a behavioral analysis of freight mode choice decisionssharif.edu/~asamimi/site_files/journal...
TRANSCRIPT
A behavioral analysis of freight mode choice decisions
Amir Samimia*, Kazuya Kawamurab and Abolfazl Mohammadianc
aDepartment of Civil Engineering, Sharif University of Technology,Azadi Avenue, Tehran 11365-8639, Iran; bCollege of Urban Planning and Public Affairs,University of Illinois at Chicago, 412 S. Peoria Street, Chicago, IL 60607-7064, USA;
cDepartment of Civil and Materials Engineering, University of Illinois at Chicago, 842 W. TaylorStreet, Chicago, IL 60607-7023, USA
(Received 3 November 2009; accepted 1 June 2011)
This paper develops a behavioral analysis of freight mode choice decisions thatcould provide a basis for an acceptable analytical tool for policy assessment. Thepaper specifically examines the way that truck and rail compete for commoditymovement in the US. Two binary mode choice models are introduced inwhich some shipment-specific variables (e.g. distance, weight and value) andmode-specific variables (e.g. haul time and cost) are found to be determinants.The specifications of the non-selected choice are imputed in a machine learningmodule. Shipping cost is found to be a central factor for rail shipments, whileroad shipments are found to be more sensitive to haul time. Sensitivity of modechoice decisions is further analyzed under different fuel price fluctuationscenarios. A low level of mode choice sensitivity is found with respect to fuelprice, such that even a 50% increase in fuel cost does not cause a significant modalshift between truck and rail.
Keywords: freight mode choice; truck and rail competition; fuel cost fluctuation;machine learning
Introduction
An efficient and reliable freight transportation system has substantial effects on
growth and sustainability of the national economy. This is because in such a system,
transportation cost will be reduced in the production process in several ways such as
decreasing inventory, labor, operating and maintenance costs (ICF Consulting and
HLB Decision-Economics 2002). Also, reducing the burden of freight traffic on the
transportation network will bring significant benefits to society through savings
in travel time, fuel consumption, pollution and diminishing other negative
consequences of an overburdened transportation system. According to the US
Commodity Flow Survey (CFS), the total value of transported commodities
increased around 30% between 1993 and 2002, and by the year 2035 the 2002
number is expected to double (US Department of Transportation 2006). The
enormous increase in freight traffic flow will require appropriate actions to address
the negative impacts of freight transportation activities.In the US, trucking is the most prolific among the freight transportation modes,
accounting for 69% of the total tonnage nationally in 2007 (US Department of
*Corresponding author. Email: [email protected]
Transportation Planning and Technology
Vol. 34, No. 8, December 2011, 857�869
ISSN 0308-1060 print/ISSN 1029-0354 online
# 2011 Taylor & Francis
http://dx.doi.org/10.1080/03081060.2011.600092
http://www.tandfonline.com
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
Transportation 2010). However, there is a consequence of the over-reliance on
trucking. Forkenbrock (2001) compared the externalities of rail and truck shipments,
revealing the former to be considerably less. According to his study, the external costs
of trucking were found to be over three times that of rail. While determining the
socially optimal balance between different freight modes will require major research
efforts (and is not the purpose of this paper), it is obvious that furthering the
understanding of modal selection behavior and having a more reliable analytical toolwill facilitate the development of broad strategies.
Over the years, there have been some notable efforts to develop such a tool.
However, freight shipment decisions have been changing rapidly during the past
three decades in response to the need for leaner, more efficient supply chain systems
that was brought on by the globalization of manufacturing process. The complexity
of today’s logistics decision-making process presents a serious challenge for freight
demand modelers to provide reliable analytical tools for both policy-makers and
practitioners. This problem can be mainly attributed to the lack of appropriate
disaggregate freight data, which prevents researchers from developing realistic
behavioral models. Meanwhile, the use of operations research (OR) techniques in
the intermodal freight transport arena is still limited (Macharis and Bontekoning
2004). This is primarily due to the complexity of such systems in which many
decision-makers are involved in a multi-commodity and multi-modal network with
several constraints. Researchers are introducing new OR-based approaches by
simplifying the problem and applying heuristic methodologies to solve multi-objective (e.g. cost, speed, reliability and risk) problems of this type (Min and
Glaister 1991). Furthermore, some disaggregate data collection efforts are under way
to introduce robust behavioral freight models (Roorda et al. 2010). It is worthwhile
to note that disaggregate data are being collected regularly by the US Department of
Transportation (2006) in their CFS. Nevertheless, this information is only available in
an aggregate format to respect the privacy of the business establishment.
