Transportation Research Part E 47 (2011) 182–189

Contents lists available at ScienceDirect

Transportation Research Part E

journal homepage: www.elsevier .com/locate / t re

Preferences for alternative short sea shipping opportunities

Sean M. Puckett a, David A. Hensher a,⇑, Mary R. Brooks a,b, Valerie Trifts b

a Institute of Transport and Logistics Studies, Faculty of Economics and Business, University of Sydney, NSW 2006, Australiab School of Business Administration, Dalhousie University, Halifax, NS, Canada B3H 1W7

a r t i c l e i n f o a b s t r a c t

Article history:Received 1 May 2010Received in revised form 9 September 2010Accepted 1 October 2010

Keywords:Freight mode choiceShort sea shippingGeneralised mixed logitScale heterogeneityPreference heterogeneity

1366-5545/$ - see front matter � 2010 Elsevier Ltddoi:10.1016/j.tre.2010.10.002

⇑ Corresponding author. Tel.: +612 9351 0071; faE-mail addresses: sean.puckett@sydney.edu.au

Valerie.Trifts@dal.ca (V. Trifts).

This paper investigates the role of preference and scale heterogeneity in the mode choiceprocess of shippers in the Atlantic Canada–US eastern seaboard market. The generalisedmixed logit model is estimated to account for heterogeneity in preferences for frequencyof departure of freight transport services, along with heterogeneity in scale across respon-dents. The contributions of the paper to the literature are: the revelation of significant pref-erence and scale heterogeneity in the sample; the estimated distribution of shippers’willingness-to-pay for gains in service frequency; and the confirmation that there is meritin accounting for scale heterogeneity in future and revisited choice studies.

� 2010 Elsevier Ltd. All rights reserved.

1. Introduction

This paper builds on previous research by Brooks and Trifts (2008), which examined the mode choice processes of ship-pers in the Atlantic Canada–US eastern seaboard geographic market. That study built a descriptive model to explain howshippers purchase freight transportation services, and then, using that model, examined how they made choices about allo-cating freight between service packages in order to predict how they would make choices when faced with a new transportmode—short sea shipping, one which did not exist on the routes examined. As there was no rail option in the corridor, it wasa two-mode choice study. This study uses that data to explore further the willingness-to-pay (WTP) for particular choiceattributes, using a mixed logit model that accounts for preference and scale heterogeneity (see Fiebig et al., 2009; Greeneand Hensher, 2010). The findings of this paper are consistent with the descriptive evidence of Brooks and Trifts (2008)but, critically, extend them by estimating respondents’ willingness-to-pay for frequency of service.

As the paper’s contribution is to the literature in an area with very few contributions—freight mode choice in a short seasetting, we begin with a review of that literature. Most of the previous transport choice literature on long-distance (or non-urban) transportation activity has focused on choice in the passenger sector (e.g., Hensher, 1997; Hess et al., 2007; Balcombeet al., 2009).

Mode choice has been argued to be a function of distance (Jiang et al., 1999; Paixão and Marlow, 2002; Commonwealth ofAustralia, 2006), with shorter distances dominated by truck, medium distances dominated by rail and longer ones by ship-ping, with a contestable market in the intervening zones. Bolis and Maggi (2003) focused on the critical issue of time, a factorof substantial difference between truck and short sea (the slower but usually less expensive option). Therefore, it is impor-tant to extend the evaluation of willingness-to-pay within the context of distance captured as a transit time-related variable.

. All rights reserved.

x: +612 9351 0088.(S.M. Puckett), david.hensher@sydney.edu.au (D.A. Hensher), m.brooks@dal.ca (M.R. Brooks),

S.M. Puckett et al. / Transportation Research Part E 47 (2011) 182–189 183

Previous research in non-urban mode choice in the freight sector has predominantly focused on binary choicemodels (e.g., Bolis and Maggi, 2003; García-Menéndez et al., 2004). García-Menéndez et al. (2004) identified cost, transittime and frequency as determinants of mode choice in a freight mode choice problem. Investigating a road versus shortsea discrete mode choice in Europe, and using personal interviews with freight buyers in four industry sectors, they iden-tified the modal splits for these sectors, and found that shippers’ choice of short sea transport is more sensitive to changesin road transport prices than to changes in sea transport costs. Most important, they concluded that imposing an eco-taxon road transport could induce modal switching to short sea. This study is the only antecedent to explore willingness-to-pay.

