Transport Analysis Guidance

Road Traffic and Public Transport Assignment Modelling
TAG Unit 3.11.2

January 2006

Unit 3.11.2 (Adobe Acrobat - 387kb)

Contents

1. Road Traffic and Public Transport Assignment Modelling

1.1 Introduction

2. Estimating Changes in Road Traffic Congestion

2.1 Introduction
2.2 Modelling road traffic congestion

3. Road Traffic Assignment Models

3.1 Introduction

4. Public Transport Passenger Assignment Models

4.1 Introduction

5. Assignment Method

5.1 Introduction
5.2 Frequency and schedule-based approaches
5.3 Deterministic and stochastic assignment
5.4 Network representation
5.5 Deciding on an assignment approach
5.6 Summary

6. Generalised Cost Definition

6.1 Introduction
6.2 Waiting time
6.3 Fare structures
6.4 Capacity constraints and crowding

7. Path Building

7.1 Introduction
7.2 Identifying acceptable paths
7.3 Identifying acceptable paths: Method 1 (frequency-based methods)
7.4 Identifying acceptable paths: Method 2
7.5 Path choice
7.6 All-or-nothing assignment
7.7 Simple discrete choice
7.8 Distribution models with 'independence'
7.9 Service frequency model
7.10 Service frequency and cost model

8. Cost Skimming

8.1 Introduction
8.2 Methods

9. The Convergence of Public Transport Passenger Assignment Models

9.1 Introduction

10. The Validation of Public Transport Passenger Assignment Models

10.1 Introduction

11. Public Transport Surveys

11.1 Introduction
11.2 Movement Surveys
11.3 Face-to-face on-board surveys
11.4 Face-to-face at-stop or at-station surveys
11.5 Self-completion on-board surveys
11.6 Self-completion at-stop and at-station surveys
11.7 Electronic ticket machine data
11.8 Summary of the preferred approach
11.9 Other sources of data
11.10 Counts

12. Public Transport Passenger Trip Matrices

12.1 Introduction
12.2 Choosing the general approach to matrix building
12.3 Combined public transport mode trip matrices
12.4 Separate public transport mode trip matrices

13. Modelling the Responses of Public Transport Operators to Changes in Demand

13.1 Introduction

14. Commercial Software

14.1 Introduction

15. Further Information

16. References

17. Document Provenance

1. Introduction to Forecasting Models for Public Transport Schemes

1.1 Introduction

1.1.1 Advice on the overall structure and design of the models which should be used for the appraisal of major public transport schemes can be found in Model Structures and Traveller Responses for Public Transport Scheme (Unit 3.11.1). That Unit also contains advice on modelling of traveller responses.

1.1.2 This current TAG Unit provides detailed advice on assignment methods for public transport models and covers the following topics:

methods of estimating changes in road traffic congestion as a result of a public transport scheme, in Section 2;
where to find advice on the development of road traffic assignment models, how to achieve convergence of these models, and how they should be validated, in Section 3;
the development of public transport passenger assignment models in Sections 4;
assignment methods in Section 5;
generalised cost in Section 6;
path building in Section 7;
cost skimming in Section 8;
convergence and validation of public transport passenger assignment models, in Sections 9 and 10;
the design and conduct of travel demand surveys required for public transport model development, in Section 11;
the creation of matrices of public transport passenger trips, in Section 12; and
the responses of public transport operators to changes in demand, in Section 13.

1.1.3 In section 14 there is Guidance to the key characteristics of the main commercial public transport assignment packages available, as of November 2004.

1.1.4 Advice on the use of the models for forecasting can be found in Forecasting and Sensitivity Tests for Public Transport Schemes (Unit 3.11.4).

2. Estimating Changes in Road Traffic Congestion

2.1 Introduction

2.1.1 Public transport schemes may reduce congestion by attracting travellers from car. Some schemes may increase road traffic congestion through measures to secure priority for public transport over general traffic. And some schemes may have both effects, in which case the net impacts will be the crucial component of the net benefits of the scheme. The only cases where the potential for impacts on road congestion may be ignored are schemes aimed at serving low car ownership areas (and where modal transfers are therefore likely to be small) and which would not impact adversely on the capacity of the road system for general traffic. In all other cases, the impacts of public transport schemes on road traffic congestion should be estimated.

2.1.2 This section provides advice on the methods which should be used to model the effects of major public transport schemes on road traffic congestion.

2.2 Modelling road traffic congestion

2.2.1 As explained in Model Structures and Traveller Responses for Public Transport Schemes (Unit 3.11.1), models for the appraisal of major public transport schemes which are expected to have an impact on road traffic congestion will require a validated road traffic assignment model. The development and validation of such models should meet the criteria set out in the references given in Section 3 below.

2.2.2 A public transport scheme may be designed to take road or junction capacity away from general traffic. For example, the essence of a bus priority strategy may be to reallocate road capacity away from general traffic to buses. Similarly, a light rail scheme could involve allocating priority to the light rail vehicles over road traffic at junctions. In cases such as these, congestion on the road system would increase. In the first case^[1], a multi-class road traffic assignment model may be all that is required, while, in the second example, a road traffic assignment model would be an essential component of a multi-mode transport model.

2.2.3 In all cases such as these, the road traffic assignment model should be sufficiently detailed to model both the road capacity changes required by the public transport scheme and the effects of those changes on road traffic congestion. If the road traffic assignment model is not sufficiently detailed to enable the effects of the scheme to be modelled, it would not be regarded as 'fit for purpose'. Feeding the model output trips and costs, for the do-minimum and with-scheme case, into an economic appraisal program, such as TUBA, will provide an estimate of the impacts on road traffic congestion.

2.2.4 The public transport scheme being appraised may aim to attract car users to use the new public transport service. In these cases, the transfer should be estimated using a fully-specified multi-mode transport model. This should include a road traffic assignment sub-model and demand sub-models which would enable the induced traffic effects of the reduced congestion to be estimated.

2.2.5 Experience has shown that the impacts of the changes in demand by mode, net of induced traffic effects, can be so small, in some cases, as to be indistinguishable from the residual 'noise' in even a well-converged road traffic assignment model. In these circumstances, some approximation is required along the following lines.

2.2.6 The difference between the car trip matrices for the without-scheme and with-scheme cases will give the modal transfer impacts of the scheme, net of induced traffic. This matrix of changes in demand should be scaled up and subtracted from the without-scheme car trip matrix to give an exaggerated estimate of the demand changes on the road system. The resulting car trip matrix should then be assigned to the road network and the assignment model run to normal convergence levels. Feeding the trips and costs output by the assignment, for the without-scheme and the exaggerated with-scheme case, into an economic appraisal program will provide an exaggerated estimate of the impacts on road traffic congestion.

2.2.7 The exaggerated impacts estimated in this way need to be scaled back. However, the relationship between changes in road demand and changes in road congestion are not linear and so a straightforward pro rata approach will introduce some inaccuracy. The shape of the relationship between demand and congestion can be found by repeating the procedure with varying degrees of exaggeration of the differences in road demand caused by the scheme. This will provide a more accurate means of estimating the true scale of the impacts of the scheme on congestion, where the original estimate was indistinguishable from the residual 'noise' in the road traffic assignment model.

2.2.8 If a method for the estimating of road traffic congestion impacts is proposed which differs from that outlined in the preceding paragraphs, the Department will expect to see the justification for the alternative approach at the model design stage.

Footnote:
1. Perhaps single class with public transport pre-loads.

3. Road Traffic Assignment Models

3.1 Introduction

3.1.1 Advice on the development of road traffic assignment models can be found in Chapters 2, 3 and 4 of Traffic Appraisal in Urban Areas, which is Volume 12, Section 2, Part 1 of the Design Manual for Roads and Bridges. Sections 4.4.19 et seq of DMRB 12.2.1 contain advice on the criteria for convergence of these models and Sections 4.4.29 et seq provide advice on model calibration and validation.

