Ridership Drivers of Bus Rapid Transit Systems

Transcription

1 Ridership Drivers of Bus Rapid Transit Systems David A. Hensher Zheng Li Institute of Transport and Logistics Studies The University of Sydney Business School The University of Sydney NSW 2006 Australia Version September 2011 Abstract We have collected information on 46 bus rapid transit (BRT) systems throughout the world to investigate the potential patronage drivers. From a large number of candidate explanatory variables (quantitative and qualitative), 11 sources of systematic variation are identified which have a statistically significant impact on daily passenger-trip numbers. These sources are fare, headway, the length of the BRT network, the number of corridors, average distance between stations; whether there is: an integrated network of routes and corridors, modal integration at BRT stations, pre-board fare collection and fare verification, quality control oversight from an independent agency, at-level boarding and alighting, as well as the location of BRT. The findings of this paper offer important insights into features of BRT systems that are positive contributors to growing patronage and hence should be taken into account in designing and planning BRT systems. Keywords: Bus rapid transit, daily passenger-trips, patronage drivers, systematic variation Acknowledgment: This study is undertaken as part of the research program LS1 of the Volvo Research and Education Foundation Bus Rapid Transit Centre of Excellence. We thank Corinne Mulley for her comments.

2 Introduction Bus rapid transit (BRT) is an innovative public transport (PT) system, which is defined as a flexible, rubber-tyred form of rapid transit that combines stations, vehicles, services, running ways, and ITS [intelligent transportation systems] elements into an integrated system with a strong identity (Levinson et al. 2003, p.s-1). BRT combines some advantages of both conventional bus and rail modes, such as the flexibility and lower cost of bus, and the speed and reliability of rail. Since the early BRT systems such as the system in Curitiba (open in1974), many cities have embraced BRT as a solution to accommodate increasing public transport patronage, in particular in urban areas. In the transport literature, a number of studies have conducted reviews of BRT systems (see e.g., Hidalgo and Graftieaux 2008; Hensher and Golob 2008; Deng and Nelson 2011). Among these existing BRT review studies, only Hensher and Golob (2008) conducted a formal statistical analysis to comparatively assess BRT systems (e.g., their infrastructure costs and ridership). This paper focuses on patronage drivers, using a comparative analysis to determine which factors systematically affect BRT patronage to offer greater comparative and analytical power relative to traditional literature reviews. Compared to Hensher and Golob (2008), this study has a larger sample (46 vs. 37), includes more recent BRT systems opened between 2006 and 2010 while only BRT systems implemented before 2006 were covered in Hensher and Golob, and takes a closer look at the relationship between a number of characteristics of BRT systems and patronage (e.g., whether there is an integrated network of routes and corridors and whether the system has pre-board fare collection and fare verification). This study is a contribution to the build up of evidence on the features of BRT systems that promote patronage growth and hence should be taken into account in designing and planning BRT systems. This paper is organised as follows. The next section briefly introduces the data. This is followed by the econometric model form, random effects regression, and its advantages over simple regression. We then present the key empirical findings, and discuss how these influence BRT patronage in terms of total system passenger-trips per day. The final section presents key conclusions and recommendations for designing better BRT systems. Data To conduct a comparative analysis on BRT patronage performance, we collected information on 46 BRT systems opened between 1974 and 2010, from a large number of disparate sources including direct contact with current operators, BRT web sites and from specialist groups engaged in BRT system planning (such as Embarq and ITDP-China) from 15 countries (Latin America: Brazil, Colombia, Ecuador and Mexico; Asia: China, Indonesia, Japan, Taiwan, India, and Thailand; North America: USA and Canada; Europe: France and the Netherlands; Oceania: Australia). The number of total system passengers-trips per day has a mean of 372,464 and a standard deviation of 1,276,264. The natural logarithm of ridership is used as the dependent variable in the regression model (shown in Figure 1), with parameter estimates associated with logarithmic explanatory variable being interpreted as direct mean elasticities with respect to fare and headway. Daily ridership per corridor 1

3 (passenger-trips/number of corridors) and daily ridership per system kilometre (passengertrips/length of BRT network) for 46 BRT systems are shown in Figures 2 and 3 respectively. Figure 1: Natural logarithm of total system passengers-trips per day (46 systems) Figure 2: Daily ridership per corridor (46 systems) 2

