Identifying preferences for public transport investments under a constrained budget

Transportation Research Part A 72 (2015) 27–46

Contents lists available at ScienceDirect

Transportation Research Part A journal homepage: www.elsevier.com/locate/tra

Identifying preferences for public transport investments under a constrained budget David A. Hensher ⇑, Chinh Ho, Corinne Mulley Institute of Transport and Logistics Studies, The University of Sydney Business School, The University of Sydney, NSW 2006, Australia

a r t i c l e

i n f o

Article history: Received 15 July 2014 Received in revised form 1 December 2014 Accepted 4 December 2014

Keywords: Public transport Bus rapid transit Light rail Value for money Heteroscedastic utility Constrained budget

a b s t r a c t As urban areas face increasing demands for new transport infrastructure to promote a sustainable future with an increasing reality of constrained government budgets, the debate on whether we should focus on rail or bus-based investments continues unabated in many jurisdictions. Associated with the debate is an emotional (or ideological) bias by communities in favour of one mode, especially rail, which carries much sway at the political level as if there is no budget constraint. This paper presents a stated choice experiment to investigate this context as two unlabelled options described by 20 potential drivers of community preferences for improved public transport, where each choice scenario is conditioned on an estimated construction cost and a total annual transport infrastructure budget for the relevant geographical jurisdiction. This is followed by a labelling of each alternative to reveal whether the option is bus rapid transit (BRT) or light rail (LRT) and to establish whether this additional information influences preference revision. Data is collected in all eight capital cities of Australia in mid 2014. Mixed logit models with heteroscedastic conditioning in terms of the cost of the project infrastructure and whether the alternative is labelled BRT or LRT, provide new evidence on the nature and extent of community modal bias in a budget-constrained choice setting. The conclusions are twofold. On the one hand, if a fully compensatory choice rule is assumed (as is common in all previous modal comparison studies), LRT is predominantly preferred over BRT despite budgetary constraints, similarities in quality of service attributes and the opportunity to choose a greater network coverage for a given construction cost. However, when we allow for attribute non-attendance (a semi-compensatory choice rule), the modal bias is no longer a significant driver of preferences. Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction There is growing pressure on many governments to provide improved public transport as a sustainable alternative to the car. In many industrialised societies where the car dominates the modal share, the amount of public transport investment required to make a significant switch from the car is substantial. However budget constraints mean that governments are unable to focus on the many corridors required to have a significant impact on traffic congestion. With tight budgets, the choice is of rather limited geographical investment using a rail solution or the delivery of a more system-wide investment using bus based technology, as illustrated by the number of developing economies where budgets are significantly more

⇑ Corresponding author. E-mail addresses: [email protected] (D.A. Hensher), [email protected] (C. Ho), [email protected] (C. Mulley). http://dx.doi.org/10.1016/j.tra.2014.12.002 0965-8564/Ó 2014 Elsevier Ltd. All rights reserved.

28

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

restricted and where bus-based systems (e.g., Curitiba and Bogota) have shown the great advantage of serving a much larger catchment area than rail systems for the given budget. Whilst buses often have a major role to play in the public transport task (e.g., carrying more passengers per year than rail in Sydney, for example), the incremental investment in additional bus services in most western economies has shown some, albeit limited success, in getting car users to switch. What is clearly needed (especially where governments are not supportive of road pricing reform for cars) is to find a way to grow public transport throughout a metropolitan area to such an extent that a non-marginal switch to public transport can occur. With a fixed budget, one might suggest that this can be most economically achieved through a bus-based plan in which corridors are dedicated to bus rapid transit (or where there is a significant increase in frequency and coverage in mixed traffic1). Government budgets are under pressure not only from the transport sector but also from competing sectors such as health, education and law and order (including security2). It is therefore particularly important to identify and promote public transport investments that are able to deliver real societal benefits in respect of outcomes related to improved accessibility, safety, and environmental sustainability, serving a larger number of current and potential public transport users for a given budget outlay. This means that the observed preference by governments for rail-based solutions as the only (or dominating) way to deliver these benefits has to be questioned. In turn this opens up the relevance of at least assessing other options such as bus-based solutions. This is a challenge in many cities where emotional ideology in support of rail-based solutions (without question) dominates the debate and political decision making (see Hensher and Waters, 1994). The issue of public preference of rail over bus has been investigated in a number of earlier studies, some of which found a significant preference of rail over bus, whilst others found no evidence of such a bias towards rail service. For example, BenAkiva and Morrikawa (2002) found no evidence of such a bias towards rail services when both services had equivalent travel times and fares. However, the same study found that a bias existed when rail offered a higher quality of service. Yannes et al. (2012) also found no significant public preference for rail service over bus service. However, these studies did not consider the reality facing many cities today, of limited budgets and a pressing need for public transport investments. In these circumstances, two issues are important to understand. First, how the majority of society would temper their preference for a new modern light rail system if its geographical scope were to be limited by budget considerations? Second, and alongside this, how society would view a bus based system giving greater network coverage for the same budget outlay as a proposed modern rail system? This paper investigates these key questions using a stated choice experiment. Each experiment begins with a binary choice set of unlabelled options described by 20 potential drivers of community preferences for improved public transport, where each choice scenario is conditioned on an estimated common construction cost and a total annual transport infrastructure budget for the relevant geographical jurisdiction. This is followed by a labelling of each alternative to reveal whether the option is bus rapid transit (BRT) or light rail transit (LRT); and in the light of this additional information, individuals can review their preferences and revise them if they wish. Data were collected in eight capital cities of Australia in 2014. The empirical evidence is derived from mixed logit models with heteroscedastic conditioning, the latter defined in terms of the construction cost of the new transport infrastructure relative to the annual infrastructure budget and whether the alternative is labelled BRT or LRT, assuming in one model that all attributes are relevant as well as allowing for attribute non-attendance in a separate model. This analysis identifies the main influences on the preferences of the sample of 1018 residents, including the role of the ex post modal naming in revised preferences for each investment option and the influence of the budget commitment, and permits commentary on the two issues identified above which motivate the paper in the conclusions and synthesis section.

2. Drivers of community preferences for public transport The challenge in selecting the factors that could influence individual’s choice in the experiment was to create an experiment which included relevant attributes, but not too many as this is known to influence the ability of respondents to complete the experiment effectively (see Hensher, 2010). The selection of the key factors that influence individuals when asked about the preferences for various public transport investments has been informed by the broader literature and especially our own investigations (see Hensher et al., 2015b) into candidate attributes, particularly relating to voting choices which are of great interest to policy makers. In broad terms, the choice of attributes centred on two main sources of literature; the literature relating to the aspects of different public transport modes which appeal to users and which are important in determining the mode choice of travellers; and the literature which focuses on the potential of modal image to influence mode preferences. The former body of literature was used to inform, in detail, a phase one part of this study which developed best–worst preference experiments, one associated with design characteristics, one with service descriptions associated with BRT and LRT, and another related to 1 As an example, if in Sydney we had not proceeded with the 23 km North–West rail project (costed at approximately $500 m per km), we could increase the number of buses threefold. Even in mixed traffic this would deliver significant improvements in frequency and connectivity. If this did not deliver a nonmarginal switch to public transport, then it is doubtful that any single corridor investment where car is 80% of modal share, is likely to have little chance of effecting noticeable change, 2 See http://sydney.edu.au/business/itls/tops as one example of such evidence.

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

29

(no modal label) voting preferences. The table of attributes assessed and identified from the extensive literature is given in Appendix A, reproduced from Hensher et al. (2015b). This section highlights the contribution of the quality of service literature in identifying the attributes of public transport which makes it attractive to users. From a review of the academic, technical and grey literature, a long list of potential attributes were selected for a the phase one best–worst experiment (Hensher et al., 2015b; Mulley et al., 2014) as a way of synthesis and creating a short list of the most important quality attributes for use in this current experiment. The best–worst experiment identified attributes affecting the design and service characteristics of the mode, with the same statements being used in the context of being in favour of a rail-based mode and a bus-based mode. The literature underpinning this work came from a review of the more technical papers by Hensher (1991), Swanson et al. (1997), Cirillo et al. (2011), dell’Olio et al. (2010a,b), Eboli and Mazzulla (2010, 2008a,b), and Marcucci and Gatta (2007), and the strategic studies of HassKlau and Crampton (2002), Hensher and Waters (1994), Hensher (1999), Mackett and Edward (1996a,b), Canadian Urban Transit Association (2004), Cornwell and Cracknell (1990), Kain (1988), Pickrell (1992), and Sislak (2000). This synthesis was supplemented by the more recent literature on the ridership drivers of public transport (Currie and Wallis, 2008; Currie and Delbrosc, 2013; Hensher et al., 2014). Whilst this literature has a long history and provides much empirical evidence, there has been a movement more recently, with the input from transport psychologists, of a focus on the importance of perceptions of quality as opposed to hard measurements. Many of the authors cited above have included this in their evidence (for example Stradling et al. (2007), Cirillo et al. (2011), dell’Olio et al. (2010a,b), Eboli and Mazzulla (2010, 2008a,b)), and a more recent review by Redman et al. (2013) widens the discussion to identify which of the quality characteristics of public transport are perceived as important to car users. Perceptions are, of course, important in the second body of literature which focuses on the potential of modal image to influence mode preferences. This literature is dominated by a comparison of the ‘image’ of the car vs. public transport which is not so relevant to this study, with the exception of the way in which the poor ‘image’ of public transport has been identified as a principal cause of low take up of public transport by car users. In this context, image is defined as ‘‘the set of ideas and impressions, both rational and emotional, which major stakeholders form about the organisation or industry’’ (TCRP, 2000, p.5). Image is thus related to ideas rather than hard facts and there is evidence to suggest that BRT has gained its image indirectly from its association with the bus mode and the way that this is affected by the character of buses in mixed traffic (slow and unreliable) (Hensher and Waters, 1994; Hensher and Mulley, in press). The literature described above, together with our previous empirical work in exactly the same setting, led to attributes being included to capture differences in cost and coverage between rail and bus based mode (population serviced, % dedicated right of way, operating and maintenance costs and route length (given a fixed budget between two options)), service levels (service capacity, peak and off-peak frequency, travel time and public transport fare), features of the system described by fare payment (on or off vehicle, integrated fare or not), interchange penalty, safety and security and ease of boarding and other general factors shown to be important in voting between transport systems (the assurance of a minimum period of operation and risk of being closed down after this period, value uplift around stations, mode switch from cars and environmental friendliness of the system). These were introduced into the survey instrument as shown in Fig. 1. It should be noted that the context of the experiment is on identifying preferences for BRT and LRT for a given budget, which is the position governments find themselves in when committing to improving public transport infrastructure. This, of course, means we are not asking our respondents whether they would use public transport or either of these modes which, although a legitimate research question, is different from the one considered here (see for example Hensher and Rose, 2007).

