Transportation Research Part E 46 (2010) 764–774

Contents lists available at ScienceDirect

Transportation Research Part E

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

How different is carrier choice for third party logistics companies?

Zachary Patterson a,*, Gordon O. Ewing b, Murtaza Haider c

a Urban and Transportation Data Division, Agence métropolitaine de transport (AMT), 500, Place d’Armes, 25th Floor, Montreal, Quebec, Canada H2Y 2W2b Department of Geography, McGill University, 805 Sherbrooke St., West Montreal, Quebec, Canada H3A 2K6c Faculty of Business, Ryerson University, 350 Victoria St., Toronto, Ontario, Canada M5B 2K3

a r t i c l e i n f o

Article history:Received 17 December 2007Received in revised form 25 September 2009Accepted 3 December 2009

Keywords:Mode choice modelingThird party logistics companiesFreight modeling

1366-5545/$ - see front matter � 2010 Elsevier Ltddoi:10.1016/j.tre.2010.01.005

* Corresponding author. Tel.: +1 514 287 2464x4E-mail addresses: zpatterson@amt.qc.ca (Z. Patte

a b s t r a c t

The purpose of this paper is to test whether third party logistics companies (3PLs) are dif-ferent from other end-shippers with respect to how they choose their carriers. The resultsof carrier choice models developed in this paper suggest that 3PLs are more biased againstintermodal shipping than other end-shippers. The principal conclusions are as follows:mode and carrier choice modeling needs to take into consideration differences between3PLs and other end-shippers; and with the increasing role of 3PLs in choosing carriers, theirstronger bias against intermodal shipping will present further challenges to increasingfreight rail mode share.

� 2010 Elsevier Ltd. All rights reserved.

1. Introduction

As companies have increasingly sought to outsource activities, there has been a dramatic rise in the use of external logis-tics companies (often referred to as third party logistics companies or 3PLs) to organize freight transportation. Little isknown about whether 3PLs’ choices of carriers differ from that of traditional end-shippers. This sector is expected to exertincreasing influence on the way freight is shipped. Understanding any differences that 3PLs manifest in carrier choice is use-ful in itself, but also in evaluating the potential for rail to increase its share of freight.

Realistic models are required to evaluate government policies that encourage increasing rail freight mode share. Variousmethodologies have been used to approach the question of freight mode choice. This paper uses data from a 2005 shippercarrier choice stated preference survey to test for and quantify differences in carrier choice preferences between 3PLs andother ‘end-shippers’. The survey was designed explicitly to evaluate shipper preferences for the carriage of intercity consign-ments, and particularly their preferences for carriers that contract the services of rail companies to carry these shipments. Animportant aspect of the survey design was to include 3PLs as a distinct subgroup of shippers in order to be able to test for anydifferences that might exist between them and other shippers. The survey data included a small but representative sample of3PLs from the population.

The paper begins with a literature review of research on third party logistics providers. It then describes the survey onwhich this research is based, beginning with background on previous freight choice studies and a short description of thedataset used in this study. The paper continues by describing how differences between 3PLs and other end-shippers weretested for, and presents the results of models that were estimated for 3PLs and other end-shippers separately. The paper con-cludes with a discussion of what these imply for rail’s potential to increase its mode share, and for freight mode choice mod-eling more generally.

. All rights reserved.

480; fax: +1 514 287 2460.rson), gordon.ewing@mcgill.ca (G.O. Ewing), murtaza.haider@ryerson.ca (M. Haider).

Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774 765

2. Literature review – 3PLs

Third party logistics companies are businesses that provide a variety of logistics-related services. Services offered by 3PLscan include public warehousing, contract warehousing, transportation management, distribution management, freight con-solidation, and increasingly the management of entire supply chains. The use of 3PLs has been increasing quickly since atleast the 1980s as companies have attempted to outsource non-core activities, including transportation logistics. As a resultof this growth, there has been a great deal of interest both in the academic literature and the business press.

The latter has focused on two broad themes related to 3PLs. The more general theme encompasses the growth of theindustry (Hoffman, 2006; Quinn, 2006), the proportion of companies using 3PLs and for what purposes they are used (Ame-kudzi and Meyer, 2005; eyefortransport, 2005). It also covers how the industry has performed financially (Hoffman, 2007;Shister, 2006). A third general theme is where the industry seems to be going in terms of both future geographical markets(e.g. China) (Biederman, 2007; Dibenedetto and Diben, 2007) and the future role that 3PLs are expected to play (Aimi, 2007;Dibenedetto, 2007).

The second theme has focused more on 3PLs from a customer perspective. This has included articles aimed at helpingcompanies establish their need for 3PLs (Oliver Silver, 2005), what type of 3PLs to choose (Tompkins, 2006), and how to eval-uate particular 3PLs (Logistics Today, 2006) and being careful which they choose (Hannon, 2007; Hoffman, 2005).

The academic literature has had some overlap with the above issues. For example, there has been survey research onthe proportion of companies using 3PLs, for what types of services and for what reasons. This research has also looked attrends in the types of services used and those expected to be used in the future. These surveys have generally focused onNorth America and Europe (Lieb and Bentz, 2004; Peters et al., 1998) and more recently on fast growing, developing mar-kets (Mitra, 2006; Sohail et al., 2006)). Another area of overlap with the business press has been on methods of establish-ing the need for, and selecting 3PLs, with the academic literature being more analytical. Some have proposed morequalitative approaches (e.g. Foggin et al., 2004) and others more quantitative (e.g. Bottani and Rizzi, 2006). There has alsobeen some meta-analysis of previous 3PL research and on areas for future research (Maloni and Carter, 2005; Murphy andPoist, 2000).

Apart from these areas of overlap, there are several other themes in the academic literature. The first concerns the use oftechnology in 3PL operations (Evangelista and Sweeney, 2006; Lai et al., 2007). There has been related work in operationsresearch applied to 3PLs (e.g. Ko and Evans, 2007; Kumar et al., 2006).