Mode choice is one of the most critical parts of any freight demand modeling
framework. However, the literature on this issue is surprisingly modest mainly due to
the absence of suitable data. A direct comparison of shipment costs was the primary
method in the most early freight mode choice models (Cunningham 1982). However,
reliability, flexibility, safety and some other non-cost factors entered the analysis
when the random utility models emerged (Norojono and Young 2003). On the other
hand, implementation of supply chain management concepts along with the
deregulation of freight industries drastically affected the shipping behaviors of
companies (Rodrigue 2006). New supply chain concepts (e.g. just-in-time) were
adopted by many companies, which subsequently influenced shipping preferences(Hensher and Figliozzi 2007) and required fundamental revision in the models.
Arunotayanun and Polak (2007) found transport cost, delivery time, quality and
flexibility of service as the significant determinants of freight mode choice in
multinomial and mixed multinomial logit specifications. Although their analysis
included four commodity types, some critical information on each shipment such as
size, value and distance were missing. Evers et al. (1996) also asked shippers in
Minnesota to rate truck, rail and intermodal modes of freight transport on 17
characteristics. Six essential factors in freight modal selection were then introduced,
among which reliability and availability of each mode were ranked the highest. This
finding is in line with several other studies that found haul time and reliability to be
858 A. Samimi et al.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
more important than the cost to the shippers (McGinnis 1979). Evers et al. (1996)
tried to capture the role of shippers’ perceptions of the modes and the past
experiences in the analysis as well. A number of studies accounted for commodity
and industry heterogeneity in freight modal selection models; however, these models
are still at an early stage (Nam 1997, Arunotayanun and Polak 2007). For instance,
non-perishable food, textiles, leather and electronics were the only commodity types
that were considered by Arunotayanun and Polak (2007). Although the small
number of categories imposes some limitations on the study’s conclusions, such
restrictions are unavoidable in many freight studies. Nevertheless, based on the
review of those studies, the dominant factors impacting on freight mode choice in the
literature can be summarized as: accessibility, reliability, cost, time, flexibility and
past experience with each mode.
This paper introduces binary logit and probit models that explain how truck and
rail are chosen as the preferred mode by shippers, third party logistics providers
(3PLs) or receivers. These models specifically look into transportation cost, distance,
weight and value of commodities, and access to truck�rail intermodal facilities.
A specific sensitivity analysis is also performed to show how freight mode choice
changes with fluctuations in fuel cost. Modeling results and data used for calibration
are presented in the next section, followed by an in-depth analysis of the findings.
Finally, conclusions and recommendations are provided.
Data and model
Freight mode-choice studies are performed traditionally using OR techniques, and
more recently by utility maximization theory. The latter is becoming more common,
as the logistics decision-making process has become extremely sophisticated and
factors other than cost (e.g. speed, reliability and risk) are playing a significant role
(Arunotayanun and Polak 2007). Complex behaviors of decision-makers in the
current freight transport market, along with the extent of the freight transport
systems, necessitate several simplifying assumptions in OR-based cost-minimization
approaches. Hence, a random utility maximization approach was chosen in this
study. However, any disaggregated data on freight activities are extremely difficult to
obtain due to the scarcity of the surveys that collect such data and the concern for
violating the confidentiality of the businesses that participated in the survey. Thus, it
is not surprising that there is no disaggregate freight data at the national level that
are publicly available in the US.
Therefore, our effort to develop a freight mode choice model had to begin with a
data collection effort. An online survey was conducted at the University of Illinois,
Chicago (UIC) in April and May 2009, providing information on 881 domestic
shipments in the United States (Samimi et al. 2010a). A total of 4544 business
establishments opened the recruiting email, of which 316 firms participated � a 7%
response rate, which is a reasonable rate in such surveys. Basic information about
each establishment along with data on five recent shipments, namely origin,
destination, transportation mode, type, value, weight, and volume of the commodity,
cost and time of the entire shipping process, were obtained (Samimi et al. 2010a, b).
Some essential information about each establishment was also collected: square
footage, number of employees, industry type, location, warehousing situation and
Transportation Planning and Technology 859
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
potential use of each freight transportation mode. Table 1 shows a comparison
between value and weight of the surveyed commodities and the 2002 CFS data.
Many shippers are reluctant to participate in surveys that enquire about their
shipping decisions, which results in low response rates that can diminish the
credibility of or even invalidate the study results, if not appropriately addressed.
A comprehensive analysis of non-random selection bias was performed in a separate
study (Samimi et al. 2010a) to test whether size, location and industry type of the
firms has affected the probability of participation. The two-stage Heckman
correction (Heckman 1990) method was discussed, revealing no significant effect
of establishment size and a very minor and negligible effect of location and industry
type on the probability of participation. Brief statistics of variables that are used in
the final mode choice models are presented in Table 2.