3. Embedding sources of heteroscedastic conditioning in the choice model For over 40 years, researchers have developed a rich suite of modelling tools to understand the role of observed and unobserved influences on choice outcomes within a random utility theoretic framework in which uncertainty is intrinsically linked to the analyst’s absence of full information on the real sources of individual choice making. In recent years, an increasing number of analysts have highlighted a concern with the assumption that all attributes are traded in a fully compensatory manner, and are by implication all relevant, and that each attribute and its trade is treated by the individual decision maker as totally certain (see e.g., Hensher and Collins, 2011). In the context of the growing interest in process heuristics (see the following section), two issues that are considered to play a role in conditioning the choices made by individuals in survey settings, especially stated choice experiments, are (i) the relevance of attributes and their levels, and (ii) the overall acceptability of an alternative as described by a package of attribute levels. In the current application we have replaced acceptability with the budget constraint and the labelling of the alternative as bus-based (BRT) or rail-based (LRT). The model form is designed to recognise that the entire utility expression, defined in terms of a series of influences on the level of relative utility, might be conditioned by a respondent’s view on the role that the available budget and the labelling of a mode might play. This approach is preferred to simply adding in two extra variables in a linear addition specification, and in a sense is equivalent to interacting every attribute with the budget and modal labelling effect. We use the framework in Hensher and Rose (2012) to condition an entire utility expression on potential influences that either do not vary between alternatives in a choice set (i.e., the budget constraint) or vary across the alternatives (i.e., the revealed modal labelling of BRT or LRT). Given the standard utility expression associated with the jth alternative contained in a choice set of j = 1, . . . , J alternatives, we assume that an index defining the fixed project budget available for each of the

30

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

Fig. 1. Illustrative choice set screen.

alternatives, Bj, as well as the modal labelling of the jth alternative associated with the qth individual, conditions the utility expression (noting that different labelling will be assigned after the initial ranking of the two alternatives that may or may not lead to rank revision to each of the alternatives). The functional form of the utility expression can be denoted by

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

31

U⁄jq = ZjqUjq = Zjq(Vjq + ej), where Ujq⁄ is the standard utility expression conditioned on the available project budget and modal labelling of an alternative. This conditioning is a form of heteroscedasticity. Zjq recognises that individual-specific perceptions associated with the available budget and modal labels condition the marginal (dis)utility of each and every attribute, observed and unobserved, associated with the jth alternative in a pre-defined choice set (Hensher and Rose, 2012). An example of heteroscedastic conditioning, implemented in the empirical section below, is Z jq ¼ ð1 þ dq Bcq þ cjq M jq Þ, where Bcq is a variable denoting the project budget that is fixed within the cth choice task, and Mjq is a dummy variable indicating whether the jth alternative has a subsequent label of BRT (=1) or LRT (=0) associated with the choice set being assessed by the qth individual. dq and cjq are estimated parameters.3 Zjq can take a number of forms including the linear form above or an exponential form. The incorporation of heteroscedastic conditioning in the random error term is equivalent to an error components specification (Wjq) given in Eq. (1) in which we allow variance heterogeneity in the error components for alternatives and nests of alternatives (where h defines the budget B and modal label M variables, with a set of parameters s).

h  i2 Var½W jq  ¼ hj exp s0j hq

ð1Þ

4. Overlaying attribute non-attendance in the heteroscedastic conditioned model Individuals are assumed to be fully compensatory in the way they assess and trade-off attributes in choice making by the great majority of choice modelling research. It is also commonly assumed that all attributes are relevant in choice making. Both these assumptions are commonly embedded in stated choice experiments. One consequence of these assumptions is the resulting view that studies which require an assessment of an increasing number of attributes by respondents impose growing complexity and cognitive burden on the individual respondent which in turn risks the loss of identifiable comprehendible settings of choice making (Hensher, 2014). One notable criticism of the fully compensatory assumption is that decision makers are not homogeneous but very heterogeneous in the way that they make choices in real and hypothetical markets, and that they draw on rules (or heuristics) to assist them in making choices. These rules may vary by contexts typified, for example, by habit or variety seeking behaviour. They may also reflect the fact that individuals bring to the choice making setting their views on what attributes are the key drivers of specific choice outcomes, and that these attributes may or may not be included in the set defined by an analyst. There is a real risk that the analyst may self-impose their own views (or prejudices), or even those of a client funding a study, on the number of attributes and alternatives that are deemed comprehendible to a sample of respondents in a survey when the fully compensatory assumption is in place. These presumptions have been questioned in the broader literature on heuristics and decision making (summarised in, for example, Hensher et al., 2015a, Ch. 21) that has evolved in a number of literatures, notably, psychology, economics and marketing; however the migration of ideas from this literature, which we refer to as process heuristics, has been slow to influence the way that discrete choice modelling has been represented. This is changing now, with a growing number of studies questioning the standard fully compensatory choice paradigm, as in this current study. A popular heuristic is attribute non-attendance (ANA) which recognises that individuals in real or hypothetical choice making settings often adopt an attribute relevance rule when deciding which alternative is their most preferred. There is a growing literature (see Hensher, 2010) on ANA that uses a number of methods and data sources to inform the analyst on the nature and extent of ANA. The two main approaches involve either asking supplementary questions at the end of the choice tasks or after each choice task in a stated choice experiment as to which attributes were ignored (i.e., not attended to) or inferring ANA from an econometric specification that can establish the incidence of zero marginal disutility associated with a specific attribute for each respondent (see Hess and Hensher, 2010). Latent class models with restrictions within class of parameter estimates (set to zero) is one way to represent the ANA inference approach (see Hensher and Greene, 2010; Greene and Hensher, 2013). Both approaches can also be combined; for example, conditioning the class allocation probabilities in a latent class model on the responses to supplementary questions on ANA. All approaches are detailed in Hensher et al. (2015a, Ch. 16). In the current study we use the supplementary question approach and implement it during the processing of each choice set. This enables us to account for the differences in levels of the attributes between choice sets in contrast to the more common approach of assessing ANA after completion of all choice sets, which is unable to establish whether relevance of an attribute depends on the levels offered in a choice set (which may be dependent on the offering levels of all attributes associated with an alternative in a choice set). An important focus of the empirical results is to establish what role the budget constraint and modal labelling plays in a model that incorporates ANA over the model that assumes that all attributes are relevant to all individuals. We chose not to test for differences in relevance between the respondent’s personal choices and the respondent’s choices for society and the ANA question, asked after each screen, produced very similar responses across the preference response questions. 3 This is not strictly scale heterogeneity – see the following paragraphs – although it appears like deterministic scale as a function only of covariates. In contrast, scale heterogeneity as represented in SMNL is a stochastic treatment which may be partially decomposed via deterministic addition of covariates.

32

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

4.1. The full choice model The model form for the utility expression that encapsulates all the elements presented above is given in Eq. (2).

" # K   X U jq ¼ 1 þ dq Bcq þ cjq Mjq aj þ ð0ANAkjq ; bkjq ÞX kjq þ ejq

ð2Þ

k¼1

All terms are defined above except aj which are alternative-specific constants and 0NAjkq which is the parameter estimate under ANA. Eq. (2) is a non-linear utility function, which in its general form departs from a standard random utility model with utility functions defined over Jqc choices available to individual q in choice situation c, given in Eq. (3).

U jqc ¼ V jqc þ ejqc ;

j ¼ 1; . . . ; J qc ;

c ¼ 1; . . . ; Cq;

q ¼ 1; . . . ; Q

ð3Þ

The IID, type I extreme value distribution, is assumed for the random terms ejqc. Conditioned on Vjqc, the choice probabilities take the familiar multinomial logit (MNL) form (4).

expðV jqc Þ Probjqc ¼ PJ qc j¼1 expðV jqc Þ

ð4Þ

When we allow for heteroscedastic conditioning, Eq. (4) becomes Eq. (5), called the heteroscedastic mixed logit and error components (HMLEC) model.

h i h  i2   P  1 þ dq Bcq þ cjq Mjq aj þ Kk¼1 0ANAkjq ; bkjq X kjq þ hj exp kq Bcq þ gjq M jq  ¼ h i h   i2  PJqc PK exp 1 þ d B þ c M a þ ð0 ; b ÞX exp k B þ g M þ h q cq q cq j j j j ANAkjq kjq kjq jq jq k¼1 j¼1 q q exp

Probjqc



ð5Þ

4.2. Empirical study The centrepiece of the empirical study is a stated choice experiment designed to understand how, from the communities perspective, a government should spend money on building infrastructure and gain voter’s support, together with the service attributes of public transport infrastructure, be it a BRT system or a LRT system, that are important to the community. The service attributes are classified into four groups shown in Table 1 together with attribute levels and attribute names. The stated choice experiment is designed with the same budget for BRT and LRT systems,4 which are referred to as System A and System B in each choice scenario. Subsequent to the respondent indicating their preference for a bundle of attributes defining two unlabelled alternatives, the modal label (BRT or LRT) is revealed for the same pair of alternatives and the respondent is asked to review, and possibly revise, their choice. Such an exercise is designed to reveal whether images about bus-based and rail-based systems are relevant in preference revelation. Each respondent is asked to answer two choice tasks. Given the number of levels for each attribute and the desire to maintain attribute level balance, the survey is designed using Ngene5 (Choice Metrics, 2012) with 24 rows (i.e., choice tasks) and blocked into 12 blocks so that each respondent will be assigned a block with 2 choice tasks. A set of conditions are employed to ensure that peak-hour level of service is no worse than the off-peak level of service, so if the level of peak-hour frequency of System A is 10 (min), then the allowed levels of off-peak frequency of System A are 10, 15 and 20 (min). When the level of peak frequency is 5 min, the off-peak frequency can be any of the predefined levels, and thus no condition is required. The correlation between design attributes is provided in Appendix C, together with detailed information about the survey design. The largest correlation is between peak and off-peak frequency of services (r = 0.46) which is a consequence of the condition applied in the survey design to ensure that the peak hours have more services than off-peak hours do. All other correlations are very small (0.2 < r < 0.2). This survey is designed for estimating MNL models defined by the utility functions of the unlabelled alternatives (System A and System B). When generating an efficient design, each parameter must have a prior value, which can be fixed (e.g., MNL model) or random (e.g., mixed logit model). To obtain the priors, we undertook a pilot survey (designed using a D-error measure, see Choice Metrics (2012) and Hensher et al. (2015a) for details of such designs) of 200 respondents (each shown the two choice screens) and estimated a choice model. The parameter estimates were used to redesign the choice experiment. We then collected another 400 respondents, giving a total of 600 respondents or 1200 choice responses, and re-estimated the model to obtain more accurate priors. These priors were used to undertake a further redesign of the choice experiment and a further 400 respondents were then sought, giving a total of 1000 respondents or 2000 choice responses (although the survey firm obtained an additional 18 respondents or 36 choice sets). The Ngene designs allow for nonlinear relationships through the model utility functions with dummy coded attributes. 4 5

A complementary set of choice scenarios (an additional two choice tasks) varied the budget whilst holding the route length fixed. Full details of the Ngene syntax and efficiency outputs for this application is given in Hensher et al. (2015a, Ch. 6.6.3 Design 3: D-Efficient Choice Design).

