A model to assess public transport demand stability

Transportation Research Part A 45 (2011) 755–764

Contents lists available at ScienceDirect

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

A model to assess public transport demand stability Pablo Bass a, Pedro Donoso b, Marcela Munizaga a,⇑ a b

Universidad de Chile, Casilla 228-3, Santiago, Chile LABTUS, Universidad de Chile, Casilla 228-3, Santiago, Chile

a r t i c l e

i n f o

Article history: Received 1 February 2011 Received in revised form 28 April 2011 Accepted 17 June 2011

Keywords: Public transport Client retention Demand modeling

a b s t r a c t Transport authorities, especially those in developing countries where rising income stimulate increased car ownership rates, are often concerned with maintaining or increasing levels of public transport use. Therefore, the ability to identify clients at risk of abandoning the system can be valuable for remedial measures, allowing for more focused quality improvements. We present and apply a model that determines the probability of migrating from public to private transport at both aggregated and disaggregated levels. In application, the model predicted migration with 60% accuracy in the first preference recovery measure. The proposed model can improve the understanding of the behavior of public transport users, the analysis of demand stability and the factors influencing migration. This, in turn, can help to focus policy and management measures and increase the efficiency of public investment. Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction Transport authorities, especially those in developing countries where rising income stimulate increased car ownership rates, are often concerned with maintaining or increasing levels of public transport use. In Santiago, Chile, this effect is prevalent; car ownership is growing rapidly, while public transport ridership is decreasing at the same rate. Therefore, the ability to identify clients at risk of abandoning the system can be valuable for remedial measures, allowing for more focused quality improvements. In this context, a public transport user or customer can be defined as a person who regularly uses public transport to travel to major activities, such as work or school. By investigating disaggregated personal information, we can determine how travelers perceive the system and adapt to external conditions. In other fields, demand stability has been analyzed via data mining and other statistical methods. In transport, very detailed and valuable data from discrete choice models and experiments designed to estimate the relative importance of quality of service variables are often used (Hensher et al., 2003). Although customer satisfaction and retention are rarely studied in the public transport field, some examples exist. Brog and Kahn (2003) analyzed the level of service variables and constructed a risk map, identifying the relation between satisfaction and risk of migration, for captive and non captive users. Minser and Webb (2010) applied a structural equation model to study the relation between customer loyalty, customer satisfaction, service quality, problem experience and public image in the Chicago public transport system. Their initial findings are that service quality and customer satisfaction have a positive effect on customer loyalty, and that public image and problem experience have a strong influence on customer satisfaction. More recently, Trepanier and Morency (2010) used long-term smartcard data to analyze rider retention and identify behavior by customer type in the Gatineau (Quebec) public transport system, they found that adult users who make mainly home based trips stay in the system for a longer period than a reference user. It is worth pointing out that all these authors say their work is preliminary, and more research is required in this field. ⇑ Corresponding author. Tel.: +56 2 9784649; fax: +56 2 6894206. E-mail addresses: [email protected] (P. Bass), [email protected] (P. Donoso), [email protected] (M. Munizaga). 0965-8564/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.tra.2011.06.003

