Freight Generation Models

Comparative Analysis of Regression Models

and Multiple Classification Analysis

Carlos Bastida and José Holguín-Veras

This paper conducts a comparative analysis of two alternative approaches in the next two decades (3). Adding to the complexity of the prob-
to freight generation modeling: ordinary least square (OLS) and cross lem is that the bulk of the freight flows are carried by trucks that
classification. OLS models were estimated to identify a functional rela- move 69% of total tonnage and 71% of total value shipped in the
tionship between the number of freight deliveries per day and a set of United States (4). Despite this fact, little attention has been paid to
company attributes used as independent variables. Cross-classification freight generation (5). As a result, freight transportation forecasting
techniques aim at identifying a classification structure that provides a still remains a complex and underdeveloped activity that necessi-
good representation of the freight generation process. To that effect, tates the development of methodologies that take into account the
multiple classification analysis was used to identify groups of independent unique characteristics of freight (6).
variables explaining freight generation, which provided the basis for The main purpose of this paper is to contribute to the study of
constructing cross-classification tables. In both cases freight generation freight demand through the estimation of freight generation models
is explained as a function of company attributes. The model estimation using disaggregated data at the establishment level. The data cor-
process used data obtained from commercial establishments located in respond to approximately 700 commercial establishments located in
Manhattan and Brooklyn, New York. More than 190 different variables Manhattan and Brooklyn, New York.
were tested as predictors for the number of deliveries received or carried Throughout the analyses, the independent variable used is the
per day. Six linear regression models found to be statistically significant number of deliveries received or transported in a typical day. Two
and conceptually valid are discussed. Establishment attributes such as important observations are warranted. The first is related to the dis-
industry segment, commodity type, facility type, total sales, and number tinction between “delivery” and “truck trip.” The truck trip is the
of employees were found to be statistically significant. The OLS models physical movement of the vehicle from an origin to a destination,
indicated that industry segment and commodity type are strong predictors and a delivery refers to the transportation of a shipment to an estab-
of freight generation. It was also noticed that many of these variables lishment. For that reason, a single truck trip could be used to make
played a significant role when interacting with economic variables such multiple deliveries to different establishments. As a result, the num-
as total sales or employment. Twelve different groups of independent ber of deliveries could be larger than the number of truck trips although
variables predicting freight generation were found to be significant as there are no data to estimate by how much. The second observation
part of the cross-classification models. Results indicate that as in the case is that the number of deliveries—although providing an indication
of regression models, commodity type, industry segment, and employment of the traffic generated—does not say much about the amount of
are strong predictors for freight generation.
commodities generated as it does not take into account the shipment
size (which was not collected in the surveys). Because of this limi-
In the attempt to find solutions to transportation issues, travel demand tation of the data, it may be entirely possible that a large company
models are routinely used and improved. As a result, passenger generates less traffic than a small one simply because it handles larger
demand modeling has evolved from the traditional four-step model shipment sizes.
to more sophisticated methodologies, for example, activity based
(1). Unfortunately, in most applications the main focus is only on
passenger demand, paying little or no attention to freight. This atti- LITERATURE REVIEW
tude is not consistent with the importance of freight. For instance,
The use of linear regression techniques for passenger trip generation
between 1975 and 1997 the amount of air and truck intercity freight
has a long history in transportation. In their review, Douglas and
transportation grew by about 60% (2). The freight analysis frame-
Lewis (7) highlight that violations of the ordinary least square (OLS)
work indicates that volumes of freight are expected to grow by 70%
assumptions could lead to inaccurate parameters, especially when
using aggregated data. To avoid such problems they recommend the
use of disaggregated models.
C. Bastida, Parsons Brinckerhoff, 506 Carnegie Center, 2nd Floor, Princeton,
NJ 08540. J. Holguín-Veras, Department of Civil and Environmental Engineering, Although the number of publications on passenger generation
Rensselaer Polytechnic Institute, JEC 4033, 110 8th Street, Troy, NY 12180. is enormous, the number of papers on freight is very small. Ogden
Corresponding author: C. Bastida, [email protected]. developed freight generation models with regression techniques
using data from Melbourne, Australia (8). Chatterjee et al. estimated
Transportation Research Record: Journal of the Transportation Research Board,
No. 2097, Transportation Research Board of the National Academies, Washington,
freight trip generation rates as a function of 10 different land uses (9).
D.C., 2009, pp. 51–61. Zavattero and Weseman estimated a series of linear regression mod-
DOI: 10.3141/2097-07 els for truck trip generation in the Chicago area (10). Industry segment

