Transportation Research Part E 40 (2004) 317337

www.elsevier.com/locate/tre

A disaggregate analysis of port selection

Matthew B. Malchow *, Adib Kanafani

Department of Civil Engineering, Institute of Transportation Studies, University of California,

109 McLaughlin Hall, Berkeley, CA 94720, USA
Received 15 January 2003; received in revised form 3 March 2003; accepted 16 May 2003

Abstract
With this article we use an alternative form of the discrete choice model to analyze the distribution of
maritime shipments among US ports. We model the distribution as a function of the characteristics that
describe each shipment and each port. We assume that vessel schedules are xed in the short-term and
examine the assignment to ports for exports of various commodity-types as a function of geographic
location, port characteristics, and characteristics of vessel schedules. We nd that the most signicant
characteristic of a port is its location. We show also how the market share predicted for a port can be
expected to vary with each commodity-type and each carrier, and we show how the choice process varies for
discretionary cargo.

2003 Elsevier Ltd. All rights reserved.
Keywords: Port choice; Shipper behavior; Carrier behavior; Shipment routing

1. Introduction
Competition between ports has intensied. As a result of containerization, which standardized
the transfer process for shipments between ocean and surface transportation, and deregulation,
which allowed maritime carriers to set contracts with rail services and establish rates independent
of location, the area considered a ports hinterland disappeared. Apart from their various marketing eorts, ports compete primarily through their investment program. Ports are improving
intermodal facilities to minimize the dwell time of shipments, and they are increasing the storage

Corresponding author. Tel.: +1-510-231-9460; fax: +1-510-231-9565.

E-mail addresses: malchow@alumni.duke.edu (M.B. Malchow); kanafani@ce.berkeley.edu (A. Kanafani).
1
Tel.: +1-510-642-3585; fax: +1-510-642-1246.
1366-5545/$ - see front matter
2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tre.2003.05.001

318

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

space available to terminal operators to allow carriers to concentrate operations. Ports are also
dredging their waters so that carriers may deploy larger vessels.
By investing to attract business, ports recognize that carriers make two primary decisions, the
long-term decision being the deployment of vessels to routes and ports, and the short-term
decision being the assignment of shipments to vessels. Thus with the assignment of a shipment to a
vessel comes the assignment of a shipment to a port. In this work we apply a choice model to
model the assignment of each shipment to a vessel/port and hence to evaluate the competition
between ports. We answer three questions:
What factors inuence a carriers selection of a port for a shipment?
In what manner and across what domain do ports compete?
What strategies might a port follow to increase its market share?
Though the choice-model approach used in this research has not previously been applied to
port selection, our approach has been inuenced by earlier work. Qualitative analysis of port
competition has been done by, among others, Bardi (1973), Foster (1978a,b, 1979), Slack (1985),
Hanelt and Smith (1987), and DEste and Meyrick (1992a,b). The results have not always been
identical, but the authors often suggest that service-related factors were more important than price
factors, and that factors within the control of port authorities were often less important than those
beyond port control. Brooks (1984, 1985) noted the dierence between a characteristics importance and salience. Quantitative analysis can dier between the two, as the ranking of characteristics comes not from word-of-mouth but the results of actions.
The scheduling of carriers vessels has also been the subject of much research. Kenyon (1970)
and Al-Kazily (1979) explored the development of a carriers maritime network. Hayuth (1981)
suggested the formation of a load center by carriers, while Foggin and Dicer (1985) and Slack
(1996) evaluated the eects of load centers. Helmick (1994) sought quantitative evidence of the
formation of load centers but suggested that other factors, e.g. the presence of tramp lines in
routes abandoned by major carriers, prevented conrmation of carrier rescheduling. Lago et al.
(2001) found that the rescheduling of vessels by carriers was not drastic but did dier between
corridors. They showed how the level at which scale economies were exploited in oceanic transit
diered between corridors.
Economic models of carrier behavior have come from three directions. First, linear programming to optimize the assignment of vessels has been advanced by Benford (1981), Perakis (1985),
Lane et al. (1987), and Perakis and Jaramillo (1991). These models show how the distribution of
vessels might be aected by the distribution of trac but deal with simplied scenarios or are
dicult to apply. Second, cost models have been estimated by Jansson and Shneerson (1978),
Talley (1990), and Lim (1998) for the general cargo or container shipping industry. Garrod and
Miklius (1985) and Jansson and Shneerson (1985) emphasized the importance of shippers
inventory costs. Griths (1976a,b) measured the optimal size of ports for a given level of trac.
de Neufville and Tsunokawa (1981) conrmed that scale economies existed at ports. Bendall and
Stent (1987, 1988), Talley (1988a,b, 1994), and Tongzon (1995) all suggested, or measured,
characteristics of the cost function faced by terminal operators. Finally, Allen (1977), Daughety
and Inaba (1978) and Daughety (1979) advanced the economic modeling of the decisions made by
carriers, in contexts other than the assignment of shipments to ports. Winston (1981a,b) and Nam

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

319

(1997) applied discrete choice models to freight transportation decisions. Malchow and Kanafani
(2001) made an initial application of the choice model to the assignment of shipments to individual ports. The present article extends the work of Malchow et al. by incorporating additional
variables and applying a form of the choice model that accounts for the correlation within the
multiple decisions made by each decision maker.