756

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

Air transport is most likely the transport sector for which the most research on customer retention has been conducted because firms offer rewards and additional services to frequent users (Thomas, 1997). In general, firms that operate in a competitive market are interested in customer retention and have developed methods, such as data mining, to study the issue (Smith et al., 2000). In the public transport sector, the problem, typically considered from a more macroscopic view, is often simplified to consider, for example, the percentage of public transport users and their relation with policy measures. Goodwin (2007) produced an in-depth analysis of the impact of several policy measures designed to reduce car travel and thereby increase public transport use. However, the discrete mode choice models traditionally used in transport modeling contain disaggregated individual information that could be used to detect individuals who rarely use public transport and may migrate to other transportation modes unless they cannot afford to buy a car or use a more expensive alternative. A reduction in the market share of public transport is a result of aggregated individual decisions; thus, if these decisions were better understood, then policy measures could be designed to be more effective. Therefore, in this paper, we analyzed public transport demand stability as a function of the quality of service variables using discrete choice models. A case study is presented for Santiago, Chile, where a constant and significant loss of the market share of public transport over the last decades has prompted Chilean authorities to modify the public transport system. The new, substantially different system was introduced in February 2007. The older system, known as ‘‘yellow buses’’, was of poor quality, but it was efficient; travel times were reasonably low, and the city was well covered by a door-to-door network. However, the buses were old, and the drivers behaved aggressively. The new system, called Transantiago, which operates in a trunk–feeder structure, has significantly reduced the age and number of buses employed. It incorporates technological advances, such as automatic fare collection (AFC) and automatic vehicle location systems. However, during its first months in operation, the system experienced major problems, and the quality of service deteriorated in most areas. Authorities reacted by allocating more resources to system improvement. Additionally, a four-part panel survey, conducted by a research team from the Pontificia Universidad Católica de Chile and led by Professor Juan de Dios Ortúzar, observed 300 people before, during and after implementation of the new system. Furthermore, a quality of service study was conducted by the Universidad de Chile to identify the most relevant attributes from the user’s point of view. Using this information, we constructed a discrete choice model to predict the likelihood of individuals abandoning the public transport system. 2. Background A very interesting data source was available for this research: the Santiago panel (Yáñez et al., 2010), which included an experimental sample of Catholic University employees surveyed before and after the public transport system change. The advantage of this database is that the public transport system changed significantly during the observation period, allowing the observation of how people react to these changes. As illustrated in Fig. 1, the first wave of the survey was conducted two months before the introduction of the new system; the second, third and fourth waves were conducted three, ten and twenty months, respectively, after the introduction, and these later waves represented ‘‘crisis’’, ‘‘transition’’ and ‘‘stationary’’ situations, respectively. The sample started with 303 participants, and the subsequent waves had 286, 279 and 258 participants, respectively. The information gathered included socioeconomic characteristics of the participants, detailed information about travel to work (origin, destination, mode and level of services variables) and some attitudinal variables related to the perception of the Transantiago system. Although the new system was designed to improve travel conditions for passengers, after its introduction, 56% of the individuals surveyed in the panel declared that their trip to work was worse, 22% declared that it was similar, and 22% declared

Fig. 1. Description of the waves of the Santiago panel.

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

757

that it was better. According to the perceptions of the panel participants, the attributes that worsened the most were comfort (41%) and waiting time (28%). The attributes that reportedly improved the most were nothing (40%) and travel time (15%). In Fig. 2, the evolution of the level of service variables measured for the bus and metro, the most important public transport modes, is shown for different times. As shown in this figure, the total and travel times for bus users increased significantly in wave 2 (crisis) and later recovered, reaching values slightly below the pre-Transantiago values in the last wave. Metro users did not experience significant differences in time variables between waves, but they reported a significant loss in comfort due to an increase in the occupancy rates, which reached values of 6 pax/m2 during peak hours. The majority of participants continued traveling in the same mode as before Transantiago, while some used other modes to travel to work, modified their mode availability (purchased a car or bike) or moved to a new residence. It was not possible to isolate these effects to determine whether the changes were due to the service quality of public transport or to other reasons, but these general observations provide information for the modeling stage. In Fig. 3, the market shares and transfers between modes at each of the four waves are represented by circles of proportional size and arrows, respectively. All public transport modes were included in the public transport category. Public transport users are divided only into those who have other modes available (non captive) and those whose only possible mean of transport to travel to work is public transport (captive users). The car category includes drivers and companions, and the other category includes walking, cycling and combinations of modes. Despite the declared objective of Transantiago, after its implementation, the market share of cars

Fig. 2. Level of service variables for the bus and metro.

Fig. 3. Market shares and mode migration over time.

