and overestimates the one for large businesses. This poses a problem because several FTG models reported in the literature are in the form of constant trip rates. Figure 12 shows an example of an establishment-based FTG model estimated with data collected by the team. Two different models are shown. The first model is the constant trip rate model that goes through the origin; while the sec- ond is the regression model with an intercept and a slope. The theory of OLS (regression) indicates that both models intercept each other, exactly at the midpoint of the data (at the average values of the independent and dependent variables). The fact that both models intercept at the midpoint (X â , Y â ) provides the basis for a simple correction proce- dure. For those industry segments expected to follow the FTG pattern with an intercept and a slope, and for which constant trip rates are available, the synthetic correction procedure is: * A pivot around the midpoint (*) will improve accuracy Figure 12. Constant freight trip generation rate model vs. regression model with intercept.
69 case study number codification enables the reader to locate these models in the FG/FTG model relational database developed by the team. As shown in Table 54, the corrected models present concep- tually valid coefficients. Only two cases were taken out because, after applying the synthetic correction, a negative slope was found. For the presented models, the intercept shows that small businesses produce between 1 and 4 daily truck trips. In terms of employment, the slope b represents the expected change in daily truck trips associated with a unit change in employment. For the models corrected, a unitary change in employment is associated with an increase in the number of daily truck trips that varies from 0.05 to 2.5. According to pre- vious findings, the models where the employment parameter presents values higher than 0.5 tend to overestimate FTG for large establishments. Seattle Region and Manhattan Grocery Stores: A Comparison of FTG Patterns Across Regions This section gives the results from the analyses con- ducted by company freight trip attraction from a sample of grocery stores in Manhattan and the Seattle Region. In the case of Manhattan, grocery stores from the receivers sample were complemented with the information of the Whole Foods stores. The OLS method was used to estimate freight trip attraction. Table 55 shows the estimated model. As shown, a combined model with intercept and slope was obtained. To compare the FTG patterns across regions, the effect of geographic location of the establishments in the attraction of deliveries was explored. In doing so, the research team com- bined the grocery store data from Manhattan and the Seattle Region and also created variables related to the geographic location (SOUPUGR and interaction term EMP_SPR). The model obtained is shown in Table 56. As shown, the geo- graphic location variables are not statistically significant (low t-value) and the adjusted R2 is low. Furthermore, the estimated attraction models for the Manhattan sample were applied to the Seattle Region gro- cery stores. The industry-based model shown in Table 55 for Manhattan, and the LBCS Function model for grocery stores (0.217 deliveries per employee) shown in Table 49, were used. The estimated delivery trips for the Seattle Region are shown in Table 57. The observed deliveries per day are estimated as the average of the survey results and the two manual counts shown in Table 33. Implications and Directions Key implications and suggested directions are discussed next. 1. Use the constant trip rate and an estimate of the average business size (X â ) to compute the average trip generation (Y â ) as: Y Rate X= ( )4 2. Estimate the number of trips produced by a small (less than five employees) business in the same industry sector (cÂ´). 