As an area of transport research, freight mode choice is a complex one to study. Unlike passenger travel, fewer, moresophisticated buyers take decisions, and many of the decisions are either outsourced to third parties or moved under nego-tiated supply arrangements (Brooks, 1998). This means that a relatively small number of decision-makers can account for avery large volume of shipments, and research costs can be high. Response rates can be low as decision-makers are concernedabout competitive intelligence leakage. This is a challenge that Rich et al. (2009) set to solve with a different approach; ratherthan using stated or revealed preference interviews with the actors making the decisions, they focused on ‘‘decouplingagents and shipments’’ and used ton and commodity data from existing origin–destination matrices to develop a theoreticalsolution and a value of time result.

We do not find this approach useful in a situation where choice is being tested between an existing mode option and anew (but not-yet-in-service) option. Moreover, their approach is shipment-based. Within supply chains, the same stream-lining has occurred; carrier choice today is about choosing suppliers for a contracted period and allocating volume basedon negotiated supply arrangements. The single shipment is no longer a valid unit of measurement if investment decisionsin new services are to be adequately evaluated for their attractiveness and market adoption. As pointed out by Brooks(1998), very few transportation decisions in North America are transactional in nature as the industry has moved to nego-tiated supply for transportation services, and supply chains have been reconfigured based on these arrangements. Therefore,discrete choice model focussing on a single shipment is not behaviourally plausible. The Brooks and Trifts (2008) data offersthe opportunity to consider choice and willingness-to-pay within a framework of allocation decision-making.

Given the range of attributes driving the short sea mode choice decision, the willingness-to-pay estimates of interest in-clude more than the traditional value of time travel savings but also the value of flexibility in departure schedules and thevalue of other attributes associated with the mode, not necessarily identified (but perceived to benefit the decision-makersuch as reliability) in the choice set offering.

In the next section, we begin by reviewing the Brooks and Trifts (2008) methodology so as to explain the derivation of thedata being used. We then develop a choice modelling framework, discussing two approaches—a multinomial logit and ascaled mixed logit model—to explore the decision, identify the model representation with the best fit, before discussingthe ramifications of this research.

2. Material and methods

2.1. The Brooks and Trifts (2008) data set

The analysis of the data in Brooks and Trifts (2008), collected in 2006, focussed on a comparison of frequency data acrossthe three corridors considered in the study (i.e., short-, medium- and long-distance travel from Halifax, NS, Canada to threedestinations along the US east coast—Boston MA, USA; Philadelphia, PA, USA; and Wilmington, NC, USA, respectively). Thecomparisons of propensities to utilise different mixtures of freight transport alternatives across corridors revealed useful in-sights into the nature of shipper preferences. However, the survey captured richer preference information than these com-parisons imply. Specifically, respondents made a series of choices exposing a functional relationship between level-of-servicemixes and respondents’ preferred allocation of freight transport activity across competing alternatives.

The survey questionnaire was administered via internet, and consisted of 15 choice sets, each with two alternatives, grad-ually including different attributes to a base set of cost and transit time. In the first 12 choice sets, the alternatives were unla-belled (presented as Option A and Option B), whilst the final three choice sets include labels of ‘‘truck’’ and ‘‘integrated shortsea shipping’’ for the two alternatives. The 15 choice sets contained five subsets of three choice sets each, representing sce-narios in which respondents needed to allocate their preferred proportions of freight transport activity to short-, medium-and long-distance destinations across the two alternatives. The alternatives offered mixes of levels-of-service defined interms of price, travel time, reliability and frequency of departure; the first choice set for each corridor centred on a restrictedtrade-off between price and travel time, the next two choice sets for each corridor centred on a trade-off between price, tra-vel time and reliability, and the final two choice sets for each corridor centred on a trade-off between price, travel time andfrequency (with reliability at equal levels across alternatives).

Respondents indicated the percentage of their shipping requirements they would allocate to each option on an 11-pointscale, labelled so as to allocate cargo in 10% increments. Such an approach recognises that the market reality is that choicesare seldom all one or another, but that many companies engage in ‘splitting their business among options. A response of 1indicated a 100% allocation to Option A and a response of 11 indicated a 100% allocation to Option B, with the midpoint of thescale (6) labelled as a 50% split between the two options.

