- Email: [email protected]
Measuring the externalities of urban traffic improvement programs
Habitat International xxx (2016) 1e7
Contents lists available at ScienceDirect
Habitat International journal homepage: www.elsevier.com/locate/habitatint
Measuring the externalities of urban traffic improvement programs Emilio Picasso a, b, *, 1, Mariano Bonoli Escobar a, 1, Maria Stewart Harris a, 1, Felipe Tanco a, 1 n. Av. Las Heras 2214, 1er. piso, C1127AAR, Buenos Aires, Argentina Universidad de Buenos Aires, Facultad de Ingeniería, Departamento de Gestio lica Argentina, Facultad de Ingeniería, Departamento de Ingeniería Industrial. Av. Alicia Moreau de Justo 1620, C1107AFF, Universidad Cato Buenos Aires, Argentina a
b
a r t i c l e i n f o
a b s t r a c t
Article history: Received 10 August 2015 Received in revised form 9 February 2016 Accepted 11 February 2016 Available online xxx
In this article we present a methodology to improve the economic analysis of urban transportation programs by measuring the externalities, and we demonstrate its application to a case study of a large scale park & ride facility in Buenos Aires, Argentina. The externalities accompanying changes in transportation systems can be significant, even exceeding the magnitude of the intended benefits. Nevertheless they are difficult to measure, because they involve subjective values that are not traded in the market. We measure the subjective value of lower traffic congestion by means of three metrics: travel time, traffic accidents, and noise. A perceptual scale is created to make the latter metric operational for the first time, to the best of our knowledge. We implement a generic choice experiment to elicit the preferences of the individuals, and a mixed logit model to obtain the value rates of the externalities. The generic experiment, in contrast to specific modal choice stated preferences applications frequently used, focuses on the effects, then it is applicable to a wide range of transportation programs. The case study involves a park & ride facility whose demand was determined in a previous paper. We show via stylized discounted cash flow analysis that the value for the users is not sufficient to justify the investment. However the positive externality generated by lower traffic congestion in the city, measured by means of the subjective value rates estimated in the present study, turns the net present value of the project highly positive. © 2016 Elsevier Ltd. All rights reserved.
Keywords: Park & ride Externality Choice experiment Stated preference Discrete choice
1. Introduction The city of Buenos Aires, like most other large cities in the world, faces traffic congestion problems, particularly at peak hours, but also throughout the whole range of working hours in the centre. The government has been working on traffic flow improvement. Significant changes in public transportation have been implemented, like adding 3 new subway lines to the 4 existing and concentrating the bus network in wider avenues with dedicated lanes. These changes have contributed to improve the locally generated traffic flow, however there is a second cause of traffic congestion that is yet to be addressed: the large traffic inflow from the greater Buenos Aires area.
lica Argentina, Facultad de Ingeniería, * Corresponding author. Universidad Cato Departamento de Ingeniería Industrial. Av. Alicia Moreau de Justo 1620, C1107AFF, Buenos Aires, Argentina. E-mail addresses: epicasso@fi.uba.ar, [email protected] (E. Picasso), mbonoli@fi.uba.ar (M. Bonoli Escobar), [email protected] (M. Stewart Harris). 1 Tel.: (54 11) 4514 3011/12.
There is a conflict of interest between the city district and the province, as the problem in the former is caused by lagging investment in transportation in the latter. Park & ride facilities represent a solution to this second cause of traffic congestion, as they would deter commuters from entering the city by car, by offering them convenient connection with other transportation modes, namely railway, subway, bus and even charter vans (Dijk & Montalvo, 2011). Demand estimation for a park & ride facility in the north access to Buenos Aires city at different price levels has been conducted in a ~ a, previous study (Picasso, Bonoli Escobar, Stewart Harris, Pen Mermoz et al., 2014). The proposed project achieved high public acceptance among northern suburbs residents. However park & ride facilities require large scale investments for acquiring several hectares of land, and heavy infrastructure works: a multi-level building structure, railway and subway lines extension, etc. The investment for the proposed facility would be in the order of
http://dx.doi.org/10.1016/j.habitatint.2016.02.002 0197-3975/© 2016 Elsevier Ltd. All rights reserved.
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
2
E. Picasso et al. / Habitat International xxx (2016) 1e7
magnitude of 1000 MUS$.2 Despite the high public acceptance among commuters, the value generated for them would only cover a fraction of the investment and costs as explained below. This is in agreement with the results in Risa Hole (2004) for an employee park & ride service, who finds that the value created is hardly enough to offset the reduction in flexibility for car drivers. Bos, Van der Heijden, Molin, Timmermans et al. (2004) also acknowledge that park & ride facilities in The Netherlands have not attracted the expected number of car drivers, then they implement a choice experiment to explore the potential reasons, finding that security, quality of the connection and travel times are the most influent ones. Mingardo (2013) and Duncan and Cook (2014) find that park & ride facilities may also attract transit users, especially if they are peripheral, offsetting the reduction in vehicle-km travelled. This effect would be prevented by a fee, which is the case proposed in Picasso et al. (2014). The studies mentioned above focus on the value to the car driver, however there is another source of value created by the park & ride that remains unexplored. The city residents would not be using the park & ride facility, but they would get the positive externality of a more fluid traffic. Then they would probably be willing to contribute for the project to be implemented. The government can channel this willingness to pay via taxes, making this Pareto efficient project possible by means of a subsidy. The social value of the positive externality may or may not be enough to make the project economically viable, and it should be determined to answer that question. The objective of this research study is to determine the subjective value of the reduction in traffic congestion for the individuals using the traffic network of the city of Buenos Aires. The statistical methodology employed is discrete choice modelling. These models have been created to estimate the demand for a new transportation mode in the San Francisco Bay Area: the BART (Bay Area Railway Transit) (Mc Fadden, 1973, 1975, Mc Fadden et al., 1977). Since then, discrete choice models have been widely applied to different transportation problems, reaching the scientific community recognition as the best practice (Ben-Akiva & Lerman, 1985) (Bos et al., 2004), (Risa Hola, 2004) (Holguin-Veras, Reilly, Aros-Vera, Yushimito, & Isa, 2012). The discrete choice models decipher the decision patterns of economic agents, individuals in this case, among a discrete set of alternatives. They are based on the theory of random utility, where this utility is modelled as a function of the characteristics of the alternatives and a random element accounting for unobserved variables and preferences. The empirical base to estimate discrete choice models can be either natural information (revealed preferences or RP) or elicited in a choice experiment (stated preferences or SP). The second approach was selected for this research project as a natural database is not available. In section 2.1 we describe the choice experiment implemented for this study. Then we specify the discrete choice model employed. The model estimation results are presented in section 2.3, and finally the application to the park & ride case study is discussed in section 2.4. 2. Development The present study focuses on the effect of a generic transport improvement program on the congestion in the city. The choice experiment involves hypothetical programs defined by their effect
2 US$ stands for United States dollars, and Ar$ for Argentine peso. Multiples follow the standard notation: K for thousand, and M for million.
on travel time, accidents and noise. The model estimates the subjective value of each of these metrics for the city network users, making it possible to measure the externality generated by any transport improvement program, in particular the case study of the park & ride facility mentioned above. 2.1. The choice experiment The population under study is formed by the frequent users of the traffic network of the city of Buenos Aires,3 even if they live in the suburbs. The sample was done in two stages. We have employed an Internet panel: a data base of individuals willing to participate in surveys, built with the objective to represent the population within the sampling frame of internet users, which is the first stage. The data collection instrument was distributed among a random sample of panellists, which is the second stage. Frequent traffic network users were selected by a filter in the questionnaire. 218 interviews were completed out of 443 invitations. Each interview includes 10 choice tasks, making a total of 2180 data units. The sample is composed of 54% men and 46% women. This is aligned with the population data (census 20104) taking into account that work incidence is higher for men. The geographical distribution is 62% residents in the city and 38% habitants of the suburbs. The age distribution of the sample shows 33% respondents below 30 years old, 28% between 30 and 40, and 39% above 40 years old. This is also aligned with the population. The distribution of the socioeconomic level5 of the sample shows 29% for the high class, 61% for the middle one, and 10% for the low class. This is somewhat skewed upwards compared to the population, where the lower class represents 35%, meaning that the conclusions of the study are primarily representative of the middle and high classes, with a smaller weight for the lower one. The measurement instrument is an online questionnaire including the choice experiment. The first part of the questionnaire contains the filter and an exploration of habits. The most relevant trip, due to frequency or length, is identified for each individual, as well as its frequency, transport mode, and travel time. The car tax and property tax levels of the individual are also collected. The second part of the questionnaire is the choice experiment, described below. The third part has the demographic questions. The choice experiment consists of 10 choice tasks, each one having 4 alternatives representing generic hypothetical traffic improvement programs for the Buenos Aires city. The experimental design focuses on the effects on traffic congestion, characterizing each transport program by: travel time, noise level, traffic accidents, and cost. The details of the transport programs causing them remain generic. A typical screen of the data collection software, presenting the choice task to the respondent, is shown in Fig. 1 (translated from Spanish). The average travel time of the most relevant trip is recorded for each individual and used as a realistic base value for the choice experiment. The travel time of each alternative in each choice task pivoted on the base value for each individual, varying across 5 points: 50%, 25%, 0, þ25%, þ50%. The noise levels presented are the same for all individuals. A scale of noise was developed in a previous study. This study was based on a random sample of 116 individuals living in the Buenos
3 noma de Buenos The “city” means the federal district, i.e. the “Ciudad Auto Aires”. The metropolitan area includes the city and the Greater Buenos Aires. 4 Instituto Nacional de Estadísticas y Censos. 5 The methodology to measure the socioeconomic level in Argentina was n developed by the Sociedad Argentina de Investigadores de Mercado y Opinio (SAIMO).