3. Assuming that cÂ´ is equal to the intercept of the model, compute the slope of the straight line connecting the inter- cept (0, cÂ´) and the average case (X â , Y â ) as: b Y c X= â â²( ) ( )5 The equivalent model is: Y c bX= â² + ( )6 The key element here is that, although an approximation, even a suboptimal assumption of the intercept is bound to perform better than the constant trip rate model. In the case shown in Figure 12, for instance, an assumption that the intercept is equal to one could reduce the total error with respect to the constant trip rate model by almost half. This improvement in the model performance definitely shows the potential of using this technique. Various models found in the literature are corrected in Table 54 with an aim to account for the difference in employment proportionality of business with different sizes. As shown, FTG models that use an employment trip rate benefit from a correction. The first step is to identify the models in the literature that can be corrected and used. Out of the 1,024 FTG models found in the literature and included in the database, only 241 are estimated as a func- tion of employment. The techniques used to estimate these models are OLS or simply trip rates. 107 out of the 241 employment dependent models were computed using OLS technique. In the other cases, FTG is calculated as a rate of employment. For the latter models a synthetic correction is needed. The correction proposed is based on the models by SIC estimated for the NYC case studies. As a result, it is only possible to correct models that have specifications similar to the ones developed. In essence, only the FTG models with employment dependency, and having a freight-related industry segment as categorical factor, or a land use, can be corrected. Moreover, the dimension of the dependent variable must be truck trips; multi-class models were not available and therefore not estimated. Some models were excluded from the correction procedure because they did not specify whether they are for trip attraction or for production. Finally, only 29 out of the 134 trip rates were selected for correction. Table 54 shows the models selected for correction. As shown, the source number and
70 Source Number** Case Study Number** Industry Segment / Land Use Trip Rate Synthetic Intercept (c') Synthetic Slope (b) FTG-SYN-1996-1 RU-1992-1 Retail Trade (SIC 52-59) 1.21 1.89 1.13 FTG-SYN-1976 UT-1977-1 Urban Downtown Retail- Knoxville 2.30 1.89 2.23 FTG-SYN-1976 UT-1977-2 Urban Downtown Retail- Modesto 0.66 1.89 0.59 FTG-SYN-1976 UT-1977-3 Urban Downtown Retail- Rochester 0.12 1.89 0.05 FTG-SYN-1976 UT-1977-4 Urban Downtown Retail- Saginaw 0.30 1.89 0.23 FTG-SYN-1976 UT-1977-5 Urban Downtown Retail 0.40 1.89 0.33 FTG-SYN-1976 UT-1977-6 Urban Wholesale Operations- Knoxville 0.39 1.94 0.35 FTG-SYN-1976 UT-1977-7 Urban Wholesale Operations- Modesto 0.68 1.94 0.64 FTG-SYN-1976 UT-1977-8 Urban Wholesale Operations- Rochester 0.44 1.94 0.40 FTG-SYN-1976 UT-1977-9 Urban Wholesale Operations- Saginaw 0.28 1.94 0.24 FTG-SYN-1976 UT-1977-10 Urban Wholesale Operations 0.30 1.94 0.26 FTG-SYN-1976 UT-1977-11 Urban Wholesale Operations- Dallas 0.60 1.94 0.56 FTG-SYN-1976 UT-1977-12 Urban Truck Terminals- Knoxville 1.35 2.76 1.20 FTG-SYN-1976 UT-1977-13 Urban Truck Terminals- Modesto 1.63 2.76 1.48 FTG-SYN-1976 UT-1977-14 Urban Truck Terminals- Rochester 1.15 2.76 1.00 FTG-SYN-1976 UT-1977-15 Urban Truck Terminals- Saginaw 1.91 2.76 1.76 FTG-SYN-1976 UT-1977-16 Urban Truck Terminals 1.40 2.76 1.25 FTG-SYN-1976 UT-1977-17 Urban Truck Terminals- Dallas 1.42 2.76 1.27 FGTG-BK-1992-1 BA-1982 Manufacturing Firms 0.56* 3.16 0.42 FGTG-BK-1992-1 BA-1982 Construction Firm 1.92* 2.16 1.82 FGTG-BK-1992-1 BA-1982 Wholesale/Retail/Dealer Firm 2.62* 2.27 2.49 FTG-SYN-1996-1 RU-1992-1 Retail Trade (SIC 52-59) 1.21 3.37 0.98 FTG-SYN-1995-1 BR-1977-1 Grocery Wholesale Establishment 0.56 1.71 0.46 FTG-SYN-1995-1 BR-1977-1 Hardware Wholesale Establishment 0.32 3.99 0.09 FTG-SYN-1995-1 BR-1977-1 Other Wholesale Establishment 0.48 2.27 0.35 FTG-SYN-1995-1 BR-1977-1 Total Wholesale Establishments 0.50 2.27 0.37 FTG-SYN-1995-1 BR-1977-5 Furniture Establishments 0.48 2.17 0.41 Production Models Attraction Models Notes: (1) *Rate was converted to daily trips dividing by 5 the original weekly rate (2) **Models can be found in the Database of Task 7 using these codes Table 54. Synthetic correction applied to models reported in the literature. Variable Name Coefficient t-value Regression model Intercept CONSTANT 5.731 2.133 Total employment USEDEMPL 0.087 2.726 n (establishments) 31 RMSE 4.92 R2 0.204 Adjusted R2 0.177 Table 55. Manhattan grocery stores freight trip attraction model. Variable Name Coefficient t-value Regression model Intercept CONSTANT 5.767 2.325 Total employment USEDEMPL 0.087 2.938 Geographic Location SOUPUGR -9.505 -0.6468 Employ. Seattle Region EMP_ SPR 0.811 0.528 n 38.00 R2 0.030 Adjusted R2 0.006 Table 56. All grocery stores freight trip attraction model (NYC & Seattle Region).