3.1.2 The advice provided in Traffic Appraisal in Urban Areas on forecasting should not be followed directly for public transport scheme appraisal. Instead, the advice given in Forecasting and Sensitivity Tests for Public Transport Schemes (Unit 3.11.4) should be used.

4. Public Transport Passenger Assignment Models

4.1 Introduction

4.1.1 The key role of public transport assignment models in demand forecasting is the provision of levels of service, the travel times, distances and costs associated with trips between origin-destination pairs, distinguishing components such as transfer and wait times, and, where relevant, different transport modes. These are used in scheme appraisal and demand modelling.

4.1.2 Capacity problems are generally less important than in highway assignment, but the characteristics of the public transport network make route choice in PT assignment complex.

4.1.3 In simple terms, the key issues in assessing public transport assignment models are:

Assignment method
Generalised cost definition
Path building
Cost skimming

These issues are discussed in Sections 5 to 8.

4.1.4 A key issue which will influence the form of the public transport assignment model is whether trips are allocated between public transport modes at the mode choice stage or at the assignment stage. The key factors which govern the choice of method are (a) the proportion of passengers who use more than one public transport mode and (b) the numbers of different mode combinations used to a significant extent. Although allocation as part of the assignment process will generally be less reliable than an explicit mode choice model, mixed modes are generally easier to handle at the assignment stage, especially if there are a number of mode combinations to consider.

4.1.5 If the allocation of trips between public transport modes is to be made at the mode choice stage, a separate network model is required for each public transport mode included in the mode choice process and at least an implicit network for each combination of modes. If the allocation of trips between some public transport modes is to be made at the assignment stage, a combined network is required which contains all the available public transport modes.

4.1.6 If separate network models are used, the sophistication of the assignment stage may be considerably simplified compared to the combined network approach. Outside London, networks of individual modes in England rarely offer a choice of routes. Within a light rail network, no cases exist where services are offered by different operators in competition with one another. Within the bus network, unless the allocation of passengers between services is of interest, it is usually sufficient to represent the whole network with combined frequencies.

4.1.7 Thus, if separate network models are used, the assignment process is likely to become a simple matter of loading a matrix of single mode trips onto single paths through the corresponding mode network. Some cases may be so simple (for example, a service between an airport and a city centre) that assignment software is not required.

4.1.8 If a combined network model is used, however, the assignment process may be much more complicated, especially if the model is to produce accurate allocations of passengers to the various modes. In a combined network, there is likely to be a significant amount of direct competition between modes and the analyst needs to be sure that the assignment model distributes trips between the modes in a realistic manner.

5. Assignment Method

5.1 Introduction

5.1.1 Within public transport assignment it is important to distinguish between:

Routes along which public transport vehicles travel
Services (lines) provided along those routes, and their characteristics
Passengers travelling from an origin to a destination, which need to make use of services, or combinations of services, to complete their journeys

5.1.2 The two different assignment methods that are in common use in public transport assignment software packages are frequency-based and schedule-based approaches, whilst a further distinction can be made between stochastic and deterministic assignment.

5.2 Frequency and schedule-based approaches

5.2.1 The most fundamental distinction in PT assignment is between frequency- and schedule-based approaches. In the case of frequency-based assignment scheduled times are not considered explicitly, but modellers refer to the line headways, or to their inverse (the service frequencies), from which the name of the approach derives. Therefore it is not possible to calculate explicitly attributes that users consider in relation to individual route options, but only the average values that relate to that line. Frequency-based modelling constitutes the classical approach as it is usually simpler, requiring less input data and less computational power than schedule-based approaches. It corresponds to the classical, steady state approach to user equilibrium road traffic assignment and therefore allows the use of some of the same techniques.

5.2.2 Schedule-based approaches have been developed more recently and are becoming more widely used, partly because of increasing computational power. Schedule-based approaches reflect the actual vehicle arrival/departure times at the time when users make their choices. This approach allows modellers to take into account the dynamics of supply and demand, and calculate the dynamics and variation of level of service attributes.

5.2.3 Compared with frequency-based models, the big advantage of schedule-based approaches is that vehicle loadings can be predicted for specific services at specific points in time. This does not necessarily mean that the schedule-based approach is the best in every situation. In particular, if passenger arrivals and/or vehicle departures are highly variable, frequency-based approaches may give more realistic results, whilst the extra data and calculation efforts of schedule-based approaches may be unnecessary with high-frequency systems.

5.2.4 Despite their theoretical advantages schedule-based methods suffer from a number of practical disadvantages:

Results can be very sensitive to the actual timetable specified, although, in this case, it could be argued that only a timetable-based assignment can give realistic results
It can be difficult to predict the schedule accurately a) for a scheme which does not yet exist or b) for several years into the future
The way in which unreliable services should be handled is not clear
Run times are much higher than for frequency-based approaches. These can be as much as 10 times longer.

5.2.5 Often a passenger at a stop has a choice of lines, referred to as common lines, which will take him directly or indirectly to his destination. The lines may differ in their attractiveness, perhaps due to comfort (rolling stock), the travel time to the destination, the number of changes, the probability of seat availability, perhaps even the fare, etc. and importantly these differences may relate to the operator providing the service. A dilemma frequently encountered by the PT user is whether to take the next vehicle on a relevant line or to wait for a more attractive line. This is referred to as the common line problem. The availability to passengers of precise information, for example via countdown displays or hand-held devices, can go some way to resolving the dilemma. The common line problem relates to both frequency and scheduled-based approaches.

5.3 Deterministic and stochastic assignment

5.3.1 The basic assumption underlying many assignment tools is the user equilibrium principle: 'All used paths are minimum (generalised) cost paths and all paths that are not minimum cost are not used'. In a deterministic user equilibrium (DUE) assignment it is assumed that passengers behave as if they share a perfect and equal perception of the generalised travel costs to their destination and all choose the cheapest option. In the absence of crowding, this assumption collapses into All-or-Nothing assignment.

5.3.2 In deterministic frequency-based assignment multiple-routeing across the acceptable paths between OD pairs is achieved via the service frequencies. In the absence of crowding effects, deterministic frequency-based assignment allocates passengers to acceptable paths in proportion to the relevant line frequencies. In many cases only one acceptable path is considered, leading to an All-or-Nothing assignment. This is attractive in terms of reducing run times and simplifying cost skimming but may prove difficult to validate.

5.3.3 In contrast, stochastic user equilibrium (SUE) assignment recognises individual variations in generalised cost perception. All passengers are still choosing their perceived cheapest option, but this may not be the same for all since they do not necessarily agree on what is the cheapest option. A further random element can be added to the assignment by not only assuming that passenger cost perception includes an 'error term' but also by assuming that the vehicle departure times are slightly random. Both assumptions will lead to traffic being split between more paths than in the deterministic case, which reflects better what happens in real life.

5.3.4 In many cases the level of taste variation amongst travellers and imperfect information are likely to require the use of a stochastic route choice method for frequency-based assignment. For schedule-based methods the minimum cost route will depend on departure time, leading directly to multi-routing behaviour. However, the point about taste variations and imperfect information also applies here and a stochastic method may still be appropriate. Note that not all software offers the option of stochastic schedule-based assignment.