184 S.M. Puckett et al. / Transportation Research Part E 47 (2011) 182–189

Within a mode, shippers choose carriers based on either time-based contracted arrangements or transaction-basedarrangements. They often perceive very small differences between carriers within a mode and, as a result, split the business.Such splitting of the business is not captured in studies that focus on a single transaction and, as splitting the business hasgrown more prominent in the risk management strategies of transport buyers, the existence of splitting the business mayalso occur at the mode level. Therefore, the turning of the discrete choice decision into a broader allocation decision capturesmultiple single transaction decisions as well as contracted decisions in each choice set. That is, rather than focussing onindividual shipments, which could reveal sensitivities that would not necessarily represent the strategies that shipperswould enact at the operational level, the survey focusses on the overall business strategy of allocating share to modaloptions.

The resulting proportional choice data are not only behaviourally representative, but are also well suited to a discretechoice analysis approach of proportional, as opposed to binary, choice data. We explored the data within a range of discretechoice models to identify the richest econometric representation of the choice behaviour within the study that could be ob-tained subject to the constraints of the dataset. The specification of reliability in the questionnaire was problematic, in that itonly specified the attribute in intangible terms (i.e., better, worse or equal reliability relative to the other alternative). Thisled to a preference to analyse the subset of choice sets in which reliability could be controlled for meaningfully; combinedwith a preference to quantify preferences for frequency of departure (which only appears in choice sets where reliability isspecified as equal across the alternatives), we chose to restrict the analysis to the final six choice sets. Hence the ensuingmodels are estimated in the context of choices under trade-offs between cost, transit time and frequency of departure forthe three corridors in the study.

The data offer a considerable number of total choice observations (252); however, with 42 unique respondents, the abilityto estimate relatively complex structures involving decomposition of preference heterogeneity was restricted by the rela-tively low variation in candidate contextual effects, e.g., delivery time requirements, firm demographics, and the like. Fur-thermore, the lack of variability in attribute level mixes across choice sets within a given corridor added difficulty inidentifying complex behavioural relationships relative to candidate experimental designs of the same number of choice sets(e.g., orthogonal or optimal designs).

Fortunately, respondents demonstrated clear preferences that were not only consistent with our expectations (e.g., costsensitivity, disutility of low-frequency service), but were also strong enough to allow for the estimation of advanced discretechoice models accounting for heterogeneity across respondents with respect to both preferences and hidden scale impactswithin parameter estimates due to unobserved effects. We now turn to a discussion of the discrete choice structure em-ployed in the analysis offered in the following section.

2.2. The choice modelling framework

The generalised mixed logit (GMXL) model extends the mixed multinomial logit model (MMNL) developed in Train(2003), Hensher and Greene (2003) and Econometric Software (2007), amongst others, and the ‘‘generalised multinomial lo-git model’’ proposed in Fiebig et al. (2009) to account for scale heterogeneity, a key potential source of respondent-specificunobserved heterogeneity that additional to preference heterogeneity where the latter is associated with the observed attri-butes describing each alternative.

The modelling form, summarised below, is from Greene and Hensher (2010). The MXL that embodies both observed (Dzi)and unobserved (Cvi) heterogeneity in the preference parameters of individual I is given in the following equation:

Probðchoiceit ¼ jjxit;j; zi;viÞ ¼ expðVit;jÞXJ

j¼1

,expðVit;jÞ ð1Þ

where

Vit,j

b0ixit;j, bi b + Dzi + Cvi, xit,j the K attributes of alternative j in choice situation t faced by individual i, zi a set of M characteristics of individual i that influence the mean of the taste parameters, and vi a vector of K random variables with zero means and known (usually unit) variances and zero covariances.

Structural parameters to be estimated are the constant vector, b, the K �M matrix of parameters D and the nonzero ele-ments of the lower triangular Cholesky matrix, C. A number of interesting special cases are straightforward modifications ofthe model. Specific non-random parameters are specified by rows of zeros in C. A pure random parameters MNL model re-sults if D = 0 and C is diagonal. The basic multinomial logit model results if D = 0 and C = 0.