2. Methodology
In this research, we apply a choice model to the assignment of shipments to vessels/ports in
order to evaluate the competition between ports. The fundamental assumption is that a shipper,
by choosing a carrier, implicitly chooses a port for a particular shipment. But in the competitive
market of shipping it is the attributes of the door-to-door service that a carrier oers that
inuences the shipper choice. In other words, while a shipper may be deliberately choosing a
particular port for his or her shipments, it is the attributes of the service as oered by the carrier
that inuence that choice process. Therefore the assumption is that the shippers preference for a
port is wholly subsumed within the preference for and choice of a carrier that oers a service
through that port.
To estimate the choice model, we use data that describe shipments exported from the United
States in December 1999. We classify shipments into four commodity-types using the rst two
digits of the harmonized commodity code (HS). 2 The four commodity-types are bulk materials
(HS 25, 26), foods (HS 07, 08, 10), fabrics (HS 52, 54), and manufactured goods (HS 85). The data
represent exports to eight foreign countries: Australia, Brazil, Egypt, Germany, Japan, Saudi
Arabia, South Africa, and the United Kingdom. The commodity classications provide variations
in the values (and related characteristics) of the shipments, and the destination countries provide
geographic distribution. We restrict shipments to those for which the carrier of record had a
schedule listed with the Journal of Commerce, since we use a carriers schedule to measure particular variables. Table 1 shows the distribution of shipments by country and commodity-type. 3
For each carrier, the choice set consists of eight ports: Charleston, South Carolina; Long Beach,
California; Los Angeles, California; New York, New York; Oakland, California; Savannah,
Georgia; Seattle, Washington; and Tacoma, Washington.
We select the eight ports in our choice set for two primary characteristics: (i) the volume of
trade moved through the port, and (ii) the proximity of the port to other signicant ports. If two
ports were geographically close, factors other than location may inuence the choice between
them. We want to capture the impact of such factors. Table 2 shows the distribution, among
ports, of shipments within our sample set.

The Harmonized System is an international six-digit commodity classication developed under the auspices of the
Customs Cooperation Council. Individual countries have extended it to ten digits for customs purposes, and to eight
digits for export purposes. The system classies goods by what they are, not according to their stage of fabrication, use,
or origin. The rst pair of digits represents a chapter, the next pair a heading, the third pair a subheading.
3
The data set was reduced to include only shipments that were moved from one of the 48 contiguous United States
through one of the eight ports by one of the carriers whose schedules are available from the Journal of Commerce.

320

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

Table 1
The distribution of shipments, by destination and commodity-type
HS code

Australia

07
08
10
25
26
52
54
85

127
160
7
63
1
33
25
164

All

580

Brazil

Egypt

Germany

Japan

Saudi
Arabia

2
8
1
18
4
25
17
61

1
33
0
7
0
0
2
15

13
169
7
85
2
26
35
79

722
836
70
352
13
190
12
254

6
64
2
6
0
18
7
69

136

58

416

2449

172

South
Africa

United
Kingdom

Total

2
6
0
23
2
21
18
36

57
141
12
56
7
52
28
162

930
1417
99
610
29
365
144
840

108

515

4434

Table 2
The distribution of shipments among ports
Rank

Port

Shipments

1
2
3
4
5
6
7
8
9
10

Oakland, CA
Los Angeles, CA
Charleston, SC
Long Beach, CA
New York/New Jersey
Seattle, WA
Savannah, GA
Houston, TX
Norfolk, VA
Tacoma, WA

1314
1010
675
650
618
515
462
346
290
254

Many factors aect the choice process being modeled. Before discussing these factors, let us
dene the scenario in which a carrier selects a port for a shipment. First, the carrier selects the port
and vessel for each shipment simultaneously. Remember that when modeling the short-term
decision, we assume that the long-term eet assignment has already been established. Thus, we
can represent each port by the vessel distribution rather than the characteristics of a particular
vessel. We also assume that sucient space exists for each shipment on vessels scheduled along
each route. Because we analyze exports rather than imports, this situation should hold. During
the 1990s, the ratio of imports to exports uctuated around 1.5, and carriers often transport
empty containers to ll slots on outbound vessels. Empty space suggests that each shipment is
exported through the optimal port.
We model the systematic utility, Vnj of each port as a linear function of ve variables: 4
Vnj aj b1
Onj b2
Inj b3
Hinj b4
Cinj b5
Pinj ;
4

We modeled the decisions with other variables as well. For example, we used the average number of sailing days in
place of oceanic distance or the frequency of voyages in place of the headway between voyages. In each case, the
explanatory power of the model decreased.

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

321

where, Onj is the oceanic distance to the destination of shipment n from port j (km, 1000s); Inj , the
inland distance from the origin of shipment n to port j (km, 1000s); Hinj , the average headway
between voyages by carrier i from port j to the destination of shipment n (days); Cinj , the average
size of vessels sailed by carrier i from port j to the destination of shipment n (TEUs, 1000s), and
Pinj , the probability that port j would be the last port visited by a vessel sailed by carrier i to the
destination of shipment n; a; b, coecients estimated in the model.
The variables Onj and Inj are of course independent of the carrier, with the remaining variables
being measures of the carrier as well as the port. 5 For the variable Onj , we use the shortest sailing
distance from port j to the destination of shipment n. This is an approximation of the actual
sailing distance, which could not be measured due to the complexity and variability of carriers
schedules. For Inj , we use the inland road distance, which should approximate the inland rail
distance as well.
We measure the variables Hinj , Cinj , and Pinj for each carrier through an Internet database
maintained by the Journal of Commerce. 6 For each destination, we use the schedule of all
voyages from one of eight United States ports to any port near the destination. 7 We measure the
variables Hinj and Pinj for each carrier directly. We measure the capacity of each vessel scheduled
along a corridor, in TEUs, through the website MaritimeData.com. We calculate the variable Cinj
to represent the average capacity of vessels sailing along the corridor for carrier i.
The variables in this choice function are selected to represent the common objective of the
shipper and the carrier: to get each shipment from its origin to its destination as eciently as
possible. This eciency is ultimately dependent on transit time and cost. For each shipment, four
factors inuence the transit time associated with each port:
ii(i)
i(ii)
(iii)
(iv)

the distance from the origin to the port,

the time needed to transfer the shipment from the ground to the vessel,
the time incurred as the vessel calls at other ports in transit, and
the oceanic distance from the port to the shipments destination.