A proper choice model is sensitive to attributes of both decision-maker and
choice alternatives. While characteristics of the decision-maker (e.g. number of
employees) do not change across alternatives, the attributes of choice alternatives
vary significantly from one alternative to the other (e.g. shipping time) and are
typically collected only for the observed choice. One of the critical challenges in
modeling freight modal selection is to obtain information on non-selected choices. In
our case, shipping cost and time for using either truck or rail was obtained for each
shipment in the survey. Using those data, the specifications of the non-selected
choice were imputed in a machine learning module.
Artificial neural networks (ANNs) are a class of learning algorithms that develop
learning rules based on data and construct a so-called ‘black box’ that can be used to
generate desired outputs that correspond to a new set of inputs. Although ANNs
generate very complicated rules that usually produce high levels of fit, the underlying
rules by which the output is generated are not revealed. Machine learning methods
have been implemented in the field of transportation planning in the past (Al-Deek
2001, Mohammadian and Miller 2003), and a more complete discussion on the topic
can be found in the literature (Parks et al. 1998, Principe et al. 2000).
For this study, two artificial neural networks were constructed with two hidden
layers and were trained using NeuroSolution v5.07 (Neuro Dimension Corporate
2009). The input data were divided into three parts: 60% of the data was used for
training the networks; 15% for cross validation; and the remaining 25% was left aside
as the test set to evaluate the quality of the trained network. The first network was
Table 1. Value and weight share of each mode.
Dollar value Weight Shipments
Mode CFS (%)a UIC (%)b CFS (%) UIC (%) UIC (%)
Truck 68 67 60 49 69
Rail and rail intermodalc 3 4 10 12 5
Other 1 8 4 8 5
aCommodity Flow Survey (2002) data do not include imports and exports that pass through the UnitedStates from a foreign origin to a foreign destination by any mode.bUIC (University of Illinois at Chicago) National Freight Survey.cIntermodal includes US Postal Service and courier shipments and all intermodal combinations, except airand truck.
860 A. Samimi et al.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
trained using data for the rail shipments to impute the unobserved shipping times
and costs for road shipments, while the other network used truck shipments for
training to estimate the aforementioned information for rail shipments.
The most common framework used for choice behavior analysis in recent years
has been the discrete choice modeling approach. Various forms of discreet choice
models are proposed in the literature depending on underlying assumptions
concerning the distribution of the unobserved utility. Two widely used forms of
discrete choice models are logit and probit models. While the logit model assumes
independent and identically distributed (IID) error terms in the utility function, the
probit model assumes a normal distribution for the error terms (Train 2003). Limdep
econometrics software (Greene 2002) was used to import explanatory variables such
as: shipping mode, time of each mode, cost of each mode, distance, commodity type,
weight, volume, value, access to truck�rail intermodal facilities, potential use of rail,
etc. The forward selection method was used for variable selection, and numerous
forms and combinations of variables were tested for the most appropriate fit. Table 3
Table 2. Variables used in the analysis.
Variable Definition Mean
Standard
deviation
Mode 1: rail or any combination of that with other
modes/0: truck
0.089 0.285
Distance Suggested distance between origin and
destination by Google Map (miles)
1077.940 2221.100
Weight Weight of the shipment (lbs) 22,901.100 25,275.100
Value Value of the shipment (USD) 48,101.389 130,150.251
Truck-cost Shipping cost by truck (USD) 1331.760 4093.390
Rail-cost Shipping cost by rail (USD) 2016.880 1128.160
Truck-time Shipping time by truck (days) 2.012 1.357
Rail-time Shipping time by rail (days) 7.281 6.662
Truck-cost-index �Ln (TRUCK-COST/(TRUCK-
TIME�VALUE))
�3.542 1.521
Rail-cost-index �Ln (RAIL-COST/(RAIL-
TIME�VALUE))
�3.705 1.940
Same-decision 1: if the same mode was preferred TWO years
ago for a similar shipment/0: otherwise
0.934 0.248
Access 0: firm has easy access to truck rail intermodal
facilities/1: neutral access/2: difficult access
0.780 0.415
Potential-
intermodal
1: truck�rail intermodal is considered always
or often as a potential transportation mode/0:
otherwise
0.349 0.477
Perishable 1: if the commodity is perishable/0: otherwise 0.160 0.367
Consolidation-
center
1: if the shipment has gone through a
consolidation center/0: otherwise
0.143 0.350
Distribution-
center
1: if the shipment has gone through a
distribution center/0: otherwise
0.270 0.445
Warehouse 1: if the shipment has gone through a
warehouse/0: otherwise
0.347 0.477
Decision-maker 1: if a 3PL company has made the shipping
decision/0: otherwise
0.104 0.305
Transportation Planning and Technology 861
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
shows the final probit and logit models that estimate the probability of choosing
between truck and rail.Newey and McFadden (1994) and Train (2003) include detailed discussions on
binary choice models. Akaike and McFadden values are among many goodness-of-fit
measures offered for binary choice models, which were used along with the chi-squared
values for model selection (Train 2003). The higher the McFadden value and the lower
the Akaike measure, the better the explanatory power of the model. The McFadden
value is also known as the likelihood ratio index or pseudo R-squared and has a similar
range (0�1) as R-squared has in ordinary least square (OLS) models (Train 2003).