In addition to preference heterogeneity associated with attributes of alternatives, there is the possibility of scale heter-ogeneity across choices. The preceding is modified as Eq. (2), which defines the GMXL form.

bi ¼ ri½bþ Dzi� þ ½cþ rið1� cÞ�Cvi ð2Þ

S.M. Puckett et al. / Transportation Research Part E 47 (2011) 182–189 185

where

ri

exp½�rþ d0hi þ swi�, the individual specific standard deviation of the idiosyncratic error term, hi a set of L characteristics of individual i that may overlap with zi, d parameters in the observed heterogeneity in the scale term, wi the unobserved heterogeneity, standard normally distributed, �r a mean parameter in the variance, s the coefficient on the unobserved scale heterogeneity, c a weighting parameter that indicates how variance in residual preference heterogeneity varies with scale, with

0 6 c 6 1.

The weighting parameter, c, is central to the generalised model. It controls the relative importance of the overall scaling ofthe utility function, ri, versus the scaling of the individual preference weights contained in the diagonal elements of C. Whenri is not equal to one, c will spread the influence of the random components between overall scaling and the scaling of thepreference weights.

In addition to the useful special cases of the original mixed model, some useful special cases arise in this model. If c = 0,then a scaled mixed logit model (SMXL) emerges, given in the following equation:

bi ¼ ri½bþ Dzi þ Cvi� ð3Þ

If, further, C = 0 and D = 0, a ‘‘scaled multinomial logit model (SMNL)’’ is implied;

bi ¼ rib ð4Þ

The full model, in the unrestricted form or in any of the modifications, is estimated by maximum simulated likelihood(see Econometric Software, 2007). Greene and Hensher (2010) discuss a number of assumptions that need to be imposedto identify and estimate the model. The fully specified model is given in Eq. (5). Combining all terms, the simulated log like-lihood function for the sample of data is shown in the following equation:

log L ¼XN

i¼1

log1R

XR

r¼1

YTi

t¼1

YJit

j¼1

Pðj;Xit ;birÞdit;j

( )ð5Þ

where

bir

rir[b + Dzi] + [c + rir(1 � c)]Cvir, rir exp[�s2/2 + d0hi + swir], vir and wir the R simulated draws on vi and wi, dit,j 1 if individual i makes choice j in choice situation t and 0 otherwise,

and

Pðj;Xit ;birÞ ¼exp x0it;jbir

� �PJit

j¼1 exp x0it;jbir

� � ð6Þ

3. Results and discussion

3.1. Data description

The data setting as described above is shippers sending short-, medium- and long-distance freight along the eastern sea-board of North America who were asked to allocate proportions of freight transport activity across two alternative servicesfor a given sampled corridor. The two alternatives offered different mixes of levels-of-service representing road and short seatransport, presented in terms of combinations of price, travel time, and frequency as illustrated in Table 1. The survey ques-tionnaire included five choice sets each for three corridors, representing travel from Halifax, Nova Scotia to Boston, Philadel-phia, and Wilmington, North Carolina. The alternatives were kept generic within the first 12 choice sets, whilst thealternatives were specifically labelled in the final three choice sets to determine if a labelling of mode effect would be in evi-dence. This was particularly important, as the examination of mode choice by Bolis and Maggi (2003) set up the specificationof the transport mode as a separate variable, on the grounds that shippers may have preferences for unobserved attributes ofthe mode; Brooks and Trifts (2008) found no evidence that shippers had a bias against integrated short sea shipping, a factorin early European and US studies (Paixão and Marlow, 2002; Government Accountability Office, 2005).

Our analytical task was to estimate utility functions based upon the proportional choice data from the survey and thecorresponding attribute level mixes in each alternative. Across the models we tested, respondents were strongly sensitive

Table 1The foundations for the choice sets. Source: Brooks and Trifts (2008), Table 1.