Likewise, four factors inuence the operating cost associated with each port:
ii(i) the inland distance from the origin to the port,
i(ii) the charges assessed by the port,
5
In fact, the headway between (or frequency of) voyages was found to be insignicant when included as the average
across all carriers, a result quite dierent from the model resulting from carrier-specic values. We would intuitively
expect the carrier-specic values to have more explanatory power as well.
6
The site at the time of writing was at: http://www.joc.com/scheds/index.shtml. Because data for a vessel is
maintained only until the vessels voyage has been completed, data for December 1999 were no longer available.
Instead, data for March 2000 were used to represent the variables. Comparison was made with the schedule for June
2000 (likewise, separated by three months) and a correlation coecient of 0.95 existed between the schedules, implying
that carriers schedules did not change much over three months.
7
These records were downloaded from the web and analyzed with a spreadsheet. The twelve US ports consisted of
the eight within the choice set, along with Houston, Miami, Norfolk, and Portland. The foreign ports were not
constrained to the country that was the destination of the shipment. For example, shipments destined for Germany
would also be aected by the nearby ports of Rotterdam and Antwerp. A complete list is given in Appendix A.

322

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

(iii) the oceanic distance from the port to the destination of the shipment, and
(iv) the average vessel size, representing economies-of-scale and density.
Before continuing, we must mention those variables that are not included in the model but may
be signicant on a disaggregate level. These include port charges, the cost of the transportation
services, and the intermodal transfer process at each port.
Increased port charges could make a port less attractive to a decision maker, with the most
prevalent port charge being wharfage. Because of complexities with port taris and the prevalence
of service contracts, port charges are dicult to measure accurately on a disaggregate level.
Industry representatives have, however, suggested that port charges are relatively insignicant.
Ocean freight rates are no longer publicly disclosed, as of the Ocean Shipping Reform Act
(OSRA) of May 1999 (Lewis and Vellenga, 2000). For two reasons, fortunately, rates might not
be signicant in port selection. First, inspection of the taris and service contracts available
through the Federal Maritime Commission showed that the freight rates prior to OSRA varied
little among West Coast ports. Rates also varied little among East Coast ports, with a slight
dierence between the rates for ports on dierent coasts. Second, a shipper cares little about the
intermediary points through which a shipment is moved, so long as the shipment arrives at the
destination at the expected time. Thus, the port should not aect the rate that a shipper is willing
to pay for transportation services. There is also some empirical evidence from Nam (1997), who
analyzed the selection of mode for shipments. He found that the rate charged was in most cases
insignicant, suggesting that either service characteristics were more important or rates did not
vary by alternative.
Finally, data are not available for intermodal transfer time. However, even if data were
available, two factors would complicate their inclusion into the model. First, we would have to
collect data for each terminal operated by each carrier at each port within the choice set, or
approximately 100 terminals. Second, and perhaps more importantly, a carrier could know a
shipments intermodal transfer time prior to a decision only as an expected value, which lacks the
variability that would be necessary for inclusion into a choice model.

3. An alternate formulation of the choice model

Under the traditional choice model, the utility of a port for a shipment is a linear function of
the variables describing that port. The carrier is observed as selecting one port from among the
alternatives, and we assume that the carrier has selected the port that provided the greatest utility
in the context of that shipment. By observing the decisions for multiple shipments, we can estimate the importance of the factors that describe each alternative. For each port, we estimate the
contribution of each factor to its utility, and we estimate a factor (referred to as the port-specic
constant) that represents the average utility of all unobserved factors. These estimates inuence
the likelihood of each decision by the carrier, and we estimate each factors contribution to
maximize the likelihood of the observations.
The data that we model represent panel data, with the shipments moved by each carrier representing a separate group. Correlation likely exists among the decisions of a given carrier. For
example, the intermodal transfer process could inuence each carriers selection of a port for each

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

323

shipment, and the transfer process at each port likely varies by carrier. This factor could not be
included directly since it does not vary across shipments. It thus aects the constant term associated with each port, and thus the constants should vary by carrier. 8
To account for the individual carriers, we could estimate a set of port-specic constants for
each carrier. This, however, would require the estimation of 288 constants (36 carriers, 8 ports).
For some carriers the number of shipments would be too small for estimation. Chamberlain
(1980) introduced an alternative model that accommodates panel data. In this model, rather than
modeling the selection of a port for each shipment, we model a carriers aggregate distribution of
shipments from the set of feasible distributions. The assignment of each shipment to a port would
dene each distribution.
Two fundamental properties determine the feasibility of each hypothetical distribution. First,
each shipment must be transferred through exactly one port. Second, the number of shipments
predicted by the distribution to move through each port must equal the number actually observed
as moving through that port. 9 In mathematical notation, let
xinj the vector of attributes discussed earlier that inuence the choice by carrier i of port j for
shipment n (i.e. Onj , Inj , Hinj , Cinj , and Pinj ),
winj 1 if carrier i actually sends shipment n through port j, and 0 otherwise,
sij the number of shipments moved by carrier i through port j (Rn winj ), and
dinj 1 for each feasible distribution, and 0 for all others.
The two constraints specify that
8in, Rj winj 1, and
8ij, Rn dinj sij .
A distribution that meets these constraints is considered feasible. Chamberlains method, rather
than maximizing the probability that the carrier selects the chosen port for each shipment,
maximizes the probability that the carrier selects the observed distribution from the set of feasible
distributions. The log-likelihood for all observations is then

X
i

"

X
ln exp b0
xinj winj
n;j

!,

X
d2D

X
exp b0
xinj dinj

!#
;

n;j

in which D represents all feasible distributions of the shipments.

to the logit
P In relating this model P
model, we note that the term Vnj has been replaced by b0 n:j xinj winj and Vnk by b0 n:j xinj dinj ,
summed over all feasible distributions D. With this formulation, the port-specic constants

8
If the alternative-specic constants remain constant across carriers, then the unobserved error term would be
correlated for each carriers shipments and not distributed with a mean of zero, as required for the model estimation.
9
The alternative-specic constant is estimated in the logit model such that the predicted share will be equivalent to
the observed share for each alternative.

