Bus rapid transit systems: a comparative assessment

Abstract

There is renewed interest in many developing and developed countries in finding ways of providing efficient and effective public transport that does not come with a high price tag. An increasing number of nations are asking the question—what type of public transport system can deliver value for money? Although light rail has often been promoted as a popular ‘solution’, there has been progressively emerging an attractive alternative in the form of bus rapid transit (BRT). BRT is a system operating on its own right-of-way either as a full BRT with high quality interchanges, integrated smart card fare payment and efficient throughput of passengers alighting and boarding at bus stations; or as a system with some amount of dedicated right-of-way (light BRT) and lesser integration of service and fares. The notion that buses essentially operate in a constrained service environment under a mixed traffic regime and that trains have privileged dedicated right-of-way, is no longer the only sustainable and valid proposition. This paper evaluates the status of 44 BRT systems in operation throughout the world as a way of identifying the capability of moving substantial numbers of passengers, using infrastructure whose costs overall and per kilometre are extremely attractive. When ongoing lifecycle costs (operations and maintenance) are taken into account, the costs of providing high capacity integrated BRT systems are an attractive option in many contexts.

Appendix

Table A1 Candidate variables

Given that many, but not all, of the candidate influences are nominally scale variables (such as shown in Figures in the text as yes, no, partial, etc.), a technique known as nonlinear canonical correlation analysis (NLCCA) was proposed as a way of considering the best way of scaling the range of levels. The first stage uses NLCCA as a way of quantifying mixtures of nominal, ordinal, and ratio scaled variables all at once, while determining the strength of the relationship between each optimally quantified variable and the (one, in this case) dependent variable. Given the small sample size (in terms of low category frequencies) and missing data, we had to progressively work through subsets of potential explanatory variables.

A solution to the nonlinear CCA problem was first proposed by Gifi (1981), De Leeuw (1985), Van der Burg and De Leeuw (1983). The method simultaneously determines both optimal re-scaling of the nominal and ordinal variables and explanatory variable weights, such that the linear combination of the weighted re-scaled variables in one set has the maximum possible correlation with the linear combination of weighted re-scaled variables in the second set. Both the variable weights and optimal category scores are determined by minimising a loss function derived from the concept of “meet” in lattice theory (see Gifi 1990). A nonlinear CCA solution involves, for each canonical variate, weights for all the variables, optimal category scores for all ordinal and nominal variables, and a canonical correlation.

After NLCCA identifies which variables are statistically significant, and how their categories score, we can do one of two things as stage 2: use the variables in terms of their new rescored scales, or break them into dummy coded (1,0) variables. The optimal scores often show that some categories can be combined. Although this can be explored from the outset through dummy variables (e.g., testing ‘yes’ and ‘partial’ separately and combined), the dummy variable approach results in more variables, and the effect of a single nominal variable (here both ‘yes’ and ‘partial’) is sometimes spread into two variables. Without NLCCA, in order to test all categories of all nominal variables, one would have more than 40 dummy variables (and, given missing values, a sample size much lower). Clearly this is too many variables. Using the new scales obtained with NLCCA is more elegant in terms of measuring the total effect of a single multi-category nominal variable. Importantly, even if one adopts a dummy variable specification as the final model, the guidance offered through NLCCA¹¹ is substantial in reducing the problem to a workable size.¹²

We present the NLCCA results in Table A2 and used this as the basis of selecting the appropriate dummy variable specification for traditional multivariate regression estimation (Table A2) for inclusion with ratio scaled variables. In Table A2, the categories of the nominal variables are optimal in that the resulting variables provide the best linear combination that explains the optimally recoded ordinal dependent variable. It is a closed-form eigenvalue least squares solution. We then use the rescaled variables in ordinary (log-linear) regression.

 Table A2 The NLCCA results used to establish the candidate levels of potential explanatory variables

Table A3  Dummy coded multivariate regression

We can clearly see that there are some very influential design and service delivery features that are linked to the costs of infrastructure, some of them being strictly associative such as operating subsidy, the presence of an independently operated and managed fare system, and entry restricted to prescribed operators. In one sense, these variables are beneficiaries of a particular infrastructure design that limits the number of operators, is designed to support efficient operators who do not require operating subsidy, and has a separate supplier of the managed fare system. Of particular note is the positive parameter estimate for operator entry, suggesting that systems with fewer operators (in many cases a single operator selected by competitive tendering or negotiated performance-based contracting) tend to be those that have more expensive infrastructure per kilometre.

Three design elements have an upward effect on infrastructure costs and two have a downward impact. Higher commercial speeds above 20 kph where this is always the case or partially the case, and signal priority or grade separation at intersections (distinguishing always and partially), result in substantially higher infrastructure costs per kilometre (noting that the dependent variable is a natural logarithmic transformation). At-level boarding and alighting, where it is the only facility in place or where it is partially provided, has a strong downward impact on infrastructure costs.