51
52 Transportation Research Record 2097

and type of commodity carried were discussed by Middleton et al. Establishment Size and Type
(11). Al-Deek et al. estimated truck trip generation models for the port
of Miami (12). Al-Deek compared artificial neural networks and The establishments were classified into single establishments, branches,
OLS to develop freight models at seaports (13). NCHRP Synthesis and headquarters. Most of the establishments in Manhattan, with more
298 provided a comprehensive synthesis of freight generation (4). than 60% and 78% for receivers and carriers, respectively, operate
Holguín-Veras and Lopez-Genao estimated truck trip generation as a single facility. Brooklyn receivers and carriers present a similar
models as a function of the number of container boxes handled and distribution, with single facilities being the majority. The bulk of
container capacity, and they estimated trip rates as a function of the receivers in Manhattan (59%) have on average 5 to 14 employees;
area of the container terminal (6). Stopher and McDonald proposed whereas the carriers are evenly distributed with a slight majority
using multiple classification analysis (MCA) as a calibration tech- (21%) having between 24 and 34 employees. Similarly, the majority
nique capable of overcoming the disadvantages associated with cross- of receivers in Brooklyn have on average 5 to 14 employees. Finally,
classification analyses (14). They used MCA to determine which Brooklyn carriers present a similar proportion (20%) of employees,
independent variables provided a better description of the trip-making having between 24 and 34, and 34 and 55.
process, and then they applied this methodology to a case study.
Guevara and Thomas (15) reviewed the methodology described by
Stopher and McDonald (14), discussing that it might overestimate the Commodity Types
future number of trips if the number of observations by category is not The Manhattan receivers were found to transport 24 different com-
exactly the same. Their analysis indicates that the MCA method of modity types. Food products represent the largest group, with 32%;
ordinary least squares (16) might overcome this problem. textiles and clothing capture 22%; and jewelry and art approximately
8%. Similarly, Manhattan carriers dealing with food products repre-
sent the majority with 16%; furniture and household goods various are
DESCRIPTION OF THE DATA next, with 12% and 10%, respectively. These findings are consistent with
land uses in Manhattan, residential and commercial.
The data were collected by the Rensselaer Polytechnic Institute as Not surprisingly, food products were found again to represent the
part of a research project conducted for the New York State Depart- majority commodity type received in Brooklyn, with 21%; followed
ment of Transportation. As part of the project disaggregated data at by textiles with 12%; metal with 10%; and metal, construction, tools,
the establishment level were collected. The surveys targeted carriers, and hardware with 8%. This is found to be conceptually correct
receivers, and intermediaries (companies that receive and ship cargo). because a large number of establishments using these commodities
Approximately 700 commercial establishments located in Manhattan are located there.
and Brooklyn were surveyed. Initially four data sets were constructed: Finally, as before, food products represent the most frequently
Manhattan carriers, Manhattan receivers, Brooklyn receivers, and carried commodities, with 17%; textiles and furniture are next with
Brooklyn carriers. Later, two additional data sets were created with 10% each. This is congruent with the predominant land uses found
the pooled data: Manhattan and Brooklyn receivers, and Manhattan in Brooklyn, residential and industrial.
and Brooklyn carriers. The resulting six data sets were analyzed and
coded for analyses purposes. The rest of the section provides a succinct
description of the data. For a comprehensive descriptive analysis of the ORDINARY LEAST SQUARE MODELS
data see Holguín-Veras et al. (17).
In this section the results of the OLS models are discussed. For
analysis purposes categorical variables were coded as binary vari-
Industry Segments ables. This was the case for industry segment, commodity type, and
facility type. Only linear models were estimated during the modeling
The Manhattan sample comprises 180 receivers and 192 carriers. process.
The companies interviewed were classified using standard indus- The independent variable was the number of deliveries per estab-
trial classification (SIC) codes. The largest group of receivers lishment per day. The model selection took into account statistical
consists of eating and drinking places and the wholesale trade significance and conceptual validity criteria. To test OLS basic
nondurable goods sector, each with 23% of the observations. They assumptions the analysis of variance and residual plots were used.
are followed by wholesale trade durable goods, with 19%. On the
carrier side, half (50%) of the carriers are in the motor freight trans-
Results for Pooled Data
portation and warehousing sector and one quarter (25%) are in the
(Manhattan and Brooklyn)
wholesale trade durable goods. This indicates that almost 50% of
the establishments are companies providing only transportation ser- In this section the Manhattan and Brooklyn pooled data were used to
vices, and the other 50% are private establishments that carry their estimate linear regression models and examine what variables trigger
own cargo. freight generation for both boroughs.
The Brooklyn sample comprises 200 receivers and 139 carriers.
About 14% of the receivers are in the wholesale trade durable goods
industry segment, followed by miscellaneous retail and wholesale Receivers
trade nondurable goods, with 11% and 10%, respectively. Not surpris-
ingly, the bulk of carriers are classified as motor freight transporta- Table 1 shows the best freight generation model obtained during the
tion and warehousing, which represents more than half the total modeling process. The model expresses the number of deliveries
percentage, with 52%. These are followed by transportation services received per day as a function of the company attributes. As shown,
with 9% and wholesale trade durable goods. its goodness of fit is relatively low (0.43). It can be observed that the
Bastida and Holguín-Veras 53

TABLE 1 Receivers Regression Model (Manhattan and Brooklyn)