Another academic stream of research involves analyzing the factors contributing to the success of particular 3PLs. Thisincludes analysis of well-managed 3PLs (Gunasekaran and Ngai, 2003), how 3PLs manage their relationships with clients(Knemeyer and Murphy, 2005; Sinkovics and Roath, 2004), customer perceptions of their 3PLs (Wilding and Juriado,2004), and evaluations of the efficiency and financial performance of 3PLs (Min and Joo, 2006; Yeung et al., 2006). Relatedresearch has also looked at the degree to which the use of 3PLs can improve a client’s performance (Stank et al., 2003), andhow 3PLs can market themselves (Wang et al., 2006).

Whatever the focus of the particular articles, the broad conclusions are clear: the 3PL industry continues to grow and hasa large potential. Since 1996 the US 3PL industry has grown annually by an average of 14.4% (Quinn, 2006), with a growth of17.7% in 2006 resulting in total revenues of US$110.6 billion (Hoffman, 2007). Globally, the market is estimated at US$390billion. This growth is expected to continue as more firms outsource logistics functions. One recent survey found that 69% ofcompanies already outsource some of their logistics functions, with the largest proportion (46%) outsourcing transportationlogistics functions (eyefortransport, 2005). Although larger companies have been at the forefront of logistics outsourcing, it isbecoming more common for smaller companies. 3PL use by the top 100 Fortune 500 companies grew from 73% to 89% be-tween 2001 and 2005. During the same period the 100 smallest increased their 3PL use rate from 24% to 51% (Quinn, 2006).

While research on 3PLs has grown quickly, it has not included carrier and mode choice. In particular, little research haslooked explicitly at choice differences between 3PLs and end-shippers.

Although, there have been numerous freight mode stated preference (SP) studies reported in the literature (Fowkes andTweddle, 1988; Fridstrom and Madslien, 2001; Norojono and Young, 2003; Shinghal and Fowkes, 2002; Vellay and de Jong,2003; Wigan et al., 2000), a survey of this literature reveals a silence on the issue of 3PLs’ choices of mode or carrier. Onerecent exception is a paper by Patterson et al. (2007). Within the revealed preference (RP) literature (e.g. Calzada and Jiang,1997; Jiang et al., 1999; Vellay and de Jong, 2003; Young et al., 1983), there is also no explicit consideration of 3PLs relative toend-shippers.

To conclude, it is clear that since 3PLs are being used much more, they are becoming more important as freight transpor-tation mode choosers and hence as a source of demand. If 3PLs do behave differently from other shippers with respect tocarrier and mode choice, this would have important implications for understanding future freight demand. This paper seeksto investigate the degree to which 3PLs differ with respect to their carrier and mode choice preferences.

3. The data set

The data set used in this analysis comes from a stated preference survey of ‘end-shippers’ in the busiest trade and trans-portation corridor in Canada – the Quebec City – Windsor Corridor. This data set has been described elsewhere (Pattersonet al., 2007) and so its description is kept short here.

766 Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774

The purpose of the survey was to establish the factors that affect end-shippers’ carrier choice and quantify their impor-tance. Of particular interest was whether shippers’ choices were affected by knowing that a carrier would send a shipmentintermodally. ‘End-shipper’ is used to describe a shipper that hires carriers for all of their shipments.

The decision to use intermodal services will generally be the carrier’s, since it organizes the movements of consignmentsfrom end-shipper to receiver. Although one might expect end-shippers to be indifferent to how their shipments are carriedas long as they arrive in good condition and on-time, carrier decisions about whether or not to use intermodal services willultimately be constrained by shipper preferences. In effect, the end-shipper can be seen as the true backstop for the demandfor intermodal services. As a result, while many previous mode choice studies (e.g. Vellay and de Jong, 2003) have surveyedboth end-shippers and own-account shippers, this survey focused exclusively on end-shippers.

The survey took the form of a ‘contextual stated preference’ or CSP survey. It consisted of 18 questions for each respon-dent. Each question asked the respondent to choose between three alternative carriers in the context of a particular ship-ment, whose details were described. These consisted of the origin and destination, when the shipment was to arrive,whether it was ‘by-appointment,’ whether it was of high or low value, whether it was fragile/perishable or not, and its size(truckload or less than truckload (LTL)). Information on value and fragility was provided implicitly through the type of com-modity being shipped. For example, televisions were the shipment used to represent high-value, fragile goods. For an exam-ple of one of the survey questions, see Fig. 1.

With respect to carrier attributes, five were provided: cost, on-time reliability, damage risk, security risk and whether thecarrier would send the shipment by rail for a portion of the journey. Whereas previous mode choice studies (Norojono andYoung, 2003; Shinghal and Fowkes, 2002; Vellay and de Jong, 2003), have included mode explicitly by asking respondents tochoose between alternative modal configurations for their shipments, in this survey, mode was considered a carrier attri-bute. Unlike many SP freight surveys, time required for shipping was considered a shipment’s attribute, not a carrier’s. Thisis because discussions with shippers established that shipping times in the Corridor are standardized, e.g. a Montreal to Tor-onto shipment is ‘overnight.’ As a result, shipping time is not a basis upon which carriers are chosen as all carriers can offerthe same overnight shipping.

The survey population included all Corridor end-shippers which were: manufacturing facilities with more than 50employees; wholesalers and retailers that were either head offices or single locations with more than 50 employees at thatlocation; and all third party logistics companies. The term third party logistics companies (3PLs) is used collectively for com-panies that are hired to organize shipments on behalf of other companies. The SIC code 4731 (Arrangement of Freight Trans-portation and Cargo) was used to identify these companies. What was important for this research was to distinguishbetween end-shippers who choose carriers for themselves and those who choose them on behalf of client companies. Thefirm’s shipping manager was the target respondent.

The list of companies used for the survey was Dun & Bradstreet’s Million Dollar Database (MDDI) of all companies in On-tario and Quebec with more than $1 million in sales or more than 20 employees. In total, 7229 companies belonged to thispopulation. The list was provided to a telephone survey company to contact and recruit respondents for the web-based

Fig. 1. A sample question from the survey.

Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774 767

survey. The survey was administered between mid-August and early December 2005. All companies in the list were con-tacted. Of these, 680 agreed to participate. Ultimately, completed results were obtained from 392 respondents, of which25 were 3PLs.

The respondents represented a relatively large spectrum of establishment sizes with the smallest being a 3PL of only afew employees and the largest being an electronics wholesaler with 1400 employees. Moreover, the average number ofemployees in the respondent population was very close to that for the population as a whole at 132 (standard deviation141) instead of 146 (standard deviation 357) for the entire population.