758

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

increased while that of public transport decreased. Between waves 1 and 2, users switched from the public transport category to the car and other categories and vice versa, but the net effect was negative for public transport. Between waves 2 and 3, the public transport and car market shares decreased, and the other category increased. In the last stage, the public transport share recovered to 59% as in wave 2. In wave 3 there are less public transport captive users (only 50%) than in the previous waves, showing that this sample adapted their situation to have more options. A clear income effect was visible after examining the characteristics of those who continued using and those who abandoned the public transport system; higher-income users abandoned the system more frequently than lower-income users. Some age and location effects were also observed. To widen the scope of the analysis, we considered other sources of information including a study of customer satisfaction and service quality perception of Transantiago users performance measures produced by the Transantiago agency to assess the performance of each operator, and also fare evasion measures. The customer satisfaction and service quality perception study conducted during May and November 2007, showed that global satisfaction was clearly lower in the first wave, consistent with the variables observed in the Santiago panel. It also showed differences by income and age. In the metro customer satisfaction surveys, a significant decrease in the level of satisfaction of users was observed in 2007, especially in the area of travel comfort, which decreased from values of over 60% before Transantiago to values of less than 20% after implementation. We searched for systematically measured operator performance variables that could be used for modeling. There were some regularity measures in the last years of operation, but the frequency compliance index (ICPH), calculated as the ratio between places (seats plus standing travel places) effectively provided by the operator and the number of places the operator should have provided in that period of time according to the contract, was available for a longer period of time. This index measures whether the operator (in aggregate) is providing the frequency of service specified by the operating program. This measure is a rudimentary quality index, but it is related to the operator’s performance because an operator with a low ICPH will not likely achieve frequency regularity. In Fig. 4, the ICPH values over time and by geographical sector are presented. From the initial measurements in August 2007, values continuously increased until October 2008, when they stabilized. The geographical distribution shows that the north and south sectors, which are the poorer sectors, were served by less reliable operators, while the wealthier east sector had the highest possible compliance index. Fare evasion, another relevant effect that was essentially non-existent before Transantiago, has been a problem since the implementation of the new system. Several types of fare evasion have been detected; there are some hard evaders, who a priori decide not to pay, and travel trough the system without validating in any boarding, and there are soft evaders who validate on some of his/her transactions but not in all of them. This latter behavior may be due to lack of credit in the

Fig. 4. ICPH variation over time and by geographical sector.

Fig. 5. Fare evasion rates by time and by geographical sector.

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

759

bip! card, to follow peers (typical in students) or because they could not reach the validation point in an overloaded bus. In Fig. 5 we present fare evasion rates calculated as an aggregate measure: passengers who board minus passengers who pay as a percentage of passengers who board. The evolution of evasion rates over time and the distribution by geographical sector show how this problem is growing, and that its distribution is not homogeneous throughout the city. The wealthier east sector had the lowest evasion rate, while the poorer south and north sectors had higher rates. 3. Proposed model The modeling of migration and retention of public transport users is a special case of market share forecasting. The decision by a public transport user to continue using the current transport mode or to change can be analyzed as a joint decision with two dimensions. The first is to migrate or to remain in the current mode and to determine the preferred mode. This is an individual decision, with mode-specific and common unobserved effects (correlation between modes). To represent migration and retention, two time periods must be observed. The traditional approach to represent mode choice is the Logit model, initially proposed by McFadden (1974) as an ‘‘econometric model of population choice behavior from distributions of individual decision rules’’. Using the same basic idea of random utility maximization, many contributions have been made since then, to account for correlation between alternatives (e.g. Williams, 1977), taste heterogeneity, heteroskedasticity and other effects. For example, Gaudry and Dagenais (1979) proposed a formulation to account for certain degree of captivity to specific alternatives. More recently, Cantillo et al. (2007) proposed a model to capture inertia and time correlation effects; with a formulation that includes state dependence. They applied the proposed model to a database containing Revealed Preferences (RP) and Stated Preference (SP) observations, and the SP (t = 1) choices are assumed to depend on the RP (t = 0) preferences. Following these ideas of mode complementarity/substitution, and time dependence, we propose a model based on a nested logit structure (Williams, 1977) and introduce the temporal effect as required. A simplified diagram of a model is shown in Fig. 6, where a bus user must decide in period t to stay in the public system (the alternatives are the bus and subway) or to migrate to another mode. Three aspects of the model must be considered: first, the hypothesis that every public transport user has at least one public transport mode available in period t; second, that some people may have only public transport modes available (public transport captive users); and third, that there is a chance that private modes are not correlated as shown in Fig. 6 (migration nest). One of the advantages of nested logit models is their ability to create different formulations depending on how groups or nests are formed. After initial testing of the original model with two nests (migration and no migration), we decided to focus on the migration nest and not to consider correlation among private modes. Therefore, the migration nest was disaggregated into the private mode alternatives available to users. Fig. 7 shows the final diagram of the model in which M and NM represent migration and no migration, respectively. A public transport user (q) who regularly uses mode bk must decide among the available modes in period t. The decision can be analyzed in two dimensions: staying in the public system (choose nest NM) or migrating (choose one of the available private alternatives M). The second level depends on the decision of the first level. If the user decides to select the nest NM, t t t then the choice set is Btq ¼ fb1 ; b2 ; . . . ; bp g. If the user migrates, then the private choice set is available. In Fig. 7, the private alternatives are represented in n nests Ml, each with only one mode available. The choice probabilities are given by the following equation:

Ptc =bt1 ¼ Pcl =M l  PMl ¼ 1  PMl l

ð1Þ

k

In Eq. (1), the probability that a regular public transport user of mode bk migrates to mode cl in period t is equal to the probability that the user chooses mode cl given migration to nest Ml multiplied by the probability of selecting nest Ml. In this case, t1 P cl =M l is equal to one because Ml has only one element. The notation =bk is omitted in the remainder of the article because all choices are conditional on the choice of mode bk in the period t  1.

Fig. 6. Structure of the proposed model.

760

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

Fig. 7. Structure of the proposed model.

expðU Ml Þ expðU Mi Þ þ expðU NM Þ i¼1

PMl ¼ Pn

ð2Þ

Eq. (2) shows the probability of choosing nest Ml, where U Mi is the utility of migrating to a private alternative i and UNM is the utility of remaining in the public system (NM).The utility U Mi is given by Eq. (3), where V ci is the representative utility of choosing ci and XMi is a function that represents characteristics of the user, the time period, or the transport system that contribute to the migration from public to private modes (mode i).

U Mi ¼ V ci þ XMi

ð3Þ

The utility of remaining in the public transport system UNM is given by Eq. (4), where EMUNM (Eq. (5)) is the expected maximum utility of nest NM. XNM is a function that represents features concerning the user, the time period, and the transport system that contribute to remaining in the public transport system. /NM is the structural parameter associated with the NM nest.

U NM ¼ EMU NM þ XNM EMU NM ¼ /NM  ln

" p X

ð4Þ  exp

j¼1

1  V bj /NM

# ð5Þ

Using Eqs. (2)–(4), Eq. (6) is obtained.

PMl ¼ Pn

i¼1

expðV cl Þ exp½V ci  ðXMl  XMi Þ þ exp½EMU NM  ðXMl  XNM Þ

ð6Þ

The same ‘‘migration’’ function (XMi ) was assumed to apply to all private modes XMi ¼ XM 8i. To simplify the notation, the ^ M ¼ XM  XNM was defined to capture the net effect of variables contributing to migration and retention. ‘‘new’’ function X With this notation, Eq. (7) is as follows:

expðV cl Þ

Ptc =bt1 ¼ Pn l k

^ i¼1 expðV ci Þ þ expðEMU NM  XM Þ

ð7Þ

^ M > 0, then the user is more likely to migrate, and if X ^ M < 0, then the user is more likely to remain in According to Eq. (7), if X the public transport system.The probability of choosing a public mode is given by Eq. (8). The probability that a user who regularly uses public mode bk will continue using public transport in period t is equal to the probability of choosing mode ba given that the user has not migrated (NM nest) multiplied by the probability of selecting nest NM.