324

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

Table 3
Results of the standard multinomial logit model estimation (ignoring panel eects)
Variable

Estimate

Standard error

Z-statistic

P -statistic

Oceanic distance (O)

Inland distance (I)
Sailing headway (H )
Vessel capacity (C)
Prob. of last (P )

)0.09
)0.67
)0.04
)0.11
0.01

0.01
0.02
0.00
0.04
0.00

)12.6
)35.8
)25.3
)3.1
13.9

0.00
0.00
0.00
0.00
0.00

A_Charleston
A_Long Beach
A_Los Angeles
A_New York
A_Oakland
A_Savannah
A_Seattle
A_Tacoma

0.05
0.07
0.47
)0.24
0.38
)0.16
)0.23
0

0.10
0.09
0.08
0.10
0.08
0.10
0.10

0.5
0.8
5.8
)2.4
4.8
)1.5
)2.4

0.64
0.45
0.00
0.02
0.00
0.13
0.02

Log-likelihood
Log-likelihood, constants only
Log-likelihood, no coecients

)6242.2
)8650.6
)9220.2

disappear from the equation. (More precisely, the sum of the constants across the feasible
distributions does not vary and thus can not inuence the choice.)
4. Choice-model estimation
We rst estimate a standard multinomial choice model, in which the probability of choosing
port j is given by
eVinj
Pinj P V ;
ink
ke

with aij , the port-specic constant that is estimated for each carrier-port combination, constant
across carriers. Table 3 shows the results of the estimation.
The estimate for each of the ve coecients is statistically signicant at a level beyond 99%.
However, only four of the ve estimates have the expected sign. The negative coecient of vessel
capacity is counterintuitive. Why might this be? Perhaps there is no immediate advantage in
placing a shipment aboard a larger vessel if space is available. Given the trade ows in and out of
the US, space should always be available for exports. Therefore, our expectation about the impact
of vessel capacity is not denite. We learn later that the impact of vessel capacity, when modeled
alone, is actually positive.
5. The Chamberlain model
To examine these results further, we estimate the Chamberlain model. The set of feasible
distributions for some carriers would be computationally cumbersome, so we instead use sample

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

325

Table 4
Results of the multinomial logit model estimation, accounting for panel data characteristics
Variable

Coecient estimate

Standard error

Z-statistic

P -statistic

Oceanic distance (O)

Inland distance (I)
Sailing headway (H )
Vessel capacity (C)
Prob. of last (P )

)0.12
)0.77
)0.03
)0.02
0.003

0.01
0.02
0.00
0.05
0.001

)15.6
)34.7
)16.2
)0.4
2.5

0.00
0.00
0.00
0.70
0.01
)2980.4
)6525.8
)7533.6

Log-likelihood
Log-likelihood from constants
Log-likelihood, no coecients

Table 5
Results of the multinomial logit model estimation, accounting for panel data characteristics and ignoring vessel
capacity
Variable

Coecient estimate

Standard error

Z-statistic

P -statistic

Oceanic distance (O)

Inland distance (I)
Sailing headway (H )
Prob. of last (P )

)0.12
)0.78
)0.03
0.003

0.01
0.02
0.00
0.001

)15.9
)34.9
)17.1
2.5

0.00
0.00
0.00
0.01

Log-likelihood
Log-likelihood from constants
No coecients

)2980.5
)6525.8
)7533.6

sets of the shipments to represent each carrier. 10 We select ve-shipment samples for each of the
nineteen carriers that had more than 50 shipments, with 100 samples collected for each of these
carriers. Table 4 shows the coecients estimated with the 1840 observations that were retained. 11
Each coecient is signicant with the exception of vessel capacity. To understand why, we
examine the signicance of each variable individually with the Chamberlain model. Inland distance provides the greatest explanatory power. The sign of each variables coecient is consistent
with the sign estimated in the multinomial model for each variable except vessel capacity. Due to
this inconsistency, as well as the fact that the capacity of a vessel might not aect port selection,
given that space is available, we remove vessel capacity from further models. Ignoring vessel
capacity, we estimate the model as shown in Table 5. The likelihood-ratio test conrms that vessel
capacity is not a signicant variable (Ben-Akiva and Lerman, 1985). We also use Hausmans test
to conrm that the Chamberlain model, in which the alternative-specic constants dier between
carriers, describes the shipment-decisions better than the traditional model.

10
For example, Evergreen represented 527 shipments within our dataset. These shipments could be distributed
among the eight ports in one of 8527 distributions, a number too large for us to begin with for consideration.
11
We discarded sixty observations because the number of feasible distribution was too large for LIMDEP to handle.
LIMDEP is the statistical package used to estimate the choice model.

326

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

6. Discretionary cargo
Here we postulate that the decision made for discretionary cargo, cargo originating in a region
that does not contain a port, diers from that made for cargo originating in a ports hinterland.
To see how the decision process might dier for discretionary cargo, we estimate a model using
only shipments that originated in the central United States. For these shipments, inland distance
would not vary as much among ports, allowing the impact of other variables to increase. At rst
glance, the results of this model appear similar to those from the model for all shipments.
However, closer examination of the variables reveals that the variables do play dierent roles for
the discretionary cargo. Table 6 shows the results.
For discretionary cargo, the probability of being the last port visited is signicantly more
important. We compare the impact of this to the impact of other variables in Table 7 and nd that
Table 6
The Chamberlain logit model estimated for discretionary cargo
Variable

Coecient estimate

Standard error

Z-statistic

P -statistic

Oceanic distance (O)

Inland distance (I)
Sailing headway (H )
Prob. of last (P )

)0.29
)1.75
)0.05
0.019

0.03
0.10
0.01
0.004

)11.3
)16.7
)8.9
5.0

0.00
0.00
0.00
0.00
)3820.0
)2275.9
)681.2

Log-likelihood
Log-likelihood from constants
No coecients

Table 7
The importance of being the last port visited for discretionary cargo
Ratio