5.4 Network representation

5.4.1 The structure of the network of existing heavy and light rail services can be derived by scaling the track layout from large-scale OS maps (or from a GIS). Details of the services, including frequency, in-vehicle time, stops served, and fares, can be obtained from the operators. For proposed light rail services, the network structure can be obtained by scaling from engineering layout plans, and details of the services, including their frequency, stops served and fares, can be postulated by the system design team. Of considerable importance, though, will be the operating speed. This should be determined by means of a simulation model which takes account of track geometry (curvature and gradients), stop spacing, acceleration and deceleration performance, and cruising speed of the light rail vehicles. For street-running sections of light rail systems, the interactions between road traffic and the light rail vehicles should be reflected in the coding.

5.4.2 The structure of the network covered by existing bus services should be available from the Local Authority or Passenger Transport Executive. The geometry of the network should be derived from the road network coded for the road traffic assignment model. However, the number of nodes may need to be increased in order for the bus stops to be represented adequately. This may cause difficulties if the effects of changes in road traffic congestion are to be transferred from the road network (see Section 2). Consideration should be given to the introduction of dummy nodes in the road traffic model network to ensure compatibility, bearing in mind that this may increase the run time significantly. Details of the services, including frequency, scheduled times, stops served, and fares, can be obtained from the operators.

5.4.3 Of particular importance will be the accuracy with which in-vehicle times are represented. Schedules may not be adhered to, of course, and for bus services in direct competition with the proposed new public transport service, in-vehicle times should be obtained by direct observation.

5.4.4 Once networks and services have been defined, zone centroids should be connected to stops or stations in a way that realistically represents how people access the available public transport services. Model Structures and Traveller Responses for Public Transport Schemes (Unit 3.11.1) offers advice on the design of zoning systems suitable for the modelling of public transport schemes. In that Unit, it is recommended that zones adjacent to rail services reflect the expectation that people would walk to the nearest stop or station whereas for zones further away some other mode of access, principally car, would be used. The times and costs allocated to the centroid connectors should, in principle, reflect the average times and costs for the people in the zone who travel by public transport.

5.4.5 The use of a car to access a public transport service may take two general forms:

kiss-and-ride, where the public transport user is driven to the station and picked up again on the return journey; and
park-and-ride, which may involve either the use of a designated park-and-ride site, parking at stations, or informal parking in streets surrounding stations.

5.4.6 In a base year model, it will generally be acceptable to represent these complexities through the allocation of average access times and costs to the centroid connectors. It will be important, therefore, to ensure that the survey data contain sufficient information for these averages to be calculated. Advice on the modelling of formal park-and-ride schemes as part of the forecasting for the scheme being appraised can be found in Models of Traveller Responses to Major Public Transport Schemes (Unit 3.11.1).

5.5 Deciding on an assignment approach

5.5.1 The choice between frequency- and schedule-based approaches, and between deterministic and stochastic models is driven by practical questions:

Is the PT system operating with high or low frequency?
How punctual is the system?
How regular is the system?
What kind of passenger information is available?
Does the demand vary significantly over the modelled period?
How detailed is the demand information (by day, by hour, or even more specific)?
Does the system experience capacity problems?
How big is the network to be modelled?
Is the network complex, so that regular users behave differently compared to occasional users?
How homogenous is the likely user group? For example, is there a large difference in perception or valuation of travel time?
What are the levels of interchange between services?
How many different sub modes are there?
What fare structures are used and do they differ between services/modes/operating companies?
Is the necessary data available for schedule-based modelling?

5.5.2 High and low frequency services. If services operate with a low frequency, public transport assignment modelling should consider the difference between desired time of travel for passengers and actual vehicle departure and arrival times. If services operate with a high frequency, it is generally sufficient to reflect the desired departure time in the dynamic OD matrix and differences between the desired departure time and vehicle departure time can be treated as a constant, for example half the headway. A commonly accepted threshold to distinguish high and low frequency services is 10-15 minutes. If the service operates with a higher frequency the passenger arrival can be assumed to be uniform, because passengers will often not check the timetable before they start their journey (if one is available in the first place). However, if the service operates with a frequency less than this suggested 10-15 min threshold^[2], travellers will turn up at the station for specific services. From this it follows directly that frequency-based models are less suitable for services that operate with headways larger than the threshold.

5.5.3 Passenger information and service punctuality. The more information a traveller has, and the more reliable this information is, the more the choice will be service- rather than line-based and hence the more valid a schedule-based approach will be. Frequency-based models will be more suitable if services operate with low punctuality and/or a low level of user information. Delays and irregularity have to be treated implicitly or explicitly in schedule-based models. An implicit treatment is possible by adding error terms to the path choice model. A Monte Carlo technique allows the explicit treatment of delays, but as far as we know this is not possible in commercial packages. Where very precise information is available to passengers (e.g. from on-line or real-time services), this can only be modelled with schedule-based approaches.

5.5.4 Service regularity. Service regularity is a separate issue from punctuality. In this case it is the scheduled intervals between the arrivals of the vehicles rather than unplanned delays. Frequency-based models assume an equal share of passengers between the runs of this service. If a service is not scheduled to arrive with regular headways, say 00, 15, 30, 45 after the hour, but say 10, 15, 40, 45 after the hour, this might lead to line loading errors in frequency-based models. Further, a schedule-based approach might be required if there is a major influx of passengers during a certain period (like an underground station connected to a train station that brings a large number of passengers to the underground network once every hour) in order to show overloading of certain services. An additional consideration in such a case is that a schedule-based approach is better equipped to estimate the correct average wait times.

5.5.5 Capacity problems. Congestion in highway assignment and capacity problems in transit assignment are not the same. This is for two reasons. Firstly, the cost function is not increasing continuously, but the finite capacity of public transport vehicles will lead to a step function; either a traveller can board the arriving vehicle or not, in which case the waiting time will increase by one headway. Secondly, capacity problems will only be experienced by boarders. Passengers on-board have priority and do not perceive the same increase in cost, although they may experience some increase in discomfort due to crowding. In frequency based-models it is possible to handle capacity problems implicitly through a concept referred to as effective frequency. The idea is to increase the perceived costs of boarders through a local reduction in service frequency, reflecting the fact that the passenger may not be able to board a vehicle at a particular point because of overcrowding. This approach is implemented in EMME/2 but can be criticised for two reasons: a) A cost increase based on the number of passengers wanting to board and spaces available is still a continuous cost function; b) an increase in cost does not prevent line capacities being exceeded, leading to inaccuracies elsewhere in the network. Additionally, it is not clear how the correct wait time can be extracted for demand response modelling and appraisal. Scheduled based models can treat capacity problems explicitly and the modeller can see which runs suffer from capacity problems.

5.5.6 Scale of network. Because of the more detailed network description and because of the dynamic representation of supply and demand, schedule-based approaches are computational more demanding. Run times may be up to 10 times higher.

5.5.7 Variation in user behaviour. If the variation in user behaviour is an important issue, models using Stochastic User Equilibrium (SUE) assignment are needed. The dispersion factor can be used to model the different cost perception of different travellers. SUE assignment can be applied to schedule- as well as frequency-based models. SUE assignment should also be applied if one wants to reflect the behaviour of occasional users in complex networks. Occasional users might not know about all available routing options and therefore the route choice might not be restricted to the least generalised cost path only. For low frequency services it is of less importance to distinguish frequent and occasional users, firstly because the route choice is in most cases not as complex and secondly because in low frequency services passengers will not often change their path en-route. The SUE models differ in their assignment assumptions. Logit, nested logit or probit models are most common. Where paths overlap significantly and hence path utilities are positively correlated (in practice usually the case), it is advisable to use the nested logit, C-logit or probit model. Logit models tend to be more tractable than probit models.

5.6 Summary

5.6.1 The main 3 types of service to distinguish are:

Low frequency service
High frequency, reliable service, good passenger information
High frequency, unreliable service or low passenger information

5.6.2 Only in the last case is a frequency-based model fully sufficient. However, schedule-based models are very data-intensive both for the base year and for future years. If the service is not yet operating, crude assumptions will have to be made which reduce the advantages a schedule-based approach might have. Table 1 below summarises which assignment models are best, depending on network characteristics and (to a lesser extent) passenger behaviour and the options to be modelled.