Ptba =bt1 ¼ Pba =NM  PNM

ð8Þ

k

For completeness, Eq. (9) is as follows:

PNM ¼ 1 

n X l¼1

P Ml ¼ 1 

n X l¼1

P cl

ð9Þ

761

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

Substitution of Eq. (7) in Eq. (9) gives Eq. (10):

PNM ¼

1þ

1

Pn

ð10Þ

^ MÞ expðV ci  EMU NM þ X

i¼1

By definition, Eq. (11) is defined as:

exp Pba =NM ¼ P p

j¼1



1 /NM

exp



 V ba

1 /NM



 V bJ



ð11Þ

Substitution of Eq. (5) after manipulating Eq. (11) gives Eq. (12):

Pba =NM ¼ exp



1  ðV ba  EMU NM Þ /NM

 ð12Þ

Finally, by replacing Eqs. (10) and (12) in Eq. (8), we obtain Eq. (13):

Ptba =bt1 ¼ k

1þ

exp Pn

h

i¼1

i  ðV  EMU Þ NM b a /NM 1

ð13Þ

^ MÞ expðV ci  EMU NM þ X

^ M is set to zero so these users do not affect the likelihood of the M/NM comCaptive users are given a special treatment; X ponent of the model. 4. Results In this section, the calibration sample is described, and the results of the calibration process and future scenario simulations are presented. The Santiago panel had four data waves and an original sample size of 303 participants. Because the proposed model requires information for two time periods, data were ordered in consecutive pairs (t  1; t). After eliminating observations with validation filters, a sample of 443 paired observations was obtained (Table 1). There were twelve different transport modes in the calibration sample; six were classified as public transport (bus, metro, bus–metro, shared taxi, shared taxi–bus, shared taxi–metro), and six were classified as private (car driver, car companion, walk, bicycle, car driver-metro, car companion-metro). To be consistent with the model, all observations corresponded to users who chose a public transport option in (t  1). Public transport users mainly used the bus (44%), metro (23%) and bus–metro (19%). Of the 443 observations, 44 (10%) migrated from public transport to other modes. Captive public transport users accounted for 54% of the sample. In Table 2, the estimation results for the proposed model and for the classic MNL model are presented. Variables a.1–a.7 are present in the utility function of alternatives, and variables b.1–b.10 are present in the migration function of the proposed model. Variables a.1–a.5 are the classic mode choice model variables, including travel, walking and waiting time, and cost/wage. Number of transfers is also present, as this is a key factor of the Transantiago public transport system. The MNL model with classic variables had reasonable results, but the proposed specification was significantly better. In addition to the nest structure, the proposed model includes additional variables to account for sociodemographic characteristics of the individuals, the time use of respondents, and other variables related to availability and level of service. Variable a.6 accounts for an inertia effect observed for captive shared taxi users. The proposed model includes interaction effects as well, such a.7 that accounts for the fact that loosing the availability of the regularly used mode, affects the utility function of non captive users depending on the relation between mode cost and wage rate (larger disutility effect for poorer users on expensive modes). Table 1 Calibration sample. Mode

Before (t  1) Number

Bus Metro Bus – Metro Shared taxi – Metro Shared taxi Shared taxi – Bus Car driver Car companion Car companion – Metro Walk Car driver – Metro Bicycle Observations

196 103 85 37 14 8

443

After (t) % 44.2 23.3 19.2 8.4 3.2 1.8

100

Number

%

140 94 109 30 14 12 15 9 8 7 3 2 443

31.6 21.2 24.6 6.8 3.2 2.7 3.4 2.0 1.8 1.6 0.7 0.5 100

762

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

Table 2 Modeling results.* Variable_Period_Type of user

*

MNL

Proposed model

Value (t-stat)

Value (t-stat)

a. Alternative utility function variables 1. Travel time, period t, all users 2. Walking time, period t, all users 3. Waiting time, period t, all users 4. Number of transfers, period t, all users 5. Cost/wage, period t, all users 6. Shared taxi users, period t  1, captive users 7. Mode used in t  1 not available(Cost/wage)t, non-captive users

0.026(3.2) 0.058(3.9) 0.042(2.6) 1.064(5.6) 0.023(3.0) – –

0.017(2.1) 0.042(2.3) 0.029(2.0) 0.694(2.4) 0.014(2.0) 1.875(2.4) 0.274(2.0)

b. Migration utility function variables 1. Age > 49, period t  1, non-captive users 2. Income [Ch$105], period t  1, non-captive users 3. Activity, period t  1, non-captive-age < 35 4. Reduction in weekly working hours, Dt, non-captive 5. New availability of private mode, Dt, non-captive 6. Parking availability at destination, t, non-captive 7. Increase in waiting time, Dt, non captive-metro users 8. ICPH > 93%, period t, non captive-bus users 9. /NM