Variable Name Coefficient t-Value

Regression model, intercept CONSTANT 1.4951 22.000

Industry segment
Building materials, hardware, garden supply, HARDW 2.0049 4.387
and mobile homes
Apparel and accessory stores APPAR 1.6478 4.731
Industry segment and number of employees
Food stores FSTEMP 0.0479 3.763
Eating and drinking places EATEMP 0.0404 7.227
Wholesale trade durable goods DUREMP 0.0397 6.625
Wholesale trade nondurable goods NONEMP 0.0292 6.734
Fabricated metal products FABEMP 0.0290 3.399
Industry segment and sales per year
Home furniture, furnishings, and equipment stores HOMSAL 8.55E-07 5.7653
Miscellaneous retail RETSAL 3.72E-07 5.1721
Food stores FOOSAL 1.08E-07 5.0326
R2 0.452
Adjusted R2 0.430

industry segments of building materials, hardware, garden supplies Carriers

and mobile homes, and apparel and accessory stores have large coef-
ficients in comparison with the intercept. This makes sense because This section discusses the trip generation models estimated for carri-
these groups are expected to generate a significant number of deliv- ers (see Table 2). The model’s statistical goodness of fit is better than
eries. Furthermore, this is consistent with the fact that Manhattan and in the previous case, with 59.40% of all variation being explained. All
Brooklyn have an old infrastructure in which there is always new coefficients of the independent variables are positive and statistically
construction, reconstruction, and maintenance that require building significant. The coefficient of the intercept is lower than the coefficient
materials and hardware supplies. of the receivers although it is still high. The coefficient of number of
In contrast, some industry segments are found to be significant truck drivers, which also can be used as an indicator of employment,
only when interacting with socioeconomic variables, such as num- clearly indicates that this variable is an important factor driving trip
ber of employees. All variables correspond to the most important generation for carriers.
industry segments currently in the area. Three additional variables Machinery was the only commodity type found to be significant.
were significant. They represent the interaction between the industry Other commodity types such as jewelry and art, petroleum and
segment and number of sales per year for various industry segments. coal, furniture, paper, metal, and alcohol were significant but only
None of the variables describing the type of commodity received and if they were interacting between number of employees or amount
type of facility were found to be statistically significant in isolation of sales per year. Wood and lumber is significant when interacting
from others. with the total number of sales per year. Likewise, “single” facilities

TABLE 2 Carriers Regression Model (Manhattan and Brooklyn)

Variable Name Coefficient t-Value

Regression model
Intercept CONSTANT 1.1243 13.078
Number of truck drivers DRDEL 0.0247 2.974
Type of commodity carried
Machinery COM14 1.0043 3.574
Type of facility and number of employees
Single SINEMP 0.0057 2.229
Type of commodity and number of employees
Jewelry and art COM20EMP 0.2288 4.896
Petroleum and coal COM10EMP 0.0647 4.663
Furniture COM7EMP 0.0357 6.155
Paper COM9EMP 0.0290 6.330
Metal COM13EMP 0.0235 6.000
Alcohol COM4EMP 0.0116 9.040
Type of commodity and sales per year
Wood and lumber COM8SAL 4.25E-08 8.194
R2 0.611
Adjusted R2 0.594
54 Transportation Research Record 2097

TABLE 3 Receivers Regression Model (Manhattan)

Variable Name Coefficient t-Value

Regression model, intercept CONSTANT 1.5160 14.484

Type of commodity received, medical supply and drugs COM22 0.8817 2.317
Industry segment and number of employees
Building materials, hardware, garden supply, HAREMP 0.2659 5.676
and mobile homes
Eating and drinking places EATEMP 0.0376 5.938
Wholesale trade nondurable goods NONEMP 0.0260 4.817
Industry segment and number of sales
Food stores FSTSAL 2.33E-06 3.225
Home furniture, furnishings, and equipment stores HOMSAL 1.14E-06 5.075
Miscellaneous retail RETSAL 1.24E-07 2.639
Type of commodity and number of employees, jewelry and art M20EMP 0.1003 3.710
R2 0.462
Adjusted R2 0.429

became statistically significant when interacting with the number of number of sales. This obviously indicates that socioeconomic vari-
employees. ables, namely number of employees and number of sales, have a trip
generation effect only for some industry segments.
The industry segments with the highest coefficient value are
Results for Manhattan building materials, hardware, garden supplies, and mobile homes.
Overall, the industry segments that were found to be statistically sig-
In this section the regression models for establishments located in nificant in the Manhattan area are eating and drinking places, whole-
Manhattan are presented. Two models are shown. As in the case of sale nondurable goods, jewelry and art, and miscellaneous retail,
the pooled data both models were found to depend on variables such which are known to be overrepresented in Manhattan. Furthermore,
as commodity type, employment, and total number of sales per year. the number of sales per year confirms the statistics showing that
such industries generate and attract a large number of trips annually
to Manhattan.
Receivers

As shown in Table 3 the trip generation model for Manhattan receivers Carriers
has a relatively low statistical goodness of fit (0.43). The intercept is
positive and has a high coefficient value. Overall, the coefficients of Table 4 shows the trip generation model for Manhattan carriers. As
the model are low. Medical supplies and drugs were the only com- shown in the table, this model has the best statistical goodness of
modity types that were found to be significant without interacting with fit among all the regression models estimated, with 79.60% of the
another variable. Jewelry and art commodities were significant when variance being explained. This is also reflected in the value of the
interacting with the number of employees. Six industry segments coefficient of the intercept, which is the lowest of all models.
have an influence on trip generation, three of them interacting with The variables representing the number of vehicles in the fleet and
the number of employees, and the other three interacting with the the number of truck drivers in the company were found to be signif-