4 Figure 3: Daily ridership per system kilometre (46 systems) A descriptive profile of the key data items is given in Table 1. In addition to a number of continuous explanatory variables such as fare and total length of the BRT network, we have investigated the role of a number of categorical variables. These include whether the BRT system has segregated busways for bus-only roadways, an integrated network of routes and corridors, enhanced station environment, pre-board fare collection and fare verification, atlevel boarding and alighting, competitively-bid and transparent contracts and concessions, quality control oversight from an independent entity/agency, signal priority or grade separation at intersections, distinctive marketing identity for system, quality control oversight from an independent entity/agency, high-quality customer information, and modal integration at stations (e.g. bicycle parking, taxi stations, and easy transfers between public transport systems). All categorical variables are coded as dummy variables (yes or no) in the regression model. Table 1: Profiles of candidate variables 1 Variable Quantitative variables Unit Mean Standard deviation Fare US$2006^ Total length of BRT network Kilometres Number of existing trunk corridors Number Number of stations Number Average distance between stations Metres Average commercial speed Kilometres per hour Average peak headway Minutes Trunk vehicle length Metre Qualitative variables: whether the BRT system has Percentage of Yes Segregated busways for bus-only roadways 78.3% An integrated network of routes and corridors 52.2% Enhanced station environment 71.7% 1 Only variables that are available in all 46 observations are reported and used in the model, given some variables have missing data, e.g., vehicle capacity, average non-peak headway, total length of existing feeder routes. 3

5 Pre-board fare collection and fare verification 47.8% At-level boarding and alighting 54.3% Competitively-bid and transparent contracts and concessions 26.1% Signal priority or grade separation at intersections 47.8% Distinctive marketing identity for system, 71.7% Quality control oversight from an independent entity/agency 41.3% High-quality customer information 76.1% Modal integration at stations 23.9% ^: All fares are converted into a common currency (US$) and period (2006) Methodology In Hensher and Golob (2008), ordinary least squares (OLS) regression is used to investigate potential sources of systematic variation in BRT patronage. A key assumption of OLS regression is that all observations are independent. However, in this study, there are some cases where multiple BRT systems are located within one country (e.g., 11 systems in China, six systems in the US, five systems in France, four systems in Brazil). Given this, observations within one of those countries could be correlated to some extent, given some common characteristics of a country. To capture this, instead of an OLS regression model, a random effects regression model (equation 1) is used. y a β x u (1) ' it it i it x is a vector of regressors associated with the i th country and t th BRT system; it is a random 2 error term, with E[ it ] 0 and Var[ it ] ; ui is a country-specific disturbance with 2 Eu [ i] 0 and Var[ u i ], also Cov[ it, ui ] 0 ; i represents a country (in this paper, i=1, 2 15), and t is the number of BRT systems located within each country. The way that a random effects regression model works is that it allows each i th country to have a unique disturbance u ; hence within a set of observations drawn from the same i country, the disturbances are no longer independent. The model is estimated by generalised least squares. In this random effects regression model, the number of sampled BRT systems within each country varies from one (e.g., Canada) to 11 (i.e., China). We also investigated the possibility of multicollinearity which can often be a concern when using mean estimates from each of the sampled studies. A popular way to analyse multicollinearity is in terms of the effect of the inter-correlation of the regressors on the variance of the least squares parameter estimates. The variance inflation factor (VIF) (equation 2) is a measure of this effect. VIF k 2 1/(1- Rk) (2) 2 k R is the overall explained variance (R 2 ) obtained when the k th regressor is regressed on the remaining variables. The optimal value for this statistic is 1.0, which occurs when R 2 is zero or this variable is orthogonal to the other variables. There is no consensus on what values of the variance inflation factor merit attention, or on what one should do with the results. Some authors (e.g., Chatterjee and Price 1991) suggest that values in excess of 10 are problematic. 4