– – – – – – – – –

1.413(2.2) 0.115(2.1) 1.627(1.5) 1.919(3.2) 1.484(2.8) 1.495(2.9) 1.885(2.9) 2.008(1.8) 0.544(2.5)

Performance indicators Mode first preference recovery (FPR) Migration/retention FPR LR(C) vs. v2(95%, d.f.) q2(c) Log-likelihood Sample size

57% 20% 87.93 > 11.08 0.09 440.19 443

65% 57% 202.14 > 26.3 0.21 383.09 443

Preference parameters only; mode constants were omitted.

According to the model definition, a negative sign for variables in the alternative utility function implies a reduction in the utility function and, therefore, a reduction in the choice probability of that alternative. In the case of variables in the migration function (b.1–b.9), a positive sign implies that the variable contributes to the probability of migrating from public to private modes. The sociodemographic variables that contributed to migration included b.1 age (over 49) and b.2 income. Among variables related to time use, b.3 indicates that users under 35 who do extra work activities are more likely to migrate, and b.4 shows a significant effect of a reduction on the weekly working hours with the same sign. Finally, for mode characteristic variables, b.5 new availability of a private mode, b.6 parking availability at one’s destination and b.7 an increase in waiting time for metro users had significant effects. This variable captures the effect of metro overcrowding that forces users to wait for the next train. Finally, b.8 captures the effect of the frequency compliance index for non captive bus users and b.9 is the structural parameter of the model. Table 2 presents the first preference recovery (FPR), calculated as the percentage of cases in which the chosen option was the most likely option according to the model. The FPR was computed both at the elementary alternative level (mode FPR) and at the nest level (migration/retention FPR). Comparing the simple MNL to the proposed model, the proposed model behaved slightly better at the more detailed (mode) level. However, at the upper level (migration/retention), there was a large difference between models; the proposed model had nearly 60% achievement, while the simple MNL had only 20%. For goodness of fit, the new model, with a q2(c) of 0.21, was much better than the MNL, with a q2(c) of 0.09. To provide a more meaningful interpretation of the calibrated model, predictions for simulated scenarios in which the explanatory variables were varied are presented. Fig. 8 shows the migration rate (%) against the variation (%) of the corresponding explanatory variable. The graph on the left shows the migration % against the compliance index ICPH. The first point is the migration rate in the current scenario, and the points to the left correspond to different levels of increase in the ICPH, representing a more reliable operation of the public transport system, with a progressive increase of 5%. A 35% increase in the ICPH decreased migration by almost three points (9.9–7.0%). The graph on the right in Fig. 8 shows the relationship between migration and fare increase for six scenarios with fare rises of 5%, 8%, 13%, 18%, 58% and 163% for buses and 10%, 12%, 17%, 21%, 57% and 152% for the metro. In the first four simulated steps, the increase in migration was not significant (0.17%). Only very large fare variations had a significant effect on the migration rate, showing low fare elasticity. Additionally, a scenario in which the car ownership rate increased (every public transport user whose income was above the average income of the medium-income users had access to a car) was built. With this assumption, users with access to a car rose from 25% to 30%, and migration to private modes increased by 1.6%.

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

763

Fig. 8. Migration sensibility to ICPH and fare increase.