33

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46 Table 1 Predefined attributes and attribute levels for the choice experiment. Attributes

Name

Attribute level

# Levels

2, 4, 7 bn$

3

cost time pop roway opcost

0.5, 1, 3, 6 bn$ 1, 2, 5, 10 5, 10, 15, 20 25, 50, 75, 100 2, 5, 10, 15 m$ 10, 20, 30 kms

4 4 4 4 4 3

Service levels Service capacity in one direction (passengers/h) Peak frequency of service, every. . . Off-peak frequency of service, every. . . Travel time (door-to-door) compared to car Fare per trip compared to car-related costs (fuel, tolls, parking)

capa pfreq ofreq tcar fare

5 k, 15 k, 30 k 5, 10, 15 min 5, 10, 15, 20 min 10%, 10%, 15%, 25% ±20%, ±10%

4 3 4 4 4

Features of the system Off-vehicle prepaid ticket required Integrated fare Waiting time incurred when transferring On-board staff for passenger safety and security Ease of boarding public transport vehicle

prepaid ticket wait staff board

Yes, no Yes, no 1, 5, 10, 15 min Present, absent Level boarding, steps

2 2 4 2 2

General characteristics of investment Operation is assured for a minimum of Risk of it being closed down after the assured minimum period Attracting business around stations/stops % car trips switching to this option within first 3 years of opening Overall environmental friendliness compared to car The two systems described above are actually

yearop close buss shiftcar env brt

10, 20, 30, 40, 50, 60 years 0%, 25%, 50%, 100% Low, medium, high 0%, 5%, 10%, 20% ±25%, 10%, ±5%, 0% BRT, LRT

6 4 3 4 6 2

Budget levels (held constant in each experiment for System A and System B) Average annual transport infrastructure budget Description of investment Construction cost of project Construction time % metropolitan population serviceda % route dedicated to this system only and no other means of transport Operating and maintenance cost per year (millions) Route length

a A referee questioned why this attribute is not endogenous and suggested the use of an instrumental variable in place of this attribute. Like all the attributes in the choice experiment, this attribute is predefined by a number of levels that are then used to construct scenarios. As the attribute levels are predetermined, it cannot be treated as endogenous. This specific attribute, however, was not statistically significant in the final model reported in Table 3.

The screenshot in Fig. 1 is an example of a choice set shown to a respondent. Each respondent was first asked to evaluate the two alternatives (System A and System B) associated with each of the two choice tasks without any knowledge of which was BRT or LRT. That is, the part of the screen that revealed the modal name of the system was suppressed as was the column relating to whether an attribute was ignored or not (i.e., deemed relevant or irrelevant). Following the initial choice, the modal label was revealed and the opportunity to revise preferences offered, and then the attribute non-attendance (ANA) column was revealed, and respondents were asked to indicate which attributes were processed as irrelevant (i.e., ignored). In this study, two Australian panels were used to obtain a sample, the Pure Profile panel (www.pureprofile.com) and the SSI panel http://www.surveysampling.com/), both of which have many thousands of participants in the chosen study areas. Ethics approval was received for the experiment and each respondent received a small incentive for a completed survey. Data was collected over the period April to June 2014 in all capital cities in Australia. The sample sizes were set to reflect the relative size of the populations of each city. In total, 509 of the surveys were collected through SSI and the balance through Pure Profile.6 The socioeconomic profile and experience in using public transport of the sample together with the city-specific sample sized is summarised in Table 2. We focus on the response to the question ‘‘which investment would you personally prefer?’’.7 In an influential paper, Cameron and DeShazo (submitted for publication) conclude that any biases associated with consumer panels are minimal in respect of a number of tests of representativeness, summarising that ‘‘Overall, our results can probably be characterised as reassuring news for researchers who have used (or who contemplate using) the Knowledge Networks panel for policy-oriented research.’’ These comments by Cameron and DeShazo (submitted for publication) support the growing evidence that a consumer panel can deliver a representative sample if appropriate quota criteria are applied (see Hatton MacDonald et al., 2010; Lindhjem and Navrud, 2011). We report the results for the choice model when the modal label is revealed as well as when accounting for ANA. A comparison of the choice response in the absence and presence of the mode name tells us that close to 90% of the choice tasks (i.e., 1796 choice tasks) assessed by respondents resulted in no revision of the preferred alternative.

6 In a further study we are undertaking an assessment of the results based on each of the two data sources; however preliminary evidence suggests that there are no statistically significant differences. 7 We found that the choice responses were predominantly the same for each of the choice response questions shown in Fig. 1.

34

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

Table 2 Sample sizes by city and descriptive profile of sample. Location

Actual (target)

Sydney Melbourne Canberra Brisbane Adelaide Perth Darwin Hobart Total sample

271 (270) 241 (240) 100 (100) 201 (200) 80 (50) 70 (50) 21 (50) 34 (50) 1018 (1010)

Socioeconomic profile

Mean (std. dev.)

Age (years) Proportion full time employed Proportion part time employed Proportion students Working hours per week Number of adults in household Number of children in household Personal income in 1000$ Number of cars in household Member of PT association (proportion) Member of environment association (proportion)

43.84 (15.5) 0.41 0.19 0.17 20.75 (16.99) 2.11 (0.89) 0.66 (1.03) 62.47 (40.47) 1.66 (0.98) 0.09 0.06

Trip profile

Mean (std. dev.)

Number of times using bus/BRT in the last month Number of times using light rail/tram in the last month Number of times using train/metro in the last month Travel time last bus trip Travel time last train trip

5.85 (9.13) 1.73 (4.35) 6.49 (9.92) 28.62 (18.11) 35.87 (21.93)

4.3. Model results A large number of nonlinear mixed logit models were estimated as we investigated which attributes demonstrated preference heterogeneity through statistically significant random parameters, as well as the functional specification of the heteroscedastic conditioning elements defined by the construction budget, and whether an alternative was labelled BRT or LRT in the preference revision scenarios. We also ran additional models in which we accounted for attribute non-attendance.8 The final models are summarised in Table 3 together with the mean and standard deviation of each attribute. The form of the heteroscedastic conditioning (with fixed parameters) selected is Z jq ¼ ðexpðdj propBcq Þ þ cj M jq Þ where propBcq is the proportion of the government transport infrastructure budget allocated to this project. Regardless of whether we defined the budget as the project construction budget, or the proportion of the overall transport budget in a linear or exponential form, it remained statistically non-significant in the presence of the statistically significant attributes describing the investment, service levels and other features. It should be noted too that a positive sign for the attribute ‘‘travel time (door to door) compared to car (% quicker or slower)’’ is as expected since quicker values are coded as positive (see Appendix C), and thus this identifies that the quicker the public transport system, the more attractive it is to respondents. In both models M1 and M2, the heteroscedastic decomposition of the error variances were not statistically significant, and hence the specification in Eq. (1) is not included, with both final models becoming standard mixed logit models with heteroscedastic conditioning only on the observed effects. The overall statistical fit of the final models is low, despite there being a number of statistically significant attributes; however the likelihood ratio test (LRtest statistic = 16.12) suggests that the HMLEC model without ANA is (marginally) statistically superior to the model with ANA, supported also by the information criterion defined as the AIC measure. When all attributes are treated as fully attended to, on the assumption that this suggests their relevance in processing the alternatives (M1), we find that the modal identifier has a mean parameter estimate of 1.247. The statistical significance suggested by a t-value of 7.7 (given the null is zero for the parameter estimate), indicates that, after controlling for preference heterogeneity on eight statistically significant descriptors9 of each project, the modal labelling is still capturing some of the reasoning in support of LRT over BRT. What is very interesting, however, is that once we allow for attribute non-attendance in the processing of the descriptions of each project,10 noting that ANA across the

8 In addition, we investigated models in which we allowed for scale heterogeneity with fixed parameters (known as Scale MNL), and where we allowed such scale heterogeneity in the presence of preference heterogeneity (referred to as generalised mixed logit – GMX). See Hensher et al. (2015a, Chapter 15) for details of SMNL and GMX model forms. 9 The percent of route dedicated to this system only is not statistically significant but we retained it since it was the best of all the attributes that were excluded. 10 In estimating the ANA model, we revisited the full set of project descriptions, and interestingly none of the attributes excluded from Model M1 were statistically significant under ANA in Model M2.