Meaning

bO =bP

41.0
The increase in the probability of being
the last port that would be equivalent
to a reduction of 1000 km in oceanic transit
258.3
The increase in the probability of being
the last port that would be equivalent
to a reduction of 1000 km in inland transit
11.0
The increase in the probability of being
the last port that would be equivalent to a
reduction of one day, expected headway
0.16
The decrease in inland distance (km) that
would be equivalent to a reduction of one km,
oceanic transit
3.7
The decrease in headway (days) that would be
equivalent to a reduction of 1000 km in
oceanic transit
The decrease in headway (days) that would be 23.5
equivalent to a reduction of 1000 km in inland
transit

bI =bP

bH =bP

bO =bI

bO =bH

bI =bH

Value, all shipments

Value, Midwest shipments

15.1

91.9

2.8

0.16

5.4

32.9

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

327

the relative importance of other variables decreases signicantly. Many industry analysts suggest
that discretionary cargo in particular is sent through the port visited last by a vessel to minimize
transit time. This evidence of shifting values supports this idea. The importance of being the last
port visited is three times as large, when compared to other factors, for discretionary shipments as
it is for all shipments.
Though we wish to emphasize the magnitude of this variables changing importance, we must
recognize that uncertainty exists in our estimated values. Each estimated coecient has a standard
error. A ratio between two estimates has an even larger standard error, and the ratio between two
ratios has a still-larger error. We conrm the statistical dierence of these probabilities with the
construction of density functions in Malchow (2001b).
From the results in Table 7 we also see that a ports share of shipments from the Midwest is less
aected by the headway between voyages, relative to the distances. The additional day or so of
headway becomes less signicant when combined with an additional 48 h of inland transit.

7. Commodity-specic models
We now consider the proposition that the importance of attributes varies with commodity-type.
The commodity groups examined and their characteristics are shown in Table 8.
The shipment size for each shipment corresponds to that led with the shipments customs
form. Because the declared value of each shipment is condential, the Journal of Commerce
estimated the value of each shipment according to the trade route and commodity code associated with each shipment. We nd that the average shipment size for a commodity decreases as
its average value increases, likely to minimize shippers inventory cost. We expect the importance of dierent attributes of each port to vary with the characteristics that describe each
shipment.
The negative impacts of distance are the associated transit time and operating costs, and we
expect the transit time to be less important relative to the operating costs for lower-valued
commodities. Inland distance is covered by modes (rail, truck) that are faster and more expensive
than water-based transportation; thus, shippers of lower-valued goods would place a lower priority on oceanic distance than on inland distance. Carriers would more likely send lower-valued
commodities through nearby ports. For example, a low-valued commodity being sent from

Table 8
Characteristics of the shipments from dierent commodity groups
Commodity

# Records

Shipment size (metric tons)

Average value ($/metric ton)

Bulk
Fruits and vegetables
Fabrics
Manufactured

610
2347
509
840

53.9
30.6
27.7
9.8

285
1198
4287
11,087

All

4434

30.5

1885

328

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

Table 9
Results of the Chamberlain model as estimated for the dierent commodity-types
Commodity

Variable

Coecient
estimate

Standard
error

Z-statistic

P -statistic

Fruits and vegetables

(HS 07, 08)
(1500 simulated shipments)

Oceanic distance (O)

Inland distance (I)
Sailing headway (H )
Prob. of last (P )

)0.38
)1.05
)0.02
)0.01

0.04
0.05
0.00
0.00

)10.0
)22.2
)7.4
)3.3

0.00
0.00
0.00
0.00

Bulk
(HS 25)
(1100 simulated shipments)

Oceanic distance (O)

Inland distance (I)
Sailing headway (H )
Vessel capacity (C)
Prob. of last (P )

)0.12
)0.79
)0.05
0.58
)0.03

0.01
0.04
0.00
0.16
0.00

)8.4
)20.7
)11.4
3.7
)6.7

0.00
0.00
0.00
0.00
0.00

Fabrics
(HS 52, 54)
(800 simulated shipments)

Oceanic distance (O)

Inland distance (I)
Sailing headway (H )
Prob. of last (P )

)0.12
)0.51
)0.04
0.01

0.01
0.04
0.01
0.00

)9.1
)13.5
)6.3
5.7

0.00
0.00
0.00
0.00

Manufactured
(HS 85)
(1300 simulated shipments)

Oceanic distance (O)

Inland distance (I)
Sailing headway (H )
Vessel capacity (C)
Prob. of last (P )

)0.26
)0.47
)0.01
0.43
0.00

0.01
0.03
0.00
0.12
0.00

)23.3
)17.5
)4.0
3.5
)2.7

0.00
0.00
0.00
0.00
0.01

California to the United Kingdom would be loaded at a California port and sent on an extended
ocean voyage, whereas a higher-valued commodity might be transshipped via landbridge to a
waiting vessel on the East Coast. Table 9 shows the results of the estimation for each commodity,
and the results do agree with our expectations.
Table 10 shows the marginal rate of substitution between oceanic distance and inland distance
for each of the dierent commodities.
As expected, the marginal rate of substitution between inland and oceanic transit generally
increases with the commodity value, a minor exception perhaps due to the perishability of fruits
and vegetables. This condition becomes important when examining the shares for certain commodities from carrier-specic models. With carrier-specic models, we can examine how a ports
share is aected by distance and how this impact varies with the value of the commodity. Again
using the likelihood-ratio test, we nd that the model estimated for each commodity is signicantly more explanatory than the generic model. This suggests that trac forecasting models use
commodity-specic data to enhance accuracy of predictions. 12

12
Certain studies (Heaver, 1972; Bryan, 1974; Brooks and Button, 1996) have used the classication of shipments by
commodity-type to show that the transit freight rate, on a unit basis, tends to vary with the density of each commoditytype. However, we noted earlier that the rates for commodity-types tend not to vary between individual ports within
coastal ranges.