71 Y â is the average of deliveries observed in the external vali- dation dataset, YËi is the number of deliveries estimated using FTG models for each establishment on the external validation dataset, and YË â is the average of deliveries estimated using FTG models for the external validation dataset. A model predicting observed data perfectly produces a straight line plot between observed Yi and predicted values YËi, and a correlation coefficient of 1.0. Conversely, linear corre- lation coefficients of 0 suggest no linear association between observed and estimated values. Table 58 and Table 59 present the results of this procedure. As shown, for SIC 52 and SIC 56 (building materials and apparel/accessory stores) predicted and observed val- ues for FTG have a linear correlation coefficient very close to one. For these industry segments, the models work the best according to the external validation. For food stores (SIC 54); retail trade (NAICS 44); retail trade (LBCS Function: Retail, LBCS Activity: goods); and grocery stores (LBCS Function: grocery) there is a linear correlation, which shows that these models are externally valid. In terms of classifications related to restaurants (SIC 58; NAICS 72; LBCS Function: Food; and LBCS Activity: Res- taurants) the Pearson Product-Moment Correlation Coef- ficients show that there is not a strong linear correlation between observed and estimated number of deliveries received. Therefore, restaurants and food service establish- ments are industry segments that need a closer examination when applying the models developed in this project to differ- ent contexts. An a priori conclusion might be that restaurants External Validity of FTG Models The objective of an external validation is to assess the predictive ability of a statistical model. The motivation is that statistical methods make use of fitting routines that can lead to over-fitting or spurious fitting. In these cases, the FTG models may fit the data used for estimation, but they might not be predictive of FTG in a different context. To validate the models estimated in the previous section, a new collection data effort was performed in the Capital Region of New York State using the questionnaire form provided in Appendix H. The establishments surveyed pro- vided information about daily FTG, employment, industry segment, service trips, and number of vehicles operated from the establishment, among others. Using the data col- lected from the establishments targeted, only the attrac- tion models using SIC, NAICS, and LBCS as categorical factors can be validated. (The models using NYC Zoning Resolution as categorical factors cannot be validated with this data because NYCZR land use classification is exclusive for NYC.) The goodness-of-fit (GOF) measure implemented to assess the statistical model performance on validation data was the Pearson Product-Moment Correlation Coefficients. This correlation coefficient denoted by r, measures the linear association between two variables Y1 and Y2 that have been measured on ratio scales. The Pearson Product-Moment Correlation Coefficients is defined as: r Y Y Y Y Y Y Y Y i i i i = â( ) â( ) â( ) â( )ï£®ï£° ï£¹ï£» â ââ Ë Ë Ë ( ) 2 2 1 2 7 Where, Yi is the observed number of deliveries observed for each establishment on the external validation dataset, QFC Wallingford 80 19 13 17 QFC Kirkland 70 14 12 15 QFC Mukilteo 70 15 12 15 QFC Capitol Hill 100 14 14 22 QFC Lynnwood 72 13 12 16 Albertsons Kent 60 14 11 13 PCC Issaquah 95 22 14 21 RMSE 4.28 3.32 Estimated del/day using LBCS Function: Grocery Observed del/day Estimated del/day using SIC model Store and Location Emp. Table 57. Estimated freight trip attraction for Seattle region grocery stores based on Manhattan models. SIC/NAICS Description Obs r SIC 52 Building Materials 5 0.94 SIC 54 Food Stores 8 0.85 SIC 56 Apparel and Accessory Stores 8 0.94 SIC 58 Eating and Drinking Places 5 0.47 NAICS 44 Retail Trade 21 0.85 NAICS 72 Accommodation and Food 5 0.47 Table 58. External validation using Pearson product-moment correlation coefficientsâ SIC & NAICS. LBCS Description Obs r F. Retail Function Retail 13 0.76 F. Grocery Function Grocery 8 0.85 F. Food Function Food Service 5 0.47 A. Goods Activity Goods 21 0.63 A. Rest Activity Restaurants 5 0.47 Table 59. External validation using Pearson product-moment correlation coefficientsâLBCS.