Table 1. Summary of recommendations for PT assignment model applicability

		Schedule-based (SB) or frequency-based (FB)	SUE or DUE
Service frequency	High		SUE
Service frequency	Low	SB	DUE
Passenger information & service punctuality	High	SB
Passenger information & service punctuality	Low	FB	SUE
Transfer choice-making by travellers	Pre-trip	SB
Transfer choice-making by travellers	En-route	FB
Regular schedule	Yes
Regular schedule	No	SB
Crowding / congestion	Yes	SB
Crowding / congestion	No
Capacity problems	Yes	SB
Capacity problems	No
Scale of network	Large	FB
Scale of network	Small
Day-by-day variations	Yes	SB
Day-by-day variations	No
Significant dispersion of behaviour	Yes		SUE
Significant dispersion of behaviour	No		DUE

N.B. Blanks indicate that either option is appropriate.

5.6.3 Two other aspects which should be considered in the design of the assignment stage are:

crowding on public transport; and
the effects of road traffic congestion on public transport levels of service.

5.6.4 In principle, if crowding is, or is expected to be, so severe that demand for the mode concerned is, or would be, constrained, some means of representing the costs of the crowding for use in the demand model would be required. In practice, crowding is more likely to be of importance in the allocation of trips between alternative routes through a combined network model than in models of separate networks. Section 6.4 offers advice on how to model crowding.

5.6.5 In principle, the effects of road traffic congestion may be so severe as to affect the operation of, and hence demand for, the public transport. In these instances, in the process of iterating between the demand and supply models, changes in public transport generalised costs should be modelled as well as changes in car generalised costs. In practice, a light rail system may run on segregated track and therefore may be insulated from rising congestion while buses may be caught up in traffic, affecting the relative speeds of the competing modes. A linkage between the road traffic and public transport assignment models is therefore often a useful feature of a combined network model, which can be reflected in an equivalent way in a model of a separate network.

Footnote:
2. Combined headway of all used services for the OD movement.

6. Generalised Cost Definition

6.1 Introduction

6.1.1 Generalised costs are used in the calculation of the utility of paths as perceived by travellers and therefore in determining the assignment of passenger flows to the paths. It is a combination of a number of different attributes of a path with each attribute being given its own weight or coefficient. The coefficients convert components to common units (often monetary) and are chosen to ensure that the relative importance of each component for passengers is reflected. These attributes will normally be a subset of the following list:

Walk access time (from trip origin to PT stop)
Walk egress time (from PT stop to trip destination)
Walk transfer time (between PT stops)
Origin wait time (time spent waiting for first service on path)
Transfer wait time (time spent waiting for subsequent services)
In vehicle time (weighting may vary by mode/vehicle type)
Fare
Transfer penalty (based on number of transfers times a fixed penalty, possibly differentiating between different transfer types)
Distance
Overcrowding

6.1.2 Ideally a PT model would have the capability to model and build each of these measurements into the generalised cost function. However, this is not always the case and commercial software differs widely in the extent to which these are incorporated - see Table 2 (Summary of Commercial Software) in Section 14. The calculation of some specific components of generalised cost is considered below.

6.2 Waiting time

6.2.1 The simplest assumption for the calculation of the mean wait time is to assume that it is half the headway. This assumes that passengers arrive randomly at the stop and that the service is reliable. This may be a reasonable assumption for services with short headways but for long headways it is more realistic to assume that passengers will try to time their arrival at the stop to minimise waiting time. For this reason some packages allow the definition of 'wait curves'. These define the waiting time as a function of headway. An example, taken from the CUBE Voyager help file, is shown in Figure 1. This gives the waiting time as half the headway for headways up to 15 minutes, after which the wait time is capped at 7.5 minutes.

6.2.2 Often it is only appropriate to use wait curves for the first service that is boarded, i.e. the one where the passenger has the most control over the time they arrive at the stop. For subsequent services on the path it may still be appropriate to use half the headway, however long that may be - in effect, this assumes random arrivals by passengers.

Figure 1. Example wait curve from CUBE Voyager

Figure 1. Example wait curve from CUBE Voyager

6.2.4 Recommendation: It is important to distinguish between the wait at the first stop on the path and the wait at subsequent stops. Passengers have some control over when they arrive at the first stop on the path, but their arrival time at subsequent stops is under less control. For service headways up to around 10-15 minutes it is acceptable to estimate waiting time using half the headway for the first and subsequent waits. For longer headways at the first stop it will be necessary to use some kind of wait curve to cap the wait time. It should be noted that when services are irregular (either planned or a result of poor punctuality), half the mean headway is actually an underestimate of the mean waiting time. In this situation it is worth considering using wait curves where the waiting time is greater than half the headway. Reduced waiting times for a given headway can be used to model the effect of improved reliability and better passenger information, although both effects can be hard to quantify.

6.3 Fare structures

6.3.1 There are a number of fare schemes that are used in PT modelling. Fares are converted to time using a value of time parameter in the assignment process. The various types of fare structures include:

Distance-based

6.3.2 This usually consists of a fixed amount (boarding penalty/cost) plus a cost per km:

fare=a + b x distance

where a and b are user-defined. A variation of this definition allows b to vary by distance band.

Stop-to-stop fare

6.3.3 The fare is defined explicitly for each stop-to-stop combination.

Stage-based

6.3.4 For stage based fares, the fare depends on the number of fare stages passed by a trip.

Zone-based

6.3.5 This is similar to stage-based, but depends on the number of zone boundaries crossed.

Travel cards

6.3.6 A travel card typically permits unlimited travel within the area and for the time that the card is valid. Typically travel cards offer significant discounts to off-peak travellers or travellers who need to make many trips during a day. For assignment purposes travel card holders can be modelled as not having to pay a fare. However, a separate model may be required to forecast how many users will choose to purchase the card.

6.3.7 The fare schemes may also be mixed together in some models. In some cases different structures and/or parameters can be used for different modes or submodes in the model. Note here that, if traveller types pay different fares (such as concessionaries) they may need to be treated as different user classes. Also, if fares differ between periods (e.g. off-peak fares) different models need to be constructed for each fare period. The most appropriate commercial package will depend on its ability to reflect the fares structure for the system in question - refer to table 2.

6.3.8 Recommendation: Fares need not be included in the assignment, provided that they do not influence route choice; in these cases matrices of fares can be added to the generalised cost after the assignment and before passing cost matrices to a demand model or appraisal package. Where fares can influence route choice then it is essential to include them in the assignment. It is accepted that the complexity of some fare systems may prevent them from being represented exactly in the assignment model, but the model representation needs to be 'acceptable'. Acceptability can be gauged from whether the assignment model validates or not (see Section 10 below for validation criteria). Fares may not be important for holders of some kinds of travel passes, in which case they would need to be assigned as a separate user class. This is in addition to any other segmentation (e.g. by purpose or car availability) that might otherwise be required by demand modelling or appraisal.

6.4 Capacity constraints and crowding

6.4.1 Section 3 discussed how the assignment package can deal with capacity constraints by locally reducing the effective frequency. Additionally, an overcrowding factor can be applied to the in-vehicle time. For instance the actual journey time might be 20 minutes but with an overcrowding factor of 1.5 this would become 30 minutes. The overcrowding factor represents the additional discomfort and inconvenience to passengers - all passengers and not just boarders. It makes the crowded service less attractive to travellers, and will reduce the general attractiveness of PT in a mode choice model. Care must be taken when skimming for demand modelling and appraisal to ensure that the appropriate times (i.e. with or without the factor) are used.