Respondents came from all of the industries in the initial survey in the approximate proportion of the original company,although ‘‘Manufacturers” were slightly overrepresented at the expense of both ‘‘Wholesalers and Retailers” and 3PLs. 3PLsmade up 11% of survey population, but 6% of respondents. That said the 6% of respondents are representative of the 3PL pop-ulation. In the population there were 781 3PLs. The largest had 600 employees and the average size was 26 employees with astandard deviation of 50. There were 25 3PLs among the respondents. The largest had 400 employees with an average size of27 and standard deviation of 78. For more detailed information about the survey or the data used, see Patterson et al. (2007).As will be seen in Sections 4 and 5, this sample of 3PLs is sufficient to provide convincing statistical evidence for differencesbetween 3PLs and other end-shippers.

3.1. A random-effects mixed-logit

This subsection draws mostly on Train (2003). The workhorse for the vast majority of discrete choice modeling is the con-ditional multinomial logit (MNL).

According to the random utility framework, the decision-maker (indexed n) will choose the alternative (indexed j) yield-ing the highest utility (Unj in Eq. (1)). Vnj represents the deterministic portion of utility, and is a linear combination of alter-native and decision-maker characteristics, i.e. xs (Eq. (2)).

Unj ¼ Vnj þ enj8j ð1ÞUnj ¼ b0xnj þ enj8j ð2Þ

It also includes random aspects of utility that cannot be observed. These ‘errors’ are denoted as enj. The assumption aboutthe distribution of the error term determines the model that results. In the case of the MNL, the errors are assumed to beindependently and identically distributed (iid).

This assumption allows for the derivation of the well-known, closed-form expression of the MNL (Eq. (3)). Pni is the prob-ability that individual n chooses alternative i.

Pni ¼eb0xniPJj¼1eb0xni

ð3Þ

The formulation implies that preferences are constant across individuals (bs are fixed across individuals) and that errorsare not correlated across observations. These consequences are not only limiting from a behavioral perspective, but oftenlikely not to be realistic. The use of a mixed-logit model can obviate these limitations by allowing for random taste variation,as well as for correlation across observations. These characteristics of the mixed-logit make it particularly attractive in thecontext of panel data, i.e. data containing several responses from each individual.

To show how the mixed-logit can overcome these limitations, the first step is to rewrite Eq. (2) to further decompose util-ity to include another random term as well as multiple observations for the same person.

Unjt ¼ a0xnjt þ b0nznjt þ �njt 8j ð4Þ

Here as represent fixed (across individuals) coefficients for variables xnj and bns are random coefficients with zero means forvariables znjt. The utilities are also indexed across t-this represents the multiple observations for each individual. The bns arefixed for an individual across his choices.

When xnjt and znjt overlap (i.e. some variables enter both terms) the coefficients of these variables are considered to varyrandomly with mean a and to have the same distribution as bn around that mean, with the remaining error term enjt beingiid. In the case of a random-effects mixed-logit (used here), an alternative-specific constant is estimated for each alternative.It is possible to estimate a choice model including this extra flexibility by using the logit model while simulating the distri-bution of the individual-specific random part of utility (bn). In the context of panel data, where preferences are allowed tovary across individuals, but not across the choices of the same individual, we get the following choice probabilities:

LniðbÞ ¼ PTt¼1

ea0xnitþb0nznitPJt¼1ea0xnjtþb0nznjt

" #ð5Þ

Since the enj is iid, Eq. (5) represents the probability of an individual making a sequence of choices i = i1, . . . , iT conditionalon b. The unconditional probability is the integral of this product over all values of b:

Pni ¼Z

Lnif ðbÞ@b ð6Þ

768 Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774

The mixed-logit takes its name from the fact that the first part of the function (Lni) is the standard logit, and f (b) is themixing function. BIOGEME (Bierlaire, 2007) was used to estimate these models. BIOGEME is a statistical package designedspecifically for discrete choice model estimation. It is capable of estimating many different types of discrete choice models,including models requiring the use of simulation techniques for solving integrals for which closed-form solutions do notexist.

4. Modeling approach

The modeling approach adopted was as follows. First, for all of the data (all observations), a global mixed-logit was esti-mated. This model was arrived at by beginning with a more general form of the model and removing insignificant variablesiteratively. In other words, the more specific global model was developed by ‘‘testing down” from a more general form of themodel to the more specific model presented below.

In the second stage, testing was conducted to see if 3PLs had statistically different utility functions from other end-ship-pers, and hence whether they ought to be modeled separately. This was accomplished using a version of the so-called Chowtest (see for example Greene, 2000). It is an econometric test of whether the coefficients in two linear regressions on differentdata are equal. It is most commonly used in time-series analysis to test for the presence of a structural break. Model coef-ficients are allowed to vary independently for a subset of the data by ‘interacting’ the explanatory variables with a dummyvariable identifying the subset of interest. A test for the joint-insignificance of all of these interacted variables is then under-taken. This amounts to testing whether the subgroup observations are statistically significantly different from the others.More precisely, it tests whether, by allowing each of the explanatory variables to be estimated separately for the subgroup,there is a statistically significant increase in the explanatory power of the model.

In this case, the test was between 3PLs and other end-shippers. As such, a test was performed for the joint-insignificanceof variables interacted with a dummy variable indicating whether a respondent was from a 3PL. If the test turned out to bestatistically significant (i.e. the null hypothesis was rejected), it would amount to saying that there is a statistically signifi-cant increase in explanatory power by estimating a model for 3PLs independently from other end-shippers, and thereforethat separate models ought to be estimated.

In a least squares environment, an F-test is used to test for the joint-insignificance of the subset-specific coefficients. Asthis is a discrete choice analysis involving maximum likelihood estimation, a likelihood-ratio test is appropriate. The likeli-hood-ratio test is performed as follows (see Train, 2003).

First, the test implies the existence of a null hypothesis (H). The null hypothesis in this case is that the 3PL-specific coef-ficients all equal zero. The likelihood-ratio test statistic is:

�2 LL b̂H� �

� LL b̂� �� �

ð7Þ

LLðb̂Þ is the value of the constrained (assuming 3PL-specific coefficients are zero) log of the likelihood function. LLðb̂Þ is theunconstrained value. This statistic is distributed chi-squared with the degrees of freedom equal to the number of restrictionsimplied by the null hypothesis. In this context it is equal to the number of variables in the constrained model.