Table 3 Summary of HMLEC model results. HMLEC with ANA (M2)

ANA incidence

0.0591 (0.79) 1.247 (7.7)

0.1326 (0.86) 0.0018 (0.48)

– –

Random parameters Route length (kms) Construction time (year) Percent of route dedicated to this system only (%) Service capacity in one direction (‘000s passengers/h) Peak service frequency (every x mins) Travel time (door to door) compared to car (% quicker or slower) Risk of being closed down after the assured minimum period (%) Overall environmental friendliness compared to car (%)

0.0231 (2.79) 0.1111 (4.84) 0.0211 (1.25) 0.0184 (2.84) 0.0416 (2.68) 0.0158 (3.43) 0.0034 (2.18) 0.0179 (3.60)

0.0066 (2.29) 0.0442 (5.29) 0.0006 (0.76) 0.0058 (2.64) 0.0107 (1.79) 0.0086 (4.24) 0.0014 (2.14) 0.0080 (4.20)

0.147 0.111 0.111 0.104 0.121 0.124 0.122 0.166

Summary statistics Log-likelihood at zero Log-likelihood at convergence McFadden pseudo-R2 Info. criterion: AIC Sample size

1411.25 1363.93 0.034 1.350 2036

1411.25 1371.99 0.028 1.358 2036

– – – – –

Sample mean and standard deviation Route length (km) Construction time (year) Percent of route dedicated to this system only (%) Service capacity in one direction (‘000s passengers/h) Peak service frequency (every x mins) Travel time (door to door) compared to car (% quicker or slower) Risk of being closed down after the assured minimum period (%) Overall environmental friendliness compared to car (%)

20.2 (8.16) 4.56 (3.52) 63.26 (27.8) 16.56 (10.23) 9.07 (3.98) 10.19 (12.66) 42.68 (36.86) 1.21 (15.4)

17.17 (10.34) 4.07 (3.63) 56.05 (32.9) 14.88 (10.98) 7.98 (4.78) 8.88 (12.29) 37.53 (37.25) 1.13 (14.03)

t-values in brackets. 500 Halton drawsa, with panel structure accommodated in all models. Random parameters are constrained triangular distributions. a We considered Halton draws of 50, 100, 150, 250, 500, 1000 and the parameter estimates were very similar over the range 150–1000.

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

HMLEC without ANA (M1) Non-random parameters Bcq = Budget constraint imposed on project (as proportion of total annual transport budget) Mjq = Modal identification (BRT = 1, LRT = 0)

35

36

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

eight attributes ranges from 10.4% to 16.6% of the sample (see last column of Table 3), the modal labelling is no longer statistically significant (t-value = 0.48) suggesting some amount of modal misrepresentation may be present when it is assumed that respondents act under a fully compensatory processing rule as in M1. If we stay with M1, we would conclude that there is a modal bias in favour of LRT when we assume that all attributes are relevant to some degree, but not so under the ANA model, M2. One possible interpretation is that when we recognise what really matters to influence each respondent’s preferences for public transport investments, the modal image expressed through the name is, on average, of no consequence. This is a potentially very important finding since almost without exception, all previous studies that contrast BRT with LRT have assumed full attribute relevance and a fully compensatory preference rule. Hence we might suggest from this evidence that once what really matters (differentially) to each individual is narrowed down, the LRT–BRT distinction blurs into a domain of non-relevance. Given that less than 10% of the sample revised their preferences when the modal label was revealed, this is an encouraging and potentially crucial piece of evidence. Table 3 identifies the relevant drivers of preferences as being intuitively plausible. Specifically, when we look at the full set of 20 candidate drivers in Fig. 1, most if not all of the statistically non-significant attributes might be argued to be of lesser relevance in a particular jurisdiction (e.g., fare, attracting business around stations/stops, and off peak service frequency). Spatial coverage, as proxied by the population serviced, was surprisingly not significant although it may be that service capacity in one direction may be a proxy for this. The statistical significance of construction time, directional route length, service capacity, peak period frequency,11 travel time compared to car, future closure risk, and environmental friendliness are encouraging and plausible drivers of preferences for public transport options. The statistical insignificance of the right of way, represented by the percent of a route dedicated to either BRT or LRT, is a surprising finding, given that much of the literature identifies the role of mixed traffic in causing delay and reliability, with the result that dedicated pathways are a key strength of BRT in particular although also a strength of LRT despite the growing number of LRT systems that operate in mixed traffic. In summary, the empirical evidence suggests that there are some statistically significant drivers of preferences for BRT and LRT; however the overall goodness of fit is small, and hence there are likely to be other almost idiosyncratic forces at play that dominate the preferences of the community. We investigated interacting each of the attributes (including the two heteroscedastic conditioning factors) with the socioeconomic characteristics of the sample, but did not find any statistically significant influences. Similarly, city-specific dummy variables were also found to play no role (see Appendix B for additional analysis). Given the random parameter distributions (using conditional estimates), the relative preference for LRT over BRT is shown in Fig. 2 for the sample. Using M1 evidence (since the modal distinction is not significant in M2), the utility associated with LRT is on average more than twice that of BRT (0.30 cf. 0.12) with standard deviations of 0.27 and 0.25 respectively for LRT and BRT. The LRT distribution is approximately normal (chi-squared normality test is 0.1554); in contrast the BRT distribution is far from normal (test = 17.2), although the range is very similar. Another profile of interest is the comparison between the preferences of the sample in the presence (UTOTL) and absence (UTOTU) of knowledge of the specific mode associated with the offered set of attributes in the choice experiment (see Fig. 3). Fig. 3 related to M1, is very informative. In the absence of knowing whether the project is a BRT or LRT investment, the mean utility is negative (0.503); however when the mode is revealed, the mean becomes positive (0.262),12 suggesting that knowing the modal modifies preferences for a bundle of infrastructure attributes and results in a higher level of positive utility. This would be expected if there are biases existing in the community for specific public transport modes. Clearly, the BRT vs. LRT mode-specific effect is an important influence under the fully compensatory preference rule as confirmed in Table 3 (M1), although this is not the case under ANA (M2). All of the above evidence from the formal modelling supports the position of a strong net preference for LRT over BRT under a fully compensatory preference rule but not under the semi-compensatory ANA rule, even in the presence of a budget constraint. It appears that the budget constraint is of little interest to the community even if such a constraint has implications for how much of the new infrastructure can be built. This evidence reinforces a view that it is an ongoing challenge to identify ways to inform the population at large about the relative merits of bus-based public transport infrastructure solutions when budgets are constrained. Marginal rates of substitution (MRS) between the statistically significant attributes in M1 and M2 are summarised in Table 4. We have chosen to present the MRSs for each attribute as a trade against the route length, also defined as a willingness to pay estimate. Although many more trade-offs can be obtained from the preference model (in Table 3), after identifying the full distribution of each random parameter, the trade-off between the route length and each of the other attributes is of particular interest since it reveals how much of the route length an individual is willing to forego in order to obtain a gain in another attribute and this reaches to the core difference between LRT and BRT for a given level of budget.

11 We note that it has a t-value of 1.79 for the ANA model, which is statistically non-significant at the 95% level of confidence, but all other attributes retain the statistical significance they exhibit in Model 1. 12 The range (respectively 2.67 to 0.93 under the unlabelled scenario and 1.103 to 2.12 under the labelled scenario), as well as the standard deviation (respectively 0.523 and 0.538), of the utility estimates are almost identical, albeit with a higher degree of skewness under the labelled scenario (1.188) compared to the unlabelled scenario (0.349). The chi-squared normality test shows clearly that deviation from normal is greater for the labelled distribution (17.99) compared to the unlabelled distribution (1.839).

37

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

2.83 ULRT

UBRT

Density

2.26 1.70 1.13 .57 .00 -.500

-.250

.000

.250

.500

.750

1.000

1.250

Fig. 2. Profile of utility associated with BRT (UBRT) and LRT (ULRT).

1.30 UTOTU

UTOTL

Density

1.04

.78

.52

.26

.00

-3

-2

-1

0

1

2

3

Fig. 3. Utility distributions in presence (labelled) and absence (unlabelled) of modal naming.

Table 4 Summary of selected willingness to pay estimates for Models 1 and 2. Standard deviation of MRS in parentheses. Variables

M1 (w/o ANA) BRT

Route length (km) Construction time (year) Percent of route dedicated to this system only (%) Service capacity in one direction (‘000s passengers/h) Peak service frequency (every x mins) Travel time (door to door) compared to car (% quicker or slower) Risk of being closed down after the assured minimum period (%) Overall environmental friendliness compared to car (%)

– 6.93 0.09 0.72 2.43 0.62 0.22 0.71

M2 (w/ANA) LRT

(4.59) (0.08)* (0.71) (1.70) (0.66) (0.15) (0.76)

– 5.55 0.11 0.89 1.95 0.78 0.17 0.88

Sample mean (std. dev.)

BRT = LRT (5.27) (0.11)* (0.89) (2.11) (0.83) (0.18) (0.95)

– 6.63 (6.58) 0.005 (0.004)* 0.90 (0.88) 1.47 (1.45) 1.26 (1.24) 0.20 (0.19) 1.12 (1.10)

20.0 (8.2) 4.48 (3.5) 62.50 (27.8) 16.65 (10.2) 9.16 (3.9) 14.97 (6.1) 43.77 (37.0) 11.85 (9.9)

Note: MRS (x, y) = |ux/uy| = y/x, where x = route length, and u = marginal utility. * Not significant at the 95% level of confidence.

Model M1 has marginal rates of substitution that vary by BRT and LRT, whereas M2 did not have a statistically significant modal label and hence the estimates are generic across BRT and LRT. In M1, the MRS between construction duration and route length for BRT is 6.93 and for LRT it is 5.55. That is, on average a sampled respondent is willing to give up 5.55 km of LRT route length but 6.93 km of BRT route length to save one year of construction time. This indicates an appreciation of the way LRT infrastructure takes longer to implement with the associated disruption. This effect is non-marginal given the average route length for the experiments is 20 km and the sample average construction time is 4.48 years, and so the interpretation is that a sampled respondent is willing to give up, on average, between a quarter and a third of the route length to save approximately 20% of the construction time. For service capacity in one direction, on average a sampled respondent is willing to give up 0.89 LRT route km in return for an increase in LRT service capacity of 1000 passengers per hour; whereas they are only willing to