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
E. Picasso et al. / Habitat International xxx (2016) 1e7
3
Fig. 1. Typical choice task (translated).
Fig. 2. Noise scale.
Aires metropolitan area. They were asked to partially rank 12 noise description statements. They first selected the 5 loudest noises and then the 5 softest ones according to their perception. Both rank orders were recorded. Thurstone case V scaling method was employed to generate an interval scale for noise (Thurstone, 1927). Results are shown in Fig. 2. Three noise statements were selected for the choice experiment. The top of the scale has five statements with similar scale values. The statement: “Noise that forces you living with windows closed” was selected to represent a high noise level because it is the most appropriate to the traffic issue. The statement: “Noise in a full restaurant” was selected to represent the medium noise level. The bottom item of the scale (“Noise in the countryside”) was discarded for unrealistic in a congested city. Instead, the next lowest statement: “Noise in a family meeting” was selected to represent a low noise level in the choice experiment. The intensity of traffic accidents is measured by the number of injured persons in the Buenos Aires city and surroundings. The same three levels are used for all respondents. The base level is set around the actual figure informed by the Ombudsman of Buenos Aires6 for the city district, extrapolating it to the whole metropolitan area proportionally to population. This amounts to 100 injured
6
Defensoría del pueblo de la ciudad de Buenos Aires. Structural parameters: Mean, standard deviation and correlation of the logarithm of the Log-Normal law. The sign of the variables was reversed to negative so that the partial utilities are positive and hence suitable for the Log-Normal law. The sign of the structural parameters is negative when the median partial utility is below one. 7
persons per day. The high and low values are set at 400 and 25 injured persons per day. The monetary vehicle used in the experiment to introduce economic valuation is a local tax. The rationale is that the government is considering different traffic management programs which would impact the traffic flow, noise level and traffic accidents, and they would also have different costs which would be paid by the population through a tax increment. The base value of the local tax increment is individual specific. The individuals owning a car are asked about their automobile tax amount. Besides that, individuals are asked about their real estate property tax amount. The base value is set at the current automobile tax amount if available, or the current property tax amount if available, or 200 Arg$/month otherwise. The last value represents an average of the real estate tax in the city, obtained from the Buenos Aires Government Budget. The local tax values of each alternative in each choice task pivoted on the base value for each individual, varying across 5 points: 30%, 15%, 0, þ15%, þ30%. The experimental design of the first alternative is created as an orthogonal plan (Federov, 1972), by optimally blocking the 5 3 3 5 full factorial into groups of 5 tasks (Meyer, Nachtsheim, 1995) and assigning them in pairs to each individual. This algorithm maximizes the information richness of the data. The design is cyclically developed into the 4 alternatives following the principles recommended by Huber and Zwerina (1996). The use of lexicographical processes for decision-making by respondents brings problems when estimating the discrete choice model (Louviere, Hensher, Swait at al., 2000), (Ortuzar & Willumsen, 1994). Lexicographic behaviour was identified in 47 respondents, mainly focused on traffic accidents. These respondents were set aside from the sample, leaving 171 individuals, i.e. 1710 data units, for parameter estimation. Once the survey was finished, the database was processed and formatted in an appropriate way to input the choice modelling software: The “mlogit” package in R language, and NLOGIT 5.0.
2.2. The discrete choice model The population under study is the set of people facing traffic congestion problems in the city of Buenos Aires in Argentina, at least once a week. This includes the residents of the city moving by private or public transport, and the people living in the Greater Buenos Aires area that work in the city. The discrete choice model is based on the following random utility specification:
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
4
E. Picasso et al. / Habitat International xxx (2016) 1e7
Ujit ¼ b1 x1;jit þ b2 x2;jit þ b3 x3;jit þ b4 x4;jit þ ejit
(1)
Where: Ujit: Utility (random) of alternative j for the individual i in the choice task t. x1,jit: Travel time in the most frequent itinerary of the individual i, established in the hypothetical situation of alternative j in choice task t. x2,jit: Noise level established in the hypothetical situation of alternative j in choice task t for individual i. x3,jit: Square root of the number of fatal accidents per year prevailing in the city established in the hypothetical situation of alternative j in choice task t for individual i. The transformation is justified in the results section. x4,jit: Tax level established in the hypothetical situation of alternative j in choice task t for individual i. bk: partial utilities (or importance) of the aforementioned variables. They are modelled as random with Log-Normal law. Hence they are positive, preceded by a negative sign to reflect the disutility nature of the four variables. εjit: Random component of utility, distributed as Gumbel type I. Alternative specific constants are omitted because the experiment is generic, i.e. alternatives are unlabelled traffic congestion situations. The random utility structure can be stated more concisely:
Ujit ¼ bT xjit þ ejit
(1a)
Where the 4-dimensional vector of partial utilities includes the negative signs. The Logit kernel of the model provides a closed form for the probability of choice for each alternative (Train, 2009):
Pjit
exp bT xjit ¼P J T exp b x lit l¼1
(2)
Pji ¼
t¼1
exp bT xjit f ðbÞdb PJ exp bT xlit l¼1
(3)
Where the Log-Normal is selected for the mixing distribution: f(b). This probability law ensures the coefficient have the right sign, and provides room for the typical asymmetry in the preferences of the population. The MXL model is more realistic than the basic Multinomial Logit model as it makes room for heterogeneity in the population and it is not restricted to the independence of irrelevant alternatives (IIA) property. It can be shown that the MXL model can represent any random utility model under general regularity conditions (Mc Fadden, Train, 2000). The likelihood function of the model is:
ln ℒ ¼
J I X T X X
yjit ln Pjit
Z bi ¼ Z
bPðyi =xi ; bÞ fðb=qÞdb (5) Pðyi =xi ; bÞ fðb=qÞdb
Where yi is the set of choices done by individual i across all tasks, xi are the variables characterizing the alternatives, and q represents the structural parameters previously determined. Discrete choice models enable the estimation of the subjective value of time (SVT) for the population under study. According to Ben-Akiva and Lerman (1985) and Mc Fadden (1997), the subjective value of time is defined as the marginal rate of substitution between travel time and the monetary vehicle:
SVT ¼
vV=vx1 b1 ¼ vV=vx4 b4
(6)
Where V is the systematic part of the utility function. Both partial utilities are random variables following a LogNormal law. Despite the difficulties brought by this long tailed distribution, as described by Hensher and Greene (2003), it has a convenient property: The Log-Normal distribution has null density at zero, and it is preserved by division, then the distribution of SVT is also Log-Normal with the following structural parameters:
8 > 2 > : N m ; s lnb > 1 1 1 > < lnb4 : N m4 ; s24 > > > > : ln SVT ¼ ln b1 ln b4 : N m1 m4 ; s21 þ s24 2r14 s1 s4 (7)
A Mixed Logit (MXL) model is formulated to take into account the heterogeneity of the population. The probability of choice for each alternative is mixed across individuals (Train, 2009):
Z Y T
The estimation of these models proceeds in two phases. The structural parameters (mean, standard deviation and correlations of b) are estimated in a first phase, via simulated maximum likelihood or Markov Chain Monte Carlo methods (Train, 2009). Then, individual parameters are estimated by taking advantage of the Bayesian structure of the model, by calculating the following integrals via Monte Carlo methods (Train, 2009):
(4)
i¼1 t¼1 j¼1
Where yjit ¼ 1 when alternative j was chosen by individual i in task t, and yjit ¼ 0 otherwise.
where r is the coefficient of correlation between the logarithms of the individual parameters. The mean of SVT can be severely affected by the long tails of the Log-Normal law. Therefore we describe the SVT by the median and two quantiles as lower and upper bounds: 10% and 90%, which can be obtained for a Log-Normal law from the following expression with t ¼ 0.5, 0.1, and 0.9 respectively:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi SVTt ¼ exp m1 m4 þ s21 þ s24 2r14 s1 s4 zt
(8)
Where zt is the corresponding Normal quantile. In this analysis we are assuming that the sampling error of the structural parameter estimates is negligible compared to the dispersion of the individual parameters. The sampling error could be introduced in the analysis via simulation, as discussed in Hensher and Greene (2003). The subjective value of noise can be defined similarly, taking into account that it is an indicator variable:
SVN ¼
DV=Dx2 b2 ¼ vV=vx4 b4
(9)
The subjective value of accidents takes into account the square root transformation:
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
E. Picasso et al. / Habitat International xxx (2016) 1e7
qffiffiffiffiffiffi x3 ¼ x0 3
Table 2 Estimation of the MXL model.