When testing to see whether 3PLs ought to be modeled separately from end-shippers, a chi-square statistic of 46.51 with20 degrees of freedom resulted. This suggests a very small probability (0.0007%) of differences in the estimated coefficientsbetween 3PLs and other end-shippers arising by chance. As such, it was concluded that 3PLs should be modeled separately.The results of two models (one for all shippers, 3PLs only and other end-shippers only) are presented below.

5. Modeling results

Table 1 presents the two models. The next section describes the ‘other end-shippers’ model while the following sectiondescribes the 3PLs model and how it compares with the ‘other end-shippers’ model.

5.1. The ‘other end-shippers’ model

Altogether, three different types of variables were used in the models. First, carrier attributes included cost, on-time reli-ability, damage and security risk, and the variable indicating the shipment as intermodal. Second, shipper attributes such astheir size and geographic location were also interacted with carrier attributes. Finally, interactions of shipment and shipperattributes with carrier attributes were included, e.g. interactions between cost and shipment type, such as by-appointment,high-value or perishable goods. Shipment distance was also interacted with carrier attributes.

Shipment attributes were included in the survey to test for and capture interactions between carrier characteristics andshipments that are expected from logistics theory – in particular, the concept of total logistics costs (see Ballou, 2004). Accord-ing to logistics theory, firms trade-off between different costs with different (sometimes opposing) cost structures. Ideally,firms make decisions that minimize their total logistics costs. A typical example is that of the trade-off that firms make be-tween transport costs and costs associated with lost sales. In assuring that its customers obtain the goods they are supposedto when they are supposed to, they improve the likelihood that they attract new customers and that current customers willrepeat orders. If customers do not receive their orders when they need them, or in the state they expect them, they may stop

Table 1Model results: other end-shippers and 3PLs.

Coefficient Other end-shippers 3PLs

Value Standard error 95% Conf. int. Value Standard error 95% Conf. int.

Cost(ln) �4.000 0.576 �2.87 ? �5.13 �7.830 1.530 �4.82 ? �10.84Cost(ln) � By-appt. 1.680 0.374 2.41 ? 0.95Cost(ln) � Fragile 4.080 1.850 7.72 ? 0.44Cost(ln) � High-value 1.480 0.374 2.21 ? 0.75 4.660 1.550 7.71 ? 1.61Cost(ln) � Perishable 4.030 1.890 7.74 ? 0.32Cost(ln) � Dist. �2.360 0.499 �1.38 ? �3.34On-time reliability 0.090 0.010 0.11 ? 0.07 0.113 0.017 0.15 ? 0.08On-time reliability � Perishable 0.053 0.007 0.07 ? 0.04On-time reliability � By-appt. 0.048 0.006 0.06 ? 0.04 0.087 0.026 0.14 ? 0.04On-time reliability � High-Value 0.012 0.006 0.02 ? 0.00On-time reliability � Dist. �0.023 0.008 �0.01 ? �0.04On-time reliability � Emp. 0.076 0.023 0.12 ? 0.03Damage risk �0.394 0.025 �0.35 ? �0.44Damage risk � Fragile �0.210 0.044 �0.12 ? �0.30 �0.545 0.150 �0.25 ? �0.84Security risk �0.100 0.037 �0.03 ? �0.17Intermodal �0.632 0.042 �0.55 ? �0.71 �1.420 0.156 �1.11 ? �1.73Intermodal � btw Railheads �0.205 0.074 �0.06 ? �0.35Intermodal � Ontario shipper �0.267 0.080 �0.11 ? �0.42ASC1 0.502 0.051 0.60 ? 0.40 0.669 0.310 1.28 ? 0.06ASC2 0.527 0.049 0.62 ? 0.43 0.900 0.272 1.43 ? 0.37BETA1 �0.641 0.059 �0.53 ? �0.76 �1.280 0.288 �0.71 ? �1.85BETA2 0.591 0.059 0.71 ? 0.47 1.100 0.291 1.67 ? 0.53

Number of observations 6624 450Number of individuals 368 25Null log-likelihood �7277 �494Final log-likelihood �5365 �341Adjusted rho-square 0.26 0.29Total coefficients 20 12

Units of measurement of continuous variables:Cost: natural logarithm of $CAD (range: 4.9–7.6).On-time reliability: % of shipment delivered by the carrier that are on-time (range: 85–95%).Damage risk: % of shipments delivered by the carrier that suffer from damage (range: 0.5–3%).Security risk: % of shipments delivered by the carrier that suffer from theft (range: 0.5–1.5%).Shipment characteristics that are interacted with carrier attributes are the following:Dist.: 1000 km between shipment origin and destination (range: 0.555–1.462).By-appt.: identifies a shipment as being a by-appointment shipment.High-value: identifies a shipment as being a high-value shipment.Perishable: identifies a shipment as being perishable.Fragile: identifies a shipment as being fragile.Shipper characteristics that are interacted with carrier attributes are the following:Emp.: represents the number of employees (1000) of the respondent’s company.Ontario shipper: identifies the respondent company as being located in Ontario.btw Railheads: identifies respondent company as being located between the Montreal and Toronto Expressway railyards.

Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774 769

ordering from their provider. This results in lost sales for the supplying firm. Supplying firms can reduce their lost sales costsby investing in better transportation services. Improved transportation, however, implies higher transportation costs. Sup-plying firms therefore need to trade-off between transportation costs and lost sales costs. This is only one of the many trade-offs that firms make in their logistics and transportation decisions.

Many factors affect these various costs and thereby the cost trade-offs that are made. Ballou identifies a number of ‘‘RiskCharacteristics” that play a particularly important role in influencing logistics strategy. They are ‘‘features such as perishabil-ity, flammability, value, tendency to explode and ease of being stolen” (Ballou, 2004, p. 74). These characteristics affect notonly transportation-related costs (they are more expensive to ship, store, etc.), but also lost sales costs. The late delivery ofperishable goods is likely to be subject to higher lost sales costs than non-perishable goods. Lost orders of high-value goodsresult in greater lost sales costs. As such, we would expect shippers to trade-off between shipping costs and shipment char-acteristics. The inclusion of shipment interactions allows the testing and quantification of these trade-offs.