However, in general the McFadden values tend to be lower than the R-squares for the
OLS models. Standard t-statistics, shown in Table 3, are for testing whether each
coefficient has a non-zero effect on the choice probability. All the estimated parameters
in the final models are significant with a p-value of less than 0.05; most are significant
with 99% confidence interval. Wald, Likelihood Ratio, and Lagrange Multiplier tests,
known as Neyman�Pearson tests (Greene 2002), were also carried out to evaluate the
overall significance of the final models. Both models have pseudo R-squared values of
more than 57%, and correctly predict 95% of the observations. Percentage of correctly
predicted observations is usually high in binary choice models that predict a rare event.
The high percentage of correct predictions could be misrepresented as the general
explanatory power of the model. However, when the two possible outcomes are either a
Table 3. Mode choice models.
Probit model Logit model
Item Value t-ratio Value t-ratio VIF
Coefficient Constant �5.902* �6.050 �10.808* �5.696 �Distance 0.237E-03** 2.273 0.452E-03** 2.156 2.776
Weight 0.310E-04* 4.293 0.569E-04* 4.195 1.564
Truck-time 0.622* 5.019 1.110* 4.815 1.648
Rail-time �0.094* �2.579 �0.176** �2.295 2.387
Truck-cost-
index
0.388** 2.532 0.670** 2.361 3.408
Rail-cost-index �0.659* �3.474 �1.188* �3.331 1.099
Potential-
intermodal
1.214* 3.468 2.270* 3.265 2.776
Fit Measures Log likelihood �47.141 � �47.780 � �Model Chi-
squared
128.577 � 127.300 � �
Akaike I.C. 0.296 � 0.300 � �Pseudo R-
squared
0.577 � 0.571 � �
Correctly
predicted (%)
95.430 � 95.699 � �
Correctly
predicted (%) �only rail
72.727 � 72.727 � �
* Significant at 99% confidence interval.**Significant at 95% confidence interval.
862 A. Samimi et al.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
rare or common event, binary models tend to over-predict the latter, resulting in high
rates of correct predictions at the expense of largely ignoring the rare event outcomes.
For example, if 99 out of 100 choices are common and only 1 is a rare event, the model
can attain 99% accuracy by simply predicting all cases to be the former. Thus thepercentage of rare events that are correctly predicted is a more valuable measure of
predictive power for such models. In our case, choosing rail over truck could be
considered as a rare event with only around 9% chance of occurrence in this data. Both
models predicted more than 72% of rail shipments correctly, which is quite impressive
especially for a freight mode choice model.
Since the shipping cost and time of unobserved modes were imputed in a machine
learning module, it seemed necessary to control for potential multicollinearity
between explanatory variables. Although collinearity is unlikely to be a serious issuewhen all the coefficients are statistically significant in a binary choice model, very
large off-diagonal values were searched for in the variance-covariance matrixes as the
primary effect of multicollinearity. Variance inflation factors (VIF) were also
estimated for all the independent variables. Kutner et al. (2004) suggested a VIF of
5 as the threshold that indicates a presence of serious multicollinearity. For our
models, none of the variables had a VIF in excess of 3.5 (Table 3).
Analysis of results
During the model fitting process, many different combinations of the independentvariables were tested. We found that a broad range of variables influence the mode
selection process. This includes establishment-specific variables (e.g. establishment
size, location, access to rail and road network, decision-making unit in the supply
chain, etc.), shipment-specific variables (e.g. commodity type, value, weight, special
handling needs, etc.), and shipping mode-specific variables (e.g. cost, velocity,
reliability, safety, flexibility, etc.). These variables not only have significant impacts
on the mode choice, but also are interdependent. For instance, commodity type and
cost of shipment are correlated. Also, shipping cost and velocity are interdependent.This, in some ways, constrains the specification of the mode choice model, and
requires a close attention to address potential collinearity issue, which was discussed
in the previous section. This part of the paper analyzes and interprets the effect of
each explanatory variable in the final models, along with some other variables that
were found not to be significant in the final models but were shown to have
considerable effects on the choice of mode. A sensitivity analysis of mode choice on
fuel cost, which is a topic of considerable interest to the researchers and policy-
makers alike for obvious reasons, is also provided in this section.