(10)
vV=vx0 3 vV=vx3 vx3 =vx0 3 1 b3 ¼ ¼ pffiffiffiffiffiffi SVA ¼ vV=vx4 vV=vx4 2 x0 3 b4 The expressions corresponding to (8) are immediate.
2.3. Results The objective of the analysis in this article is to measure the positive externality of traffic reduction produced by an improvement like the park & ride facility studied in Picasso et al. (2014). For that purpose we have estimated a MNL model as a first step. Results are shown in Table 1. Performance indicators of the model are log likelihood, Akaike information criterion, and Mc Fadden R2. Travel time and tax enter utility equation directly. Noise is represented by two indicator variables: one for medium level and the other one for high level, being low level the reference. In the case of the variable “accidents”, it was found that transforming the variable by the square root produced better results, what is not surprising as the scale used in the experiment was not linear. Accidents entering the model linearly produced a worse AIC: 2812. The risk of accidents may have decreasing marginal influence on utility, or perhaps the high level of accidents was too high for individuals to form a perception. All parameters are statistically significant and they have the expected sign for disutility factors. This model provides a first approximation to the problem, although it still assumes homogeneous preferences among individuals. A MXL model was estimated to release this assumption. A Log-Normal probability law was specified for the parameters for travel time, high noise, accidents, and tax. This enforces their negative sign. The results of the estimation of the MXL model are shown in Table 2. A higher maximum likelihood was reached. The lower Akaike Information Criterion (AIC) for the MXL model establishes its superiority over the MNL model, confirming the heterogeneity of the population. An exception was medium level noise which parameter was kept fixed, as its variance was found not statistically significant. The population is homogeneous in the sensitivity to medium level noise, whereas sensitivity to high level noise is heterogeneous. The structural parameters specified for the Log-Normal random parameters are the mean, the standard deviation and the correlation of their logarithm. Structural means of the random parameters are:
Table 1 Estimation of the MNL model. Variable
Travel time Noise-med Noise-high Accidents Tax ln Likd AICe R2f a b c d e f
5
Variable
Travel time Noise-med Noise-high Accidents Tax r time-tax r NoiseH-Tax r Acc-Tax ln Likb AICc R2d a b c d
Partial utility7 Prob. Law
Mean
pa
S.D.
pa
LogNormal Fixed LogNormal LogNormal LogNormal
3.432 0.799 0.949 1.431 5.796 0.26 0.22 0.53
<0.001 <0.001 <0.001 <0.001 <0.001 0.03 0.60 0.29
0.847 e 1.075 0.593 1.776
<0.001 e <0.001 <0.001 <0.001
1214 2446.1 0.4561
Risk of rejecting null hypothesis. Natural logarithm of the likelihood function. Akaike information criterion. Mc Fadden goodness of fit indicator.
0.046 min1 for travel time, 4.60 for high noise level, 0.29 acc½ for accidents (square root), and 0.015 Ar$1 for the tax. These figures resemble the pattern of the parameters in the MNL model, however at a higher scale, due to the lower variance of the error terms in the more accurate MXL model. The subjective value for the attributes can be calculated by means of (6), (9), and (10). Expression (8) is used to calculate the 10%, 50% and 90% quantiles representing the population heterogeneity. The median subjective value for travel time is: exp(3.432 þ 5.796) ¼ 10.6. Taking into account that the times presented in the experiment are in minutes per one way trip, the taxes are in Ar$/month,8 and the mean stated travel frequency is 14.2 trips/month, the subjective value of time results: 0.38 Ar$/min. The median subjective value of medium noise is: 0.799/ exp(5.796) ¼ 263 Ar$/M, and for high noise is: exp(0.949 þ 5.796) ¼ 850 Ar$/M. The extrapolation of these figures would require to know the distribution of the exposure to different noise levels in the population of the city. This could be measured with our noise scale in another survey. Accidents were operationalized as the daily number of injured victims in traffic accidents in the metropolitan area of Buenos Aires. The Buenos Aires Statistics data bank indicated that there were 8935 injured victims of traffic accidents in 2012. Extrapolating this figure to the whole metropolitan area by population (census 20109) it gets to 115 injured/day. The actual figure enters the calculation of the median subjective value of traffic accidents in (10): 0.12 Ar$ per injured. This can be extended to the whole population of the metropolitan area of Buenos Aires: 532 kAr$ per injured person. Hence the society is willing to pay 61 MAr$/year to eliminate the problem of injuries caused by traffic accidents. The Table 3 shows the heterogeneity of the subjective values calculated via expression (8):
Partial utility Coefficient
S.E.a
tb
pc
0.0236 0.778 1.922 0.183 0.00631 1390.1 2790.2 0.3689
0.00139 0.0786 0.916 0.00724 0.00080
17.06 20.97 25.29 25.29 7.88
<0.001 <0.001 <0.001 <0.001 <0.001
Standard error. Student t statistic. Risk of rejecting null hypothesis. Natural logarithm of the likelihood function. Akaike information criterion. Mc Fadden goodness of fit indicator.
2.4. Case study The subjective values determined in the present study are useful to evaluate traffic programs impacting the city district. We evaluate ~ a, and the park & ride project studied in Picasso, Bonoli, Pen Mermoz (2012) and Picasso et al. (2014). The authors found that the concept was well accepted by the habitants of the northern suburbs of the city, showing high intended use. A choice experiment was implemented and modelled to estimate the demand
8 9
Exchange rate prevalent in Sep 2012: 6.26 Ar$/US$. Instituto Nacional de Estadísticas y Censos.
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
6
E. Picasso et al. / Habitat International xxx (2016) 1e7
Table 3 Subjective value of Attributes. Attribute
Units
Quantile 10%
Median
Quantile 90%
Travel time Noise-med Noise-high Accidents
Ar$/min Ar$/M Ar$/M Ar$/acc
0.024 27 46 0.017
0.38 263 850 0.12
5.93 2560 15500 0.87
curve for the park & ride facility. This showed that a large facility would be required, in the order of a one million square metre building plus high capacity transportation to the city centre, like subway or train line extension. The high investment implied makes a thorough economic evaluation of the project necessary to ensure the social value is commensurate with the costs. Such a study is beyond the scope of the present article, however we make a high level economic evaluation with the intention to understand the orders of magnitude. The demand estimated via discrete choice modelling for the park & ride at the base price10 was 14% of the trips, representing 29.2 Kveh/day.11 The commuters’ willingness to pay (WTP) for the park & ride can be readily calculated. On the other hand the investment is estimated at 1150 MUS$,12 and the operating costs are estimated at 50 MUS$/year.13 The net present value of the cash flow equates to 400 MUS$.14 The negative sign demonstrates that the population of commuters is not willing to bear the full cost of the project. Sensitivity analysis can be performed by moving on the demand curve or fine tuning the costs, however it is clear that the economic evaluation would hardly be positive. This analysis means that the park & ride project is not worth the investment based on the commuters sole perspective. Nevertheless there is value in the project for other people as well. The city habitants would face a positive externality: lower traffic congestion, lower noise, and possibly lower rate of accidents. This fact motivates the present study, where we have measured the value of this externality. In order to link both studies we need two additional pieces of knowledge. First we need to estimate the number of people benefiting from the externality, including city habitants moving around by car or bus, and also people coming from the suburbs and using the same network. Second we need to estimate the reduction in travel times in the city due to the traffic headed to the park & ride. We first make a conservative estimate of the number of people affected. The commuters from the north access represent 546 Kpax/ day, inflow plus outflow informed by the national office of transport, conservatively assuming one passenger in each car. The rest of the accesses account for 1400 Kpax/day from the same source and with the same assumptions. The car population in the city reaches 1323 Kveh, as informed by the national office of car property.15 The fraction using their cars daily are estimated at 662 Kpax/day. The number of taxis registered in the city is 38 Kveh, as informed in the Buenos Aires Government statistics data bank. They are
10 The base price for the park & ride facility was 25 Ar$, including the transportation fare to the city centre. This was equivalent to 6.5 US$. 11 n Nacional de Vialidad is 273 Kveh/ The traffic inflow informed by the Direccio day, 75% of which enters the city at peak time. 12 Land value is estimated at 350 MUS$ for 23 ha. The building cost is estimated at 500 MUS$ for 970 Km2, including access roads. The subway extension is estimated at 300 MUS$ for 2.5 km. 13 The salaries are estimated at 8 MUS$/Y for 430 workers. Maintenance is estimated at 4% of the investment per year. Other costs are estimated at 20% of total. 14 The cash flow extends over 3 years for investment and over 10 years for operation. The net present value is calculated at a social discount rate of 3%. An annuity is added to account for future periods. 15 n Nacional del Registro de la Propiedad Automotor. Direccio
conservatively estimated to transport 190 Kpax/day. The city district buses carry 1409 Kpax/day, as informed by the national office of transport.16 Overall, at least 4209 K persons are affected by traffic congestion in the city of Buenos Aires every day. Motorcycles are not considered as they have the ability to avoid congestion to a large extent by sneaking through the traffic, and truck drivers are ignored in this conservative estimate because the number is not as significant. The link between the number of cars headed to the park & ride and the travel times would require a large scale traffic assignment model of the whole metropolitan area of Buenos Aires, that is unavailable unfortunately. Nevertheless, as a preliminary approximation, we use a classic formula:
4 t q ¼ 1 þ 0:15 t0 qMAX
(11)
Where t is the travel time, t0 is the free flow travel time, q is the traffic flow, and qMAX is the capacity of the arc. We are assuming that the north city access behaves essentially like an arc in the network. The capacity cannot be determined directly because it is not a true single arc, however we have calibrated the formula with actual data. We have collected travel time data for 13 itineraries representing key traffic channels in the city, both at rush hour and at decongested time. The mean traffic speed was calculated for each itinerary in both conditions and averaged. The ratio of off peak to peak time speed is 1.69. Substituting this time ratio into expression (11) produces an estimate of the pseudo capacity of the north access. The remaining traffic in the north access is calculated by deducting the park & ride demand from the inflow, and input into the formula (11) for an improved time ratio calculation, which compared to the calibration time ratio provides a time saving index. The time saving at the base price of the park & ride results 19%. This figure multiplied by the average length of trip measured in the interviews represents almost 11 min, and 686 MUS$/year when applied to the population previously determined and employing the subjective value of time estimated in the present study. This amount represents the externality on the city network users. When both sources of value are considered in the cash flow (WTP of commuters and externality on city network users) the net present value turns highly positive: 25700 MUS$, even ignoring the positive externalities in terms of noise and traffic accidents. This means that the park & ride project in the north access of Buenos Aires is interesting and deserves the attention and thorough study from the authorities. In more general terms, this shows the importance of the knowledge produced in the present article for rational public policy making. The subjective value of traffic accidents determined in this study should be further assessed. On the one hand it is consistent with similar preliminary valuations performed by the authors in other studies. On the other hand there are signs suggesting that the range of rates used in the experiment was too wide. Some individuals may have had a hard time forming a perception about the high rate of accidents, what could explain the decreasing marginal influence on utility. Some other individuals behaved lexicographically meaning that the high rate of accidents overrode the effect of the other variables for them. The square root employed in the model protects the valuation from bias, however it is probably conservative and we would recommend to narrow the range of the rate of accidents in future studies.