Overall, the model performs well with each of the direct carrier attribute coefficients being significant and having the ex-pected sign. Increases in cost, damage risk and security risk decrease the probability that a carrier is chosen, while greateron-time reliability increases the probability.

Before continuing, it is worth mentioning how logit coefficients are interpreted. Logit coefficients describe the effect ofthe independent variable on the odds of choosing a given alternative. If P represents the pre-existing probability of choosinga given alternative, the odds of choosing it are (P/(1 � P)). In particular, the logit coefficient describes by how much the nat-ural logarithm of the odds change with a unit change in the value of the independent variable. For example, for a coefficient

770 Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774

of b = 0.14, a unit increase in x results in an ebDx = 1.15-fold increase in the odds of choosing the alternative. This is the sameas saying the odds of choosing the alternative increase by 15%. Note that this does not mean the probability is multiplied by1.15.

If the explanatory variable is logarithmic, the coefficient can be directly interpreted for small changes in the explanatoryvariable. That is, if a logarithmic explanatory variable (e.g. ln(price)) has a coefficient of 0.1, a 10% increase in price will in-crease the odds of choosing a given alternative by about 1%.

Because the cost variable is in natural logarithms, the coefficient of �4.00 suggests that a 1% increase in cost would resultin a 4% decrease in the odds that a carrier would be chosen, and a 10% increase would decrease the odds by 32%. This figure isin the range of other similar studies. In Fridstrom and Madslien’s (2001), ‘‘shipment level” model they report an estimate of�2.21 (half the magnitude of our estimate), whereas Wigan et al. (2000) report coefficients from �0.049 to�0.298. The latterare based on nominal figures (i.e. the explanatory variable for cost is expressed in dollars and not natural logarithms). Table 2provides the estimated coefficients (as well as the effect that the coefficient implies for the odds of choosing a given alter-native) of variables in comparable carrier choice models.

The 0.09 coefficient for on-time reliability suggests a similarly strong effect on carrier choice as cost. Its value suggeststhat if a carrier’s on-time reliability were to improve by 1%, the odds of choosing that carrier would increase by almost10%, and would increase two and a half times with an increase of 10%. In more intuitive terms, supposing the initial prob-ability of a carrier being chosen was one half, a 10% increase in on-time reliability would increase the probability to 0.7. It isless straightforward to compare this coefficient with other studies, since other studies have tended to quantify on-time reli-ability in terms of percentage late as opposed to percentage on-time. That having been said, this seems a reasonable estimateand indicates extremely high sensitivity to on-time reliability in the choice of a carrier.

The damage risk coefficient, �0.39, indicates a 1% increase in damage risk would decrease the odds of choosing a carrierby about a third. This would reduce a probability of 0.5 to about 0.4. The coefficient is within the range of other studies, withFridstrom and Madslien and Wigan et al. reporting coefficients of �0.25 and ca. �500 respectively. The extremely large coef-ficients reported in Wigan et al. may have partly to do with a stricter definition of damage risk.

While other studies have not reported on security risk, the coefficient reported here,�0.10, seems reasonable. An increaseof 1% in security risk will reduce the odds of choosing a carrier by 10%. This would result in a decreased probability of choos-ing a carrier from an initial probability of a third to a quarter.

The coefficients for the continuous variables in the model seem quite strong and reasonable. The most remarkable result,however, is the value of the intermodal coefficient, �0.63. It implies that the odds of choosing a carrier that uses intermodalservice is almost halved. If for example the probability of choosing a particular carrier were one half, knowing that a carrierused intermodal services would reduce its probability of being used to a third. This result while strong should not be sur-prising given business press commentary (see for a recent example Luczak, 2005), and more generally, trucking’s growingmode share. It likely reflects general shipper perceptions of rail performance versus truck-only transportation services. Sincemode was one of the characteristics of alternative carriers it was included in the factorial design. As a result, any one of thealternatives (and sometimes two) could be identified as intermodal. The intermodal coefficient was independent of alterna-tive so that each alternative was equally likely to be characterized as intermodal.

Table 2Comparison of carrier choice model coefficients.

Variable Other end-shippers Fridstrom and Madslien (2001) Wigan et al. (2000)

Price(ln)Coefficient �4.000 N/A N/A N/A1% Increaseb �3.902

PriceCoefficient N/A �2.21 �0.049 �0.2981$ Increasea 0.110 0.952 0.742

On-timeCoefficient 0.09 N/A N/A N/A1% Increasea 1.094

Damage riskCoefficient �0.39 �0.25 �500 N/A1% Increasea 0.677 0.779 0.000

Security riskCoefficient �0.1 N/A N/A N/A1% Increasea 0.905

IntermodalCoefficient �0.63 N/A N/A N/A1 if Intermodala 0.533

a Amount by which current odds multiplied for unit increase in variable.b Amount by which current odds are increased for a 1% increase in price.

Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774 771

The coefficients ASC1 and ASC2 are the alternative-specific constants. They represent the position of a particular alterna-tive relative to the others in a choice task. ASC1 is the alternative-specific constant identifying the first, or left-most, alter-native in a choice task. ASC2 identifies the second alternative. Both suggest that respondents were more likely to choose thefirst and second alternative as compared to the third. In fact, the odds of choosing the first and second alternatives relative tothe third were 65% and 70% higher, respectively. It was necessary to include these variables to account for the effect that theposition of alternatives within the choice set plays. In other words, it removes positional bias. It was also necessary to includethem because the use of a random-effects model requires the presence of ASCs since the estimated BETA1 and BETA2 areactually the standard errors of the distribution of the ASCs. Without the ASCs it is not possible to estimate their distributionand therefore include the added flexibility that the random-effects model allows.

In addition to the five carrier attributes, there were another 11 shipment and shipper interaction variables. There are twosignificant distance interaction terms: cost and on-time reliability. The negative distance-cost interaction term means ship-pers are more sensitive to cost for longer distance shipments. Since cost also increases with distance, this is consistent witheconomic theory that shows that the elasticity of transportation demand increases with cost (see Wilson, 1980).