6.4.2 The overcrowding factor is typically a continuous function of the flow to capacity ratio on the service. The two effects of crowding/discomfort/standing and physically not being able to board a service are not considered separately in commercial packages, i.e. there is no step function to deal with hard capacity constraints of the kind discussed in section 3. This can lead to undesirable biases in the results when capacity constraints are active.

Figure 2. Typical crowding curve

Figure 2. Typical crowding curve

6.4.3 Vehicle capacity, i.e. the number of vehicles that can use a road/track, becomes important when capacity limitations cause delays to PT services. Current commercial software can only model this effect for road-based modes (bus and possibly LRT) but not heavy rail. They usually require a link to a highway assignment model in which PT and private vehicles compete for road space. Congested travel times for PT vehicles can then be fed back into the passenger model. Modelling the effect of capacity constraints in public transport has the effect of making the generalised cost of travel time dependent on passenger flows (for example, more passenger flow, less car traffic, less congestion, faster travel time). This then introduces the problem of iteration and convergence between flows and costs^[3] in exactly the same way as happens with highway assignment models. Hence, unless capacity and crowding are serious issues in the PT system modelled, they are best ignored, although the linkage to road traffic congestion is important in virtually all urban cases.

6.4.4 Recommendation: Crowding. The introduction of crowding has significant practical problems for PT assignment, namely the need for assignment to be an iterative procedure with a consequent impact on run times, the need to achieve convergence, and the need to calibrate overcrowding curves. For these reasons crowding should only be modelled where it is likely to have a significant effect on traveller behaviour or where an effect on crowding is one of the objectives of the scheme. Where crowding is not modelled it is still important to monitor volume to capacity ratios when forecasting to determine whether crowding will become a problem in the future.

6.4.5 Recommendation: Effect of road congestion. For submodes that run on-street and share road space with other vehicles (mainly bus, but some LRT schemes) it is important that journey times in the PT assignment model are consistent with the level of traffic congestion. This will require some linkage of on-street PT mode link times in the PT network to assigned journey times from a highway assignment model (and possibly vice versa). In congested urban situations it may not be appropriate to take bus times from a published timetable. In any case this will not be possible when forecasting and a link to a highway assignment model will be necessary to estimate PT on-street journey times for forecast years. The presence of bus or no-car lanes needs to be taken into account.

Footnote:
3. Currently no guidance exists on the required level of convergence for public transport assignment models. Until experience has been gained with sufficient numbers of congested PT assignment applications, we suggest that the same criteria for convergence are used as in highway assignment (DMRB Volume 12). It might also be useful to relate the changes in generalised cost between iterations to the scale of the impact of the scheme - further discussion can be found in VDM Convergence Realism and Sensitivity (Unit 3.10.4).

7. Path Building

7.1 Introduction

7.1.1 PT path identification consists of three stages:

Identification of least cost paths between specific OD-pairs
Establishing connectivity between PT paths
Selection of acceptable PT paths

7.1.2 The objective of path identification is to find all potentially attractive paths and calculate their cost.

7.2 Identifying acceptable paths

7.2.1 There are two common methods for identifying acceptable paths.

7.3 Identifying acceptable paths: Method 1 (frequency-based methods)

7.3.1 In this method the first step is to identify the shortest path excluding the waiting time for the first service from the generalised cost function. This represents the best option if the passenger can time their arrival at the stop/station to avoid any waiting. In practice however the passenger may well have to wait for this 'best path' and it may be possible to get to the destination sooner by taking an earlier service, albeit one which has a slightly higher journey time. In other words the passenger may choose to take a path with a longer journey time, provided that in return they benefit from a short waiting time. Any path whose generalised cost excluding wait time is higher than the generalised cost of the best path plus the headway of the best path will not be used, i.e. the passenger will reach the destination sooner by waiting for the service on the best path^[4].

7.3.2 The process is illustrated in Figure 3. The best path (excluding the origin wait time) is Path 1. On the other hand the path with the lowest maximum generalised cost (generalised cost excluding waiting time plus headway (headway=maximum waiting time)) is Path 2. Paths 2-4 all have a generalised cost (excluding the initial wait) that is less than the generalised cost of best path plus the headway of the best path; they are therefore considered acceptable paths, i.e. in some circumstances it would be better to take one of these paths rather than wait for the service on Path 1. Path 5 is not acceptable because the generalised cost is too high - it would always be preferable to wait for the service on Path 2.

Figure 3. Identification of acceptable paths (Method 1)

Figure 3. Identification of acceptable paths (Method 1).

7.4 Identifying acceptable paths: Method 2

7.4.1 Minimum generalised cost paths are identified between OD pairs to establish base path costs. A set of paths between each OD pair is then generated using simple network connectivity rules. These represent sets of links a traveller could use to get from an origin to a destination. These may be constrained to some computationally practical maximum number.

7.4.2 These are then constrained further to a set within some cost range, i.e. all paths with a cost a certain amount higher (expressed either as an absolute or percentage difference) than the minimum cost path will be discarded. This may also involve limiting number of transfers and interchanges for example (although these may already be built into the generalised cost function) and the total length of walk segments. This path set may be further reduced at the path choice stage, for instance if less than 1% of flow between the respective origin and destination uses a path it may be discarded for computational reasons.

7.5 Path choice

7.5.1 Where multiple paths are identified some mechanism for allocating flow to each path is required, usually as a function of the generalised cost on each path. Ideally explicit consideration would be given to common/overlapping and parallel paths. The different methods used in commercial software are considered below. Path choice is governed by calculations of 'probability of use' of each of the acceptable paths between OD pairs. As noted earlier a useful distinction can be made between deterministic and stochastic methods.

7.6 All-or-nothing assignment

7.6.1 In an all-or-nothing assignment all flow is loaded onto the single minimum cost route for each OD pair. With frequency-based methods there is therefore no multi-routing. This may be an adequate reflection of reality in some cases, particularly in schedule-based models. In others, e.g. complex urban networks, there is likely to be observed multi-routing which would require a more complex assignment method to model accurately. The all-or-nothing assignment is a deterministic method.

7.7 Simple discrete choice

7.7.1 In these stochastic methods no consideration is made of whether paths are overlapping or in parallel. Only the generalised cost on each path is considered. The following discrete choice functions are used:

Logit: the most commonly used discrete choice model where passengers are distributed over a set of paths according to the absolute difference in cost between them.
Power function (Kirchoff): passengers are distributed over paths according to the ratio of the costs of alternative paths.
Box-Cox: a flexible model form that includes power and logit as special cases.
Lohse: uses the ratio of path costs relative to the minimum cost.
Probit.

7.7.2 In each case the 'spread' of the path choice can be controlled by a user-defined parameter. This determines how strong the preference is for the minimum cost path. This will depend on the level of taste variation among passengers and how complete their knowledge of services is ('errors in perception').

7.7.3 With the exception of probit (which is not actually used in commercial packages) all of the above have theoretical shortcomings regarding their ability to deal with a choice between correlated alternatives. Path utilities will be correlated if, for example, they share a common segment.

7.8 Models with 'independence'

7.8.1 The choice models given above in their basic form do not cater adequately for schedule-based stochastic assignment. Temporal factors are therefore incorporated into the models in order to make them more suited to schedule-based PT routeing. In order to do this, interactions between different connections are defined:

The temporal proximity of the connections with regard to departure and arrival
Perceived journey cost differences between connections
Fare differences between connections

7.8.2 These factors are combined to derive an independence of connection factor which defines the attractiveness of a particular connection relative to all others. They ensure that identical alternatives are assigned same volumes of traffic if no other connections with temporal proximity have an effect.