TABLE 4 Carriers Regression Model (Manhattan)

Variable Name Coefficient t-Value

Regression model
Intercept CONSTANT 0.4750 4.270
Number of vehicles on the fleet VEHFL 0.0860 12.601
Number of truck drivers DRDEL 0.0589 4.731
Type of commodity carried, household goods and various COM16 0.7854 3.608
Type of facility and number of employees, single SINEMP 0.0075 2.280
Type of commodity and number of employees
Jewelry and art COM20EMP 0.3048 6.895
Furniture COM7EMP 0.0380 7.117
Wood and lumber COM8EMP 0.0328 6.090
Paper COM9EMP 0.0227 4.523
Metal COM13EMP 0.0207 6.039
R2 0.809
Adjusted R2 0.796
Bastida and Holguín-Veras 55

TABLE 5 Receivers Regression Model (Brooklyn)

Variable Name Coefficient t-Value

Regression model, intercept CONSTANT 1.4774 16.283

Industry segment
Food stores FSTORE 2.5226 6.896
Home furniture, furnishings, and equipment stores HOME 1.0304 2.551
Type of commodity received
Furniture COM7 1.4610 3.618
Office supplies COM19 1.1377 2.235
Industry segment and number of employees
Wholesale trade durable goods DUREMP 0.0558 4.916
Wholesale trade nondurable goods NONEMP 0.0373 4.609
Fabricated metal products FABEMP 0.0360 4.066
Industry segment and sales per year
Eating and drinking places EATSAL 3.33E-06 3.792
Miscellaneous retail RETSAL 3.77E-07 3.201
Apparel and accessory stores APPSAL 2.63E-07 5.579
Food stores FOOSAL 1.09E-07 5.245
R2 0.577
Adjusted R2 0.540

icant, reinforcing the previous findings that indicate that employ- As in previous sections the industry segment (food stores and home
ment is a good predictor for freight generation. However, only one furniture, furnishings, and equipment stores), commodity type,
industry segment was found to be significant, that is, household employment, and total number of sales per year were found to be
goods. In addition, jewelry and art, furniture, wood and lumber, significant. Food stores have a high coefficient value with respect to
paper, and metal were found to be significant when interacting with the intercept, indicating a strong impact on freight generation. More-
number of employees. over, this variable was found to be significant only when interacting
with the total number of sales per year. This industry segment is also
being reflected by the eating and drinking places segment, which
Results for Brooklyn also was found to be significant when interacting with total number
of sales per year. The binary variables representing furniture, office
In this section, the models for Brooklyn receivers and carriers are supplies, wholesale trade durable goods, wholesale trade nondurable
described and analyzed. Similarly, variables such as commodity goods, and fabricated metal were found to be significant.
type, number of employees, and total number of sales per year were
found to be strong predictors of freight generation.
Carriers

Receivers Among all the models estimated, the model for Brooklyn carriers
(Table 6) has the smaller number of explanatory variables. All inde-
The model for Brooklyn receivers is shown in Table 5. All indepen- pendent variables are positive and statistically significant. The
dent variables are positive and statistically significant. The coeffi- coefficient of the intercept is relatively low and statistically signif-
cient of the intercept is relatively high and statistically significant. icant. In this model only the carriers of machinery were found to be

TABLE 6 Carriers Regression Model (Brooklyn)

Variable Name Coefficient t-Value

Regression model, intercept CONSTANT 0.5748 3.928

Type of carrier, motor freight transportation and warehousing FRGHT 0.7847 4.379
Type of commodity carried, machinery COM14 3.8848 8.153
Type of carrier and employees, shipper SHIPEMP 0.0267 5.587
Type of facility and number of employees
Branch BRAEMP 0.0187 3.839
Industry segment and sales per year
Miscellaneous retail RETSAL 6.80E-06 4.215
Type of commodity and number of employees
Petroleum and coal COM10EMP 0.0766 5.711
Don’t know or other COM23EMP 0.0057 1.969
R2 0.691
Adjusted R2 0.665
56 Transportation Research Record 2097

TABLE 7 Number of Deliveries Received per Day per Establishment (Manhattan receivers)

Commodity Type

Agriculture,
Number of Forestry, Nonalcoholic Textiles and Wood and
Employees Fishing Food Beverages Alcohol Tobacco Clothing Furniture Lumber Paper Chemicals

0–15 1.1 3.5 2.1 1.1 1.1 2.5 6.6 5.6 1.1 1.1
15–30 2.4 4.8 3.4 2.4 2.4 3.8 7.9 6.9 2.4 2.4
30–45 0.9 3.4 1.9 0.9 0.9 2.4 6.4 5.4 0.9 0.9
45–60 2.4 4.9 3.4 2.4 2.4 3.8 7.9 6.9 2.4 2.4
60–75 1.8 4.2 2.8 1.8 1.8 3.2 7.3 6.3 1.8 1.8
> 75 6.5 9.0 7.5 6.5 6.5 8.0 12.0 11.0 6.5 6.5

NOTE:
Multiple R-square adjusted 0.247 Average 5.0 Min 0.9
Multiple R-adjusted 0.497 Median 3.4 Max 28.0
Trip rates for commodities not reported, here referred to as DK/Ref, were estimated with a relatively small number of observations. Hence, caution is advised if used.

significant without interacting with another explanatory variable. eters are important; the first defines the proportion of variance explained
Other variables and attributes are significant when interacting with by the predictor, and the second is an approximate representation of
total number of sales per year or the total number of employees, the relationship between a predictor and the dependent variable (16).
owing to the intrinsic characteristics of industry segment captured At the end of the process, 10 groups of independent variables were cho-
by the commodity type. sen. In all cases the dependent variable is the number of deliveries
received or carried by the company in a typical day.