6 Sources of Systematic Variation in BRT Ridership The final random effects regression model is reported in Table 2. This model explains 90 percent of the variation in daily passengers-trips of 46 BRT systems, where all parameter estimates are statistically significant at or over the 95 percent confidence level, with the exception of average distance between stations which is statistically significant at the 90 percent confidence level. VIF values are well below 10 on all regressors (see last column of Table 2), and hence we can confidently reject the presence of multicollinearity to ensure the robustness of parameter estimates. The descriptive statistics and correlation matrix for these explanatory variables and the dependent variable (i.e., natural logarithm of total system passengers-trips per day) are given in Appendix A. Table 2: Random effects regression model (Dependent variable: natural logarithm of daily passengers-trips) Explanatory variable Parameter t-ratio VIF Continuous variable Nature logarithm of fare (US$2006) Nature logarithm of headway (minutes) Number of existing trunk corridors Total length of BRT network (kilometres) Average distance between stations (metres) Dummy variable Existence of an integrated network of routes and corridors (Yes) Modal integration at stations (Yes) Pre-board fare collection and fare verification (Yes) At-level boarding and alighting (Yes) Quality control oversight from an independent entity/agency (Yes) Latin America (Location of BRT) Constant Disturbance term effects Country-specific disturbance ( u i ) Random error term ( it ) Sample size 46 Adjusted R In this model, the natural-logarithmic transformation is applied to two continuous variables: the fare variable and headway variable, and given that the dependent variable (ridership) is already in natural logarithm, the double-logarithmic form directly delivers the mean estimates of direct fare elasticity and headway elasticity, which measure the impacts of fare and headway on daily passenger-trips. The estimated fare elasticity is which is substantially higher than the estimate of Hensher and Golob (2008) (i.e., -0.12), but which is closer to common estimates of fare elasticities associated with conventional and bus and rail systems. Hensher (2008) in a meta analysis of 241 observations reports a mean estimate of for fares, which is close to reported in Holmgren (2007) for 81 observations and other reviews such as Goodwin (1992), Oum et al. (1992) and Litman (2002). Price elasticities of PT demand for European countries reported in Nijkamp and Pepping (1998) are between 0.4 and

7 This study estimates a headway elasticity of which is close to , the mean estimate reported in Hensher (2008), calculated based on 21 observations ranging from to This headway elasticity is equivalent to a frequency per hour (=60/headway) elasticity of 0.299, suggesting that a 100 percent increase in frequency would increase ridership by nearly 30 percent, holding other factors constant. Improved frequency has an important role in promoting public transport by reducing waiting time and dwelling time, and consequently reducing uncertainty and anxiety. 2 In addition to Fare and Headway (or frequency per hour), we also indentify other systematic sources significantly influencing ridership. The number of existing trunk corridors represents the capacity of a BRT system. A positive parameter estimate suggests that the increased number of corridors would lead to an increase in ridership, which is as expected, given that increased capacity in terms of trunk corridors would stimulate demand. The length of the BRT network is another dimension of the capacity of a BRT system, and it also has a positive impact on ridership given a positive parameter estimate. The average distance between stations has a negative parameter estimate, which in turn suggests that ridership would be increased through reducing distance between stations (i.e., adding more stations). This finding shows the importance of connectivity in encouraging patronage, i.e., the shorter distance between BRT stations (or the more stations) would improve access and egress. This also translates into a cost effective potential advantage of BRT over other mass transit such as heavy rail, as it is much easier to add a new station in a BRT system both for a relatively low cost and also in terms of design constraints. A number of categorical variables are found to have a statistically significant influence on ridership, providing further insights into the design and planning of BRT systems. Our statistical model suggests that two levels of integration are crucial to ridership, namely between systems (existence of an integrated network of routes and corridors), and at stations (modal integration at stations). A BRT system needs to be integrated with other PT routes to allow for more convenient transit (e.g., door-to-door service) so as to attract more users to public transport. Integration at stations is also important, such as bicycle parking, taxi stations, and easy transfers between public transport systems. At the planning stage, these two levels of integration have to been carefully considered. We found that, all other things being equal, a BRT system equipped with pre-board fare collection and fare verification would attract more ridership. Pre-board fare collection and fare verification would significantly reduce the boarding time, and hence contribute to the reduction in total journey time and time variability, as well as less crowding at stations and reduced congestion amongst buses. These improvements would substantially improve user benefits and consequently increase public transport patronage. This finding is in line with Tirachini and Hensher (2011) who found that the pre-board system is the optimal choice for bus fare collection from a cost-effective perspective. At-level boarding and alighting is negative to the passenger-trip numbers. Although at-level boarding/alighting improves the service level, it reduces seating capacity of a vehicle. Operators (in personal communications in August 2011) have advised us that up to 3 seats are often lost in low floor buses compared 2 An OLS regression model is also estimated (see Appendix B), which delivers a fare elasticity of and a headway elasticity of , both higher than the estimates of the random effects regression model. Given that the random effects regression model (Table 2) is econometrically appealing given the data than OLS regression, the elasticities reported in Table 2 are preferred. 6