72 important role as categorical factors. The question remaining is: Should one create new models by state, or estimate models based on land use characteristics (such as store location), ignor- ing transferability concerns? Accordingly a RMSE analysis was performed to assess the two alternatives. Table 61 shows the resulting estimation errors for the observations when apply- ing the best model found by state and by store location. As shown, attraction models using âland useâ as a categorical factor perform slightly better than the ones using âstate.â This is an interesting finding because, although the state where a store is located was found to be statistically signifi- cant, the best results were found using store location (land use) as the key factor. In essence, this suggests that land use models have a good performance, even when no distinction is made among states. In addition, results for the comparison of freight trip attraction of grocery stores (see Table 56 and Table 57) have great implications for FTG modeling. They also sug- gest the transferability of FTG models; the sample, grocery stores exhibit similar FTG patterns across regions. Results also indicate that the models provide good estimates con- sidering the low RMSE found for the sample. As previously shown in Table 57, the LBCS model performs better than the industry-based model. This suggests that implementing a standard land use classification system such as the LBCS could improve FTG modeling. However, past performance is not necessarily a guarantee of future performance, there- fore further research is needed. The example discussed in the Seattle Region and Manhattan grocery stores case shows the potential benefits of applying LBCS models for FTG purposes, but it is based on a small sample and should not be generalized. have different logistic choices in big cities like NYC than in other urban/rural contexts. Although the Pearson Product-Moment Correlation Coefficients provide metrics to measure objectively the per- formance of the models developed, measuring the perfor- mance of a benchmark model can set the basis for compari- son. The Quick Response Freight Manual (U.S. Department of Transportation 1996) was chosen as the benchmark. The methodology proposed in this Manual uses SIC as a cat- egorical factor, and estimates FTG using employment rates from a study in Phoenix, Arizona. While the approach in the Quick Response Freight Manual focuses on traffic gener- ated by establishments, the models developed herein dif- ferentiate between freight trips attracted and freight trips produced. This refinement allows a different specification for productions and attractions, even when the establish- ments are in the same industry sector. For example, for building materials stores (SIC 52) trips attracted are better estimated using an employment rate, however trips pro- duced are better estimated using a constant generation per establishment. This characteristic (of the models developed in this project) allows smaller aggregation errors. Table 60 compares the performance of the models developed in this project to the benchmark (Phoenix Models). The metric used for comparison is the Pearson Product-Moment Cor- relation Coefficients. As shown, for most of the industry sectors, both models have a similar performance. However, for the building mate- rials stores (SIC 52) NYC Models perform significantly bet- ter. As previously explained, for this type of establishment trips attracted are better estimated using an employment rate, while trips produced are better estimated using a con- stant generation per establishment. This difference is not considered in the Phoenix case study, producing a loss in the accuracy of the estimates. Transferability of FTG Models Considering the results and models found for the furni- ture store chain, both state location and store location play an SIC Description Obs NYC Models: r Phoenix Models: r SIC 52 Building Materials 5 0.94 -0.66 SIC 54 Food Stores 8 0.85 0.85 SIC 56 Apparel and Accessory Stores 8 0.94 0.91 SIC 58 Eating and Drinking Places 5 0.47 0.47 Table 60. Quick Response Freight Manual performance metrics-Pearson Product-Moment Correlation Coefficients. Classification System Deliveries Attracted Pallets Attracted State 0.392 4.692 Store location 0.386 4.517 RMSE Table 61. Total estimation error for each classification system (state and store location).