38

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

give up 0.72 BRT route km for the same increase in BRT service capacity; however this effect is relatively small in the context of the sample means. For peak service frequency, on average a sampled respondent is willing to give up 1.95 LRT route km to gain a minute reduction in peak service frequency, but would be willing to give up 2.43 BRT route km to gain the same minute reduction. As with construction time above, this is also a sizeable impact given the peak service frequency mean of 9.16 min, highlighting how important peak service frequency appears to be for respondents. For travel time compared to the car time, on average a sampled respondent is willing to give up 0.78 LRT route km for a 1% quicker door-to-door travel time relative to car, but would be willing to give up only 0.62 km of BRT for a 1% quicker travel time relative to car. For the risk of being closed down after the assured minimum period, on average a sampled respondent is willing to give up 0.17 LRT route km to gain a 1% improvement in the risk of being closed down after the assured minimum period, but would be willing to give up 0.22 BRT route km to gain a 1% improvement in the risk of being closed down after the assured minimum period. Finally, for overall environmental friendliness compared to the car, on average a sampled respondent is willing to give up 0.88 LRT route km to obtain a 1% gain in overall environmental friendliness compared to the car, but would only be willing to give up 0.71 BRT route km to obtain a 1% gain in overall environmental friendliness compared to the car. These last three sets of MRSs are non-trivial, but are not as substantial as the MRS for construction time and peak frequencies. When we account for attribute non-attendance (M2), the MRS evidence shows some differences from the fully compensatory decision rule Model M1. The main differences in MRS between M1 and M2 are for peak service frequency, door-todoor travel time compared to car (% quicker or slower), and overall environmental friendliness compared to car, with the latter two attributes having a much higher MRS under ANA. That is, under M2, the results suggest that on average a sampled respondent is willing to give up 1.28 route kms for a 1% gain in door-to-door travel time relative to car. This is compared to 0.78 route kms of LRT and 0.62 route kms of BRT under M1. Similarly, the MRS suggests that, on average, a sampled respondent is willing to give up 1.12 route kms to obtain a 1% gain in overall environmental friendliness compared to the car, which is higher than 0.71 for BRT and 0.88 for LRT under the fully compensatory decision rule discussed above under M1. These differences in MRS derived from the two models are statistically significant with t-values greater than 1.96 (see Table 5). Hence we can conclude that attribute non-attendance does impact on the findings at the 95% level of confidence. This evidence reinforces the need to recognise not only the choices made (i.e., preference ranking) but also the attribute processing strategy adopted in establishing which attributes matter to each sampled individual. The MRS associated with risk of being closed down after the assured minimum period (%) and service capacity in one direction, remain almost unchanged under both decision rules. In accounting for ANA, we are able to elevate the relative importance of public transport travel time (door to door) compared to car (% quicker or slower) and overall environmental friendliness of public transport compared to the car, and discount peak service frequency. This has clear implications on priorities in marketing the attractiveness of LRT and BRT. We obtained a series of elasticities to determine which attributes have the greatest impact on changing preferences. These are summarised in Table 6 for each model as probability weighted averages using full sample enumeration. It should be noted that the BRT and LRT estimates are very similar. This is a consequence of generic parameter estimates and the way in which

Table 5 Statistical significance of differences in marginal rates of substitution (in Table 4) between Models 1 and 2 (H0: MRSM1 = MRSM2). Attributes

BRT

LRT

Construction time Percent of route dedicated to this system only Service capacity in one direction Peak service frequency Travel time (door to door) compared to car Risk of being closed down after the assured minimum period Overall environmental friendliness compared to car

1.66 42.60 7.27 19.40 20.59 2.29 13.99

5.56 43.10 0.17 8.41 14.76 5.08 7.45

Table 6 Summary of direct elasticities for each model. Model 1

Route length (km) Construction time (year) Percent of route dedicated to this system only (%) Service capacity in one direction (‘000s passengers/h) Peak service frequency (every x mins) Travel time (door to door) compared to car (% quicker or slower) Risk of being closed down after the assured minimum period (%) Overall environmental friendliness compared to car (%)

Model 2 (ANA)

BRT

LRT

BRT

LRT

0.038 0.058 0.018 0.028 0.030 0.023 0.019 0.004

0.039 0.066 0.017 0.026 0.032 0.025 0.022 0.009

0.212 1.350 0.077 0.150 0.331 0.263 0.031 0.333

0.230 1.340 0.009 0.158 0.338 0.270 0.028 0.351

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

39

the design of the experiment, so as to answer the research question, gave the same set of attribute levels to each mode, albeit assigned differently in the design for each choice scenario for each respondent. Model 2 results indicate the influence on an attribute being attended to or not, and shows not only numerically different mean direct elasticities but also sign changes for all attributes except environmental friendliness. The sign change reflects the fact that the range, and hence the percentage of estimates in each sign domain associated with the distribution of a random parameter, changes when ANA is taken into account. For example, when it is assumed that route length is relevant for all respondents, the positive mean estimate suggests that a 10% increase in route length (i.e., 2.02 kms, see Table 3) of BRT increases the probability of this system being chosen by 0.38%. In contrast, when ANA is taken into account, the same percentage increase in route length of BRT (1.72 kms) decreases its probability of being chosen by 2.1%. The largest behavioural response (and the only one that is relatively elastic) is the construction time (with a mean of 4.07 years in Table 3) under ANA, which suggests that a 10% increase in construction time, ceteris paribus, increases the probability of choosing a package of attributes by 13.5%. This suggests that when construction time matters, there might be a sense of getting a better outcome if it takes longer to deliver, although the extra time on average is only 5 months (i.e., 10% ⁄ 4.07 ⁄ 12 = 5 months). A close examination of all attributes suggests that despite the statistical significance of the attributes, the values of their associated direct elasticities are generally very inelastic.

5. Conclusions and synthesis We know that many urban areas are keen to provide new public transport infrastructure for their citizens as a way of providing higher quality services which could tempt car drivers out of their cars and contribute to a more sustainable future. This is in line with the literature which identifies good quality of service (broadly defined) as the most important attribute in achieving mode switch from car to public transport. Recent developments throughout the developed world have concentrated, with rare exceptions, on the provision of this new infrastructure in the form of LRT, in contrast to the predominant implementation of BRT in the developing world, despite the way in which developing economies have significant budgetary constraints for transport with many other important calls on their overall budgets. With this as background, this paper sets out to investigate the apparent belief that rail based modes and bus based modes are substitutes for a particular solution or to take account of the way in which community debate behaves as though there is no budget constraint. The paper sought to see whether society would temper its preference for a rail based solution to a bus based solution if a budget constraint meant that the former was limited and the latter could contribute more significantly to a network development because of its lower cost. A discrete choice experiment was conducted that included important potential drivers of community preferences for improved public transport where a budget constraint in the form of construction cost conditioned the options of choice for the respondents. The experiment was initially unlabelled but followed by a labelling of each alternative as either BRT or LRT. The analysis and results confirm the existing literature in terms of a number of key attributes which drive the choice of a public transport mode, with the marginal rates of substitution (willingness to pay) results highlighting the importance of construction time, perhaps also being related to the disruption associated with building, and the relevance of peak frequencies in relation to route length of service. Whilst the literature identifies perceptions as being very important, we did expect the budget constraint and the unlabelled nature of the experiment to show more indifference between alternatives emulating BRT and LRT given that these two modes can offer very similar quality of service attributes. However, despite the number of statistically significant attributes, there is still a limited explanation of the overall preferences of the sampled respondents. It would seem that, however appealing the quality of service attributes, LRT is still preferred to BRT when the common preference assessment approach is adopted that assumes all attributes are relevant to all individuals (the fully compensatory decision rule). However, when a semi-compensatory decision rule that accounts for attribute non-attendance is invoked, LRT is no longer unambiguously preferred to BRT. This is a significant finding and suggests a pathway to the way that public transport options need to be marketed. Surprisingly, under both decisions rules, societal preferences do not seem to be affected by the quantum or existence of the budget constraint and also appear immune to the way in which a lower cost bus based solution can provide equivalent quality of service attributes and greater geographical coverage. The literature also suggests that the ‘image’ of bus may have been tainted by its association of the bus in mixed traffic which is slow and unreliable. It is also possible that respondents’ experience of this sort of bus might be contaminating their responses to BRT. We do not feel that this is a particular concern otherwise we would have identified differences in the respondents from Brisbane (Queensland) where BRT has been successfully implemented as the backbone of the public transport network. Using the traditional way in which preferences are revealed would encourage us to conclude that in a developed country context, LRT is preferred over BRT despite budgetary constraints and similarities in quality of service attributes. We would be concluding that maybe it is choice vs. blind commitment (Hensher, 1999) that wins out or that there is more to the anecdotal evidence which continues to reinforce the idea that LRT is more ‘comfortable’ than any bus mode could be. We would also need to point to the observed success of BRT in developed country settings where BRT has been a choice in the context of providing greater network coverage in the face of budgetary constraints, or to Brisbane, Australia, where there is an extensive BRT network and where this experiment was also conducted. What appeared to make a difference in these cases was the existence of a champion who drove through an implementation package that turned out to be successful as in Brisbane, Curitiba (Brasil) and Bogota (Columbia).

40

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

However, using the more realistic semi-compensatory decision rule, we can find more understanding of the observed success of BRT in developed economies. This relates to the results from assuming non-homogeneity in respondent processing, namely that once the subset of attributes that are the real drivers of each individual’s preferences (which vary across the population) are identified, the modal bias is no longer a significant driver of preferences. This tempers our conclusion to identify the challenge as being to find ways to ensure that these dominating drivers of preferences are properly incorporated in studies that evaluate public transport modal options, such that superfluous or non-relevant attributes are not confounding the evaluation in ways that introduce modal bias (especially in favour of LRT). This is the challenge for ongoing research. Another challenge is how we break the emotional ideological cycle of a commitment to rail regardless of the benefits of bus-based systems. Schwanen et al. (2012) offer some very useful suggestions on breaking the habit cycle and specifically the way in which durable behaviour change is unlikely to result from individual behaviour change. Behaviour change is argued to come from a very broad view of stakeholder influence in changing views. This is aligned with the growing literature on nudging (behavioural insights) in which a wider community of stakeholders can be launched to effect change. In the context we are looking at in this paper of preferences for modes under a constrained budget, the stakeholders go well beyond ‘users’ of transport systems – transport industry, retailers and advertising companies, the media, lobby groups, employers and non-transport government agencies and departments – whose practices contribute to the collective sense-making and normative coding of different forms of mobility. ‘‘The informational, financial, symbolic or affective signals regarding which travel practices are desirable from across government agencies to citizens and other stakeholders should be made more coherent for extended periods of time.’’ (Schwanen et al., 2012, page 530). Finally, it is possible that the softer variables are perceived differently in different modes and in future research, we should continue to explore other possible influences on LRT vs. BRT such as the comfort of seats, Wi-Fi availability (ability to be ‘productive’ on journey), status, fit with streetscape, neighbourhood design, and the streetscape to see or to rule out these aspects as influencing modal preferences. Acknowledgments This paper is a contribution to the research program of the Volvo Research and Education Foundation Bus Rapid Transit Centre of Excellence. We acknowledge the Foundation for funding support and discussions with Juan Carlos Muñoz as part of our ongoing collaborative research program. We are especially thankful to Jun Zhang for his contribution in programming the survey instrument. Authors contributed equally and are placed alphabetically. The suggestions by three referees have materially improved the paper. Appendix A. Attributes investigated in Phase 1 to identify those that are important candidates for the choice experiments Source Hensher et al. (2015b) (see Table A1).