8 to higher floor buses, in large due to the design to accommodate wheel chair access and positioning onboard. If there is quality control oversight from an independent entity/agency, the ridership number would be higher, holding other influences constant. This finding highlights how important it is to ensure the service quality of BRT. Our model also finds that the BRT systems operating in Latin America have significantly higher ridership than BRT in other locations, all other factors remaining unchanged. Among the top 10 BRT systems in terms of daily passenger number sampled in this paper, seven systems are located in Latin America. We speculate that an important underlying reason for high ridership of Latin American BRT systems is the relatively higher population density and lower car ownership. For example, urban density in Bogota (Columbia) is around 14,000 persons per square kilometre, compared to approximately 800 persons per square kilometre in Pittsburgh (USA); while the car ownership in Bogota is around 150 cars per 1,000 people, significantly lower than the ownership in Pittsburgh (nearly 700 cars per 1,000 people). Asian cities also have high population density 3 and low car ownership; however Asian BRT systems have a much shorter history (the majority started after 2008), and lower capacity relative to Latin American BRT systems. Most BRT systems in Asia (in particular in China) are still expanding or undergoing active expansion. For example, the initial BRT network in Hangzhou (Line 1 opened in 2006) is 27.2 kilometres 4, and doubling in length to 55.4 kilometres (Line 2 opened in 2008 and Line 3 opened in 2010). The plan is to expand the Hangzhou BRT network to 395 kilometres by Given the findings in this study that the capacity of a BRT system is positive to patronage, BRT patronage in Asian cities is expected to grow substantially in the future. Given the expectation that the economic and spatial base of a metropolitan area has an influence on BRT patronage, we investigated the role of GDP per capita and population density. 5 We ran an additional random effects regression model to examine the influence of GDP per capita (in thousand US$2006) and population density (see Appendix C). As expected, we found that the higher is GDP per capita (linked to higher car ownership), the lower is BRT ridership; and the higher population density supports potentially higher demand for BRT ridership. However these macroeconomic and geographical effects could not be included in the same model as the significant system and locational effects shown in Table 2 because of the high levels of partial correlation (see Table 3), supporting our view that we have adequately captured the macroeconomic and aggregate spatial influences 6. For example, the correlation between GDP per capita and the nature logarithm of fare (LnFare) is , stronger than the correlation between GDP per capita and the natural logarithm of 3 We are only able to identify population density at the city level, not at the BRT corridor or catchment area level. We did investigate the potential role of population density in the ridership model shown in Table 2, however, this very aggregate measure, defined as both continuous and dummy variables (low, medium, high population density), is highly correlated with a number of system variables and consequently was found to be statistically insignificant. We also tested GDP per capita at the city level, which was also not statistically significant, in terms of continuous or dummy variables for the same correlation logic. 4 The passenger-trip number of Hangzhou BRT was collected in Therefore, only Line 1 (27.2 kilometres) is used as its total length. We also used the year that the passenger-trip number was collected in the mode to address the difference in the data collection period, but it is not statistically significant. 5 We have not been able to obtain data on car ownership at the city level for all of the BRT system locations. Also GDP per capita will be highly positively correlated with car ownership, and hence including both would be problematic. 6 The model estimated is not a demand model in the fuller sense of accounting for competing modes and the influence of the socio-economic and spatial context; rather it is a representation of a model designed to identify the potential influence of BRT design, service and fares on passenger trips per day, holding all other possible influences constant at an average level that is captured by the model constant. 7

9 daily passengers-trips (LnRidership). Population density also has a much higher correlation with the length of a BRT network than with the dependent variable. Table 3: Correlation matrix of selected system and context variables LnRidership LnFare BRT network length GDP per capita Population density LnRidership LnFare BRT network length GDP per capita Population density 1 Conclusions We have collected information on 46 BRT systems from 15 countries to investigate the potential patronage drivers. A number of sources of systematic variation are identified which have a statistically significant impact on daily passenger-trip numbers. These sources include: (1) price sensitivity: the estimated fare elasticity in this study is ; (2) frequency of service: the estimated headway elasticity is ; (3) capacity of a BRT system: we found that the length of BRT network and the number of corridors are significantly positive to total system passenger-trips per day; (4) connectivity: the shorter average distance between stations would stimulate the demand for BRT; (5) integration: it is crucial to have integration between systems (e.g., with other PT routes) and at stations (allowing for more convenient transfers), and hence BRT must be treated as a part of system, not an individual corridor; (6) equipment: for example, pre-board fare collection and fare verification would increase the use of BRT, given that it can reduce travel time, variability, crowding and congestion; and (7) quality control to ensure the service level. These findings offer valuable advice on what characteristics of BRT systems really matter to users, which can be used to assist in planning and designing BRT systems to attract more users, especially from cars to public transport 7. BRT has great potential as a sustainable transport system, which can deliver high levels of frequency, regularity, connectivity and visibility for a relatively lower cost than other fixed rail systems, resulting in an attractive value for money outcome for an entire metropolitan area. References Chatterjee, S. and Price, B. (1991) Regression Analysis by Example, John Wiley and Sons, New York. Deng, T. and Nelson, J.D. (2011) Recent developments in Bus Rapid Transit: A review of the Literature, Transport Reviews, 31(1), Goodwin, P. (1992) A review of new demand elasticities with special reference to short and long run effects of price changes, Journal of Transport Economics and Policy, 26, This also requires some appropriate pricing mechanisms such as congestion pricing. 8