Evidence from the relevant interaction terms suggests that in the case of by-appointment and high-value shipments,shippers are more sensitive to on-time reliability than price. Since these characteristics represent higher-risk, these findingsare expected from, and consistent with, logistics theory. Moreover, carrier choice for fragile goods is shown to be more sen-sitive to damage risk (b = �0.21). Since fragile goods are ‘‘higher-risk”, this is also consistent with logistics theory.

The BETA coefficients are simply the standard error of the individual-specific term around the alternative-specific con-stants. The fact that they are statistically significant confirms that it is appropriate to include random error componentsin the estimation.

5.2. Comparison of models for 3PLs and other end-shippers

The most important aspect of this research is the difference between the preferences of 3PLs and other end-shippers. Thecolumn labeled ‘3PLs’ in Table 1 shows the carrier choice model based solely on the 3PL data.

The 3PL model has fewer variables than the other two, which is to be expected given the smaller number of observationsavailable for estimation. Despite fewer 3PL observations, it was possible to test statistically for differences between 3PLs andother shippers as reported in Section 5. The hypothesis that the subgroups had similar preferences was rejected at the0.0005% level. That is, there were enough 3PL observations and sufficient difference for it to be clear that they should bemodeled separately. Moreover, enough variables were statistically significant in the 3PL model to render the analysis useful.At first blush the coefficient estimates of the variables common to both models show quite different results for 3PLs andother end-shippers. Analysis begins by looking at the differences in coefficient estimates and is nuanced afterwards withan examination of the confidence intervals of the coefficients.

For cost, it appears that with ‘regular’ shipments (neither fragile, nor high-value nor by-appointment) 3PLs are moreprice-sensitive than other shippers. The coefficient of �7.83 suggests that a 1% increase in shipment cost reduces the oddsof a 3PL choosing a carrier by almost 8%. In order to compare with other end-shippers, it is proper to compare the cost coef-ficients for an average shipping distance (1000 km). This results in the other end-shipper coefficient being �6.36 (�4.000–2.360). This suggests that a 1% increase in shipment cost reduces the odds of an other end-shipper choosing a carrier by 6%,much closer to the 3PL figure of almost 8%.

This conclusion changes when shipment characteristics are considered. The high-value shipment/cost interaction coeffi-cients are 1.48 for other shippers and 4.66 for 3PLs. This suggests that for other end-shippers the cost coefficient for high-value goods is �3.12 (�4.00 + 1.48) compared to �3.17 (�7.83 + 4.66) for 3PLs. This would seem to suggest that while 3PLsare more price-sensitive for ‘regular’ shipments, this difference almost disappears in the case of high-value shipments. In-deed, it seems that 3PLs trade-off more between cost and transportation quality for ‘‘higher-risk” (see Section 5) shipmentsthan other end-shippers, or that they are less sensitive to cost for higher-risk shipments.

A similar pattern is observed in the case of on-time reliability. The on-time reliability coefficient for other end-shippers is0.09 compared to 0.11 for 3PLs, suggesting the latter are generally more sensitive to on-time reliability than other end-ship-pers are. Moreover, if one considers the coefficients for on-time reliability for by-appointment shipments, we see this resultmagnified. The interaction coefficient between on-time reliability and by-appointment shipments is 0.05 for other end-ship-pers and 0.09 for 3PLs. This gives combined on-time/by-appointment coefficients of 0.14 for other end-shippers and 0.20 for3PLs. If a carrier’s on-time reliability were to improve by 1%, it would increase the odds of other end-shippers choosing thatcarrier by 15% but that of 3PLs by 22%.

A pattern of particularly acute sensitivity to carrier quality for higher-risk shipments is also suggested in the case of dam-age risk. For 3PLs, damage risk was not considered significant for non-fragile goods, whereas for other end-shippers it was,with a coefficient of �0.39. However, damage risk was important to 3PLs for fragile shipments with a coefficient of �0.55.This is a little less than the other end-shippers’ combined coefficient of damage risk for fragile goods (�0.60). This suggeststhat 3PLs are less sensitive to damage risk, even for fragile goods (although not by much in the latter case). At the same time,knowing that a shipment is fragile prompts a bigger change in the carrier choice of 3PLs than of other end-shippers. For frag-ile shipments, 3PLs seem to react particularly strongly to the damage risk of their carriers.

A final variable significantly affecting 3PL carrier choice is the intermodal variable. The coefficient on this variable alsomarks a significant departure from other end-shippers. The intermodal coefficient for 3PLs is �1.42, larger than for the most

772 Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774

sensitive of other end-shippers (Ontario end-shippers located between the railyards of the existing service), and larger stillthan for the remaining other end-shippers (�0.63). This implies that knowledge that a carrier is intermodal reduces its oddsof being chosen by a 3PL by three quarters as compared to about a half for other end-shippers, not particularly for higher-riskshipments, but for shipments generally. As suggested above, this reluctance is likely caused by the bad reputation that railhas with shippers in terms of its service characteristics.

The analysis so far has concentrated on the relative size of the coefficients of the 3PL and ‘other end-shippers’ models.This, of course, is incomplete without considering the standard errors and confidence intervals of the coefficients. Overall,an analysis of the confidence intervals of the coefficients (apart from the intermodal coefficient) weakens the conclusionsbased only on an analysis of the coefficients. The reason is the relatively large standard errors for the 3PL coefficients thatare generally 2–4 times larger than those for the other end-shippers. These result in large confidence intervals that oftenoverlap with the confidence intervals of the other end-shipper coefficients. Statistically, this suggests that the null hypoth-esis that the coefficients are the same cannot be rejected.

Consider the cost variable. The 3PL cost coefficient is �7.83 and its confidence interval is �4.82 to �10.84. This overlapswith the other end-shipper cost + distance coefficient (�6.36) confidence interval of �5.71 to�7.01, suggesting no differencebetween the two models. This is the case for most of the other coefficients that can be easily compared in Table 2. The sameis also true for the 3PL damage risk for fragile goods coefficient. With a confidence interval of �0.25 to �0.84, this overlapswith the confidence interval of the damage + fragile coefficient (�0.60) of �0.53 to �0.68.

This is not the case, however, for the intermodal variable. The confidence interval for the 3PL intermodal variable(�1.11 ? �1.73) is far outside the interval for other end-shippers (�0.55 ? �0.71). It is only for those end-shippers thatare the most averse to using intermodal carriers that there is an overlap in the confidence intervals. The confidence intervalfor these carriers is �0.97 ? �1.24 which overlaps a little with the lower end of the confidence interval for the 3PLs.