Market Truck Short sea

Short distance (1)(Halifax, NS–Gloucester, MA) 695 road miles 380 nautical milesPrice of the service (in USD) (2) 1774 1690Total transit time in hours (3) 30 30Frequency Daily Twice a week

Medium distance (1)(Halifax, NS–Philadelphia, PA) 1000 road miles 743 nautical milesPrice of the service (in USD) (2) 2559 1739Total transit time in hours (3) 34 58Frequency Daily Every 5 days

Long distance (1)(Halifax, NS–Wilmington, NC) 1529 road miles 1007 nautical milesPrice of the service (in USD) (2) 3899 1644Total transit time in hours (3) 56 72Frequency Daily Once a week

Notes: (1) The distances chosen are specific to the three largest volume marketsand so are not chosen to optimise the variation between the three profiles.Because of the shape of the landmass in this region, the sea (nautical) miles aresubstantially less than the land (road) miles.(2) The truck price is based on a mileage rate plus the current fuel surcharge of27–30%. The short sea price is grossed up from cost, including port costs andindustry average margin.(3) Both should be seen as door-to-door. The trucking times to market are basedon distance, normal speed and regulated service conditions for the route. Thetotal transit time for short sea includes the local truck haul, which is consideredto be within 50 miles of the port each end. It is based on a traditional short seavessel as opposed to a high-speed option as Brooks et al. (2006) concluded ahigh-speed option could not be offered at a market-acceptable price.(4) An important outlier in the above data is that the price of the short seaservice to Wilmington, the longest corridor, is lower than the price of short seaservice along the shorter corridors. The numbers are accurate primarily becauseWilmington’s port costs are substantially below other ports on the coast,making the total door-to-door cost less all in. The prices are built up using astandard proportional profit margin and actual costs that would be incurred forall activities in executing the move.

186 S.M. Puckett et al. / Transportation Research Part E 47 (2011) 182–189

to cost and frequency of departure. Respondents did not reveal significant sensitivities to travel time trade-offs with respectto the other attributes under consideration. Rather, cost and frequency of departure appear to have driven allocation pref-erences across the two alternatives for each of the three corridors within the study; that is, whilst transit time may be asource of disutility, in general, the variations in transit time across the two alternatives did not appear significant enoughto impact allocation preferences. These findings will be considered in greater detail in the discussion of the model results,to which we now turn.

3.2. Model results

The final discrete choice models are of shippers’ preferences across the three distance-varied corridors. Each choice setwithin the analysis offers mixes of level-of-service attributes, and prompts the respondent to allocate a preferred proportionof freight transport activity across two alternatives that are either unlabelled or mode-specific (i.e., ‘‘truck’’ and ‘‘integratedshort sea shipping’’). The set of attributes includes price (in US dollars), travel time (in hours), and frequency of service (indepartures per week). The attribute levels in the choice sets are fixed for each corridor, yielding specific combinations of cost,frequency and transit time for each corridor. We tested for corridor-specific effects and for distinct sensitivities across cor-ridors, but a linear specification (i.e., generic to corridor type) provided the best explanatory power.

We controlled for the presence of correlations in preferences and unobserved effects due to the panel nature of the choiceset data. That is, the choice observations in the dataset are not independent, but rather represent repeated choices by thesame individuals across multiple choice tasks. The SMXL model presented here is estimated with respect to the panel natureof the data, estimating distributions of preferences at the individual level rather than at the choice set (i.e., fully indepen-dent) level.

Table 2 presents the model outputs for our preferred model of shipper preferences, in which shippers’ utility of freighttransport alternatives is a function of price (in hundreds of US dollars), departures per week, and alternative-specific con-stants representing general preferences for the unlabelled and labelled alternatives with attributes for truck transport.The choice variable is the proportion of total freight transport activity allocated across the two alternatives in a given choiceset. Candidate models are presented in order of complexity, beginning with the multinomial logit (MNL), followed by the

Table 2Model results.

Attributes MNL SMXLa

Parameter t-ratio Parameter t-ratio

Random parameter in SMXLDepartures per week – Mean 0.1646 0.29 0.3968 2.22

Fixed parameters in SMXLUnlabelled truck alternative �0.5334 �0.19 �1.5204 �1.62Service by truck (Labelled) �0.9148 �0.33 �2.3368 �1.70Price (US$ in hundreds) �0.0241 �0.68 �0.0410 �1.97

Parameter for scale effects in SMXLVariance parameter s – – 0.4486 1.98

Model fitLog likelihood at zero �349.35Log likelihood �172.73 �167.40Pseudo R-squared 0.505 0.521Akaike information criterion 1.403 1.376Number of observations 252 252

a Random parameter distributed triangular with a spread equal to the mean. 500 Halton draws were used in the estimation of the parameter distributionsfor the SMXL model.