General discussion
Marginal effect analysis was performed on the final models to provide a better
understanding of each explanatory variable’s impact on freight modal selection � andshown in Table 4. Although the values are similar for logit and probit models,
marginal effects have higher levels of significance in the logit model and thus the
discussion in this section will focus on it. Distance, weight, truck shipping time, rail
shipping time, truck cost index and rail cost index (see Table 1 for their definition)
turned out to be significant in the final model. DISTANCE has a positive sign
Transportation Planning and Technology 863
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
indicating that rail is more likely to be chosen for long hauls. This finding is
intuitively interpretable and was also confirmed in former studies (Oum 1979). One
explanation for this trend is that rail shipments have a higher base price compared to
truck, which is diminished in long hauls. Weight of the shipment is another
significant variable in the model with a positive coefficient, indicating that larger
shipments are more likely to be transported by rail. This observation is also in line
with past studies. As indicated by Evers et al. (1996), past experiences with each
mode plays a determining role in the selection of mode. POTENTIAL-
INTERMODAL variable shows such effect in the models with positive coefficients,
indicating that firms that always or often consider truck�rail intermodality as a
possible option are more likely to select rail. Since in our model, the mode ‘rail’
includes shipments by rail alone or in combination with any other mode including
trucks, this finding is intuitive. This finding may seem trivial at first glance, but from
the modeling perspective the inclusion of such a variable makes other coefficients
more meaningful. For instance, shipping behavior of a firm preferring truck over rail
may be mistakenly attributed to the differences in cost and/or haul time, while the
real reason may have been that the shipper is unfamiliar with the rail mode in terms
of its service quality, cost and other factors. Therefore excluding such variables that
capture the effects of shippers’ knowledge or prejudice from the models may result in
erroneous interpretation of the coefficients.
Cost and haul time of each transportation mode are other significant factors in
mode selection. Having such mode-specific indicators enhances the explanatory
power of the model, especially when modeling freight transport behaviors.
A comparison between the coefficients of truck and rail transit time reveals that
the choice probability for truck is more sensitive to haul time than for rail. The
elasticities of truck and rail haul time, shown in Table 4, indicate that the effect of
truck travel time is almost 20 times greater for the truck mode. This shows that time
is a crucial factor especially when truck is preferred to rail. The cost index � which is
defined for each mode as the log of shipping cost divided by the product of haul time
and value of shipment � shows that the choice of rail is sensitive to cost. Shipping
cost is normalized over the shipping distance and value of the shipment in the
proposed cost index. Also, the log of this ratio conveys a non-linear behavior with
the attractiveness of each mode. Rail shipments’ sensitivity to the cost index is
around 1.7 times greater than that for truck shipments. An interesting observation in
the coefficients of time and cost variables is that shippers preferring truck are mainly
Table 4. Marginal effect analysis of the mode choice models.
Probit model Logit model
Variable
Marginal
effects Elasticity t-ratio
Marginal
effects Elasticity t-ratio
Distance 0.151E-05 1.244E-10 0.989 2.361E-06 1.943E-10 1.376
Weight 0.198E-06 7.665E-13 1.004 2.968E-07 1.150E-12 1.549
Truck-time 0.397E-02 1.749E-04 1.028 5.792E-03 2.554E-04 1.608
Rail-time �0.599E-03 �7.303E-06 �1.042 �9.183E-04 �1.119E-05 �1.457
Truck-cost-index 0.247E-02 �6.194E-05 0.964 3.497E-03 �8.759E-05 1.36
Rail-cost-index �0.420E-02 1.006E-04 �1.107 �6.200E-03 1.484E-04 �1.722
864 A. Samimi et al.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
concerned about shipping time and, in general, not overly sensitive to cost. On the
other hand, decisions on rail shipments are more sensitive to cost, but not to time.
This suggests that rail shipments are generally quite sensitive to cost and easily react
to changes in price. A more complete discussion on the effects of fuel pricefluctuation on modal selection is provided later in this section.