Taken all together the results of the 3PL and ‘other end-shippers’ models suggest the following. First, 3PLs are more biasedagainst the use of intermodal carriers than are other end-shippers. Second, there is evidence that 3PLs are less price-sensitiveand more sensitive to other carrier characteristics than their counterparts for higher-risk shipments. This is suggested by thefact that the interaction coefficients between higher-risk shipments and carrier characteristics are much larger than for otherend-shippers. While the large standard errors and thereby confidence intervals for the 3PLs, weaken such a conclusion, itdoes not rule it out as a possibility. If more observations were available, the standard errors would be smaller. If they wereas small as for the other end-shippers, it is quite possible that the confidence intervals of many of the 3PL coefficients wouldfall outside those of the other end-shippers. Since this is not the case, it would not be justified to draw such strong conclu-sions. At the least, however, these results provide the first evidence of such a difference between 3PLs and other end-ship-pers. They also suggest the need for further research to establish the differences in preferences for carriers between the twotypes of shippers.

6. Discussion of results and conclusions

The contributions of this paper to the literature are the following. First, it represents the first attempt to establish andquantify differences in preferences for carrier and mode choice between shippers that choose their own carriers, and thosethat choose carriers on behalf of shippers. Second, it shows clearly that 3PLs have a particularly strong bias against inter-modal transportation. Third, it provides evidence that 3PLs may behave differently from other end-shippers in other respectsas well.

In particular, it suggests 3PLs to be more sensitive to cost for ‘regular’ (non high-risk) goods. It also suggests them to bemore sensitive to carrier attributes for the high-risk shipments that they organize with their shipment type/carrier attributecoefficients being larger for cost, on-time reliability and damage risk. It is not justifiable (statistically speaking) to be too cat-egorical about this last observation since an analysis of the confidence intervals of the coefficients in the two models does notbear out the conclusions. The fact remains, however, that with more 3PL observations and thereby narrower confidenceintervals, these differences may well be born out. As a result, this represents an important avenue of further research giventhe paucity of research looking directly at 3PLs. The results also beg the question ‘‘why do 3PLs behave differently from otherend-shippers?”

Taken together, the results yield an interesting interpretation, despite the lack of literature or theory to rely on for thistopic. Perhaps increased sensitivity to shipment attributes for high-risk goods is due to the fact that 3PLs are serving clients.Since risks are more substantial for high-risk goods, shippers in general are more concerned about them arriving on-time andin good condition. For other end-shippers a botched shipment can be dealt with by changing carriers. For a shipper using a3PL, a botched shipment can be dealt with by changing their 3PL. From the 3PL’s perspective the fact that a botched shipmentcan result in the loss of a client appears to make them more sensitive to carrier attributes, particularly in the case of high-riskgoods. For the most part, 3PLs appear to behave with their clients’ best interests at heart. As such, they are more sensitive tocost and on-time reliability. They are also particularly sensitive to damage risk if they are shipping fragile goods, more sen-sitive to on-time reliability if a shipment is by-appointment, etc. That 3PLs appear to be acting in their clients’ interest couldalso explain why they have become so popular.

Because of the rapid increase in the importance of 3PLs in freight transportation, this research has two important impli-cations. First, that 3PLs appear to be very reluctant to use intermodal carriers presents an even greater challenge to increas-

Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774 773

ing rail mode share. Second, this has important implications for carrier and mode choice modeling more broadly. In partic-ular, it suggests that failure to capture differences between 3PLs and other end-shippers will result in the use of inadequatemodels for describing and predicting carrier and mode choice – certainly in the case of carrier mode, and perhaps in otherrespects as well. As a result, future models ought to seek to capture differences between 3PLs and other end-shippers, andfuture research to better understand the differences between 3PLs and other end-shippers.

References

Aimi, Greg., 2007. Will 3PLs grab the global opportunity? Logistics Management (June), 12S–14S.Amekudzi, A., Meyer, M., 2005. Consideration of Environmental Factors in Transportation Systems Planning. Transportation Research Board, Washington,

DC. p. 541.Ballou, Ronald H., 2004. Business Logistics/Supply Chain Management: Planning, Organizing, and Controlling the Supply Chain. Pearson Prentice Hall, Upper

Saddle River, New Jersey.Biederman, David., 2007. Champing at the bit: 3PLs weigh the risk and rewards of expanding into China. Journal of Commerce 4 (June), 26–30.Bierlaire, M., 2007. BIOGEME [1.4]. Lausanne, Switzerland, Transport and Mobility Laboratory, Ecole Polytechnique Fédérale de Lausanne.Bottani, Eleonora, Rizzi, Antonio, 2006. A fuzzy TOPSIS methodology to support outsourcing of logistics services. Supply Chain Management – An

International Journal 11 (4), 294–308.Calzada, C., Jiang, F., 1997. Modelling the influences of the characteristics of freight transport on the value of time and the mode choice. In: Seminar E of the

PTRC European Transportation Forum, Brunel University, England, 1–5 September 1997, first ed. PTRC, London.Dibenedetto, Bill., 2007. Third-party logistics: signed, sealed, delivered. Journal of Commerce 25 (June), 38–39.Dibenedetto, Bill, Diben, 2007. 3PL services expand in Latin America despite political, economic and infrastructure challenges. Journal of Commerce 7 (May),

25–28.Evangelista, Pietro, Sweeney, Edward, 2006. Technology usage in the supply chain: the case of small 3PLs. The International Journal of Logistics

Management 17 (1), 55–74.eyefortransport, 2005. 2nd Annual Outsourcing Logistics Report: Fortune 500 Shipper’s Views on their 3PL Partnerships, London. <eyefortransport.com>.Foggin, James H., Mentzer, John T., Monroe, Carol L., 2004. A supply chain diagnostic tool. International Journal of Physical Distribution & Logistics

Management 34 (10), 827–855.Fowkes, A.S., Tweddle, G., 1988. A computer guided stated preference experiment for freight mode choice. In: Proceedings of Seminar D, Transportation

Planning Methods, Held at the 16th PTRC Transport and Planning Summer Annual Meeting, University of Bath, England from 12–16 September 1988.PTRC Education and Research Services, Ltd., London, pp. 295–305.