5. Conclusions This article presented a method for analyzing migration from public transport to other transport modes. Although customer retention has been widely studied in marketing and other areas, there have been very few applications to public transport. The proposed model is a discrete choice model calibrated using two levels of aggregation: the elementary alternative level and the public-transport-private level. This model can be used to determine the reasons underlying migration from public transport to other (private) modes and to propose remedial measures. Using the Santiago panel database, we applied the model to determine the probability of migration from public transport to private modes. The model predicted migration with 57% accuracy in the first preference recovery measure. The variables relevant to predicting migration were income, age (over 49), changes in mode availability, quality of service and time use. These results can help to focus policy and management measures and to increase the efficiency of public investment. In the case study, there was more sensitivity to quality of service (ICPH) than to fare changes. Additionally, lower income users received the worst service but migrated less than higher-income users. On the other hand, higher evasion rates are observed for low-income users. This kind of user does not have the option to change to another mode, they are captive of the pubic transport system. Their evasion can be seen as their own way of migration towards a virtual mode of public transport. This particular mode has the same quality of service but ‘‘zero’’ fare, considering they are not paying to use the service. The proposed model is an alternative to traditional customer retention models and has the advantage of being a slight variation of the traditional discrete mode choice model with calibration information readily available in many instances. In future research, to properly study customer retention, the traditional view of mode choice must be widened to consider, for example, that people who decide between using public transport and cars adapt their environment to their decision. Public transport users account for the accessibility of public transport in their activity location decisions, while car users prefer locations where parking areas are provided. Furthermore, structural decisions, such as renting a parking place and buying or selling a second car, are related to the mode decision and are a means of altering the attributes and constraints traditionally used to model mode choice decisions. There is a certain endogeneity in the process of deciding (e.g., where to live, how many cars to have and how to move around), and this was observed in the Santiago panel. The incorporation of all such decisions into the migration model is a research challenge. Also, the availability of Automatic Vehicle Location and Automatic Fare Collection systems provides an excellent data source that could be used with the purpose of monitoring user retention/migration and model it as a function of level of service variables such as bus speed (Cortés et al., 2011). Acknowledgements This research was partially funded by Fondecyt Grant 1090204, PBCT Redes Urbanas and the Complex Engineering Systems Institute (ICM P-05-004-F, CONICYT FBO16). We especially thank Professor Juan de Dios Ortúzar and María Francisca Yáñez for providing the Santiago panel data, Transantiago for providing ICPH data and Metro for providing customer satisfaction data. Preliminary results of this research were presented at IATRB 2009 and TRB 2011; comments from the audience and reviewers helped us to improve our work. Remaining errors are of course our responsibility. References Brog, W., Kahn, T., 2003. Customer satisfaction surveys for public transport companies – greater efficiency trough more demand oriented methods. In: Paper presented at ECOMM, Karistad, May. Cantillo, V., Ortúzar, J. de D., Williams, H.C.W.L., 2007. Modeling discrete choices in the presence of inertia and serial correlation. Transportation Science 41, 195–205. Cortés, C., Gibson, J., Gschwender, A., Munizaga, M.A., Zúñiga, M., 2011. Commercial bus speed diagnosis based on GPS-monitored data. Transportation Research 19C, 695–707. Gaudry, M., Dagenais, M., 1979. The Dogit model. Transportation Research 13B, 105–111.

764

P. Bass et al. / Transportation Research Part A 45 (2011) 755–764

Goodwin, P., 2007. Practical evidence on the effectiveness of transport policies in reducing car travel. In: Garling, T., Steg, L. (Eds.), Threats from car traffic to the quality of urban life. Elsevier, Amsterdam, pp. 401–424. Hensher, D., Stopher, P., Bullock, P., 2003. Service quality – developing a service quality index in the provision of commercial bus contracts. Transportation Research 37A, 499–517. McFadden, D., 1974. Conditional logit analysis of qualitative choice behavior. In: Zarembka, P. (Ed.), Frontiers in econometrics. Academic Press, New York, pp. 105–142. Minser, J., Webb, V., 2010. Quantifying the benefits: application of customer loyalty modeling in public transportation context. Transportation Research Record 2144, 111–120. Smith, K.A., Willis, R.J., Brooks, M., 2000. An analysis of customer retention and insurance claim patterns using data mining: a case study. Journal of the Operational Research Society 51, 532–541. Thomas, J.S., 1997. Econometric analysis of customer retention in an aviation trade organization. Transportation Research Record 1567, 33–40. Trepanier, M., Morency, C., 2010. Assessing transit loyalty with smart card data. In: 12th WCTR. Lisbon, Portugal. Williams, H.C.W.L., 1977. On the formation of travel demand models and economic evaluation measures of user benefit. Environment and Planning 9A, 285– 344. Yáñez, M.F., Mansilla, P., Ortúzar, J.deD., 2010. The Santiago panel: measuring the effects of implementing Transantiago. Transportation 37, 125–149.