7.9 Service frequency model

7.9.1 Passengers are assigned to a path according to the frequency of services along available paths, i.e. the probability of using a path is proportional to its frequency. This is a simple approach where travellers are assumed to possess no knowledge of timetables or journey times and take the first reasonable service from the stop.

7.10 Service frequency and cost model

7.10.1 In this extension of the service frequency model the path choice probability is modified to reflect the difference in costs. Passengers are assumed to have some knowledge of the frequencies and journey times of alternative services and will decide whether to take the first feasible service from the stop or wait for a faster one.

7.10.2 Recommendation: in all but the simplest public transport networks travellers between certain OD pairs are likely to be split between different routes and services. Therefore a multi-routing algorithm must be used to reproduce this behaviour. Most path identification methods are acceptable; the crucial part of the algorithm is how the flow is allocated to the used paths. Methods that take into account generalised costs, rather than just frequencies are likely to produce better-validated results. Where there are overlapping routes methods that consider the degree of independence between competing routes should, ideally, be used. However, at the time of writing the best method for doing this (probit) is not available in the commercial assignment packages that have been reviewed. Ultimately the best test of the adequacy of a particular algorithm is its ability to reproduce observed routing behaviour.

Footnote:
4. There is an assumption here that passengers are fully aware of the timetable.

8. Cost Skimming

8.1 Introduction

8.1.1 The skimming of costs from a public transport assignment is important as skimmed costs are used in demand modelling and in economic appraisal (for instance as input into TUBA). Calculating costs along a particular path is straightforward. However, packages differ in the way that the costs on individual paths are combined to provide a single skimmed cost for each origin-destination pair.

8.2 Methods

8.2.1 A number of different methods of skimming costs are available, with some packages offering more than one option:

Costs on minimum cost path; usually available as either total generalised cost or for individual components of cost
Straight average over all used paths; usually available as either total generalised cost or for individual components of cost
Flow-weighted average costs; usually available as either total generalised cost or for individual components of cost
Frequency-weighted average costs; usually available as either total generalised cost or for individual components of cost
Composite cost (logsum where path choice is based on a logit model); available only for total generalised cost, not individual components

8.2.2 The most appropriate skimming method depends on the subsequent use. TUBA recommends that flow-weighted average costs are used. The user should be aware that the skimmed costs may differ substantially between methods and in some cases may not be consistent with route choice. For example, a logit model may be used in route choice but cost skims may be weighted by frequencies.

8.2.3 For input into TUBA it is important to be able to skim the individual components of generalised costs separately, particularly travel time and fares. Moreover, times for business trips will need to be unweighted total OD travel times; for consumer trips they need to be weighted OD travel times, using the weights for waiting and walking time recommended in Values of Time and Operating Costs (Unit 3.5.6).

8.2.4 For input into TUBA and the demand model costs need to be skimmed for each demand segment (e.g. combination of purpose, car availability and/or income). It is worth noting that it is possible to have fewer segments in assignment than in the demand model, provided the demand segments map unambiguously onto the assignment segments.

8.2.5 The results of different skimming procedures may lead to rather different results. Earlier work by Mott MacDonald for the TUBA project suggested that the method used for skimming costs should be consistent with that used to split flow between routes. For instance, if a logit model of route choice is used then the skimmed cost should be the logsum measure. This is particularly important if the scheme being appraised involves the introduction of a new route, either of an existing or a new sub-mode. However, many packages that use a logit model cannot provide logsum skimmed costs. The current TUBA recommendation is therefore to use flow-weighted average route costs. For input to TUBA these need to be separated between fares and time-related costs.

8.2.6 The skims also need to feed the demand model which itself may require skims of individual cost components and apply coefficients that vary by purpose and/or person type. This can be a problem if in the skimming procedure the model aggregates over routes. Any inconsistency here can lead to counter-intuitive results. As with the TUBA requirements, this interface requires further research.

8.2.7 Overall, we recommend that the assignment package's skimming capabilities are assessed before committing to its role in the modelling structure, to ensure that a robust interface between assignment and demand model can be achieved.

8.2.8 A final issue in skimming is the need for, and implications of biased networks. In section 5 we have discussed that, to enable the skimming of levels of service for each of the public transport sub-modes as input to the choice model, the assignment networks may need to be manipulated or biased to favour one or more of the modes. This is also done to obtain consistency between the sub-modal split in the demand and assignment models.

8.2.9 If biased networks cannot be avoided, the analyst should ensure that the amount of bias introduced is as little as possible and recorded. A quantitative assessment should be made of the level of bias introduced, and its acceptability. Care should be taken that the biased networks used in application are also used for model estimation, and in future forecasting, the biases should be retained.

8.2.10 At the moment biased networks are the only alternative to handle the possible inconsistencies between the choice and assignment models. Future research is required to identify how alternative path building algorithms could overcome this problem.

9. The Convergence of Public Transport Passenger Assignment Models

9.1 Introduction

9.1.1 Certain special features of public transport passenger assignment models, as described in Section 6 of this Unit, may change the generalised costs used to build paths through the network and therefore change the levels of demand assigned to the routes, services and vehicles in the assignment process. Where these features are used in a public transport assignment model, a process will be required to ensure that the costs used in the iterative process of path loading and generalised cost estimation converge.

10. The Validation of Public Transport Passenger Assignment Models

10.1 Introduction

10.1.1 The validation of a public transport passenger assignment model should involve three kinds of check:

validation of the trip matrix;
network and service validation; and
assignment validation.

10.1.2 The validation of the trip matrix is covered in Section 12 of this Unit and the other two validation strands are discussed below.

10.1.3 Validation of the network should involve checks on the accuracy of the coded geometry. This can be achieved by overlaying the coded network on a map base, either in paper or electronic form.

10.1.4 Validation of the services should involve comparing the modelled flows of public transport vehicles with roadside counts.

10.1.5 Validation of the assignment should involve comparing modelled and observed:

passenger flows across screenlines and cordons, usually by public transport mode and sometimes at the level of individual bus or train services; and
passengers boarding and alighting in urban centres.

10.1.6 Across modelled screenlines, modelled flows should, in total, be within 15% of the observed values. On individual links in the network, modelled flows should be within 25% of the counts, except where observed flows are particularly low (less than 150).

10.1.7 The validation of assignment models of separate modes should be comparatively straightforward if the network, services and trip matrices have validated satisfactorily. The validation, and subsequent recalibration, of an assignment model of a combined network may be considerably more problematic.

10.1.8 In all cases, wherever possible, a check should be made between the annual patronage derived from the model and annual patronage derived by the operator from revenue records. Precise comparisons may be difficult but may be sufficiently accurate to provide a cross-check on the general scale of patronage.

10.1.9 If the validation fails to meet the required standards, the assignment model may be calibrated by one or more of the following means:

adjustments may be made to the zone centroid connector times and costs;
adjustments may be made to the network detail, and any service amalgamations in the interests of simplicity may be reconsidered;
the in-vehicle time factors may be varied;
the values of walking and waiting time may be varied;
the interchange penalty may be varied;
the parameters used in the trip loading algorithms may be modified;
the path building and trip loading algorithms may be changed; and
the demand may be segmented by person (ticket) type.

10.1.10 The above suggestions are generally in the order in which they should be considered. However, it would be hard to argue that the priority between suggestions, which are adjacent in the above list, should be maintained in all instances. In all cases, any adjustments must remain plausible.

11. Public Transport Surveys

11.1 Introduction

11.1.1 Many major public transport schemes in areas of established development will derive a major part of their patronage by extraction from existing public transport services. Sound information about the patterns of movements served by existing services can therefore be very important for the robustness of the appraisal of the proposed scheme.