CROSS-CLASSIFICATION MODELS
Results for Manhattan
This section describes the cross-classification models developed as
part of this research. Throughout the analyses conducted here, con- This section describes the cross-classification models developed for
tinuous variables were converted into interval scales, in which the Manhattan. Four of the 10 groups found to be significant are pre-
number of bins and width for each interval class were selected as in sented here. As in the regression models, employment and the indus-
Scott (18). After data transformations, a two-step process was fol- try segment were found to be good predictors for trip production.
lowed to select the best group of independent variables describing Trip rates for some combinations between classes are not provided
trip generation. First the authors conducted a conceptual analysis to because of the low number of observations.
identify the potential explanatory variables, which resulted in the
identification of 40 pairs of independent variables. Subsequently, a
multiple classification analysis was performed for each of the groups. Receivers
Finally, the best groups were selected on the basis of the multiple R2,
multiple adjusted R2, eta (also called correlation ratio) and beta, and The analyses indicated that only one pair of independent variables
the 95% interval confidence for parameter. The eta and beta param- is statistically significant, the group involving employment versus

TABLE 8 Number of Delivery Trips per Day per Establishment (Manhattan carriers)

Company SIC
Line of Business
Wholesale Wholesale
Number Motor Freight Trade and Trade and Die; Third-Party
of Truck Transportation Transportation Die; Durable Nondurable Business Logistic
Drivers and Warehousing Services Goods Goods Services Shipper Manufacturer Consignee Provider

0–15 2.2 — 0.8 3.9 — 2.3 1.7 0.0 2.2

15–30 6.0 2.8 4.5 7.7 2.8 6.1 5.4 3.8 6.0
30–45 8.3 5.1 6.9 10.1 5.1 8.4 7.8 6.1 8.3
45–60 5.2 2.0 3.8 7.0 2.0 5.3 4.7 3.0 5.2
60–75 3.0 — 1.5 4.7 — 3.1 2.4 0.8 3.0
> 75 25.0 21.8 23.5 26.7 21.8 25.1 24.4 22.8 25.0

NOTE:
Multiple R-square adjusted 0.670 Multiple R-square adjusted 0.677
Multiple R-adjusted 0.449 Multiple R-adjusted 0.458
Average 8.3 Min 0.8 Average 7.9 Min 0.0
Median 5.1 Max 26.7 Median 5.4 Max 26.0
Trip rates for establishments with more than 75 truck drivers and more than 75 number of vehicles in the fleet were estimated with a relatively small number of
observations. Hence, caution is advised if used.
 Bastida and Holguín-Veras 57

Household
Computers and Goods and Stone and Office Jewelry Printed Medical
Metal Machinery Electronics Various Concrete Supplies and Art Material Supplies DK and Ref

2.8 3.1 2.1 1.5 1.1 1.3 2.5 6.3 2.3 22.6
4.1 4.4 3.4 2.8 2.4 2.6 3.8 7.6 3.6 23.9
2.7 2.9 1.9 1.3 0.9 1.1 2.4 6.1 2.2 22.4
4.1 4.4 3.4 2.8 2.4 2.6 3.8 7.6 3.6 23.9
3.5 3.8 2.8 2.1 1.8 1.9 3.2 7.0 3.0 23.3
8.3 8.5 7.5 6.9 6.5 6.7 8.0 11.7 7.8 28.0

commodity type (Table 7). The average trip rate is five trips per day. deliveries received. Finally, it was observed that number of trucks
The lowest trip rate found is approximately one trip per day, and the in the fleet and number of truck drivers are correlated and that both
highest is 28 trips per day. Establishments receiving printed materi- SIC code and line of business can capture the characteristics of the
als have the highest trip rates. For most commodities the trip rates industry segment.
remain in a relatively narrow range although caution is advised in
the case of Categories 15 to 30 and 45 to 60 owing to the low num-
ber of observations. Similarly, attention should be paid to categories Results for Brooklyn
with more than 75 employees.
As in the previous section the MCA analyses identified employ-
ment, commodity type, and industry segment as strong predictors
Carriers for freight generation. Six of the 10 pairs of independent variables
found significant are described. As in the previous section, trip rates
The MCA results indicated that three different pairs of variables are for some combinations between classes are not provided owing to
statistically significant: number of truck drivers versus establish- the low number of observations.
ment SIC code, number of truck drivers versus line of business, and
number of vehicles in the fleet versus SIC code. These findings are
similar to those obtained from regression analyses. Receivers and Intermediaries
As shown in Table 8, the wholesale trade industry segment has
the higher trip rates. In most cases the trip rates increase with the In Brooklyn some commercial establishments play two roles: they
number of truck drivers, reaching a maximum value in the 30 to act as receivers and as shippers of products. These establishments
45 truck driver interval, and then decrease. This may indicate that are referred to here as “intermediaries.” For instance, manufactur-
larger companies receive larger shipments, which reduces the total ers in Brooklyn may receive raw materials and then transform or