Table A1

Public transport voting preferences

Public transport service levels

Public transport design levels

1

Systems with comfortable vehicles

2

Smart vehicles

3

Quick journey times

Travelling by bus is safer than travelling by light rail (tram) Bus travel times in a bus lane or dedicated corridor are faster than light rail (tram) Crowded buses are less horrible to travel in than crowded light rail (trams)

4

Some corridors with good service levels, even if other corridors had less good service levels New rail links, even if these are shorter than a package of investments with good bus-based services Value for money for the taxpayer

Buses in a bus lane or dedicated corridor are more reliable than light rail (trams)

There are less light rail (tram) stops than bus stations so people have to walk further to catch light rail Bus systems provide better network coverage than light rail (tram) systems A new bus route in a bus lane or dedicated corridor can bring more life to the city than a new light rail (tram) line A bus service in a bus lane or dedicated corridor looks faster than a light rail (tram) service

The greatest length of high quality corridors, irrespective of whether train, tram or bus A network that is cost effective to operate

A bus journey in a bus lane or dedicated corridor is more comfortable for passengers than a light rail (tram) journey Buses are more modern looking than light rail (trams) and hence have more appeal in urban

5

6

7

8

Buses look cleaner than light rail (trams)

Buses are cleaner than light rail (trams)

Bus routes are fixed, so bus stops provide more opportunity for new housing than a light rail (tram) line which can be changed very easily New bus stops or a new bus route in a bus lane or dedicated corridor will improve surrounding properties more than new light rail (tram) stops Buses in a bus lane or dedicated corridor are more environmentally friendly than light rail (trams) More jobs will be created surrounding a bus route in a bus lane or dedicated corridor than a light rail

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46 Table A1 (continued)

9

10 11 12

13

14

15 16

17

18

Public transport voting preferences

Public transport service levels

Public transport design levels

Low fares

settings Bus journeys require less transfers than light rail (tram) journeys

(tram) route A bus service in a bus lane or dedicated corridor is more likely than a light rail (tram) to still be in use in 30 years time Bus services stop nearer to more people than light rail (trams) services Bus services are less polluting than light rail (trams)

Higher fares to pay for higher quality services Frequent services Fast overall journey time to destination, including getting to and from the station or stop A network with few interchanges

Interchanges between services and modes (bus, train, ferry) if this makes overall journey times quicker The package that is quickest to implement Slow implementation is not a problem if the package delivers the right public transport system High quality bus routes on dedicated roads (so that they do not suffer from delays from cars) Systems that give wide network coverage

Buses have cleaner seats than light rail (trams) Buses are cleaner on the outside than light rail (trams) Bus stops are cleaner than light rail (tram) stops Bus services in a bus lane or dedicated corridor are more frequent than light rail (tram) services Bus stops are safer than light rail (tram) stops

Bus services in a bus lane or dedicated corridor do not get delayed like light rail (tram) services Buses provide a better comfort level than light rail (tram) services Buses provide easier boarding than light rail (trams)

19

Packages which offer good safety for the passenger

20

Packages which give an outcome that will last for many years

Car drivers are more likely to transfer to bus services in a bus lane or dedicated corridor than to light rail (tram) services Buses in a bus lane or dedicated corridor provide a better quality of service than light rail (trams) Buses provide better personal security for travellers than light rail (trams)

21

Bus based systems of public transport

Buses are sexy and light rail (trams) are boring

22

Easy to use fare system

A public transport network with bus rapid transit (BRT) will provide a greater network coverage than one with light rail (trams)

23

Investment package most likely to benefit your city

24

The package of investments most likely to benefit you

25

The package of investments most likely to get car drivers out of their car and onto public transport The package of investments least likely to increase taxes The package of investments giving the highest capacity for travellers The package of investments which allows the city to grow sustainably The package of investments which allows housing to be built around stations.

26 27 28 29

30

31

32

33

Bus services are more likely to have level boarding (no steps up or down to get on the vehicle) than light rail (trams) Buses are quieter than light rail (trams)

Bus services in a bus lane or dedicated corridor services have been more successful for cities than light rail (trams) Buses in a bus lane or dedicated corridor are more permanent than light rail (trams) Buses in a bus lane or dedicated corridor provide more opportunities for land redevelopment than light rail (trams) Buses in a bus lane or dedicated corridor provide more focussed development opportunities than light rail (trams) Buses in a bus lane or dedicated corridor are more likely to be funded with private investment than light rail (trams) Buses in a bus lane or dedicated corridor support higher population and employment growth than light rail (trams) Building bus lane or a dedicated roads and buying buses makes a bus system cheaper than putting down rails and buying light rail (trams) Bus services provided in a bus lane or dedicated corridor have lower operating costs than light rail (tram) systems Bus services provided in a bus lane or dedicated corridor have lower operating costs per person carried than light rail (tram) systems Building a new bus route in a bus lane or dedicated corridor will cause less disruption to roads in the area than a new light rail (tram) line Overall, buses in a bus lane or dedicated corridor have lower maintenance costs than light rail (trams) and light rail (tram) track Bus stops have greater visibility for passengers than light rail (tram) stops Buses in a bus lane or dedicated corridor have lower accident rates than light rail (trams) Buses in a bus lane or dedicated corridor provide a more liveable environment than light rail (trams) Buses in a bus lane or dedicated corridor have greater long term sustainability than light rail (trams) Buses provide more comfort for travellers than light rail (trams) Bus systems are quicker to build and put in operation than light rail (tram) services in a light rail (tram) lane or dedicated corridor The long term benefits of a new bus route in a bus lane or dedicated corridor are higher than a new light rail (tram) line House prices will rise faster around new bus associated with a bus lane or dedicated corridor stops than light rail (tram) stops Buses in a bus lane or dedicated corridor provide better value for money to taxpayers than light rail (trams)

41

42

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

Table B1 Summary of binary logit results. Constant Gender (1 = female) Number of cars in household Number of times using bus/BRT in the last month Number of times using light rail/tram in the last month Interaction of most recent trip train travel time (mins) and getting a seat (1) Interaction of most recent trip bus travel time (mins) and getting a seat (1) Personal income ($000s) Age (years) Work full time (1, 0) Work part time (1, 0) Sydney dummy (1, 0) Melbourne dummy (1, 0) Canberra dummy (1, 0) Adelaide dummy (1, 0) Perth dummy (1, 0) Brisbane dummy (1, 0) Darwin dummy (1, 0)

0.8678 (3.38) 0.0392 (0.51) 0.0103 (0.27) 0.0002 (0.05) 0.0003 (0.32) 0.000037 (0.32) 0.000067 (0.62) 0.0006 (0.56) 0.0027 (0.67) 0.0672 (0.67) 0.0553 (0.56) 0.1047 (0.49) 0.0679 (0.32) 0.0960 (0.43) 0.0661 (0.28) 0.0690 (0.29) 0.0758 (0.36) 0.0369 (0.12)

Log-likelihood at zero Log-likelihood at convergence McFadden pseudo-R2 Info. criterion: AIC Sample size

2319.02 2314.16 0.002 1.145 2036

t-values in brackets. Dependent variable is BRT (1) vs. LRT (0).

Table B2 Summary of binary logit results for no change in preferences. Constant Gender (1 = female) Member of a public transport association (1, 0) Interaction of train travel time (mins) and getting a seat (1) Age (years) Work full time (1, 0) Melbourne dummy (1, 0)

1.4150 (8.12) 0.3470 (3.85) 0.8143 (6.37) 0.00019 (2.05) 0.0063 (2.08) 0.1712 (1.88) 0.3834 (3.91)

Log-likelihood at zero Log-likelihood at convergence McFadden pseudo-R2 Info. criterion: AIC Sample size

1781.15 1721.88 0.033 0.849 2036

t-values in brackets. Dependent variable is no change (1) vs. change (0).

Appendix B. Assessing the influence of socioeconomic characteristics, experience in using public transport, and location on choice between BRT and LRT We were unable to find any statistically significant socioeconomic characteristics, public transport experience effects and city location dummy variables when interacting them with the choice experiment attributes. To investigate further we ran some simple logit models in which the choice variable was BRT (1) vs. LRT (0) (denoted BRTCH). We were unable to find any statistically significant influences as shown in Table B1 which is one of the numerous models performed. The partial correlation coefficient matrix in Table B3 reaffirms this. We also ran a logit model in which the dependent variable was defined as no change (=1) in preference when the modal name was introduced vs. change (0) 84.14% of the sample stayed with their initial preference. The results are summarised in Table B2. The overall fit is still very small but there are six variables that are statistically significant influences.13 Individuals who are female and older tend to stay with their preferences when the modal name is revealed, but full time employees tend to change as do individuals whose most recent trip train travel time (mins) interacted with getting a seat increases; members of a public transport association tend to change their modal preference (possibly because they belong to a rail association). Finally, Melbourne residents tend to make fewer changes than residents of the other capital cities, possibly due to the presence of relatively good train and tram services throughout much of the metropolitan area. 13

We also investigated the role of the attributes in the choice experiment but not one was statistically significant.

43

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46 Table B3 Correlation matrix on Table A1 variables. Cor. mat.