Other influential factors on mode choice
A variety of explanatory variables were considered in the modeling step. So far, we
have discussed the variables that were included in the final model. This does notnecessarily mean that other variables have no effect on modal selection. In most
cases, they cannot be in the model mainly due to interdependencies with other
variables. Two different tests of independence, Chi-square test and G-test, were
performed between shipping mode and other explanatory variables. Table 5 shows a
list of variables that were found to be dependent on transportation mode, according
to Chi-squared and G-squared values (Greene 2002).
Table 5 indicates the perishability of the commodity affects the choice of mode at
the 80% confidence level. This result has also been observed in past studies (Oum1979) and is mainly attributed to the effect of transit time on such commodities. Rail
shipments are more likely to go through a consolidation center, distribution center or
a warehouse, as suggested by 2nd, 3rd and 4th variables in Table 5. This could be
explained by size and distance of such shipments. Not a significant benefit is
obtained by sending small and short haul shipments to a consolidation or
distribution center (Higginson and Bookbinder 1994) resulting in such associations.
Two other variables, DECISION-MAKER and SAME-DECISION, were deployed
to capture the role of knowledge and previous experience of shippers about eachmode in the selection process. Shipments that are planned by a 3PL company, which
Table 5. Test of independence.
Mode
Chi-square
independence test
G-test of
independence
Variable Category 0 (%) 1 (%) Chi-squared p-value G-squared p-value
Perishable 0 77.1 6.9 1.847 0.174 1.663 0.197
1 13.8 2.2
Consolidation-center 0 79.9 5.8 15.532 0.000 12.050 0.001
1 10.8 3.5
Distribution-center 0 67.6 5.4 3.322 0.068 3.082 0.079
1 23.3 3.7
Warehouse 0 61.2 4.1 3.153 0.076 3.014 0.083
1 30.6 4.1
Decision-maker 0 83.4 6.2 4.452 0.035 3.566 0.059
1 8.6 1.8
Same-decision 0 4.9 1.6 8.463 0.004 6.100 0.014
1 86.3 7.1
Access 0 28.7 5.0 10.658 0.005 13.134 0.001
1 40.4 3.9
2 21.7 0.3
Transportation Planning and Technology 865
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
usually have a broader knowledge of available modes and perform a more
comprehensive analysis for mode selection, are more likely to be shipped by rail.
This indicates that some shippers are not aware of rail benefits, and their decision
may be different if they had complete information about their alternatives. Also,3PLs may be able to combine shipments into a load that is large enough for a rail
shipment. Another variable showing a significant association with the choice of
mode is SAME-DECISION. Interestingly, shippers preferring the same mode they
would have used two years ago for a similar shipment are less likely to choose rail
over truck. This finding is in line with the aforementioned observation that persons
who had chosen truck were found not to consider rail as a potential transportation
mode in some cases. Simply put, trucking seems to have better customer loyalty. This
is perhaps a surprising finding considering the increase in the price of fuel that tookplace in the summer of 2008, less than a year before the survey. The last variable of
interest is the accessibility to intermodal terminals. Obviously, as the level of access
to rail or truck�rail intermodal facilities decreases, shippers prefer to use trucks.
Table 5 confirms such association at the 99% confidence level.
Mode choice and fuel cost fluctuation
Fuel price is an important component of freight transportation cost and has gonethrough large fluctuations in recent years. Its effects on the shipping behaviors are of
interest in many disciplines and are specifically looked into in this part of the paper.
Road freight demand is often considered much more inelastic to shipping cost than
passenger traffic, and there is a wide variation in the fuel cost elasticities estimated in
the past (Graham and Glaister 2004), although the variation is mainly due to the
difference in scope and method of studies.
Figure 1 illustrates changes in the share of rail freight transportation, when the
fuel price is increased by four different amounts. The binary logit model in Table 3 isused for this part of the analysis. In each case, 16 different possibilities are explored
in which the share of fuel price in the total shipping cost varies. Depending on
shipping distance, congestion level, fuel consumption of the fleet, topography, etc.,
the share of the fuel cost within the total cost of the shipment, and thus its possible
influence, varies. For instance, a long haul truck shipment, traveling through an
uncongested corridor, may be expected to be more affected by fuel price increases.
However, such conclusions require more investigation, since labor cost is by far the
largest part of trucking and is complicated to estimate, since some truckers are paidby the hour, some are paid by load, and some by miles driven. Therefore, estimating
share of the fuel cost in total shipping cost, just based on congestion level and
shipping distance, is not accurate.
The results of the analysis, shown in Figure 1, suggest that freight modal
decisions are very much inelastic to fuel cost and do not change significantly with
even a 50% increase in fuel cost. When the fuel price doubles, however, shippers start
shifting to rail when fuel cost accounts for a large portion of the total cost. This may
happen in long haul shipments that experience a relatively low level of congestion.Figure 1 also explores two other scenarios with 150 and 200% increases in fuel cost.