73 attraction, and NAICS provided better models for freight trip production. The findings indicate that the results of FTG estimates are somewhat impacted by the type of industry classification system used in the analysis. The overall results indicate that the replacement of SIC by NAICS codes would lead to more accurate models for freight production models. In the case of freight attraction, SIC is more suitable to capture these freight behaviors. A 3-digit aggregation in NAICS would reduce the internal heterogeneity of the resulting group- ings and improve NAICS performance overall, specifically in the area where it is currently lacking, freight attraction. It is important to note that the results shown here are representa- tive of this specific dataset. Chief Findings for Land Use Based FTG Modeling As discussed throughout the case chapter, the best models were selected based on t-statistics and RMSE. For 73% of the NYCZR (local) models, a constant coefficient produces the best FTG models; for 18% of these models, FTG depends on employment; and for 9% of these models, the best FTG esti- mate is a combined model, with a constant coefficient plus an employment term. For 67% of the LBCS Function models, a constant coefficient also produces the best results; for 11% of these Function models, FTG depends on employment; and for 22% of these Function models, FTG is a combined model, with a constant coefficient plus an employment term. For the LBCS Activity models, this breakdown is 80%, 0%, and 20%, respectively. In terms of RMSE, the analysis indicates that NYCZR is somewhat better than LBCS, as the RMSE (4.21) is lower than the one for LBCS (4.53). What is most surpris- ing is that FTG is a constant value in a preponderance of the best performing models, and employment, therefore, plays no role. For the small number of models where employment is an important factor, MCA was used. The RMSE of the MCA models is 4.43 for NYCZR and 4.60 for LBCS. This indicates that MCA performs slightly better than regression analysis for almost every land use classification where FTG is employ- ment dependent, even though it is a very small difference with respect to the overall error. This is not surprising because MCA has more degrees of freedom than regression analysis, at the expense of higher data requirements. When the NYCZR and LBCS models are compared to the ITE trip rates, results indicate that NYCZR and LBCS models give more accurate FTG estimates than ITE rates. When con- sidering the complete sample, the total RMSE of NYCZR and LBCS models are about 30% lower than ITEâs. The local and standardized national land use classifica- tion code models clearly provide a better alternative to ITE Use of SIC System and NAICS for FTG Modeling The SIC system and the NAICS differ in the level of detail that each offers. SIC uses a four-digit code while NAICS employs a six-digit code. The increase in the number of digits from SIC to NAICS allows NAICS to cover a larger number of sectors for a more disaggregated industrial classification system. This characteristic of NAICS provides the advantage of a more detailed system and more flexibility when catego- rizing subsectors. These differences were evidenced in the estimations of the FTG models for both freight attraction and productions. Results from the linear regression analysis indicate that some differences in detail between the two classifications systems may reflect different types of models. While all the other industry groupings in the analysis derived from the same model types, the retail industry differed. The freight attraction (receivers) estimation showed that the retail indus- try estimated a constant FTG per establishment model (Type S) when using SIC, but when using NAICS the estimation resulted in a combined model (Type C). Therefore, retail trade is not dependent on business size when using SIC, but business size is a factor when using NAICS. Freight production estimation also revealed differences in model types for the retail trade industry. SIC derived a Type C model, consistent with freight attraction results, but NAICS resulted in a FTG rate per employee (Type E) model, which is completely dependent on business size. The results indicate some disparities between SIC and NAICS models as well as differences between freight attraction and production models when using NAICS. The difference in model types between SIC and NAICS reflects the differences in details at the two-digit level between SIC and NAICS. The dissimilarity for NAICS between freight attraction and production may result from: (1) differences in the number of observations in each sample; or (2) NAICS captures the distinction between freight attraction and pro- duction within the retail industry. Regardless of the reasons, the results indicate that disparities exist when using SIC and NAICS in freight transportation modeling. Multiple Classification Analysis (MCA) was used to estimate trip rates for the industries that were found to be employment dependent (Type E & C models) from the lin- ear regression analyses. The estimated trip rates were used to calculate the RMSE, which serves as the measure of fit of the model, for each two-digit SIC or NAICS code as well as the total RMSE for the entire carriers and receivers samples for both industry classification systems. The models with the best fit using the individual two-digit codes fluctuated between the two systems. However, the overall RMSE val- ues indicated that SIC offered better models for freight trip