16
n Nacional de Regulacio n del Transporte. Comisio
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
E. Picasso et al. / Habitat International xxx (2016) 1e7
3. Conclusions We present in this article an assessment of the value generated by transport improvement programs in the city of Buenos Aires. The subjective value of lower traffic congestion is measured in terms of travel time, noise, and traffic accidents, by means of a choice experiment. The target population is the people using the traffic network of the city, both habitants of the city and habitants of the suburbs frequently coming to the city. Respondents were stimulated with hypothetical transport improvement programs with specific values for travel time, noise, accidents, and cost, for them to choose. The data was analysed by means of a MXL discrete choice model. The median WTP (willingness to pay) was estimated for each for each variable, as well as its heterogeneity in the population. The subjective value of travel time in the city is 0.38 Ar$/min. The subjective value of reducing noise from high to low level in our perceptual scale is 850 Ar$/Month. The subjective value of eliminating traffic accidents in the metropolitan area is 61 MAr$/Year. These subjective value rates are useful to properly assess the impact of transport improvement programs. We employ this knowledge to complete the evaluation of a park & ride facility, studied in Picasso et al. (2012, 2014). In that opportunity, the commuters WTP for a park & ride facility located in the north access to Buenos Aires city was determined via discrete choice modelling. A stylized discounted cash flow analysis shows that the cost of the project is higher than the amount commuters would be willing to pay. However, there is value for city habitants due to lower traffic congestion, overlooked in that study. Actually, all city network users would benefit from that. We use the subjective value rates measured in the present study to assess the value of the lower traffic congestion. We determine the number of city network users from different sources of information. We estimate the impact on travel time in the city as a consequence of reduced congestion due to the flow headed to the park & ride. This enables the calculation of the value of the project for city network users by means of the value rate previously determined in this study. The net present value of the project turns highly positive when the externality is taken into account, demonstrating the importance of the results obtained in the present study. References Ben-Akiva, M., & Lerman, S. (1985). Discrete choice analysis: Theory and application to travel demand. Cambridge, MA: MIT Press. Bos, I. Ilona D. M., Van der Heijden, Rob E. C. M., Molin, E. J., & Timmermans, H. J. P. (2004). The choice of park and ride facilities: an analysis using a contextdependent hierarchical choice experiment. Environment and Planning A, 36(9), 1673e1686. Dijk, M., & Montalvo, C. (2011). Policy frames of park-and-ride in Europe. Journal of Transport Geography, 19(6), 1106e1119. Duncan, M., & Cook, D. (2014). Is the provision of park-and-ride facilities at light rail stations an effective approach to reducing vehicle kilometers traveled in a US
7
context? Transportation Research part A, 66, 65e74. Federov, V. V. (1972). translated and edited by. In W. J. Studen, & E. M. Klimko (Eds.), Theory of experiments. New York: Academic Press. Hensher, D. A., & Greene, W. H. (2003). The mixed logit model: the state of practice. Transportation, 30, 133e176. Holguin-Veras, J., Reilly, J., Aros-Vera, F., Yushimito, W., & Isa, J. (2012). Park-andride facilities in New York city. Transportation Research Record, 2276(1), 123e130. Huber, J., & Zwerina, K. (1996). The importance of utility balance in efficient choice designs. Journal of Marketing Research, 33, 307e317. August. Louviere, J., Hensher, D., & Swait, J. (2000). Stated choice methods: analysis and applications. New York: Cambridge Univ. Press. Mc Fadden, D. (1973). Conditional logit analysis of qualitative choice behavior. In P. Zarembka (Ed.), Frontiers of econometrics. New York: Academic Press. Mc Fadden, D. (1975). The measurement of urban travel demand. Berkeley, USA: University of California. Mc Fadden, D. (1997). Measuring willingness-to-pay for transportation improvements. Berkeley, USA: University of California. Mc Fadden, D., Talvitie, A., Cosslett, S., Hassan, I., Johnson, M., Reid, F., et al. (1977). Demand model estimation and validation; final report, volume V, urban travel demand forecasting project. Berkeley: Institute of Transportation Statistics, Univ. of California. Mc Fadden, D., & Train, K. E. (2000). Mixed MNL models for discrete response. Journal of Applied Econometrics, 81, 447e770. Meyer, R. K., & Nachtsheim, C. J. (1995). The coordinate-exchange algorithm for constructing exact optimal experimental designs. Technometrics, 37, 60e69. Mingardo, G. (2013). Transport and environmental effects of rail-based Park and Ride: evidence from the Netherlands. Jour. Transport Geography, 30, 7e16. Ortuzar, J. D., & Willumsen, L. G. (1994). Modelling transport (2nd ed.). UK: John Wiley & Sons. ~ a, G., & Mermoz, O. L. (2014). Picasso, E., Bonoli Escobar, M., Stewart Harris, M., Pen Demand elasticity estimation for a park and ride facility by means of discrete n choice modeling. Revista de la Escuela de Perfeccionamiento en Investigacio Operativa, 35, 167e185. ~ a, G., & Mermoz, O. (2012). Estimacio n de la demanda de Picasso, E., Bonoli, M., Pen rico mediante modelos de eleccio n discreta. Anales de un estacionamiento perife n Operativa, 1e14. Article number: la Escuela de Perfeccionamiento de Investigacio 23. Risa Hole, A. (2004). Forecasting the demand for an employee park and ride service using commuters' stated choices. Transport Policy, 11(4), 355e362. Thurstone, L. L. (1927). A law of comparative judgment. Psychological Review, 34, 273e286. Train, K. E. (2009). Discrete choice methods with simulation. Cambridge Univ. Press.
Internet data bases: n Nacional de Regulacio n del Transporte http://www.cnrt.gob.ar/content/ Comisio estadisticas/. Buenos Aires Government budget: http://www.buenosaires.gob.ar/areas/hacienda/ presupuesto2012/index.php. Buenos Aires Government statistics data bank: http://www.buenosaires.gob.ar/ areas/hacienda/sis_estadistico/banco_datos/. Defensoría del pueblo de la ciudad de Buenos Aires: http://www.defensoria.org.ar/ especiales/informesespeciales.php. n Nacional de Vialidad (national transportation office) traffic statistics: Direccio http://transito.vialidad.gov.ar:8080/SelCE_WEB/intro.html. n Nacional del Registro de la Propiedad Automotor statistics: http://www. Direccio dnrpa.gov.ar/portal_dnrpa/boletines_estadisticos.php#.VT-bLCGeDGc. Instituto de Estadísticas y Censos de Argentina: http://www.censo2010.indec.gov.ar/. n (SAIMO): http://www. Sociedad Argentina de Investigadores de Mercado y Opinio saimo.org.ar/socios/Socios/NSE2006-23nov2006-Informe_final.pdf. Strategic traffic plan of the city of Buenos Aires: http://www.cedom.gov.ar/es/ legislacion/poderdepolicia/transito/index4.html.