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

329

Table 10
The importance of inland and oceanic transit for each commodity-type
Commodity

Average value ($/metric ton)

The marginal rate of substitution

between inland transit and oceanic transit

Bulk
Fruits and vegetables
Fabrics
Manufactured

285
1198
4287
11,087

0.15
0.37
0.23
0.56

1885

0.16

All

8. Carrier-specic models
To measure the eect of each variable, we could use the elasticity of choice, as described in
Train (1986). We rst estimate models with port-specic constants for each carrier, subject to the
constraint that the variable-specic coecients equal those estimated for the combined data set.
We estimate the model for two carriers, American President Lines (APL) and Maersk/SeaLand.
The estimated elasticities are given in Table 11. The italicized ports for each carrier are those
through which the carrier transported no shipment. The logit model allows innitesimally small
probabilities to exist, but the predicted share of these alternatives would be zero.
From these estimates, we see that distance inuences port selection most. The probability of
being the last port appears to have a very inelastic eect, and the eect of the headway between
voyages is largest for ports not visited at all. 13 With the likelihood-ratio test, we nd that the
generic model would be rejected for the model estimated for each carrier; in other words, the
behavior of each carrier is not consistent.

9. Discussion
One interview with a carrier suggested that the selection of a port is not entirely predictable,
implying that the group deciding often does so without much evaluation. If this were true then one
might question the worth of models of the type presented here. However, a fundamental belief
underlying economic analysis is the rational behavior principle. Carriers will in most cases make
rational decisions and it might be that some decisions do not require the same level of analysis.
We should be able to model this process to some extent, if for anything to estimate the implications of rational behavior.
The carrier-specic models allow analysis of the share of trac for each port. Port managers
could use such models in marketing. Ports can consider marketing to improve their position in an

13

Recall that to allow the variable to be assigned a nite value, ports for which the observed frequency of sailing
during March 2000 was zero were assigned an arbitrarily high headway of sixty days. Note also that the elasticity with
regard to being the last port visited would likely increase were the cargo of the discretionary sort.

330

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

Table 11
The choice elasticities for the individual ports, for two carriers
Carrier

Port

Oceanic distance (O)

Inland distance (I)

Sailing headway (H )

Prob. of last (P )

APL

Cha
LB
LA
NY
Oak
Sav
Sea
Tac

)1.62
)1.43
)1.02
)1.73
)1.06
)1.84
)1.38
)1.22

)1.97
)1.29
)1.03
)2.17
)1.08
)2.10
)1.40
)1.29

)0.15
)0.41
)0.23
)0.16
)0.26
)0.48
)0.33
)0.43

0.01
0.00
0.00
0.02
0.12
0.00
0.00
0.00

Maersk

Cha
LB
LA
NY
Oak
Sav
Sea
Tac

)1.18
)1.24
)1.53
)1.35
)1.01
)1.48
)1.67
)1.52

)1.90
)1.13
)1.22
)2.10
)1.05
)2.14
)1.48
)1.36

)0.10
)0.18
)0.22
)0.17
)0.17
)1.02
)1.11
)1.18

0.05
0.00
0.00
0.03
0.04
0.00
0.04
0.02

established market or enter a new market. Estimates of the eect of certain factors could be used
to assess the worth of port investments.
To show how this model could be used, we create simpler environment with three of four
ports: 14
(1)
(2)
(3)
(4)

Los Angeles or Long Beach,

Oakland, and
Seattle or Tacoma,
Charleston.

We examine the competition of these ports for a shipment under various scenarios. One
hypothetical shipment would be destined for Japan, to be moved by APL. 15 The independent
variable represents the shipments origin and moves inland from the Port of Oakland, and we
estimate the distance to each port geometrically. Fig. 1 shows the market share for each port in
this scenario.
The predicted market share of the Port of Oakland decreases from 64% within its hinterland to
53% as the origin of the shipment moves inland. The market share of competing ports increases as
expected to account for this lost share. Thus, each port does hold an advantage for shipments
within its hinterland, but the predicted advantage is not sucient to ignore competition.

14

The model applies regardless of the available choice set. In addition, because carriers do not operate multiple
terminals within a region, each port competes with ports in other regions for the assignment of a shipment.
15
To simplify the analysis, the variable representing the probability of being the last port selected is ignored.

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

331

100%
90%
80%
70%

Share

60%
50%
40%
30%
20%
10%
0%
0

500

1000

1500

2000

2500

3000

3500

4000

Distance inland from Oakland (km)

Oakland

Seattle

Los Angeles

Fig. 1. The impact of inland origin location on ports market shares for a theoretical shipment moved by American
President Lines to Japan.

The impact of inland distance becomes more apparent when comparing market shares of ports
along dierent coasts. In another hypothetical case, the export of a shipment by Maersk to Japan,
the market share that would be captured by each of three ports is shown in Fig. 2. In this gure,
we replace the Port of Oakland with the Port of Charleston, which is located along the opposite
coast. Initially, market share for the Port of Charleston remains insignicant, while the Port of
Tacoma attracts a small market share away from the Port of Long Beach. 16 The Port of
Charleston does not begin to steal signicant market share away from the West Coast ports until
the origin of the shipment has shifted halfway across the country, and Charleston does not steal
signicant market share until the origin has shifted even further.
Further analysis shows that the potential for competition, however, is impacted by the value
of the good being shipped. We suggested earlier that a carrier would transport higher-valued
shipments to a distant port to minimize oceanic distance relative to inland distance. Fig. 3 shows
this, as the share predicted for the Port of Charleston for a shipment bound to Japan is shown to
dier for a manufactured good relative to the share for a generic shipment. For this example, we
estimate a model for P&O Nedlloyd. Clearly, a ports competition with a landbridge-alternative
is greater for higher-valued commodities. 17 This conrms that lower-valued goods are more
likely to be loaded at a neighboring port and transited a longer distance via ocean, and highervalued, time-sensitive goods are more likely to be shipped via landbridge to a port with greater

16

The Port of Long Beach is closer to Oakland than Tacoma, so with the shifting of the origin, Long Beachs
advantage is lessened.
17
The generic commodity would be of lower value than the highest-value manufactured goods.

332

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

100%
90%
80%

Market share

70%
60%
50%
40%
30%
20%
10%
0%
0

500

1000

1500

2000

2500

3000

3500

4000

Distance inland from Oakland (km)

Tacoma

Charleston

Long Beach

Fig. 2. The impact of inland origin location on ports market shares for a theoretical shipment moved by Maersk to
Japan.