Company SIC

Wholesale Wholesale
Distributor, Number of Motor Freight Trade and Trade and Die;
Trucking Retail, and DK and Vehicles in Transportation Transportation Die; Durable Nondurable Business
Company Warehouse Mover Wholesale Ref the Fleet and Warehousing Services Goods Goods Services

2.0 2.6 1.8 3.3 — 0–15 2.5 — 1.0 3.9 —

5.8 6.4 5.5 7.0 3.3 15–30 5.5 2.3 4.1 6.9 2.3
8.1 8.7 7.9 9.4 5.6 30–45 10.0 6.8 8.5 11.3 6.8
5.0 5.6 4.8 6.3 2.5 45–60 19.7 16.5 18.3 21.1 16.5
2.8 3.4 2.5 4.0 0.3 60–75 — — — — —
24.8 25.4 24.5 26.0 22.3 > 75 22.1 18.9 20.6 23.5 18.9

Multiple R-square adjusted 0.624

Multiple R-adjusted 0.389
Average 11.7 Min 1.0
Median 10.0 Max 23.5
TABLE 9 Number of Deliveries Received and Shipped per Day per Establishment (Brooklyn receivers and shippers)

Commodity Type (received)

Agriculture,
Number of Forestry, Textiles and Wood and Petroleum Plastics and Computers and Other, (VOL)
Employees Fishing Food Tobacco Clothing Furniture Lumber Paper and Coal Rubber Metal Machinery Electronics specify: ______ DK and Ref

0–20 — 9.4 — — 1.3 9.3 — 0.8 2.8 0.9 0.1 61.8 10.4 2.8
20–40 — 7.1 — — — 7.0 — — 0.5 — — 59.5 8.1 0.5
40–60 0.1 10.7 — 1.1 2.6 10.6 0.7 2.1 4.1 2.2 1.4 63.1 11.8 4.1
60–80 — 9.9 — 0.3 1.8 9.8 — 1.3 3.3 1.4 0.6 62.3 11.0 3.3
>80 59.3 69.9 57.3 60.3 61.8 69.8 59.9 61.3 63.3 61.4 60.6 122.3 71.0 63.3

NOTE:
Multiple R-square adjusted 0.633
Multiple R-adjusted 0.4009
Average 24.9 Min 0.1
Median 8.7 Max 122.3

Commodity Type (shipped)

Number of Textiles and Wood and Petroleum Plastics and Computers and Other, (VOL)
Employees Food Tobacco Clothing Furniture Lumber Paper and Coal Rubber Metal Machinery Electronics specify: ______ DK and Ref

0–20 8.4 — — 1.3 8.3 — 5.3 2.8 0.2 0.8 61.8 10.5 2.8
20–40 6.1 — — — 6.0 — 3.0 0.5 — — 59.5 8.2 0.5
40–60 9.7 — 1.1 2.6 9.7 0.7 6.6 4.1 1.5 2.1 63.1 11.8 4.1
60–80 8.9 — 0.3 1.8 8.9 — 5.8 3.3 0.7 1.3 62.3 11.0 3.3
>80 68.9 57.3 60.3 61.8 68.9 59.9 65.8 63.3 60.7 61.3 122.3 71.0 63.3

NOTE:
Multiple R-square adjusted 0.617
Multiple R-adjusted 0.380
Average 24.4 Min 0.2
Median 8.2 Max 122.3
Trip rates for establishments with more than 80 employees were estimated with a relatively small number of observations. Hence, caution is advised if used.
Bastida and Holguín-Veras 59

repackage them, thus acting as both a receiver and a shipper. To ensure type carried versus number of employees, number of employees ver-
that the analyses were based on a sufficient amount of data, these inter- sus SIC code, and number of employees versus line of business. As
mediaries were grouped together with the set of pure receivers, which shown in this model the same independent variables as before are
are discussed here. once more driving freight generation. However, the computer and
Four groups of independent variables were found to be significant: electronics class stands out as the group with the highest trip rate
commodity type received versus number of employees, commodity among the models estimated (see Tables 9 and 10).

TABLE 10 Number of Deliveries Received and Shipped per Day per Establishment (Brooklyn receivers and shippers)

Line of Business

Third-Party Other,
Number of Logistic Trucking specify:
Employees Shipper Manufacturer Provider Company Warehouse ______ (VOL) DK and Ref

0–20 8.8 0.5 48.3 — 14.4 8.3 30.8

20–40 6.5 — 46.0 — 12.1 6.0 28.5
40–60 10.1 1.8 49.6 1.1 15.7 9.6 32.1
60–80 9.3 1.0 48.8 0.3 14.9 8.8 31.3
>80 69.3 61.0 108.8 60.3 74.9 68.8 91.3

NOTE:
Multiple R-square adjusted 0.675
Multiple R-adjusted 0.455
Average 30.6 Min 0.3
Median 15.3 Max 108.8