74 mance of model BR-1977-1, which is the only model devel- oped for grocery stores, and the corresponding corrected model. Table 63 shows the FTG estimated using the origi- nal model, and the corrected model. As shown, the RMSE is 20% lower for the corrected model. Although this is only one example, it is consistent with the theory and shows how the proposed synthetic correction helps to improve the accuracy of the models for FTG estimation. The synthetic correction procedure can be applied to any employment trip rate where the dependent variable is in terms of truck trips, and the categorical factor is either an industry segment or a land use compatible with the ones described in the NYC case study. The corrected models were added as new models in the FG/FTG model relational data- base that was developed. Summary Chapter 6 remarks are as follows: â¢ Transferability: Some industry segments such as grocery stores and furniture stores exhibit similar FTG patterns across regions. â¢ The use of the SIC system and the NAICS for FTG mod- eling: NAICS is best suited to model freight production. On the other hand, SIC better captures freight attraction behaviors. â¢ In Most Cases, FTG does not depend on employment, as it is a constant per establishment: For most of the land uses studied, the best performing models consider FTG as a constant. The same was found for FTG models based on industrial classifications. This finding suggests that as opposed to FG, FTG does not necessarily increase with business size. â¢ The land use models estimated with the Land-Based Classification Standards perform better than the corre- sponding models from the ITE Trip Generation Manual: The models based on land use estimated for this project lead to more accurate FTG estimates than the ITE rates. â¢ Use of Multiple Classification Analysis (MCA): MCA performs slightly better than regression analysis for trip rates for the following reasons: they give more accurate estimates of freight trip attraction; they cover a wider range of land use classifications; and they are developed exclusive- ly for freight trip attraction. Future research will focus on freight trip production. The main practical limitation of this research is that the findings are based on a dataset from NYC, specifically Man- hattan and Brooklyn. As suggested by the external validity and the transferability findings, the FTG models developed can be extrapolated to similarly situated cities or parts of cities that may have similar freight trip generation patterns. However, the results found are valid only in the context of this dataset. Comparison Between LBCS, NYCZR, SIC and NAICS This section discusses the performance of the different FTG models that were estimated. The models were applied to the sample data, and the estimation errors were computed. The RMSE measure was used to identify the most appropriate model for each industry, or group of industries. The analysis was also expanded to identify the best model for each land use category. Detailed tables containing the estimation errors can be found in the supplemental materials and appendices of the Task 11: Case Studies report, available online10. The best way to estimate the number of deliveries per establishment is using an industrial classification or land use as the first criteria. According to the category selected, trip generation can depend on business size or not. If it does not depend on business size, the analyst can use a constant num- ber of deliveries per establishment. If it depends on employ- ment, a trip rate table based on MCA can be used. After identifying the best model for each industry or land use, the next step was to estimate the total error for the pooled data using these individual models to the correspond- ing observations. As shown in Table 62, SIC models are better than NAICS and land use models, with land use based models giving less accurate results. In addition, the results show the improvements of using individual models; however, the ana- lyst should consider a trade-off between the quality obtained and the efforts to estimate needed inputs for the different models. Synthetic Correction Validation Because of data limitations, it is not possible to evaluate how synthetic corrections improve accuracy for every model. However, by using data for grocery stores in the Seattle Region (Section 3.5), it is possible to compare the perfor- 10Available online at http://transp.rpi.edu/~NCFRP25/downloads.shtml Classification system RMSE SIC 3.332 NAICS 3.566 New York Land Use (NYCZR) 4.205 Land Based Classification System (LBCS) 4.529 Table 62. Total estimation error when using the best land use model for each industry or land use.
75 â¢ Implementing Synthetic Correction increases the per- formance of FTG models: This procedure consists of correcting existing employment trip rates to reflect the differences in FTG patterns for small establishments and large establishments. The application of this procedure produced a significant decrease in the estimation errors of FTG models. most land use classification, where FTG is employment dependent. â¢ FTG models based on industrial classifications (SIC and NAICS) are the most accurate: Using industrial classificationâSIC and NAICSâcodes as categorical factors leads to better models than using the land use classifications, e.g., NYCZR and LBCS. FTG-SYN-1995-1* BR-1977-1* QFC Wallingford 80 19 45 39 QFC Kirkland 70 14 39 34 QFC Mukilteo 70 15 39 34 QFC Capitol Hill 100 14 56 48 QFC Lynnwood 72 13 40 35 Albertsons Kent 60 14 34 29 PCC Issaquah 95 22 53 45 RMSE 28.73 22.52 Store and Location Emp. Observed del/day Corrected Version of the model (using Synthetic Correction) *Models can be found in the Database of Task 7 using these codes Table 63. FTG estimation comparison between original model BR-1977-1 and synthetic correction for grocery wholesale establishments.