100%
90%

Market share

80%
70%
60%
50%
40%
30%
20%
10%
0%
0

500

1000

1500

2000

2500

3000

3500

4000

Distance inland from Oakland (km)

All

Manufactured

Fig. 3. The market share predicted for an Atlantic port for the shipment of dierent commodity-types by Maersk to
Japan.

access to the shipments destination. Additional models could be evaluated to get more precise
estimates.
In evaluating oceanic distance, we nd that its impact is again greatest when comparing ports
from opposite coasts. Ports adjacent along one coast would have near-equivalent oceanic distances, such that an advantage could be gained only when linking calls for the scheduling of a
vessel.

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

333

100%
90%

Market share

80%
70%
60%
50%
40%
30%
20%
10%
0%
0

10

20

30

40

50

60

Headway between voyages at the Port of Oakland (days)

Oakland

Los Angeles

Seattle

Fig. 4. The predicted impact of headway between voyages on ports market shares for theoretical shipments moved by
APL from Oregon to Japan.

Fig. 4 represents the signicance of the headway between voyages for a theoretical shipment
moved by American President Lines from Oregon to Japan. Though the market share decreases
steadily as headway increases, the actual impact is not as dramatic. To decrease the headway from
sixty days to thirty days, a carrier needs to add only one voyage per month. A second voyage
reduces the average headway again from thirty days to fteen days. Once a carrier has scheduled a
sucient number of voyages, an incremental voyage adds insignicantly to a ports market share.
The potential impact becomes more apparent when observing the predicted market share as a
function of frequency. 18 The impact of sailing frequency decreases quickly once a minimum
number of voyages (on the magnitude of one per week, or 4+ per month) have been scheduled. An
additional voyage would reduce the expected headway by an insignicant amount. A second
voyage (if the voyages were spaced evenly) would reduce the headway by fteen days, and a third
voyage by ve days, but a fth sailing would reduce the headway by only one day. Therefore, so
long as a port has voyages scheduled at the frequency of one per week or greater, additional
voyages do little but increase the capacity available for shipments. Fig. 5 shows the impact of
sailing frequency on the distribution of shipments between ports.
In a nal example, we analyze the importance of being the last port visited by a vessel. Recall
that the signicance of being visited last is greatest with discretionary cargo. Fig. 6 represents the
share predicted for the Port of Oakland, when competing with Los Angeles and Seattle, for
shipments transported by APL from Kansas to Japan. As mentioned earlier, the signicance of
being visited last is much greater for a port with discretionary cargo. Simply by convincing APL
to make all of its last calls there, the Port of Oakland could increase its predicted market share for
discretionary cargo from 24% to 85%.

18

We used headway in the choice model because headway is more linearly related to the time spent by a shipment in
transit, of which it is the decision makers objective to minimize.

334

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

100%
90%
80%

Market share

70%
60%
50%
40%
30%
20%
10%
0%
0

10

15

20

25

30

Voyages at the Port of Oakland per month

Oakland

Los Angeles

Seattle

Fig. 5. The predicted impact of sailing frequency on ports market shares for theoretical shipments moved by APL from
Oregon to Japan.

100%
90%

Market share

80%
70%
60%
50%
40%
30%
20%
10%
0%
0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

Probability of being last visited (%)

All

Discretionary

Fig. 6. The impact of being visited last on the market share predicted for the Port of Oakland, for theoretical shipments
moved by APL from Kansas to Japan.

10. Conclusion
These results have in many ways rearmed the results of earlier qualitative analysis. We have
found that the variables furthest from the control of port authorities, the oceanic and inland
distances, have the greatest impact on carriers distribution of shipments. We have found other
factors to be signicant, but more so in the context of discretionary cargo. Port managers could
recognize these factors when marketing themselves to carriers, with particular regard to the

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

335

location their port holds in a vessels schedule. In another important nding, choice behavior
varies signicantly across carriers as well as commodities. When forecasting trac for a port or
marketing to carriers, analysts should recognize these dierences. Managers might recognize that,
in the context of shipment assignment, many of the investments being made might not be necessary. Of interest with future research would be changes that have occurred to the choice process
through time as a result of technological change or the development of linked transportation
networks.
In addition, the exports analyzed here represent only half of port trac. We could expect the
decisions to be similar for imports, except for certain factors. One such factor is the desire of
importers to concentrate trac around a particular distribution center. Storage space should not
aect the decision for exports, due to imbalances and carriers desire to move shipments quickly,
but it could aect imports if shippers are using terminal space for storage. The capacity available
on vessels could be of more importance for imports, due to the trade imbalance between imports
and exports.
We must remember also that the distribution of shipments across ports is but one part of the
larger picture. An equally signicant (if not more signicant) task faced by the carriers is the
assignment of vessels to particular routes. Further examination could be made of the carriers
selection of the last port to visit or, in the context of imports, the rst port visited. The scheduling
of vessels, coupled with our analysis of shipment distribution, could aect the future predictions
for port trac.
References
Al-Kazily, J., 1979. Choice of technology for containerized shippinga study of interaction between vessels and ports.
Dissertation submitted to the University of California at Berkeley.
Allen, W.B., 1977. The demand for freight transportation: a micro-approach. Transportation Research 11 (1), 914.
Bardi, E.J., 1973. Carrier selection from one mode. Transportation Journal 13 (1), 2329.
Ben-Akiva, M.E., Lerman, S.R., 1985. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT
Press, Cambridge, MA.
Bendall, H.B., Stent, A.F., 1987. On measuring cargo handling productivity. Maritime Policy & Management 14 (4),
337343.
Bendall, H.B., Stent, A.F., 1988. Cargo handling productivity. Part 2. The ndings. Maritime Policy & Management 15
(2), 97106.
Benford, H., 1981. A simple approach to eet deployment. Maritime Policy & Management 8 (4), 223228.
Brooks, M., 1984. An alternative theoretical approach to the evaluation of liner shipping: Part 1. Situational factors.
Maritime Policy & Management 11 (1), 3543.
Brooks, M., 1985. An alternative theoretical approach to the evaluation of liner shipping: Part 2. Choice criteria.
Maritime Policy & Management 12 (2), 145155.
Brooks, M., Button, K., 1996. The determinants of shipping rates: a North Atlantic case study. Transport Logistics 1
(1), 2130.
Bryan, I., 1974. Regression analysis of ocean liner freight rates on some Canadian export routes. Journal of Transport
Economics and Policy 8 (2), 161173.
Chamberlain, G., 1980. Analysis of covariance with qualitative data. Review of Economic Studies 47, 225238.
Daughety, A.F., 1979. Freight transportation demand revisited: a microeconomic view of multimodal, multicharacteristic service uncertainty and the demand for freight transport. Transportation Research B 13 (4), 281288.
Daughety, A.F., Inaba, F.S., 1978. Estimating service-dierentiated transport demand functions. Transportation
Research Record 668, 2330.