11.1.2 As explained in Model Structures and Traveller Responses for Public Transport Schemes (Unit 3.11.1), an essential part of the development of a model for the appraisal of all major public transport schemes, except a bus priority strategy, will be a well-validated public transport passenger assignment model. A well-validated model requires a realistic representation of the network of services and good quality trip information. In most cases, the base year trip matrices should be developed from observations of travel movements with the minimum of trip synthesis, certainly in the corridors most directly affected by the proposed scheme. The following sub-section provides advice on the types of survey that may be used to obtain data on movements by public transport passengers.

11.2 Movement surveys

11.2.1 The information required about passenger movements for an assignment model is: origin and destination address; access mode to public transport, including any costs; time of travel; and preferably also type of ticket.

11.2.2 Other information, such as trip purpose and car availability for the journey, will usually be required for the development of the demand model. While other sources, such as a household interview survey, may provide this information, it will normally be cost-effective to collect this additional information in passenger movement surveys. Advice on the sources of data for demand model calibration can be found Variable Demand Modelling - Scope of the Model (Unit 3.10.2). This sub-section is primarily concerned with the data required for an assignment model but mention is made of information for demand model development where appropriate.

11.2.3 The main sources of information about passenger movements are as follows:

interviews with passengers, which may be by means of face-to-face interviews with passengers, either on-board the public transport vehicles or as they wait to board at stops and stations, or by means of self-completion questionnaires; and
electronic ticket machine data.

The relative merits of these sources in providing information the required information are as follows.

11.3 Face-to-face on-board surveys

11.3.1 Face-to-face interviews with passengers on-board, providing the sample size is adequate, should provide the best quality data. The sample can be selected by the interviewer rather than being self-selected and it is easier to ensure that all the required information is provided. This method will enable origin and destination addresses to be provided, along with all the other information that may be required, such as access mode (and trip purpose and car availability, if required). A sample of interviews should be obtained on each service in the modelled area, thereby enabling a complete picture of travel to be obtained (once the sample data have been expanded to passenger counts). However, on-board surveys can be difficult on vehicles which are over-crowded.

11.3.2 In face-to-face interviews, the data can be collected from illiterate and, in the presence of guardians, juvenile travellers, which is not the case with self-completion surveys. In addition, non-response biases are reduced compared with self-completion surveys. The method also allows the surveyor to probe for particular information so that exact details of origin and destination can be discovered. And the surveyor is able to inspect and confirm the ticket type being used by each respondent.

11.3.3 One of the biases that can arise from this method relates to length of journey - the longer the journey, the more likely that a passenger is to be sampled. To counter this bias, those travellers who are making only short hop journeys should have an equal chance of being sampled as those making longer journeys. To do this, the surveyor should select respondents by choosing a random passenger from those boarding at each stop. This can create a practical difficulty in terms of productivity, in that if no passenger boards at a stop then no interviews should be conducted. At off-peak and on rural routes, this can lead to the surveyor being idle for some time until a passenger boards the vehicle.

11.3.4 Because more passengers will board at each stop during peak hours than off-peak, surveyors will each be able to interview a smaller proportion of travellers during busy periods compared with off-peak hours. This imbalance can be addressed by conducting an increased number of survey shifts during peak hours and by weighting data according to passenger counts. Response rates can be as low as 10% during peak times and as high as 70% off peak, although these figures are highly variable depending on passenger loads. The important consideration is that the sample size during both the peak and off-peak periods should be sufficiently large to support the creation of acceptably reliable trip matrices for these periods.

11.3.5 The main practical difficulties relate to interviewing when vehicles are over-crowded. In these situations, surveyors might find it very difficult in observing who boards at each stop or station, making selection of a random respondent impossible. Where such observations are possible, then there may be difficulties in moving through the vehicle to reach the randomly selected respondent. There are health and safety concerns for surveyors in administering the interview while the vehicle is in motion - there are difficulties for the surveyor in recording answers while being able to hold tight to supports. The issues of observation, moving through the vehicle and safety are most problematic with buses, particularly double-decker vehicles. Because of these difficulties, this method may be unsuitable for data capture at peak times when a self-completion survey may be preferable.

11.3.6 Consideration must be given to the choice of bus and public transport routes to be sampled. Ideally, each public transport route serving the modelling area should be surveyed at least once. These should be surveyed over the whole period for which the model will apply. It should also be noted that it is often difficult to establish which bus routes are operating in an area at any one time, particularly for the majority of the UK where bus routes are deregulated. It can often be impractical to sample all bus and public transport routes covering an area at all times, in which case choices have to be made as to which routes to survey. This can lead to biases where respondents who use low volume services are under-represented, such as those using night services or those with low passenger numbers.

11.3.7 This method can only be conducted with the permission of the bus or train operating companies.

11.4 Face-to-face at-stop or at-station surveys

11.4.1 Face-to-face interviews with passengers at stops and stations can provide the same scope and quality of information as on-board interviews. However, interviews are required at all stops and stations for a complete picture of travel in the modelled area to be derived. This is often uneconomic and so surveys of this kind are best targeted at major generators or attractors which, in practice, means town and city centres. It is important that all stops and stations in the target area are included in the survey. If the operators will not allow surveys to be conducted on board their vehicles, the modeller will have no option but to rely on a combination of at-stop or at-station surveys and electronic ticket machine data.

11.4.2 While interviews can, in principle, be conducted with boarders and alighters at bus and light rail stops or train stations, people alighting are likely to be much less willing to be interviewed than those waiting for a service to arrive. As with on-board surveys, the data collected from each interview are of better quality compared with data collected via self-completion surveys.

11.4.3 As with on-board face-to-face surveys, greater proportions of off-peak travellers may be interviewed than peak travellers. The important consideration is that the sample size during both the peak and off-peak periods should be sufficiently large to support the creation of acceptably reliable trip matrices for these periods.

11.4.4 Unlike on-board surveys, issues of long and short journey biases do not apply in that those making short hops are as likely to be approached for interview as those making longer journeys.

11.4.5 However, the data collected by this method is likely to be biased in that those passengers that arrive at the stop or station immediately before the vehicle arrives ('runners') will be under-represented. Where there are infrequent regular services, such as on trains, then this can lead to biases between regular and in irregular passengers. Regular users of the service may time their arrival to the station more closely to the departure time. This bias may not be as prevalent on regular services where departure times cannot be easily predicted, such as for most urban bus routes.

11.4.6 Interviews at bus stops can be less productive than on board the vehicles. The number of passengers who use any one bus stop during the day is likely to be lower than the number who use a certain bus vehicle, so surveyors are more likely to be productive when travelling on the bus. Furthermore, passengers alighting from buses are relatively unlikely to stop to participate in a survey - particularly if they are on time-critical journeys such as on their way to work. Passengers waiting at bus stops are unlikely to agree to be interviewed when a bus can be seen approaching the bus stop as they may not risk missing the bus for the sake of participating in the survey. The same applies when the bus is at the bus stop picking up passengers. Once the bus has pulled away from the stop, there will not be anyone who uses the bus route for surveyors to interview. So for a substantial time of each day surveyors will be unable to conduct interviews.

11.4.7 Similar difficulties can arise when conducting surveys on train station platforms in that passengers will be unwilling to participate when trains are close to arrival or when at the platform.

11.4.8 An alternative to platform surveys are entry or exit surveys at station concourses, where every second or third passenger entering or exiting the station is interviewed. However, when passenger flows are high, such as in the morning and evening peaks, it can be impossible for surveyors to stop anyone even to ask whether or not they are willing to participate in the survey. This method is also liable to under-sample those people on time-critical journeys such as commuters or those on employer's business as they will be less willing to participate. For this reason entrance and exit surveys are highly impractical for morning peaks. One may consider capturing peak travellers only during the evening peak when their return journeys home are less time-critical and when it may be permissible to ask them about details of all journeys made that day, including morning peak journeys. Although this method may allow easier collection of data about morning journeys than during the morning peak itself, it might introduce biases against those who use different modes or routes between travel to and from work. It also introduces biases against those on time-critical journeys during evening peaks, such as those on evening shift work.