In these scenarios, around 7% of total shipments are expected to shift to rail when
the fuel price is a major component of total shipping cost. However, even when the
fuel cost is not a large factor, a significant shift of around 3% is expected.
866 A. Samimi et al.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
The low elasticities of modal decisions with respect to fuel cost that were
obtained in this study are in line with many other studies in which such decisions
were introduced as inelastic or in best cases much less elastic than passenger
transportation (Graham and Glaister 2004).
Conclusions
Behavioral freight mode choice models are of great importance for both academia
and practice (for example, policy-making). However, it has been mainly overlooked
in freight demand modeling primarily due to data limitations. In this paper, we
presented the development of binary mode choice models between truck and rail
(including intermodal modes) based on the data obtained from an online national
level freight survey conducted recently in the US by the research team (Samimi et al.
2010a). This study modeled modal selection decisions as part of a microsimulation
framework, which also shed light on modal selection behavior.
40%30%20%10%
Around 3% reduction in rail share. Around 3% increase in rail share.
Rail share is almost unchanged. Around 7% increase in rail share.
50% increase in fuel price150% increase in fuel price
200% increase in fuel price
Fuel share in total truck costFuel share in total truck cost
Fuel share in total truck cost Fuel share in total truck cost
Fuel
sha
re in
tota
l rai
l cos
tFu
el s
hare
in to
tal r
ail c
ost
Fuel
sha
re in
tota
l rai
l cos
t
100% increase in fuel price
Fuel
sha
re in
tota
l rai
l cos
t
40%30%20%10%
1%
3%
5%
10%
40%30%20%10%
1%
3%
5%
10%
40%30%20%10%
1%
3%
5%
10%
10%
3%
50%
10%
Figure 1. Freight mode sensitivity to fuel price fluctuation in different scenarios.
Transportation Planning and Technology 867
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
Two binary choice models were developed to broaden the understanding of the
mode selection behavior by shippers, 3PLs, and receivers in the US freight markets.
We used the machine-learning approach to generate the time and cost of shipments
by modes that were not chosen by the respondents. Of the shipment-specific
variables, distance, weight and value of commodity were found to be significant. It
was also found that the truck shipments were extremely sensitive to travel time and
rail shipments sensitive to cost. We found that familiarity with a mode, especially
trucking, also had a strong influence on mode choice behaviors. Other variables were
found to have significant correlation with mode choice, although they could not be
included in the model due to interdependency issues. For example, the perishability
of the commodity, access to intermodal facility, and having a 3PL as the decision-
maker all seem to affect the mode choice.
Analysis of various scenarios involving large increases in fuel price revealed that
mode choice is not particularly sensitive to fuel price � up to a point. Our analysis
showed that even a 50% increase in fuel price did not cause any significant modal
shift between truck and rail. However, when the increase reaches 150 and 200%,
around a 7% shift from truck to rail shipments can be expected.
The findings from this study are generally in line with past efforts by other
researchers. However, some of the findings, especially the effect of the variables
related to the decision-maker, such as the past experience and familiarity with modes,
are unique and provide valuable insights into the mode choice behaviors. Also, the
mode-specific characteristics of time and cost of shipments, made possible by the
application of a machine-learning technique, enabled a more comprehensive analysis
with respect to those two variables than before.
As a final note, to the best of the authors’ knowledge there is no robust and
comprehensive longitudinal study of the effect of fuel cost on freight mode choice
decisions. This can be attributed primarily to the fact that disaggregate freight data
are so rare and expensive to collect that the researchers have been forced to limit their
study to a cross-sectional data. This obviously applies to the present study. It is our
hope that we will be able to address such a gap in the near future.
Acknowledgements
The authors appreciate the assistance of the National Center for Freight, Infrastructure,Research and Education (CFIRE) at the University of Wisconsin-Madison and the IllinoisDepartment of Transportation for funding this study. All responsibility for the content of thepaper lies with the authors.
References
Al-Deek, H., 2001. Comparison of two approaches for modeling freight movement atseaports. The ASCE Journal of Computing in Civil Engineering, 15 (4), 284�291.
Arunotayanun, K. and Polak, J.W., 2007. Taste heterogeneity in freight shippers’ mode choicebehaviour. Proceedings of the 86th annual meeting of the transportation research board,21�25 January 2007, Washington, DC: Transportation research board of the nationalacademies, Paper #07-2063.
Cunningham, W.H.J., 1982. Freight modal choice and competition in transportation: acritique and categorization of analysis techniques. Transportation Journal, 21 (4), 66�75.