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
Contents lists available at ScienceDirect
Habitat International journal homepage: www.elsevier.com/locate/habitatint
Measuring the externalities of urban traffic improvement programs Emilio Picasso a, b, *, 1, Mariano Bonoli Escobar a, 1, Maria Stewart Harris a, 1, Felipe Tanco a, 1 n. Av. Las Heras 2214, 1er. piso, C1127AAR, Buenos Aires, Argentina Universidad de Buenos Aires, Facultad de Ingeniería, Departamento de Gestio lica Argentina, Facultad de Ingeniería, Departamento de Ingeniería Industrial. Av. Alicia Moreau de Justo 1620, C1107AFF, Universidad Cato Buenos Aires, Argentina a
b
a r t i c l e i n f o
a b s t r a c t
Article history: Received 10 August 2015 Received in revised form 9 February 2016 Accepted 11 February 2016 Available online xxx
In this article we present a methodology to improve the economic analysis of urban transportation programs by measuring the externalities, and we demonstrate its application to a case study of a large scale park & ride facility in Buenos Aires, Argentina. The externalities accompanying changes in transportation systems can be significant, even exceeding the magnitude of the intended benefits. Nevertheless they are difficult to measure, because they involve subjective values that are not traded in the market. We measure the subjective value of lower traffic congestion by means of three metrics: travel time, traffic accidents, and noise. A perceptual scale is created to make the latter metric operational for the first time, to the best of our knowledge. We implement a generic choice experiment to elicit the preferences of the individuals, and a mixed logit model to obtain the value rates of the externalities. The generic experiment, in contrast to specific modal choice stated preferences applications frequently used, focuses on the effects, then it is applicable to a wide range of transportation programs. The case study involves a park & ride facility whose demand was determined in a previous paper. We show via stylized discounted cash flow analysis that the value for the users is not sufficient to justify the investment. However the positive externality generated by lower traffic congestion in the city, measured by means of the subjective value rates estimated in the present study, turns the net present value of the project highly positive. © 2016 Elsevier Ltd. All rights reserved.
Keywords: Park & ride Externality Choice experiment Stated preference Discrete choice
1. Introduction The city of Buenos Aires, like most other large cities in the world, faces traffic congestion problems, particularly at peak hours, but also throughout the whole range of working hours in the centre. The government has been working on traffic flow improvement. Significant changes in public transportation have been implemented, like adding 3 new subway lines to the 4 existing and concentrating the bus network in wider avenues with dedicated lanes. These changes have contributed to improve the locally generated traffic flow, however there is a second cause of traffic congestion that is yet to be addressed: the large traffic inflow from the greater Buenos Aires area.
lica Argentina, Facultad de Ingeniería, * Corresponding author. Universidad Cato Departamento de Ingeniería Industrial. Av. Alicia Moreau de Justo 1620, C1107AFF, Buenos Aires, Argentina. E-mail addresses: epicasso@fi.uba.ar, [email protected] (E. Picasso), mbonoli@fi.uba.ar (M. Bonoli Escobar), [email protected] (M. Stewart Harris). 1 Tel.: (54 11) 4514 3011/12.
There is a conflict of interest between the city district and the province, as the problem in the former is caused by lagging investment in transportation in the latter. Park & ride facilities represent a solution to this second cause of traffic congestion, as they would deter commuters from entering the city by car, by offering them convenient connection with other transportation modes, namely railway, subway, bus and even charter vans (Dijk & Montalvo, 2011). Demand estimation for a park & ride facility in the north access to Buenos Aires city at different price levels has been conducted in a ~ a, previous study (Picasso, Bonoli Escobar, Stewart Harris, Pen Mermoz et al., 2014). The proposed project achieved high public acceptance among northern suburbs residents. However park & ride facilities require large scale investments for acquiring several hectares of land, and heavy infrastructure works: a multi-level building structure, railway and subway lines extension, etc. The investment for the proposed facility would be in the order of
http://dx.doi.org/10.1016/j.habitatint.2016.02.002 0197-3975/© 2016 Elsevier Ltd. All rights reserved.
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
2
E. Picasso et al. / Habitat International xxx (2016) 1e7
magnitude of 1000 MUS$.2 Despite the high public acceptance among commuters, the value generated for them would only cover a fraction of the investment and costs as explained below. This is in agreement with the results in Risa Hole (2004) for an employee park & ride service, who finds that the value created is hardly enough to offset the reduction in flexibility for car drivers. Bos, Van der Heijden, Molin, Timmermans et al. (2004) also acknowledge that park & ride facilities in The Netherlands have not attracted the expected number of car drivers, then they implement a choice experiment to explore the potential reasons, finding that security, quality of the connection and travel times are the most influent ones. Mingardo (2013) and Duncan and Cook (2014) find that park & ride facilities may also attract transit users, especially if they are peripheral, offsetting the reduction in vehicle-km travelled. This effect would be prevented by a fee, which is the case proposed in Picasso et al. (2014). The studies mentioned above focus on the value to the car driver, however there is another source of value created by the park & ride that remains unexplored. The city residents would not be using the park & ride facility, but they would get the positive externality of a more fluid traffic. Then they would probably be willing to contribute for the project to be implemented. The government can channel this willingness to pay via taxes, making this Pareto efficient project possible by means of a subsidy. The social value of the positive externality may or may not be enough to make the project economically viable, and it should be determined to answer that question. The objective of this research study is to determine the subjective value of the reduction in traffic congestion for the individuals using the traffic network of the city of Buenos Aires. The statistical methodology employed is discrete choice modelling. These models have been created to estimate the demand for a new transportation mode in the San Francisco Bay Area: the BART (Bay Area Railway Transit) (Mc Fadden, 1973, 1975, Mc Fadden et al., 1977). Since then, discrete choice models have been widely applied to different transportation problems, reaching the scientific community recognition as the best practice (Ben-Akiva & Lerman, 1985) (Bos et al., 2004), (Risa Hola, 2004) (Holguin-Veras, Reilly, Aros-Vera, Yushimito, & Isa, 2012). The discrete choice models decipher the decision patterns of economic agents, individuals in this case, among a discrete set of alternatives. They are based on the theory of random utility, where this utility is modelled as a function of the characteristics of the alternatives and a random element accounting for unobserved variables and preferences. The empirical base to estimate discrete choice models can be either natural information (revealed preferences or RP) or elicited in a choice experiment (stated preferences or SP). The second approach was selected for this research project as a natural database is not available. In section 2.1 we describe the choice experiment implemented for this study. Then we specify the discrete choice model employed. The model estimation results are presented in section 2.3, and finally the application to the park & ride case study is discussed in section 2.4. 2. Development The present study focuses on the effect of a generic transport improvement program on the congestion in the city. The choice experiment involves hypothetical programs defined by their effect
2 US$ stands for United States dollars, and Ar$ for Argentine peso. Multiples follow the standard notation: K for thousand, and M for million.
on travel time, accidents and noise. The model estimates the subjective value of each of these metrics for the city network users, making it possible to measure the externality generated by any transport improvement program, in particular the case study of the park & ride facility mentioned above. 2.1. The choice experiment The population under study is formed by the frequent users of the traffic network of the city of Buenos Aires,3 even if they live in the suburbs. The sample was done in two stages. We have employed an Internet panel: a data base of individuals willing to participate in surveys, built with the objective to represent the population within the sampling frame of internet users, which is the first stage. The data collection instrument was distributed among a random sample of panellists, which is the second stage. Frequent traffic network users were selected by a filter in the questionnaire. 218 interviews were completed out of 443 invitations. Each interview includes 10 choice tasks, making a total of 2180 data units. The sample is composed of 54% men and 46% women. This is aligned with the population data (census 20104) taking into account that work incidence is higher for men. The geographical distribution is 62% residents in the city and 38% habitants of the suburbs. The age distribution of the sample shows 33% respondents below 30 years old, 28% between 30 and 40, and 39% above 40 years old. This is also aligned with the population. The distribution of the socioeconomic level5 of the sample shows 29% for the high class, 61% for the middle one, and 10% for the low class. This is somewhat skewed upwards compared to the population, where the lower class represents 35%, meaning that the conclusions of the study are primarily representative of the middle and high classes, with a smaller weight for the lower one. The measurement instrument is an online questionnaire including the choice experiment. The first part of the questionnaire contains the filter and an exploration of habits. The most relevant trip, due to frequency or length, is identified for each individual, as well as its frequency, transport mode, and travel time. The car tax and property tax levels of the individual are also collected. The second part of the questionnaire is the choice experiment, described below. The third part has the demographic questions. The choice experiment consists of 10 choice tasks, each one having 4 alternatives representing generic hypothetical traffic improvement programs for the Buenos Aires city. The experimental design focuses on the effects on traffic congestion, characterizing each transport program by: travel time, noise level, traffic accidents, and cost. The details of the transport programs causing them remain generic. A typical screen of the data collection software, presenting the choice task to the respondent, is shown in Fig. 1 (translated from Spanish). The average travel time of the most relevant trip is recorded for each individual and used as a realistic base value for the choice experiment. The travel time of each alternative in each choice task pivoted on the base value for each individual, varying across 5 points: 50%, 25%, 0, þ25%, þ50%. The noise levels presented are the same for all individuals. A scale of noise was developed in a previous study. This study was based on a random sample of 116 individuals living in the Buenos
3 noma de Buenos The “city” means the federal district, i.e. the “Ciudad Auto Aires”. The metropolitan area includes the city and the Greater Buenos Aires. 4 Instituto Nacional de Estadísticas y Censos. 5 The methodology to measure the socioeconomic level in Argentina was n developed by the Sociedad Argentina de Investigadores de Mercado y Opinio (SAIMO).