Company SIC

Apparel Electronic
Building and Other and Other
Construction Finished Electrical
General- Products Fabricated Metal Equipment
Contractors Fruits Made from Stone, Clay, Products, Except and
and and Food and Fabrics and Poultry Glass, and Machinery and Components,
Number of Operative Tree Kindred Similar Dairy and Concrete Transportation Except
Employees Builders Nuts Products Material Farms Eggs Products Equipment Computer

0–20 1.8 10.5 23.0 1.3 — 2.8 0.8 — 8.8

20–40 — 8.2 20.7 — — 0.5 — — 6.5
40–60 3.1 11.8 24.3 2.6 — 4.1 2.1 — 10.1
60–80 2.3 11.0 23.5 1.8 — 3.3 1.3 — 9.3
>80 62.3 71.0 83.5 61.8 58.3 63.3 61.3 58.3 69.3

Company SIC

Building
Materials,
Hardware, Home
Wholesale Wholesale Garden Furniture,
Trade Trade Supply, Furnishings, Eating
Miscellaneous and Die; and Die; and Mobile Apparel and and and
Number of Manufacturing Durable Nondurable Home Accessory Equipment Drinking Miscellaneous
Employees Industries Goods Goods Dealers Stores Stores Places Retail

0–20 2.8 7.4 2.4 14.8 4.8 — — 35.8

20–40 0.5 5.1 0.1 12.5 2.5 — — 33.5
40–60 4.1 8.7 3.7 16.1 6.1 1.1 0.6 37.1
60–80 3.3 7.9 2.9 15.3 5.3 0.3 — 36.3
>80 63.3 67.9 62.9 75.3 65.3 60.3 59.8 96.3

NOTE:
Multiple R-square adjusted 0.694
Multiple R-adjusted 0.481
Average 23.3 Min 0.1
Median 8.8 Max 96.3
Trip rates for establishments with more than 80 employees were estimated with a relatively small number of observations. Hence, caution is advised if used.
60 Transportation Research Record 2097

TABLE 11 Number of Delivery Trips per Day per Establishment (Brooklyn carriers)

Commodity Type

Agriculture, Plastics
Number Forestry, Textiles and Wood and Petroleum and
of Trucks Fishing Food Alcohol Clothing Furniture Lumber Paper and Coal Chemicals Rubber

0–15 — 1.4 — 0.5 1.6 — — 2.9 25.9 —

15–30 0.3 2.8 1.3 1.9 3.0 0.3 1.3 4.3 27.3 0.3
30–45 — 1.9 0.4 1.1 2.2 — 0.4 3.4 26.4 —
45–60 3.6 6.1 4.6 5.2 6.3 3.6 4.6 7.6 30.6 3.6
60–75 11.7 14.2 12.7 13.3 14.4 11.7 12.7 15.7 38.7 11.7
> 75 12.2 14.7 13.2 13.8 14.9 12.2 13.2 16.2 39.2 12.2

NOTE:
Multiple R-square adjusted 0.611
Multiple R-adjusted 0.374
Average 8.5 Min 0.1
Median 4.8 Max 39.2
Trip rates for chemical commodities were estimated with a relatively small number of observations. Hence, caution is advised if used.

Carriers lishment level. It has also shown that regression models can be used
not only to forecast future demand but also to establish the dynam-
The results for Brooklyn carriers are similar to those for Manhattan ics between variables. Moreover and foremost, it is the first research
receivers and carriers. In all cases the number of trucks in the fleet, that has estimated regression models and cross-classification rates
the commodity type, and the number of truck drivers are the strongest with explanatory variables interacting with employment and sales
predictors. However, neither line of business nor SIC code are found per year. In addition, it provides insight into the freight-related land
to be significant. Both the number of trucks in the fleet and the num- use of the metropolitan area of New York City.
ber of truck drivers seem to have a direct relationship with trip rates.
The number of truck drivers and the number of trucks in the fleet
might be correlated. Chemicals have the highest trip rates; however,
ACKNOWLEDGMENTS
this category has a low number of observations, which could mislead
the results (see Table 11). The research reported in this paper was partially supported by the U.S.
As the number of trucks increases the number of delivery trips per Department of Transportation’s Integrative Freight Demand Manage-
day per establishment for a given commodity type, the number of ment in the New York City Metropolitan Area project funded by the
deliveries increases. This is most noticeable in the categories 60 to Commercial Remote Sensing and Spatial Information Technology
75 trucks and >75 trucks.
Application Program, and by the New York State Department
of Transportation’s Potential for Off-Peak Freight Deliveries to
Commercial Areas project.
CONCLUSIONS