Evers, P.T., Harper, D.V., and Needham, P.M., 1996. The determinants of shipper perceptionsof modes. Transportation Journal, 36 (2), 13�25.
868 A. Samimi et al.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11
Forkenbrock, D.J., 2001. Comparison of external costs of rail and truck freight transportation.Transportation Research Part A, 35 (4), 321�337.
Graham, D.J. and Glaister, S., 2004. Road traffic demand electricity estimates: a review.Transport Reviews, 24, 261�274.
Greene, H.W., 2002. LIMDEP Version 8.0 User’s Manual. Plainview, NY: EconometricSoftware.
Heckman, J., 1990. Varieties of selection bias. American Economic Review, 80 (2), 313�318.Hensher, D. and Figliozzi, M.A., 2007. Behavioural insights into the modelling of freight
transportation and distribution systems. Transportation Research Part B, 41, 921�923.Higginson, J.K. and Bookbinder, J.H., 1994. Policy recommendations for a shipment-
consolidation program. Journal of Business Logistics, 15 (1), 87�112.ICF Consulting and HLB Decision-Economics, 2002. Economic effects of transportation: the
freight story. Fairfax, VA: ICF Consulting.Kutner, M.H., Nachtsheim, C.J., and Neter, J., 2004. Applied linear regression models. 4th ed.
Boston: McGraw-Hill Irwin.Macharis, C. and Bontekoning, Y.M., 2004. Opportunities for OR in intermodal freight
transport research: a review. European Journal of Operational Research, 153, 400�416.McGinnis, M.A., 1979. Shipper attitude toward freight transportation choice: a factor analytic
study. International Journal of Physical Distribution & Logistics Management, 10, 25�34.Min, H. and Glaister, S., 1991. International intermodal choices via chance-constrained goal
programming. Transportation Research Part A, 25 (6), 351�362.Mohammadian, A. and Miller, E.J., 2003. Nested logit models and artificial neural
networks for predicting household automobile choices: a comparison of the performance.Transportation Research Record, (1807), 92�100.
Nam, K.C., 1997. A study on the estimation and aggregation of disaggregate models of modechoice for freight transport. Transportation Research Part E, 33 (3), 223�231.
Neuro Dimension Corporate, 2009. NeuroSolutions: premier neural network developmentenvironment. Available from: http://www.neurosolutions.com [Accessed 15 July 2009].
Newey, W.K. and McFadden, D., 1994. Large sample estimation and hypothesis testing. In:R. Engle and D. McFadden, eds. Handbook of econometrics Vol. 4. Amsterdam: NorthHolland, 2111�2245.
Norojono, O. and Young, W., 2003. A stated preference freight mode choice model.Transportation Planning & Technology, 26 (2), 195�212.
Oum, T.H., 1979. A cross sectional study of freight transport demand and rail-truckcompetition in Canada. Bell Journal of Economics, 10, 463�482.
Parks, R.W., Levine, D.S., and Long, D.L., 1998. Fundamentals of neural network modeling.Cambridge, MA: MIT Press.
Principe, J.C., Euliano, N.R., and Lefebvre, W.C., 2000. Neural and adaptive systems:fundamentals through simulations. New York: John Wiley & Sons.
Rodrigue, J.P., 2006. Commodity chains and freight transportation. In: J.P. Rodrigue,C. Comtois, and B. Slack, eds. The geography of transport systems. New York: Routledge,192�222.
Roorda, M.J., et al., 2010. A conceptual framework for agent-based modelling of logisticsservices. Transportation Research Part E, 46 (1), 18�31.
Samimi, A., Mohammadian, A., and Kawamura, K., 2010a. An online freight shipmentsurvey in us: lessons learnt and a non-response bias analysis. Proceedings of the 89th annualtransportation research board meeting, 10�14 January 2010, Washington, DC: Transporta-tion research board of the national academies, Paper #10-0230.
Samimi, A., Mohammadian, A., and Kawamura, K., 2010b. A behavioral freight movementmicrosimulation model: method and data. Journal of Transportation Letters, 2 (1), 53�62.
Train, E.K., 2003. Discrete choice methods with simulation. New York: Cambridge UniversityPress.
US Department of Transportation, 2006. Freight facts and figures. Washington, DC: FederalHighway Administration.
US Department of Transportation, 2010. DOT releases new freight transportation data.Available from: http://www.fhwa.dot.gov/pressroom/fhwa1062.htm [Accessed 20 November2010].
Transportation Planning and Technology 869
Dow
nloa
ded
by [
Uni
vers
ity o
f C
hica
go]
at 0
3:41
21
Nov
embe
r 20
11