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
E. Picasso et al. / Habitat International xxx (2016) 1e7
3
Fig. 1. Typical choice task (translated).
Fig. 2. Noise scale.
Aires metropolitan area. They were asked to partially rank 12 noise description statements. They first selected the 5 loudest noises and then the 5 softest ones according to their perception. Both rank orders were recorded. Thurstone case V scaling method was employed to generate an interval scale for noise (Thurstone, 1927). Results are shown in Fig. 2. Three noise statements were selected for the choice experiment. The top of the scale has five statements with similar scale values. The statement: “Noise that forces you living with windows closed” was selected to represent a high noise level because it is the most appropriate to the traffic issue. The statement: “Noise in a full restaurant” was selected to represent the medium noise level. The bottom item of the scale (“Noise in the countryside”) was discarded for unrealistic in a congested city. Instead, the next lowest statement: “Noise in a family meeting” was selected to represent a low noise level in the choice experiment. The intensity of traffic accidents is measured by the number of injured persons in the Buenos Aires city and surroundings. The same three levels are used for all respondents. The base level is set around the actual figure informed by the Ombudsman of Buenos Aires6 for the city district, extrapolating it to the whole metropolitan area proportionally to population. This amounts to 100 injured
6
Defensoría del pueblo de la ciudad de Buenos Aires. Structural parameters: Mean, standard deviation and correlation of the logarithm of the Log-Normal law. The sign of the variables was reversed to negative so that the partial utilities are positive and hence suitable for the Log-Normal law. The sign of the structural parameters is negative when the median partial utility is below one. 7
persons per day. The high and low values are set at 400 and 25 injured persons per day. The monetary vehicle used in the experiment to introduce economic valuation is a local tax. The rationale is that the government is considering different traffic management programs which would impact the traffic flow, noise level and traffic accidents, and they would also have different costs which would be paid by the population through a tax increment. The base value of the local tax increment is individual specific. The individuals owning a car are asked about their automobile tax amount. Besides that, individuals are asked about their real estate property tax amount. The base value is set at the current automobile tax amount if available, or the current property tax amount if available, or 200 Arg$/month otherwise. The last value represents an average of the real estate tax in the city, obtained from the Buenos Aires Government Budget. The local tax values of each alternative in each choice task pivoted on the base value for each individual, varying across 5 points: 30%, 15%, 0, þ15%, þ30%. The experimental design of the first alternative is created as an orthogonal plan (Federov, 1972), by optimally blocking the 5 3 3 5 full factorial into groups of 5 tasks (Meyer, Nachtsheim, 1995) and assigning them in pairs to each individual. This algorithm maximizes the information richness of the data. The design is cyclically developed into the 4 alternatives following the principles recommended by Huber and Zwerina (1996). The use of lexicographical processes for decision-making by respondents brings problems when estimating the discrete choice model (Louviere, Hensher, Swait at al., 2000), (Ortuzar & Willumsen, 1994). Lexicographic behaviour was identified in 47 respondents, mainly focused on traffic accidents. These respondents were set aside from the sample, leaving 171 individuals, i.e. 1710 data units, for parameter estimation. Once the survey was finished, the database was processed and formatted in an appropriate way to input the choice modelling software: The “mlogit” package in R language, and NLOGIT 5.0.
2.2. The discrete choice model The population under study is the set of people facing traffic congestion problems in the city of Buenos Aires in Argentina, at least once a week. This includes the residents of the city moving by private or public transport, and the people living in the Greater Buenos Aires area that work in the city. The discrete choice model is based on the following random utility specification:
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
4
E. Picasso et al. / Habitat International xxx (2016) 1e7
Ujit ¼ b1 x1;jit þ b2 x2;jit þ b3 x3;jit þ b4 x4;jit þ ejit
(1)
Where: Ujit: Utility (random) of alternative j for the individual i in the choice task t. x1,jit: Travel time in the most frequent itinerary of the individual i, established in the hypothetical situation of alternative j in choice task t. x2,jit: Noise level established in the hypothetical situation of alternative j in choice task t for individual i. x3,jit: Square root of the number of fatal accidents per year prevailing in the city established in the hypothetical situation of alternative j in choice task t for individual i. The transformation is justified in the results section. x4,jit: Tax level established in the hypothetical situation of alternative j in choice task t for individual i. bk: partial utilities (or importance) of the aforementioned variables. They are modelled as random with Log-Normal law. Hence they are positive, preceded by a negative sign to reflect the disutility nature of the four variables. εjit: Random component of utility, distributed as Gumbel type I. Alternative specific constants are omitted because the experiment is generic, i.e. alternatives are unlabelled traffic congestion situations. The random utility structure can be stated more concisely:
Ujit ¼ bT xjit þ ejit
(1a)
Where the 4-dimensional vector of partial utilities includes the negative signs. The Logit kernel of the model provides a closed form for the probability of choice for each alternative (Train, 2009):
Pjit
exp bT xjit ¼P J T exp b x lit l¼1
(2)
Pji ¼
t¼1
exp bT xjit f ðbÞdb PJ exp bT xlit l¼1
(3)
Where the Log-Normal is selected for the mixing distribution: f(b). This probability law ensures the coefficient have the right sign, and provides room for the typical asymmetry in the preferences of the population. The MXL model is more realistic than the basic Multinomial Logit model as it makes room for heterogeneity in the population and it is not restricted to the independence of irrelevant alternatives (IIA) property. It can be shown that the MXL model can represent any random utility model under general regularity conditions (Mc Fadden, Train, 2000). The likelihood function of the model is:
ln ℒ ¼
J I X T X X
yjit ln Pjit
Z bi ¼ Z
bPðyi =xi ; bÞ fðb=qÞdb (5) Pðyi =xi ; bÞ fðb=qÞdb
Where yi is the set of choices done by individual i across all tasks, xi are the variables characterizing the alternatives, and q represents the structural parameters previously determined. Discrete choice models enable the estimation of the subjective value of time (SVT) for the population under study. According to Ben-Akiva and Lerman (1985) and Mc Fadden (1997), the subjective value of time is defined as the marginal rate of substitution between travel time and the monetary vehicle:
SVT ¼
vV=vx1 b1 ¼ vV=vx4 b4
(6)
Where V is the systematic part of the utility function. Both partial utilities are random variables following a LogNormal law. Despite the difficulties brought by this long tailed distribution, as described by Hensher and Greene (2003), it has a convenient property: The Log-Normal distribution has null density at zero, and it is preserved by division, then the distribution of SVT is also Log-Normal with the following structural parameters:
8 > 2 > : N m ; s lnb > 1 1 1 > < lnb4 : N m4 ; s24 > > > > : ln SVT ¼ ln b1 ln b4 : N m1 m4 ; s21 þ s24 2r14 s1 s4 (7)
A Mixed Logit (MXL) model is formulated to take into account the heterogeneity of the population. The probability of choice for each alternative is mixed across individuals (Train, 2009):
Z Y T
The estimation of these models proceeds in two phases. The structural parameters (mean, standard deviation and correlations of b) are estimated in a first phase, via simulated maximum likelihood or Markov Chain Monte Carlo methods (Train, 2009). Then, individual parameters are estimated by taking advantage of the Bayesian structure of the model, by calculating the following integrals via Monte Carlo methods (Train, 2009):
(4)
i¼1 t¼1 j¼1
Where yjit ¼ 1 when alternative j was chosen by individual i in task t, and yjit ¼ 0 otherwise.
where r is the coefficient of correlation between the logarithms of the individual parameters. The mean of SVT can be severely affected by the long tails of the Log-Normal law. Therefore we describe the SVT by the median and two quantiles as lower and upper bounds: 10% and 90%, which can be obtained for a Log-Normal law from the following expression with t ¼ 0.5, 0.1, and 0.9 respectively:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi SVTt ¼ exp m1 m4 þ s21 þ s24 2r14 s1 s4 zt
(8)
Where zt is the corresponding Normal quantile. In this analysis we are assuming that the sampling error of the structural parameter estimates is negligible compared to the dispersion of the individual parameters. The sampling error could be introduced in the analysis via simulation, as discussed in Hensher and Greene (2003). The subjective value of noise can be defined similarly, taking into account that it is an indicator variable:
SVN ¼
DV=Dx2 b2 ¼ vV=vx4 b4
(9)
The subjective value of accidents takes into account the square root transformation:
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
E. Picasso et al. / Habitat International xxx (2016) 1e7
qffiffiffiffiffiffi x3 ¼ x0 3
Table 2 Estimation of the MXL model.