This paper has conducted an exploratory investigation of alternative

techniques for freight generation modeling involving the use of OLS REFERENCES
models and cross-classification techniques. The literature review
indicates that this is the first application of cross classification in 1. Ortuzar, J. D., and L. G. Willumsen. Modelling Transport. John Wiley
and Sons, Inc., New York, 2001.
freight generation. 2. Transportation in America. Eno Transportation Foundation, Washington,
Results indicate that regression models are able to identify not-so- D.C., 2002.
obvious relations pertaining to demand generation. Consider the role 3. FHWA. Freight Analysis Overview. 2004. ops.fhwa.dot.gov/freight/
of the commodity type, which is widely acknowledged to be a proxy freight_analysis/frt_analysis.htm.
4. NCHRP Synthesis of Highway Practice 298: Truck Trip Generation
for the industry segment in which the company operates. As dis-
Data. TRB, National Research Council, Washington, D.C., 2001.
cussed throughout the paper, the commodity type was routinely found 5. Ogden, K. W. Urban Goods Movement: A Guide to Policy and Planning.
to be a strong explanatory variable when interacting with other com- Ashgate Publishing Co., Brookfield, Vt., 1992.
pany attributes such as total sales per year or number of employees. 6. Holguín-Veras, J., Y. Lopez-Genao and A. Salam. Truck-Trip Generation
All of this suggests that the commodity type acts as a modifier of the at Container Terminals: Results from a Nationwide Survey. In Transpor-
tation Research Record: Journal of the Transportation Research Board,
relationship that exists between freight generation and other explana-
No. 1790, Transportation Research Board of the National Academies,
tory variables such as total sales per year and number of employees. Washington, D.C., 2002, pp. 89–96.
Results show that OLS and MCA are in general agreement about 7. Douglas, A. A., and R. J. Lewis. Trip Generation Techniques. Traffic
what are the most important variables explaining freight generation Engineering and Control, Vol. 12.7, 12.10, 1970/1971.
(i.e., commodity type, employment, industry segment, and total sales 8. Ogden, K. W. Modelling Urban Freight Generation. Traffic Engineering
and Control, Vol. 18.3, 1977, pp. 106–109.
per year). 9. Chatterjee, A., D. R. Martison, and K. C. Sinha. Trip Generation Analysis
Finally, this research has provided an important analysis of some for Regional Studies. Transportation Engineering Journal of ASCE,
of the most important factors driving freight generation at the estab- Vol. 103, No. TE6, 1977, pp. 825–841.
Bastida and Holguín-Veras 61

Household Waste Other, (VOL) Number of Truck Drivers

Goods and Stone and and specify: DK and
Metal Machinery Various Concrete Scrap _____ Ref 0–15 15–30 30–45 45–60 60–75

— 3.1 0.1 1.9 — 2.3 — 0.8 2.2 1.4 5.5 10.6

0.9 4.5 1.5 3.3 1.3 3.6 1.3 2.2 3.5 2.7 6.9 12.0
0.1 3.7 0.6 2.4 0.4 2.8 0.4 1.4 2.7 1.9 6.1 11.2
4.3 7.8 4.8 6.6 4.6 7.0 4.6 5.5 6.9 6.1 10.2 15.3
12.4 15.9 12.9 14.7 12.7 15.1 12.7 10.6 12.0 11.2 15.3 20.4
12.9 16.4 13.4 15.2 13.2 15.6 13.2 14.1 15.5 14.7 18.8 23.9

Multiple R-square adjusted 0.637

Multiple R-adjusted 0.406
Average 9.0 Min 0.8
Median 8.5 Max 23.9

10. Zavattero, D., and S. Weseman. Commercial Vehicle Trip Generation 14. Stopher, P. R., and K. McDonald. Trip Generation by Cross Classification:
in the Chicago Region. In Transportation Research Record 834, TRB, An Alternative Methodology. In Transportation Research Record 994,
National Research Council, Washington, D.C., 1981, pp. 12–15. TRB, National Research Council, Washington, D.C., 1983, pp. 84–91.
11. Middleton, D. R., J. M. Mason, Jr., and T. Chira-Chavala. Trip Gen- 15. Guevara, C. A., and A. Thomas. Multiple Classification Analysis in Trip
eration for Special-Use Truck Traffic. In Transportation Research Production Models. Transport Policy, Vol. 14, 2007, pp. 514–522.
Record 1090, TRB, National Research Council, Washington, D.C., 16. Nagpaul, P. S. Guide to Advanced Data Analysis Using IDAMS Soft-
1986, pp. 8–13. ware. UNESCO. 2001. www.unesco.org/webworld/idams/advguide/
12. Al-Deek, H. M., G. Johnson, A. Mohamed, and A. El-Maghraby. Truck TOC.htm.
Trip Generation Models for Seaports with Container and Trailer Oper- 17. Holguín-Veras, J., J. Polimeni, and N. Basu. Potential for Off-Peak
ation. In Transportation Research Record: Journal of the Transporta- Freight Deliveries to Congested Urban Areas (TIRC Project C-02-15).
tion Research Board, No. 1719, TRB, National Research Council, NYSDOT. 2006. www.rpi.edu/∼holguj2/OPD/OPD_FINAL_REPORT_
Washington, D.C., 2000, pp. 1–9. 12-18-06.pdf.
13. Al-Deek, H. M. Which Method Is Better for Developing Freight Planning 18. Scott, D. W. On Optimal and Data-Based Histograms. Biometrika, Vol. 66,
Models at Seaports—Neural Networks or Multiple Regression? In Trans- No. 3, 1979, pp. 605–610.
portation Research Record: Journal of the Transportation Research
Board, No. 1763, TRB, National Research Council, Washington, D.C., The Freight Transportation Planning and Logistics Committee sponsored publication
2001, pp. 90–97. of this paper.