Fridstrom, L., Madslien, A., 2001. Wholesalers’ freight choice: a representative stated preference survey. In: AET European Transport Conference. PTRCEducation and Research Services, Ltd., London.

Greene, William., 2000. Econometric Analysis, fourth ed. Prentice Hall, Inc., Upper Saddle River, NJ.Gunasekaran, A., Ngai, E.W.T., 2003. The successful management of a small logistics company. International Journal of Physical Distribution & Logistics

Management 33 (9), 825–842.Hannon, David, 2007. Choose wisely: not all third-party logistics providers are the same. Purchasing 14 (June), 35–36.Hoffman, William, 2005. Logistics costs: who’s getting squeezed? Traffic World 7 (November), 9–12.Hoffman, William, 2006. 3PLs maturing fast. Traffic World 27 (November), 16–17.Hoffman, 2007. 3PLs reach record revenue. Traffic World 23 (April), 18–19.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.Knemeyer, Michael A., Murphy, Paul R., 2005. Exploring the potential impact of relationship characteristics and customer attributes on the outcomes of

third-party logistics arrangements. Transportation Journal (Winter), 5–19.Ko, Hyun Jeung, Evans, Gerald W., 2007. A genetic algorithm-based heuristic for the dynamic integrated forward/reverse logistics network for 3PLs.

Computers & Operations Research (34), 346–366.Kumar, Manoj, Vrat, Prem, Shankar, Ravi, 2006. A multi-objective 3PL allocation problem for fish distribution. International Journal of Physical Distribution

& Logistics Management 36 (9), 702–715.Lai, F., Zhao, X., Wang, Q., 2007. Taxonomy of information technology strategy and its impact on the performance of third-party logistics (3PL) in China.

International Journal of Production Research 45 (10), 2195–2218.Lieb, Robert, Bentz, Brooks A., 2004. The use of third-party logistics services by large American manufacturers. Transportation Journal 43 (3), 1–7.Logistics Today, 2006. Third-party logistics providers (3PLs). Logistics Today (May), 34–36.Luczak, M., 2005. Shipper sound-off on service. Railway Age 206 (11), 37–49.Maloni, Michael, Carter, Craig, 2005. Opportunities for research in third party logistics. Transportation Journal (Spring), 23–38.Min, Hokey, Joo, Seong Jong, 2006. Benchmarking the operational efficiency of third party logistics providers using data envelopment analysis. Supply Chain

Management – An International Journal 11 (3), 259–265.Mitra, Subatra., 2006. A survey of third-party logistics (3PL) service providers in India. IIMB Management Review (June), 159–174.Murphy, Paul R., Poist, Richard F., 2000. Third-party logistics: some user versus provider perspectives. Journal of Business Logistics 21 (1), 121–133.Norojono, Olly, Young, William, 2003. A stated preference freight mode choice model. Transportation Planning and Policy 26 (2), 195–212.Oliver Silver, 2005. How 3PLs can help manage the increasing complexities of transportation. Canadian Transportation and Logistics (October), 36–37.Patterson, Zachary, Ewing, Gordon, Haider, Murtaza, 2007. Shipper preferences suggest strong mistrust of rail: the first SP carrier choice survey in the

Quebec City – Windsor Corridor. Transportation Research Record 2008, 67–74.Peters, Melvyn J., Lieb, Robert, Randall, Hugh L., 1998. The use of third-party logistics services by European industry. Transport Logistics 1 (3), 167–179.Quinn, John Paul, 2006. 3PLs hit their stride. Logistics Management (July), 3T–8T.Shinghal, N., Fowkes, Anthony, 2002. Freight mode choice and adaptive stated preferences. Transportation Research Part E: Logistics and Transportation

Review 38 (5), 367–378.Shister, Neil, 2006. Executive overviews: 3PLs. World Trade (June), 30–32.Sinkovics, Rudolph, Roath, Anthony, 2004. Strategic orientation, capabilities, and performance in manufacturer – 3PL relationships. Journal of Business

Logistics 25 (2), 43–64.Sohail, M., Bhatnagar, R., Sohal, A., 2006. A comparative study on the use of third party logistics services by Singaporean and Malaysian firms. International

Journal of Physical Distribution & Logistics Management 36 (9), 690–701.Stank, Theodore, Goldsby, Thomas, Vickeree, Shawnee, Savitskie, Katrina, 2003. Logistics service performance. Estimating its influence on market share.

Journal of Business Logistics 24 (1), 27–55.Tompkins, James A., 2006. The business imperative of outsourcing. Industrial Management (January/February), 8–12.Train, K.E., 2003. Discrete Choice Methods with Simulation. Cambridge University Press, Cambridge.Vellay, Carine, de Jong, Gerard, 2003. Analyse conjointe SP/RP du choix du mode transport de marchandises dans la Région Nord – Pas-de-Calais. RAND,

Santa Monica, CA.

774 Z. Patterson et al. / Transportation Research Part E 46 (2010) 764–774

Wang, Q., Zantow, Kenneth, Lai, F., Wang, X., 2006. Strategic postures of third-party logistics providers in mainland China. International Journal of PhysicalDistribution & Logistics Management 36 (10), 793–819.

Wigan, Marcus, Rockliffe, Nigel, Thoresen, Thorolf, Tsolakis, Dimitris., 2000. Valuing long-haul and metropolitan freight travel time and reliability. Journal ofTransportation and Statistics 3 (3), 83–89.

Wilding, Richard., Juriado, Rein., 2004. Customer perceptions on logistics outsourcing in the European consumer goods industry. International Journal ofPhysical Distribution & Logistics Management 34 (8), 628–644.

Wilson, George.W., 1980. Economic Analysis of Intercity Freight Transportation. Indiana University Press, Bloomington, IA.Yeung, Jeff, Selen, Willem, Sum, Chee-Chuong, Huo, Baofeng, 2006. Linking financial performance to strategic orientation and operational priorities an

empirical study of third-party logistics providers. International Journal of Physical Distribution & Logistics Management 36 (3), 210–230.Young, W., Richardson, A., Ogden, K., Rattray, A., 1983. An inter-urban freight mode choice model. Transportation Planning and Technology (8), 61–80.