(10)
vV=vx0 3 vV=vx3 vx3 =vx0 3 1 b3 ¼ ¼ pffiffiffiffiffiffi SVA ¼ vV=vx4 vV=vx4 2 x0 3 b4 The expressions corresponding to (8) are immediate.
2.3. Results The objective of the analysis in this article is to measure the positive externality of traffic reduction produced by an improvement like the park & ride facility studied in Picasso et al. (2014). For that purpose we have estimated a MNL model as a first step. Results are shown in Table 1. Performance indicators of the model are log likelihood, Akaike information criterion, and Mc Fadden R2. Travel time and tax enter utility equation directly. Noise is represented by two indicator variables: one for medium level and the other one for high level, being low level the reference. In the case of the variable “accidents”, it was found that transforming the variable by the square root produced better results, what is not surprising as the scale used in the experiment was not linear. Accidents entering the model linearly produced a worse AIC: 2812. The risk of accidents may have decreasing marginal influence on utility, or perhaps the high level of accidents was too high for individuals to form a perception. All parameters are statistically significant and they have the expected sign for disutility factors. This model provides a first approximation to the problem, although it still assumes homogeneous preferences among individuals. A MXL model was estimated to release this assumption. A Log-Normal probability law was specified for the parameters for travel time, high noise, accidents, and tax. This enforces their negative sign. The results of the estimation of the MXL model are shown in Table 2. A higher maximum likelihood was reached. The lower Akaike Information Criterion (AIC) for the MXL model establishes its superiority over the MNL model, confirming the heterogeneity of the population. An exception was medium level noise which parameter was kept fixed, as its variance was found not statistically significant. The population is homogeneous in the sensitivity to medium level noise, whereas sensitivity to high level noise is heterogeneous. The structural parameters specified for the Log-Normal random parameters are the mean, the standard deviation and the correlation of their logarithm. Structural means of the random parameters are:
Table 1 Estimation of the MNL model. Variable
Travel time Noise-med Noise-high Accidents Tax ln Likd AICe R2f a b c d e f
5
Variable
Travel time Noise-med Noise-high Accidents Tax r time-tax r NoiseH-Tax r Acc-Tax ln Likb AICc R2d a b c d
Partial utility7 Prob. Law
Mean
pa
S.D.
pa
LogNormal Fixed LogNormal LogNormal LogNormal
3.432 0.799 0.949 1.431 5.796 0.26 0.22 0.53
<0.001 <0.001 <0.001 <0.001 <0.001 0.03 0.60 0.29
0.847 e 1.075 0.593 1.776
<0.001 e <0.001 <0.001 <0.001
1214 2446.1 0.4561
Risk of rejecting null hypothesis. Natural logarithm of the likelihood function. Akaike information criterion. Mc Fadden goodness of fit indicator.
0.046 min1 for travel time, 4.60 for high noise level, 0.29 acc½ for accidents (square root), and 0.015 Ar$1 for the tax. These figures resemble the pattern of the parameters in the MNL model, however at a higher scale, due to the lower variance of the error terms in the more accurate MXL model. The subjective value for the attributes can be calculated by means of (6), (9), and (10). Expression (8) is used to calculate the 10%, 50% and 90% quantiles representing the population heterogeneity. The median subjective value for travel time is: exp(3.432 þ 5.796) ¼ 10.6. Taking into account that the times presented in the experiment are in minutes per one way trip, the taxes are in Ar$/month,8 and the mean stated travel frequency is 14.2 trips/month, the subjective value of time results: 0.38 Ar$/min. The median subjective value of medium noise is: 0.799/ exp(5.796) ¼ 263 Ar$/M, and for high noise is: exp(0.949 þ 5.796) ¼ 850 Ar$/M. The extrapolation of these figures would require to know the distribution of the exposure to different noise levels in the population of the city. This could be measured with our noise scale in another survey. Accidents were operationalized as the daily number of injured victims in traffic accidents in the metropolitan area of Buenos Aires. The Buenos Aires Statistics data bank indicated that there were 8935 injured victims of traffic accidents in 2012. Extrapolating this figure to the whole metropolitan area by population (census 20109) it gets to 115 injured/day. The actual figure enters the calculation of the median subjective value of traffic accidents in (10): 0.12 Ar$ per injured. This can be extended to the whole population of the metropolitan area of Buenos Aires: 532 kAr$ per injured person. Hence the society is willing to pay 61 MAr$/year to eliminate the problem of injuries caused by traffic accidents. The Table 3 shows the heterogeneity of the subjective values calculated via expression (8):
Partial utility Coefficient
S.E.a
tb
pc
0.0236 0.778 1.922 0.183 0.00631 1390.1 2790.2 0.3689
0.00139 0.0786 0.916 0.00724 0.00080
17.06 20.97 25.29 25.29 7.88
<0.001 <0.001 <0.001 <0.001 <0.001
Standard error. Student t statistic. Risk of rejecting null hypothesis. Natural logarithm of the likelihood function. Akaike information criterion. Mc Fadden goodness of fit indicator.
2.4. Case study The subjective values determined in the present study are useful to evaluate traffic programs impacting the city district. We evaluate ~ a, and the park & ride project studied in Picasso, Bonoli, Pen Mermoz (2012) and Picasso et al. (2014). The authors found that the concept was well accepted by the habitants of the northern suburbs of the city, showing high intended use. A choice experiment was implemented and modelled to estimate the demand
8 9
Exchange rate prevalent in Sep 2012: 6.26 Ar$/US$. Instituto Nacional de Estadísticas y Censos.
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
6
E. Picasso et al. / Habitat International xxx (2016) 1e7
Table 3 Subjective value of Attributes. Attribute
Units
Quantile 10%
Median
Quantile 90%
Travel time Noise-med Noise-high Accidents
Ar$/min Ar$/M Ar$/M Ar$/acc
0.024 27 46 0.017
0.38 263 850 0.12
5.93 2560 15500 0.87
curve for the park & ride facility. This showed that a large facility would be required, in the order of a one million square metre building plus high capacity transportation to the city centre, like subway or train line extension. The high investment implied makes a thorough economic evaluation of the project necessary to ensure the social value is commensurate with the costs. Such a study is beyond the scope of the present article, however we make a high level economic evaluation with the intention to understand the orders of magnitude. The demand estimated via discrete choice modelling for the park & ride at the base price10 was 14% of the trips, representing 29.2 Kveh/day.11 The commuters’ willingness to pay (WTP) for the park & ride can be readily calculated. On the other hand the investment is estimated at 1150 MUS$,12 and the operating costs are estimated at 50 MUS$/year.13 The net present value of the cash flow equates to 400 MUS$.14 The negative sign demonstrates that the population of commuters is not willing to bear the full cost of the project. Sensitivity analysis can be performed by moving on the demand curve or fine tuning the costs, however it is clear that the economic evaluation would hardly be positive. This analysis means that the park & ride project is not worth the investment based on the commuters sole perspective. Nevertheless there is value in the project for other people as well. The city habitants would face a positive externality: lower traffic congestion, lower noise, and possibly lower rate of accidents. This fact motivates the present study, where we have measured the value of this externality. In order to link both studies we need two additional pieces of knowledge. First we need to estimate the number of people benefiting from the externality, including city habitants moving around by car or bus, and also people coming from the suburbs and using the same network. Second we need to estimate the reduction in travel times in the city due to the traffic headed to the park & ride. We first make a conservative estimate of the number of people affected. The commuters from the north access represent 546 Kpax/ day, inflow plus outflow informed by the national office of transport, conservatively assuming one passenger in each car. The rest of the accesses account for 1400 Kpax/day from the same source and with the same assumptions. The car population in the city reaches 1323 Kveh, as informed by the national office of car property.15 The fraction using their cars daily are estimated at 662 Kpax/day. The number of taxis registered in the city is 38 Kveh, as informed in the Buenos Aires Government statistics data bank. They are
10 The base price for the park & ride facility was 25 Ar$, including the transportation fare to the city centre. This was equivalent to 6.5 US$. 11 n Nacional de Vialidad is 273 Kveh/ The traffic inflow informed by the Direccio day, 75% of which enters the city at peak time. 12 Land value is estimated at 350 MUS$ for 23 ha. The building cost is estimated at 500 MUS$ for 970 Km2, including access roads. The subway extension is estimated at 300 MUS$ for 2.5 km. 13 The salaries are estimated at 8 MUS$/Y for 430 workers. Maintenance is estimated at 4% of the investment per year. Other costs are estimated at 20% of total. 14 The cash flow extends over 3 years for investment and over 10 years for operation. The net present value is calculated at a social discount rate of 3%. An annuity is added to account for future periods. 15 n Nacional del Registro de la Propiedad Automotor. Direccio
conservatively estimated to transport 190 Kpax/day. The city district buses carry 1409 Kpax/day, as informed by the national office of transport.16 Overall, at least 4209 K persons are affected by traffic congestion in the city of Buenos Aires every day. Motorcycles are not considered as they have the ability to avoid congestion to a large extent by sneaking through the traffic, and truck drivers are ignored in this conservative estimate because the number is not as significant. The link between the number of cars headed to the park & ride and the travel times would require a large scale traffic assignment model of the whole metropolitan area of Buenos Aires, that is unavailable unfortunately. Nevertheless, as a preliminary approximation, we use a classic formula:
4 t q ¼ 1 þ 0:15 t0 qMAX
(11)
Where t is the travel time, t0 is the free flow travel time, q is the traffic flow, and qMAX is the capacity of the arc. We are assuming that the north city access behaves essentially like an arc in the network. The capacity cannot be determined directly because it is not a true single arc, however we have calibrated the formula with actual data. We have collected travel time data for 13 itineraries representing key traffic channels in the city, both at rush hour and at decongested time. The mean traffic speed was calculated for each itinerary in both conditions and averaged. The ratio of off peak to peak time speed is 1.69. Substituting this time ratio into expression (11) produces an estimate of the pseudo capacity of the north access. The remaining traffic in the north access is calculated by deducting the park & ride demand from the inflow, and input into the formula (11) for an improved time ratio calculation, which compared to the calibration time ratio provides a time saving index. The time saving at the base price of the park & ride results 19%. This figure multiplied by the average length of trip measured in the interviews represents almost 11 min, and 686 MUS$/year when applied to the population previously determined and employing the subjective value of time estimated in the present study. This amount represents the externality on the city network users. When both sources of value are considered in the cash flow (WTP of commuters and externality on city network users) the net present value turns highly positive: 25700 MUS$, even ignoring the positive externalities in terms of noise and traffic accidents. This means that the park & ride project in the north access of Buenos Aires is interesting and deserves the attention and thorough study from the authorities. In more general terms, this shows the importance of the knowledge produced in the present article for rational public policy making. The subjective value of traffic accidents determined in this study should be further assessed. On the one hand it is consistent with similar preliminary valuations performed by the authors in other studies. On the other hand there are signs suggesting that the range of rates used in the experiment was too wide. Some individuals may have had a hard time forming a perception about the high rate of accidents, what could explain the decreasing marginal influence on utility. Some other individuals behaved lexicographically meaning that the high rate of accidents overrode the effect of the other variables for them. The square root employed in the model protects the valuation from bias, however it is probably conservative and we would recommend to narrow the range of the rate of accidents in future studies.