BRTCH

NOCHANGE

NCHGBRT

NCHBRTSW

NCHLRTSW

WOMAN

VEHICLES

PTASS

BRTCH NOCHANGE NCHGBRT NCHBRTSW NCHLRTSW WOMAN VEHICLES PTASS

1.00000 .02419 .50273 .16815 .20036 .00365 .01807 .02228

.02419 1.00000 .41189 .66566 .66566 .04637 .00107 .13063

.50273 .41189 1.00000 .27418 .27418 .01910 .00044 .05381

.16815 .66566 .27418 1.00000 .11378 .03087 .00071 .08696

.20036 .66566 .27418 .11378 1.00000 .03087 .00071 .08696

.00365 .04637 .01910 .03087 .03087 1.00000 .08608 .07744

.01807 .00107 .00044 .00071 .00071 .08608 1.00000 .11936

.02228 .13063 .05381 .08696 .08696 .07744 .11936 1.00000

BRTCH

NOCHANGE

NCHGBRT

NCHBRTSW

NCHLRTSW

WOMAN

VEHICLES

PTASS

.00009 .05917 .11109 .01193 .05087 .01720 .07730 .03297

.00004 .02437 .04576 .00491 .02095 .00709 .03184 .01358

.00006 .03939 .07395 .00794 .03386 .01145 .05146 .02195

.00006 .03939 .07395 .00794 .03386 .01145 .05146 .02195

.05239 .01514 .05749 .03791 .22345 .18113 .14628 .12921

.04853 .03997 .00642 .07046 .18209 .12088 .09789 .06442

.03549 .20593 .04075 .00053 .20399 .21742 .12713 .00674

USEDBUS USEDRAIL TNTTST BSTTST PINCOME AGE FT PT

USEDBUS USEDRAIL TNTTST BSTTST PINCOME AGE FT PT

SYD MEL CAN ADL PER BRS DWN

SYD MEL CAN ADL PER BRS DWN

SYD MEL CAN ADL PER BRS DWN

.00235 .01218 .01964 .01036 .00520 .00694 .02894 .01525 USEDBUS

USEDRAIL

TNTTST

BSTTST

PINCOME

AGE

FT

PT

1.00000 .08281 .01668 .00070 .00691 .11236 .11382 .01937

.08281 1.00000 .19116 .14982 .01232 .03661 .05000 .02824

.01668 .19116 1.00000 .24272 .04832 .10374 .09988 .04295

.00070 .14982 .24272 1.00000 .08262 .03926 .13038 .07990

.00691 .01232 .04832 .08262 1.00000 .02629 .61062 .22580

.11236 .03661 .10374 .03926 .02629 1.00000 .14244 .03681

.11382 .05000 .09988 .13038 .61062 .14244 1.00000 .44557

.01937 .02824 .04295 .07990 .22580 .03681 .44557 1.00000

BRTCH

NOCHANGE

NCHGBRT

NCHBRTSW

NCHLRTSW

WOMAN

VEHICLES

PTASS

.00530 .00749 .03238 .02790 .00673 .00795 .03562

.02743 .03659 .00546 .04661 .02749 .03705 .03210

.01130 .01507 .00225 .01920 .01132 .01526 .01322

.01826 .02436 .00363 .03103 .01830 .02466 .02137

.01826 .02436 .00363 .03103 .01830 .02466 .02137

.03047 .02425 .10824 .04947 .04931 .00373 .04192

.02783 .06320 .10970 .00446 .04504 .11971 .01913

.07106 .09569 .04690 .02527 .01688 .02461 .02393

USEDBUS

USEDRAIL

TNTTST

BSTTST

PINCOME

AGE

FT

PT

.03402 .04240 .06916 .04430 .11325 .10163 .07570

.21572 .50982 .04782 .03579 .12439 .18917 .03729

.04814 .06232 .05318 .07809 .12639 .15042 .06075

.04391 .06407 .01259 .02449 .13420 .08106 .01235

.09457 .02846 .02302 .07140 .00352 .02294 .01798

.02961 .10191 .10113 .04462 .02069 .06097 .08336

.04480 .01054 .00000 .00427 .01770 .04467 .01931

.01855 .08226 .03607 .08550 .02945 .07421 .03059

SYD

MEL

CAN

ADL

PER

BRS

DWN

1.00000 .45337 .15578 .16712 .20762 .33352 .07077

.45337 1.00000 .14196 .15230 .18920 .30392 .06449

.15578 .14196 1.00000 .05233 .06501 .10443 .02216

.16712 .15230 .05233 1.00000 .06974 .11203 .02377

.20762 .18920 .06501 .06974 1.00000 .13918 .02953

.33352 .30392 .10443 .11203 .13918 1.00000 .04744

.07077 .06449 .02216 .02377 .02953 .04744 1.00000

Appendix C. Design considerations and attributes The survey was designed using Ngene (Choice Metrics, 2012) which has an attribute level balance property. Survey blocks were designed and assigned to each respondent in a real-time manner to maximise the balance of the blocks. The final data set has a very good balance of 12 blocks with frequencies ranging from 84 to 87 for each block. The structure of the survey showing the attribute level balance is shown in Table C1. The only attribute that does not exhibit the attribute level balance is the frequency of services during peak and off-peak hours as the experiment is designed with the condition that off-peak frequency is less prevalent than peak frequency, and thus a D-efficient design with the complete balance for this attribute might not exist. This is mirrored in the correlation matrix of stated choice design attributes shown in Table C2.

44

Table C1 Structure of the design of the choice experiment. Task Block Length cost time pop roway opcost capa pfreq ofreq tcar fare prepaid tick wait staff board yearop close buss shiftcar env brt km $000m years % % $m ’000 mins mins % % Yes = 1 Yes = 1 mins Present|absent Level|steps years % Low|med|high % % BRT = (slower|quicker+) (lower|higher+) (worse|better+) 1|LRT 10 10 30 20 20 20 10 10 10 30 30 20 20 30 20 10 20 30 30 20 10 10 30 30

6 1 1 6 6 0.5 1 6 0.5 0.5 1 3 3 6 0.5 1 3 0.5 3 0.5 3 3 1 6

2 2 1 10 2 1 10 5 1 10 5 1 2 5 10 2 10 10 2 5 5 1 5 1

20 5 10 15 20 15 15 10 15 10 10 20 15 5 10 5 20 10 5 20 20 5 15 5

100 100 25 100 25 25 75 25 100 75 50 75 75 50 25 75 75 50 50 50 25 100 100 50

5 5 10 10 2 10 5 15 2 15 10 15 5 2 5 2 2 15 2 10 5 10 15 15

5 30 15 30 5 30 30 30 15 5 5 15 30 5 5 15 15 15 30 5 30 5 15 15

5 15 15 5 5 10 10 10 5 15 5 15 10 10 5 15 15 5 5 10 10 10 5 5

5 15 20 10 15 10 15 20 15 15 10 15 20 15 5 20 20 20 5 20 10 10 20 10

15 25 10 15 15 10 10 25 10 15 10 10 25 10 25 10 10 25 10 15 15 25 10 10

20 20 10 20 10 10 10 20 10 20 20 20 10 20 20 10 10 10 10 20 10 10 20 20

1 0 1 0 0 0 1 1 1 0 1 0 1 1 1 0 0 0 0 0 1 1 1 0

1 0 1 1 0 1 0 1 0 0 0 0 0 0 1 1 1 0 1 1 0 1 1 0

10 15 10 1 10 1 15 10 1 5 10 5 1 1 15 10 15 5 1 5 5 5 15 15

1 1 0 0 0 0 0 1 0 0 1 1 0 1 0 1 0 1 1 1 1 0 1 0

1 0 0 1 0 1 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 1

40 60 50 40 20 50 20 30 50 20 40 10 10 60 20 10 50 30 10 60 40 30 30 60

0 25 0 25 50 50 0 100 100 25 100 50 0 0 25 100 100 50 25 0 100 50 50 25

1 0 1 1 1 2 2 1 1 1 0 0 0 2 0 2 0 1 0 0 2 2 2 2

10 10 10 20 10 5 0 0 5 20 10 0 20 5 5 20 0 0 5 5 20 0 20 10

25 0 25 0 5 25 5 25 10 5 25 10 5 25 5 10 0 25 5 0 25 5 5 10

0 0 0 1 0 0 1 1 1 1 0 1 0 1 1 1 0 0 0 1 1 0 0 1

System B 1 1 2 1 3 2 4 2 5 3 6 3 7 4 8 4 9 5 10 5 11 6 12 6 13 7 14 7 15 8 16 8 17 9 18 9 19 10 20 10 21 11 22 11 23 12 24 12

10 10 30 20 20 20 10 10 10 30 30 20 20 30 20 10 20 30 30 20 10 10 30 30

1 3 3 0.5 0.5 6 3 0.5 6 6 3 0.5 1 0.5 6 3 1 6 1 6 1 1 3 0.5

5 5 10 1 5 10 1 2 10 1 2 10 5 2 1 5 1 1 5 2 2 10 2 10

5 20 15 10 5 10 10 15 10 15 15 5 10 20 15 20 5 15 20 5 5 20 10 20

25 25 100 25 100 100 50 100 25 50 75 50 50 75 100 50 50 75 75 75 100 25 25 75

10 10 5 5 15 5 10 2 15 2 5 2 10 15 10 15 15 2 15 5 10 5 2 2

30 5 15 5 30 5 5 5 15 30 15 30 5 30 30 15 15 15 5 30 5 30 15 15

10 5 5 15 15 5 10 5 15 5 10 5 5 5 15 5 5 10 15 10 10 10 10 15

20 15 5 20 15 20 15 10 15 20 15 15 5 10 20 5 10 10 20 10 20 15 10 15

10 10 25 10 10 15 25 10 15 10 25 25 10 15 10 15 15 10 25 10 10 10 25 15

20 20 10 20 10 10 10 20 10 20 20 20 10 20 20 10 10 10 10 20 10 10 20 20

0 1 0 1 1 1 0 0 0 1 0 1 0 0 0 1 1 1 1 1 0 0 0 1

0 1 0 0 1 0 1 0 1 1 1 1 1 1 0 0 0 1 0 0 1 0 0 1

5 1 5 10 5 15 1 5 15 10 5 10 15 15 1 5 1 10 15 15 10 10 1 1

0 0 1 1 1 1 1 0 1 1 0 0 1 0 1 0 1 0 0 0 0 1 0 1

0 1 1 0 1 0 1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 1 0

30 10 20 30 50 20 50 40 20 50 30 60 60 10 50 60 20 40 60 10 30 40 40 10

50 50 100 50 25 25 100 0 0 50 0 25 100 100 25 0 0 50 100 100 0 25 25 50

0 2 2 2 0 0 0 2 2 0 2 2 1 0 2 0 1 0 1 1 1 1 1 1

5 5 5 0 5 10 20 20 10 0 5 20 0 10 10 0 20 20 10 10 0 20 0 5

25 5 25 5 10 25 5 25 5 10 25 5 10 25 0 5 0 25 0 5 25 0 10 5

1 1 1 0 1 1 0 0 0 0 1 0 1 0 0 0 1 1 1 0 0 1 1 0

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

System A 1 1 2 1 3 2 4 2 5 3 6 3 7 4 8 4 9 5 10 5 11 6 12 6 13 7 14 7 15 8 16 8 17 9 18 9 19 10 20 10 21 11 22 11 23 12 24 12

Table C2 Correlation matrix of stated choice design attributes. budget

brt

length

time

roway

pop

pfreq

ofreq

capa

tcar

fare

close

env

cost

opcost

prepaid

tick

wait

staff

board

yearop

buss

shiftcar

1 0.00 0.00 0.02 0.00 0.00 0.04 0.00 0.00 0.00 0.00 0.01 0.00 0.02 0.00 0.00 0.00 0.12 0.00 0.00 0.01 0.10 0.00