336

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

de Neufville, R., Tsunokawa, K., 1981. Productivity and returns to scale of container ports. Maritime Policy &
Management 8 (2), 121129.
DEste, G.M., Meyrick, S., 1992a. Carrier selection in a RO/RO ferry trade: Part 1. Decision factors and attitudes.
Maritime Policy & Management 19 (2), 115126.
DEste, G.M., Meyrick, S., 1992b. Carrier selection in a RO/RO ferry trade: Part 2. Conceptual framework for the
decision process. Maritime Policy & Management 19 (2), 127138.
Foggin, J.H., Dicer, G.N., 1985. Disappearing hinterlands: the impact of the logistics concept on port competition.
Journal of the Transportation Research Forum 26 (1), 385391.
Foster, T., 1978a. Ports: what shippers should look for. Chiltons Distribution Worldwide 77 (1), 4143.
Foster, T., 1978b. Ports: what shippers do look for. Chiltons Distribution Worldwide 77 (1), 4448.
Foster, T., 1979. Whats important in a port. Chiltons Distribution Worldwide 78 (1), 3236.
Garrod, P., Miklius, W., 1985. The optimal ship size: a comment. Journal of Transport Economics and Policy 19 (1),
8390.
Griths, J.D., 1976a. Optimal handling capacity at a berth. Maritime Studies and Management 3 (3), 163167.
Griths, J.D., 1976b. Unloading facilities at an iron-ore terminal. Maritime Studies and Management 3 (4), 237243.
Hanelt, R.L., Smith, D.S., 1987. The dynamics of west coast container port competition. Journal of the Transportation
Research Forum 28 (1), 8291.
Hayuth, Y., 1981. Containerization and the load center concept. Economic Geography 57 (2), 160176.
Heaver, T.D., 1972. Trans-Pacic trade, liner shipping, and conference rates. Logistics and Transportation Review 8
(1), 328.
Helmick, J.S., 1994. Concentration and Connectivity in the North Atlantic Liner Port Network, 19701990.
Dissertation submitted to the Department of Management and Logistics, University of Miami.
Jansson, J.O., Shneerson, D., 1978. Economies of scale of general cargo ships. Review of Economics and Statistics 60,
287293.
Jansson, J.O., Shneerson, D., 1985. A model of scheduled liner freight services: balancing inventory cost against
shipowners costs. The Logistics and Transportation Review 21 (3), 195215.
Kenyon, J.B., 1970. Elements in inter-port competition in the United States. Economic Geography 46 (1), 124.
Lago, A., Malchow, M., Kanafani, A., 2001. An analysis of carriers schedules and the impact on port selection.
Presentation at 2001 International Association of Maritime Economists Conference, Hong Kong.
Lane, D.E., Heaver, T.D., Uyeno, D., 1987. Planning and scheduling for eciency in liner shipping. Maritime Policy &
Management 14 (2), 109125.
Lewis, I., Vellenga, D.B., 2000. The ocean shipping reform act of 1998. Transportation Journal 39 (4), 2734.
Lim, S., 1998. Economies of scale in container shipping. Maritime Policy & Management 25 (4), 361373.
Malchow, M., 2001. An analysis of port selection. Dissertation submitted to the Department of Civil Engineering,
University of California.
Malchow, M., Kanafani, A., 2001. A disaggregate analysis of factors inuencing a ports attractiveness. Maritime
Policy & Management 28 (3), 361373.
Nam, K., 1997. A study of the estimation and aggregation of disaggregate models of mode choice for freight transport.
Transportation Research E: Logistics and Transportation Review 33E (3), 223231.
Perakis, A.N., 1985. A second look at eet deployment. Maritime Policy & Management 12 (3), 209214.
Perakis, A.N., Jaramillo, D.I., 1991. Fleet deployment optimization for liner shipping: Part 1. Background, problem
formulation, and solution approaches. Maritime Policy & Management 18 (3), 183200.
Slack, B., 1985. Containerization, inter-port competition and port selection. Maritime Policy & Management 12 (4),
293303.
Slack, B., 1996. Shipping lines as agents in the port industry. Maritime Policy & Management 23 (3), 289300.
Talley, W.K., 1988a. The role of US ocean ports in promoting an ecient ocean transportation system. Maritime Policy
& Management 15 (2), 147155.
Talley, W.K., 1988b. Optimum throughput and performance evaluation of marine terminals. Maritime Policy &
Management 15 (4), 327331.
Talley, W.K., 1990. Optimal containership size. Maritime Policy & Management 17 (3), 165175.
Talley, W.K., 1994. Port pricing: a cost-axiomatic approach. Maritime Policy & Management 21 (1), 6176.

M.B. Malchow, A. Kanafani / Transportation Research Part E 40 (2004) 317337

337

Tongzon, J.L., 1995. Determinants of port performance and eciency. Transportation Research A 29A (3), 245252.
Train, K.E., 1986. Qualitative Choice Analysis. The MIT Press, Cambridge, MA.
Winston, C., 1981a. A disaggregate model of the demand for intercity freight transportation. Econometrica 49 (4), 981
1006.
Winston, C., 1981b. A multinomial probit prediction of the demand for domestic ocean container service. Journal of
Transport Economics and Policy 15 (3), 243252.