10 Hensher, D. A. and Golob, T. F. (2008) Bus Rapid Transit systems: a comparative assessment, Transportation, 34(4), Hensher, D.A. (2008) Assessing systematic sources of variation in public transport elasticities: some comparative warnings, Transportation Research Part A, 42(7), Hidalgo, D. and Graftieaux, P. (2008) BRT systems in Latin America and Asia: results and difficulties in 11 Cities, Transportation Research Record: Journal of the Transportation Research Board, No. 2072, Holmgren, J. (2007) Meta-analysis of public transport demand, Transportation Research Part A, 41(10), Levinson, H., Zimmerman, S., Clinger, J., Rutherford, S., Smith, R. L. and Cracknell, J. (2003) Bus Rapid Transit: Case Studies in Bus Rapid Transit, Transportation Research Board of the National Academies, Washington, DC. Litman, T. (2005) Transit price elasticities and cross-elasticities, Journal of Public Transportation, 7(2), Nijkamp, P. and Pepping, G., (1998) Meta-analysis for explaining the variance in public transport elasticities, Journal of Transportation and Statistics 1, Oum, T.H., Waters, W.G. and Yong, J. (1992) Concepts of price elasticities of transport demand and recent empirical estimates - an interpretative survey, Journal of Transport Economics and Policy, 26(2), Tirachini, A. and Hensher, D.A. (2011) Bus congestion, optimal infrastructure investment and the choice of a fare collection system in dedicated bus corridors, Transportation Research Part B: Methodological, 45(5),

11 Appendix A: Descriptive Statistics and Correlation Matrix for Variables in Table 2. Descriptive Statistics Variable Definition Mean Std.Dev. Minimum Maximum Cases LNPAST Nature logarithm of daily passenger-trips LNFARE Nature logarithm of fare (US$2006) LNHWAY Nature logarithm of headway (minutes) TLETC37 Number of existing trunk corridors NUETC36 Total length of BRT network (kilometres) AVDBS46 Average distance between stations (metres) EXIINYES Existence of an integrated network of routes and corridors (Yes) MODISYES Modal integration at stations (Yes) PRFCVYES Pre-board fare collection and fare verification (Yes) ATLBAYES At-level boarding and alighting (Yes) QCOIEYES Quality control oversight from an independent entity/agency (Yes) LATIN Latin America (Location of BRT) Correlation Matrix for Listed Variables LNPAST LNFARE LNHWAY TLETC37 NUETC36 AVDBS46 EXIINYES MODISYES LNPAST LNFARE LNHWAY TLETC NUETC AVDBS EXIINYES MODISYES LNPAST LNFARE LNHWAY TLETC37 NUETC36 AVDBS46 EXIINYES MODISYES PRFCVYES ATLBAYES QCOIEYES LATIN PRFCVYES ATLBAYES QCOIEYES LATIN PRFCVYES ATLBAYES QCOIEYES LATIN

12 Appendix B: Ordinary Least Squares (OLS) Regression Model Explanatory variable Parameter t-ratio VIF Nature logarithm of fare (US$2006) Nature logarithm of headway (minutes) Number of existing trunk corridors Total length of existing trunk corridors (kilometres) Average distance between stations (metres) Trunk vehicle length (metre) Existence of an integrated network of routes and corridors (Yes) Pre-board fare collection and fare verification (Yes) Modal integration at stations (Yes) At-level boarding and alighting (Yes) Quality control oversight from an independent entity/agency (Yes) Latin America (Location of BRT) Constant Adjusted R Appendix C: Random Effects Ridership Regression Model with GDP per capita and population density Explanatory variable Parameter t-ratio GDP per capita (Thousand US$2006) D Population density (persons/km²) Constant Disturbance term effects Country-specific disturbance ( u i ) Random error term ( it ) 1.0 Sample size 46 Adjusted R