Table 3Willingness-to-pay for gains in frequency of departure (all values in US dollars).

MNL SMXL

Mean Mean Std. deviation

Value of frequency gains ($ per departure per week) $682.99ns $1113.06 $320.72

S.M. Puckett et al. / Transportation Research Part E 47 (2011) 182–189 187

scaled mixed logit (SMXL) model. The model fit estimates improve (with the 2 degrees of freedom difference) when account-ing for preference and scale heterogeneity; furthermore the MNL model has no statistically significant effects, highlightingthe importance of accounting for preference and scale heterogeneity in the behaviourally more plausible SMXL model. Wealso estimated a full GMXL model; however c, the weighting parameter that indicates how variance in residual preferenceheterogeneity varies with scale (see Eq. (2)) was not statistically significantly different from zero (t-ratio of 0.42), and hencethe preferred model is of the SMXL form.

The effects of preference and scale heterogeneity are both significant in the models tested. After accounting for hetero-geneity in sensitivities to frequency of departure, the SMXL model identifies a significant sensitivity that the MNL modeldoes not. Furthermore, the SMXL model identifies a statistically significant scale effect. The general implications of the SMXLmodels are intuitive, with respondents demonstrating preferences for price and frequency that are of the expected sign. Mostsurprisingly, a marginally significant labelling effect was found, with respondents demonstrating a preference for short seashipping (i.e., a negative utility associated with alternatives labelled as ‘‘truck’’) when they were presented with the modalnature of the alternatives in the final three choice sets; this effect was more pronounced than the disutility associated withthe unlabelled alternatives representing truck transport.

The estimates of shippers’ sensitivities to freight transport attributes in the SMXL model are further enlightening whenconsidered directly. The SMXL model confirms that shippers are cost sensitive, and prefer high departure frequencies. How-ever, one major advantage of discrete choice modelling is the capability to identify estimates of willingness-to-pay relatingto attributes within the analysis. Table 3 summarises the estimates WTP for the SMXL model, together with the MNL model,although we note that the MNL model estimate is not statistically significant.

The estimates of WTP are found by taking the ratio of the estimated marginal utility parameter for the attribute of interest(measured in utils per focal unit) to the estimated marginal utility parameter for price (measured in utils per monetary unit);this yields a value in focal units per monetary unit. These estimates are derived conditional on the choices made by respon-dents, leading to the estimation of distributions of respondent-specific WTP measures for each attribute of interest. Takingthese ratios within the models shown in Table 2, we can derive WTP measures for improvements in frequency of departure.Whereas the MNL model offers only mean parameter estimates, the SMXL model offers parameter estimates for each respon-dent in the sample. Hence, we are able to derive a distribution of WTP measures across the respondents in the sample withinthe SMXL model. The distribution is graphed in Fig. 1.

3.3. Discussion

Shippers demonstrated a strong WTP for higher frequencies of departure, with a mean value of over $1100 per additionaldeparture per week corresponding to a given service; this represents a considerable premium in level-of-service for current

0200400600800100012001400160018002000

0 5 10 15 20 25 30 35 40 45

Fig. 1. WTP distribution ($ per departure per week, in ascending order across respondents).

188 S.M. Puckett et al. / Transportation Research Part E 47 (2011) 182–189

or potential high-frequency transport services. Shippers in the sample demonstrated that services with a high frequency ofdeparture offer important value, whether in monetary terms or in terms of other level-of-service attributes (i.e., improve-ments in frequency of departure could offset slow travel times that cause a significant degree of disutility). The spread ofshippers’ values associated with departures per week shows that there may be a significant proportion of shippers whoare willing to pay a particularly strong premium for high-frequency service, such as would be desired in a just-in-time sup-ply chain situation. Indeed, the distribution indicates the presence of a key (upper-tail) segment of the sample that has adistinct preference for frequency of service above the majority of the sample. However, it is important to acknowledgethe limitations of the data with respect to sample size when drawing conclusions pertaining to the population at large. Thatis, the level of confidence in the specific magnitude of willingness-to-pay should be tempered, given the sample size ana-lysed here. Still, we are confident that the analysis confirms the presence of both a significant value of frequency gainsfor shippers, and heterogeneity in such preferences across shippers.