1 0.02 0.07 0.04 0.01 0.02 0.01 0.06 0.00 0.02 0.01 0.06 0.01 0.04 0.04 0.03 0.08 0.04 0.02 0.02 0.01 0.02

1 0.02 0.02 0.01 0.01 0.01 0.04 0.02 0.03 0.01 0.03 0.01 0.03 0.03 0.01 0.02 0.03 0.00 0.00 0.03 0.01

1 0.09 0.05 0.02 0.00 0.03 0.11 0.05 0.08 0.07 0.04 0.01 0.00 0.01 0.03 0.01 0.01 0.02 0.00 0.00

1 0.01 0.03 0.02 0.02 0.06 0.02 0.01 0.07 0.03 0.00 0.05 0.04 0.05 0.00 0.01 0.03 0.01 0.03

1 0.01 0.01 0.08 0.04 0.02 0.05 0.05 0.02 0.01 0.00 0.01 0.05 0.01 0.04 0.01 0.00 0.01

1 0.46 0.04 0.02 0.01 0.02 0.03 0.03 0.04 0.02 0.02 0.02 0.02 0.00 0.02 0.03 0.04

1 0.02 0.00 0.03 0.01 0.01 0.03 0.02 0.01 0.01 0.09 0.03 0.01 0.00 0.03 0.01

1 0.04 0.01 0.02 0.03 0.05 0.00 0.01 0.01 0.02 0.00 0.00 0.05 0.07 0.02

1 0.04 0.03 0.08 0.01 0.01 0.03 0.01 0.07 0.00 0.05 0.02 0.03 0.02

1 0.04 0.02 0.01 0.00 0.04 0.03 0.00 0.05 0.03 0.01 0.02 0.00

1 0.05 0.03 0.00 0.00 0.02 0.06 0.02 0.00 0.01 0.01 0.00

1 0.03 0.02 0.04 0.01 0.07 0.02 0.02 0.04 0.00 0.03

1 0.01 0.02 0.02 0.05 0.02 0.02 0.00 0.02 0.03

1 0.03 0.00 0.06 0.01 0.01 0.02 0.01 0.03

1 0.06 0.00 0.00 0.03 0.01 0.00 0.04

1 0.02 0.01 0.02 0.01 0.01 0.01

1 0.04 0.02 0.10 0.08 0.03

1 0.03 0.01 0.01 0.02

1 0.01 0.03 0.02

1 0.02 0.03

1 0.00

1

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

budget brt length time roway pop pfreq ofreq capa tcar fare close env cost opcost prepaid tick wait staff board yearop buss shiftcar

45

46

D.A. Hensher et al. / Transportation Research Part A 72 (2015) 27–46

References Ben-Akiva, M., Morrikawa, T., 2002. Comparing ridership attraction of rail and bus. Transp. Policy 9 (2), 107–116. Cameron, T.A., DeShazo, J.R., 2014. A comprehensive assessment of selection in a major internet panel for the case of attitudes toward government regulation. Surv. Methodol. (submitted for publication). Canadian Urban Transit Association, 2004. Bus Rapid Transit: A Canadian Perspective, Issues Paper #10, CUTA, Toronto. Choice Metrics (2012) NGene, Choice Metrics, Sydney. Cirillo, C., Eboli, L., Mazzulla, G., 2011. On the Asymmetric user Perception of Transit. Cornwell, P., Cracknell, J., 1990. The case for busway transit. In: PTRC 18th Summer Annual Meeting, 1990 (This paper is a summary of TRL Research Report 329 and Overseas Road Note 12 of the Transport Research Laboratory, Berkshire, UK). Currie, G., Delbrosc, A., 2013. Exploring comparative ridership drivers of bus rapid transit and light rail transit routes. J. Public Transport. 16, 47–66. Currie, G., Wallis, I., 2008. Effective ways to grow urban bus markets – a synthesis of evidence. J. Transp. Geogr. 16, 419–429. dell’Olio, L., Ibeas, A., Cecìn, P., 2010a. The quality of service desired by public transport users. Transp. Policy. http://dx.doi.org/10.1016/ j.tranpol.2010.08.005. dell’Olio, L., Ibeas, A., Cecìn, P., 2010b. Modelling user perception of bus transit quality. Transp. Policy 17, 388–397. Eboli, L., Mazzulla, G., 2008a. An SP experiment for measuring service quality in public transport. Transport. Plann. Technol. 31 (5), 509–523. Eboli, L., Mazzulla, G., 2008b. Willingness-to-pay of public transport users for improvement in service quality. Eur. Transp. 38, 107–118. Eboli, L., Mazzulla, G., 2010. How to capture the passengers’ point of view on a transit service through rating and choice options. Transp. Rev. 30 (4), 435– 450. Greene, W.H., Hensher, D.A., 2013. Revealing additional dimensions of preference heterogeneity in a latent class mixed multinomial logit model. Appl. Econ. 45 (14), 1897–1902. Hass-Klau, C., Crampton, G., 2002. Future of urban transport: learning from success and weaknesses: light rail. Environment and Transport Planning, Brighton, UK. Hatton MacDonald, D., Morrison, M., Rose, J.M., Boyle, K., 2010. Untangling differences in values from internet and mail stated preference studies. In: Fourth World Congress of Environmental and Resource Economists, Montreal, Canada, June 28–July 2. Hensher, D.A., 1991. Hierarchical stated response designs and estimation in the context of bus use preferences. Logis. Transport. Rev. 26 (4), 299–323. Hensher, D.A., 1999. Bus-based transitway or light rail? Continuing the saga on choice versus blind commitment. Roads Transp. Res. 8 (3), 3–21. Hensher, D.A., 2010. Attribute processing, heuristics and preference construction in choice analysis. In: Hess, S., Daly, A. (Eds.), State-of Art and State-of Practice in Choice Modelling. Emerald Press, UK, pp. 35–70. Hensher, D.A., 2014. Special issue on process heuristics in choice analysis – an editorial. J. Choice Modell. 11 (1), 1–4. Hensher, D.A., Collins, A.T., 2011. Interrogation of responses to stated choice experiments: is there sense in what respondents tell us? J. Choice Modell. 4 (1), 62–89. Hensher, D.A., Greene, W.H., 2010. Non-attendance and dual processing of common-metric attributes in choice analysis: a latent class specification. Empirical Econ. 39 (2), 413–426. Hensher, D.A., Mulley, C.M., 2014. Candidate drivers of preference differences for physical images of BRT and LRT. BRT Centre Perceptions Study Paper #3, Transportation (in press). http://dx.doi.org/10.1007/s11116-014-9516-7. Hensher, D.A., Rose, J., 2007. Development of commuter and non-commuter mode choice models for the assessment of new public transport infrastructure projects: a case study. Transp. Res. Part A 41 (5), 428–433. Hensher, D.A., Rose, J.M., 2012. The influence of alternative acceptability, attribute thresholds and choice response certainty on automobile purchase preferences. J. Transp. Econ. Policy 46 (3), 451–468. Hensher, D.A., Waters II, W.G., 1994. Light rail and bus priority systems: choice or blind commitment? In: Starr Macmullen, B. (Ed.), Research in Transportation Economics, vol. III. JAI Press, Greenwich, Connecticut, pp. 139–162. Hensher, D.A., Rose, J.M., Greene, W.H., 2015a. Applied Choice Analysis, second ed. Cambridge University Press, Cambridge. Hensher, D.A., Mulley, C., Rose, J.M., 2015b. Understanding the relationship between voting preferences for public transport and perceptions and preferences for bus rapid transit versus light rail. J. Transp. Econ. Policy (in press). Hensher, D.A., Mulley, C., Li, Z., 2014. Drivers of bus rapid transit systems - influences on ridership and service frequency. In: Paper Presented at the 13th International Conference on Competition and Ownership of Land Passenger Transport (Thredbo 13), Oxford, September 15–19, 2013, Research in Transportation Economics 48, 159–165. Hess, S., Hensher, D., 2010. Using conditioning on observed choices to retrieve individual-specific attribute processing strategies. Transp. Res. Part B 44 (6), 781–790. Kain, J.F., 1988. Choosing the wrong technology: or how to spend billions and reduce transit use. J. Adv. Transport. 21, 197–213. Lindhjem, H., Navrud, S., 2011. Using internet in stated preference surveys: a review and comparison of survey modes. Int. Rev. Environ. Resour. Econ. 5, 309–351. Mackett, R.L., Edwards, M., 1996a. Guidelines for planning a new urban public transport system. Proc. Inst. Civ. Eng.: Transp. 117, 193–201. Mackett, R.L., Edwards, M., 1996b. An expert system to advise on urban public transport technologies. Comput. Environ. Urban Syst. 20, 261–273. Marcucci, E., Gatta, V., 2007. Quality and public transport service contracts. Eur. Transp. 36, 92–106. Mulley, C., Hensher, D.A., Rose, J.M., 2014. Do preferences for BRT and LRT vary across geographical jurisdictions? A comparative assessment of six Australian capital cities. Case Stud. Transp. Policy 2 (1), 1–9. Pickrell, D.H., 1992. A desire named streetcar: fantasy and fact in rail transit planning. J. Am. Plann. Assoc. 58, 158–176. Redman, L., Friman, M., Garling, T., Hartig, T., 2013. Quality attributes of public transport that attract car users: a research review. Transp. Policy 25, 119– 127. Schwanen, T., Banister, D., Anable, J., 2012. Rethinking habits and their role in behaviour change: the case of low-carbon mobility. J. Transp. Geogr. 24, 522– 532. Sislak, K.G., 2000. Bus rapid transport as a substitute for light rail: a tail of two cities. In: Paper Presented at 8th Joint Conference on Light Rail Transit. Transportation Research Board and American Public Transportation Association . Stradling, S.G., Anable, J., Carreno, M., 2007. Performance, importance and user disgruntlement: a six-step method for measuring satisfaction with travel modes. Transp. Res. Part A 41, 98–106. Swanson, J., Ampt, L., Jones, P., 1997. Measuring bus passenger preferences. Traffic Eng. Control (June), 330–336. TCRP (Transit Cooperative Research Program), 2000. TCRP Report 63: Enhancing the Visibility and Image of Transit in the United States and Canada, TRB, National Academy Press, Washington, DC. Yannes, C.D., Lownes, N.E., Garrick, N.W., Johnston, R.J., 2012. Operationalizing placemaking in a choice experiment context. Transport. Res. Rec. 2144 (2), 121–129.