16
n Nacional de Regulacio n del Transporte. Comisio
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002
E. Picasso et al. / Habitat International xxx (2016) 1e7
3. Conclusions We present in this article an assessment of the value generated by transport improvement programs in the city of Buenos Aires. The subjective value of lower traffic congestion is measured in terms of travel time, noise, and traffic accidents, by means of a choice experiment. The target population is the people using the traffic network of the city, both habitants of the city and habitants of the suburbs frequently coming to the city. Respondents were stimulated with hypothetical transport improvement programs with specific values for travel time, noise, accidents, and cost, for them to choose. The data was analysed by means of a MXL discrete choice model. The median WTP (willingness to pay) was estimated for each for each variable, as well as its heterogeneity in the population. The subjective value of travel time in the city is 0.38 Ar$/min. The subjective value of reducing noise from high to low level in our perceptual scale is 850 Ar$/Month. The subjective value of eliminating traffic accidents in the metropolitan area is 61 MAr$/Year. These subjective value rates are useful to properly assess the impact of transport improvement programs. We employ this knowledge to complete the evaluation of a park & ride facility, studied in Picasso et al. (2012, 2014). In that opportunity, the commuters WTP for a park & ride facility located in the north access to Buenos Aires city was determined via discrete choice modelling. A stylized discounted cash flow analysis shows that the cost of the project is higher than the amount commuters would be willing to pay. However, there is value for city habitants due to lower traffic congestion, overlooked in that study. Actually, all city network users would benefit from that. We use the subjective value rates measured in the present study to assess the value of the lower traffic congestion. We determine the number of city network users from different sources of information. We estimate the impact on travel time in the city as a consequence of reduced congestion due to the flow headed to the park & ride. This enables the calculation of the value of the project for city network users by means of the value rate previously determined in this study. The net present value of the project turns highly positive when the externality is taken into account, demonstrating the importance of the results obtained in the present study. References Ben-Akiva, M., & Lerman, S. (1985). Discrete choice analysis: Theory and application to travel demand. Cambridge, MA: MIT Press. Bos, I. Ilona D. M., Van der Heijden, Rob E. C. M., Molin, E. J., & Timmermans, H. J. P. (2004). The choice of park and ride facilities: an analysis using a contextdependent hierarchical choice experiment. Environment and Planning A, 36(9), 1673e1686. Dijk, M., & Montalvo, C. (2011). Policy frames of park-and-ride in Europe. Journal of Transport Geography, 19(6), 1106e1119. Duncan, M., & Cook, D. (2014). Is the provision of park-and-ride facilities at light rail stations an effective approach to reducing vehicle kilometers traveled in a US
7
context? Transportation Research part A, 66, 65e74. Federov, V. V. (1972). translated and edited by. In W. J. Studen, & E. M. Klimko (Eds.), Theory of experiments. New York: Academic Press. Hensher, D. A., & Greene, W. H. (2003). The mixed logit model: the state of practice. Transportation, 30, 133e176. Holguin-Veras, J., Reilly, J., Aros-Vera, F., Yushimito, W., & Isa, J. (2012). Park-andride facilities in New York city. Transportation Research Record, 2276(1), 123e130. Huber, J., & Zwerina, K. (1996). The importance of utility balance in efficient choice designs. Journal of Marketing Research, 33, 307e317. August. Louviere, J., Hensher, D., & Swait, J. (2000). Stated choice methods: analysis and applications. New York: Cambridge Univ. Press. Mc Fadden, D. (1973). Conditional logit analysis of qualitative choice behavior. In P. Zarembka (Ed.), Frontiers of econometrics. New York: Academic Press. Mc Fadden, D. (1975). The measurement of urban travel demand. Berkeley, USA: University of California. Mc Fadden, D. (1997). Measuring willingness-to-pay for transportation improvements. Berkeley, USA: University of California. Mc Fadden, D., Talvitie, A., Cosslett, S., Hassan, I., Johnson, M., Reid, F., et al. (1977). Demand model estimation and validation; final report, volume V, urban travel demand forecasting project. Berkeley: Institute of Transportation Statistics, Univ. of California. Mc Fadden, D., & Train, K. E. (2000). Mixed MNL models for discrete response. Journal of Applied Econometrics, 81, 447e770. Meyer, R. K., & Nachtsheim, C. J. (1995). The coordinate-exchange algorithm for constructing exact optimal experimental designs. Technometrics, 37, 60e69. Mingardo, G. (2013). Transport and environmental effects of rail-based Park and Ride: evidence from the Netherlands. Jour. Transport Geography, 30, 7e16. Ortuzar, J. D., & Willumsen, L. G. (1994). Modelling transport (2nd ed.). UK: John Wiley & Sons. ~ a, G., & Mermoz, O. L. (2014). Picasso, E., Bonoli Escobar, M., Stewart Harris, M., Pen Demand elasticity estimation for a park and ride facility by means of discrete n choice modeling. Revista de la Escuela de Perfeccionamiento en Investigacio Operativa, 35, 167e185. ~ a, G., & Mermoz, O. (2012). Estimacio n de la demanda de Picasso, E., Bonoli, M., Pen rico mediante modelos de eleccio n discreta. Anales de un estacionamiento perife n Operativa, 1e14. Article number: la Escuela de Perfeccionamiento de Investigacio 23. Risa Hole, A. (2004). Forecasting the demand for an employee park and ride service using commuters' stated choices. Transport Policy, 11(4), 355e362. Thurstone, L. L. (1927). A law of comparative judgment. Psychological Review, 34, 273e286. Train, K. E. (2009). Discrete choice methods with simulation. Cambridge Univ. Press.
Internet data bases: n Nacional de Regulacio n del Transporte http://www.cnrt.gob.ar/content/ Comisio estadisticas/. Buenos Aires Government budget: http://www.buenosaires.gob.ar/areas/hacienda/ presupuesto2012/index.php. Buenos Aires Government statistics data bank: http://www.buenosaires.gob.ar/ areas/hacienda/sis_estadistico/banco_datos/. Defensoría del pueblo de la ciudad de Buenos Aires: http://www.defensoria.org.ar/ especiales/informesespeciales.php. n Nacional de Vialidad (national transportation office) traffic statistics: Direccio http://transito.vialidad.gov.ar:8080/SelCE_WEB/intro.html. n Nacional del Registro de la Propiedad Automotor statistics: http://www. Direccio dnrpa.gov.ar/portal_dnrpa/boletines_estadisticos.php#.VT-bLCGeDGc. Instituto de Estadísticas y Censos de Argentina: http://www.censo2010.indec.gov.ar/. n (SAIMO): http://www. Sociedad Argentina de Investigadores de Mercado y Opinio saimo.org.ar/socios/Socios/NSE2006-23nov2006-Informe_final.pdf. Strategic traffic plan of the city of Buenos Aires: http://www.cedom.gov.ar/es/ legislacion/poderdepolicia/transito/index4.html.
Please cite this article in press as: Picasso, E., et al., Measuring the externalities of urban traffic improvement programs, Habitat International (2016), http://dx.doi.org/10.1016/j.habitatint.2016.02.002