Ultimately, the distribution of WTP for gains in the frequency of departure reveals opportunities for freight transport ser-vice providers to be aware of the existence of real trade-offs being made across market alternatives (i.e., that frequency ofservice does matter), and that these trade-offs are evaluated in considerably different ways by different shippers (i.e., someshippers are particularly sensitive to certain attributes). For example, a new short sea shipping service could exploit thisknowledge by targeting its service to compete strongly on attributes over which it has an advantage, offering sufficient valuefor those attributes to overcome any systematic disadvantages short sea shipping may have relative to road, like departurefrequency. This information is not only relevant at the general level (e.g., offering sufficient price savings to offset travel timeand frequency shortcomings), but also for subgroups of shippers that are particularly sensitive (or insensitive) with respectto frequency of departure.

4. Conclusions

Brooks and Trifts (2008, p. 156) concluded that splitting business does indeed occur and speculated that:A decision by a shipper to split their business between services could reflect diverse service needs across the supply chain,

different needs of end-market buyers, strategic or transactional considerations, or a desire to mitigate route or carrier risk.(Route risk mitigation has become more prevalent since port capacity issues have become more apparent while carrier risk,and in particular the risk of carrier bankruptcy, has long been a factor.)

Therefore, it is worthwhile considering, as part of the examination of the allocation problem, the willingness-to-pay thatis attached to the various attributes within the mode choice decision and what this means from a managerial perspective of amodal operator in the marketplace. This paper illustrates how that may be done.

These findings support strategic management thinking that companies may compete in many ways and that there isroom in any transport market for competition based on more than just price. As noted by Porter (1980), there can be onlyone cost leader but there may be many who choose to compete with a focus or differentiation strategy. Our findings showthat there is the willingness-to-pay a premium for departure frequency. As for the scale of the willingness-to-pay, it shouldbe noted that this research was undertaken in a period of buoyant economic growth prior to the financial crisis of 2008. Suchpremiums may have been reduced in the period after the study as cargo interests re-examined their supply chain routingsand modal choices to capture greater value from the transport choices.

While earlier European and US studies showed a bias against short sea shipping, a labelling effect was found (albeitweakly significant) indicating a preference for integrated short sea shipping over options labelled ‘‘truck’’. This bias in theopposite direction as seen in other studies was explained by Brooks and Trifts as reflecting the market reality; anecdotal

S.M. Puckett et al. / Transportation Research Part E 47 (2011) 182–189 189

evidence from those answering the survey indicated respondents were frustrated with the level of road congestion on thecorridor, and particularly in the Hudson River Lower Manhattan vicinity. In addition to confirming the Brooks and Trifts find-ings, we can only say that these results may not be generalisable to other corridors or routes, and that situation-specific cir-cumstances will alter preferences for a particular mode from what might otherwise be expected.

A new attribute of interest for future research will be that of willingness-to-pay for emissions reductions. Vanherle andDelhaye (2010) have begun the process of documenting emissions and social costs for road versus short sea options; futureresearch should incorporate the added variable of emissions so that willingness-to-pay for GHG mitigation is wellunderstood.

This research contributes to the literature by offering an empirical advancement in the analysis of the competitiveness ofintegrated short sea shipping through the use of advanced discrete choice analysis. A discrete choice framework offers apowerful means of analysing trade-offs made by shippers in their choices of preferred freight transport strategies, enablingresearchers to quantify behavioural measures of strategic and policy interest, including willingness-to-pay. Furthermore, thegeneralised mixed logit model presented in this paper is capable of identifying distributions of preferences across a sample.These distributions represent powerful information about subsets of shippers who may benefit from, and show support for,freight transport alternatives that offer novel mixes of level-of-service attributes; transport providers could utilise this infor-mation to offer strategic alternatives that distinguish between cost- and service-quality-sensitive carriers.

Furthermore, the SMXL model presented in this paper advances the literature by offering evidence of the power ofaccounting for heterogeneity in scale across respondents. This is of importance not simply within freight transport studies,but in discrete choice studies, in general. Without accounting for scale effects, we were unable to identify distributions ofpreferences without sacrificing model fit and precision in parameter estimates. However, the SMXL model identified a highlysignificant scale effect, yielding a model that both is behaviourally consistent with the simpler MNL model, and identifiesheterogeneity in preferences for frequency of departure. These results support the merit of re-examining extant data withinGMXL models to test the robustness of implications from models that do not account for scale heterogeneity. Likewise, theseresults confirm the value of accounting for scale heterogeneity in future choice studies.

Acknowledgements

Mary R. Brooks thanks the Faculty of Economics and Business for the opportunity to do research at the University of Syd-ney as a Visiting Scholar in 2010. The authors would like to thank the Halifax Port Authority for the use of their shipper data-base in identifying exporters and importers in the target geographic area. The detailed comments of the referees are alsoappreciated and have aided us in the revision.

References

Balcombe, K., Fraser, I., Harris, L., 2009. Consumer willingness to pay for in-flight service and comfort levels: a choice experiment. Journal of Air TransportManagement 15 (5), 221–226.

Bolis, S., Maggi, R., 2003. Logistics strategy and transport service choices: an adaptive stated preference experiment. Growth and Change 34 (4), 490–504.Brooks, M.R., 1998. Performance evaluation in the North American transport industry: user’s views. Transport Reviews 18 (1), 1–16.Brooks, M. R., Hodgson, J.R.F., Frost, J.D., 2006. Short sea shipping on the east coast of North America: an analysis of opportunities and issues, Halifax:

Dalhousie University. Available from: <http://management.dal.ca/Research/ShortSea.php>.Brooks, M.R., Trifts, V., 2008. Short sea shipping in North America: understanding the requirements of Atlantic Canadian shippers. Maritime Policy

Management 35 (2), 145–158.Commonwealth of Australia, 2006. Demand Projections for AusLink Non-Urban Corridors: Methodology and Projections, Working Paper 66, Bureau of

Transport and Regional Economics, Canberra. <http://www.bitre.gov.au/publications/72/Files/AusLinkDP_FullReport.pdf> (accessed 15.02.10).Econometric Software, 2007. Nlogit 4, Econometric Software, New York and Sydney.Fiebig, D., Keane, M., Louviere, J., Wasi, N., 2009. The generalized multinomial logit: accounting for scale and coefficient heterogeneity. Marketing Science.

doi:10.1287/mksc.1090.0508 (published online before print July 23).García-Menéndez, L., Martinez-Zarzoso, I., Pinero De Miguel, D., 2004. Determinants of mode choice between road and shipping for freight transport:

evidence for four Spanish exporting sectors. Journal of Transport Economics and Policy 38 (3), 447–466.Government Accountability Office, 2005. Short Sea Shipping Option Shows Importance of Systematic Approach to Public Investment Decisions, 05-768, July,

United States Government Accountability Office, Washington.Greene, W.H., Hensher, D.A., 2010. Does scale heterogeneity matter? A comparative assessment of logit models. Transportation 37 (3), 413–428.Hensher, D.A., 1997. A practical approach to identifying the market for high speed rail in the Sydney-Canberra corridor. Transportation Research Part A 31

(6), 431–446.Hensher, D.A., Greene, W.H., 2003. Mixed logit models: state of practice. Transportation 30 (2), 133–176.Hess, S., Adler, T., Polak, J.W., 2007. Modelling airport and airline choice behaviour with the use of stated preference survey data. Transportation Research

Part E: Logistics and Transportation Review 43 (3), 221–233.Jiang, F., Johnson, P., Calzada, C., 1999. Freight demand characteristics and mode choice: an analysis of the results of modeling with disaggregate revealed

preference data. Journal of Transportation and Statistics 2 (2), 149–158.Paixão, A.C., Marlow, P.B., 2002. The strengths and weaknesses of short sea shipping. Marine Policy 26, 167–178.Porter, Michael E., 1980. Competitive Strategy. Free Press, New York.Rich, J., Holmblad, P.M., Hansen, C.O., 2009. A weighted logit freight mode-choice model. Transportation Research Part E: Logistics and Transportation

Review 45 (6), 1006–1019.Train, K., 2003. Discrete Choice Methods with Simulation. Cambridge University Press, Cambridge.Vanherle, K., Delhaye, E., 2010. In: Road versus Short Sea Shipping: Comparing Emissions and External Costs, Proceedings, July, International Association of

Maritime Economists, Lisbon.