11.4.9 Because of these variations in survey rates according to passenger flows, response rates could be vary from very few at peak times to almost all passengers waiting at quiet stations off-peak.

11.4.10 As with on-board surveys, data collection should be conducted for the entire period for which the transport model will apply.

11.4.11 Whereas permission is often not required for surveys at bus stops, it will be required for platform or station entry or exit surveys.

11.5 Self-completion on-board surveys

11.5.1 Self-completion surveys should be regarded as a means of last resort. While, in principle, the scope of the information which may be gathered by this means is as comprehensive as that obtained by face-to-face interview, the quality is likely to be poorer in a number of respects. First, some information may not be provided. Secondly, some information may be inaccurate if the questionnaire is completed some time after the journey was made. And, thirdly, the sample of respondents will be self-selected and may therefore not be adequately representative of all travellers.

11.5.2 The method involves the distribution of self-completion questionnaires to passengers as they board vehicles. Often these questionnaires would be collected from passengers as they alight from the vehicle.

11.5.3 Although forms should be distributed with reply-paid envelopes, in order to maximise response rates, it is recommended that measures be taken to ensure that forms are completed during the journey and returned to the surveyors when alighting. For this reason survey forms should be very brief and uncomplicated. Pens and pencils should be distributed to boarders along with the survey forms.

11.5.4 The main advantage of the self-completion survey compared with face-to-face interviewing is that all passengers have an equal likelihood of being given a survey form, regardless of whether they are making long or short journeys and whether travelling during peak or off peak times. However, return rates do vary due to self-selection biases.

11.5.5 Another principal advantage is that self-completion forms can be distributed and collected even on crowded vehicles. For buses with separate entrance and exit doors, it is often recommended that two fieldworkers are used - one to distribute forms to boarders and the other to collect forms from alighters. For trains, different methods of distribution and collection will be necessary depending on the internal design of rolling stock. On very crowded vehicles, self-completion surveys may be the only option available by which the data capture could be undertaken.

11.5.6 There are two main disadvantages with this method - the first is that it introduces self-selection biases and the second that the amount and accuracy of information collected for each respondent is limited.

11.5.7 Self-selection biases arise from some passengers being unwilling to record their details on survey forms. As a result, self-completion surveys may suffer from biases towards older respondents, those with higher social segmentation standing and those with higher levels of education. There may also be biases whereby those most interested in the survey topic will be more willing to participate - in the case of a public transport survey it is likely that those with the keenest interest in public transport such as frequent users will be more likely to respond. In order to minimise these self-selection biases then methods should be used to maximise response rates, such as by simplifying survey forms and maybe offering incentives.

11.5.8 If incentives are used, it should be recognised that these could introduce their own biases. Incentive prizes of free travel cards will encourage frequent travellers to participate more than infrequent travellers. Cash prizes will be more attractive to those with lower incomes.

11.5.9 One particular self-selection bias that can be prevalent is that those on short hop journeys are less likely to complete a questionnaire than those on longer journeys. In order to reduce this it is necessary that forms are kept as brief as possible.

11.5.10 Because survey forms should be kept simple in order to encourage participation, there is a limited amount of information that can be collected. Survey forms should not exceed more than one side of A4. This should be sufficient to establish origin and destination details, journey purpose, ticket type used, access and egress mode, age and gender. Where greater detail is required for complicated segmentation, a self-completion survey may not be appropriate.

11.5.11 It should also be remembered that individuals may not accurately transcribe the required data. For example, several respondents will record that origin or destinations in general terms such as "home" or "school". Ticket type details will often be misrecorded.

11.5.12 Where self-completion surveys are collected from, as well as distributed to, passengers, return rates of forms of up to 95% can be achieved. However, large proportions of these forms will be incomplete. Response rates to individual questions in bus surveys can vary from between 40% and 70% of all respondents answering a particular question. Response rates to questions may be higher on trains.

11.5.13 The same issues regarding which routes and times of day to survey apply to the self-completion on-board survey as to the face-to-face interview on-board survey. Similarly, permission is necessary from operating companies for surveyors to work on board vehicles.

11.6 Self-completion at-stop and at-station surveys

11.6.1 This method is conducted by distributing self-completion forms to passengers as they enter or exit stations or as they wait at or alight at bus stops.

11.6.2 The main advantages of this method compared with an at-stop face-to-face interview is that more passengers are likely to accept survey forms than would be willing to participate in an interview. This will apply to both those who arrive at stops or stations just before the train/bus arrives and also to those on time-critical journeys. Bus passengers waiting at stops would be more likely to accept a survey form even if their bus is approaching or already at the bus stop.

11.6.3 Even so, there will be times of day at the most popular transport interchanges when it will be extremely difficult for surveyors to distribute survey forms. At certain levels of passenger flow, individuals will develop a mindset in which they will not be distracted from their task of moving through the station, even to collect a survey form.

11.6.4 As with on-board surveys, at stop and at station surveys will suffer from inaccuracies and limitations to the data that can be collected. Because respondents will be required to return self-completion forms by post, rather than having them collected at journey's end, response rates will be much lower than for an on-board self-completion survey. As a consequence, self-selection biases will be much stronger than for an on-board survey. Depending on the survey materials, return rates of distributed questionnaires could be as low as 15% to 25%. Furthermore, some questions will not be answered.

11.6.5 Again, sampling biases may occur depending on which stations and stops are surveyed and what times of day are covered. Litter generation of discarded survey forms can be a significant problem. Permission must be sought from station managers to distribute survey forms within stations.

11.7 Electronic ticket machine data

11.7.1 Electronic ticket machine (ETM) data provide information on all journeys rather than a sample. ETM data provide time of travel and can be obtained for long periods of time, thereby avoiding day-to-day variations. They can also relate to the network of services as a whole. However, ETM data will only provide trip records in terms of fare stages at which passengers board or alight, and fare stages may differ between different operators. Moreover, trips involving an interchange may be recorded for each leg separately if separate tickets are bought for each leg. ETM data include no information about the traveller or the purpose of journey - while the former may be important for an assignment model, the latter is unlikely to be required for that purpose. In areas where the use of travel cards, concessions and other pre-paid tickets are prevalent, ETM data may provide a less accurate picture of passenger movement, despite drivers being required to record all passengers using their bus. Further information is available in Data Sources (Unit 3.1.5).

11.8 Summary of the preferred approach

11.8.1 The following table sets out the main advantages of each survey method, along with biases that might occur, the practical difficulties of the method and likely response rates. Details of when each different type of method are most appropriate are also shown.

Method	Advantages	Potential Bias	Practical Difficulties	Response Rates	Appropriate Use
On-board face-to-face survey	Individual interviews of high quality	Fewer short-journey passengers, fewer peak passengers	Impossible to administer on crowded vehicles	From 10% on crowded vehicles to 90% on quiet services	Not busy buses, particularly in peak times
At bus stop surveys; on-platform surveys	Individual interviews of high quality	Fewer 'runners', fewer peak passengers, no passengers who use those stops not surveyed	Unproductive, expensive if administered at all bus stops	From 5% at peak times to 60% off peak of those visiting survey bus stops	When there is no permission for on-board interviews, where it is acceptable that some stops are not surveyed. Where attitudes towards stop facilities required
Entry/exit surveys at stations	Individual interviews of high quality	Fewer time-critical journeys, fewer 'runners', fewer peak passengers	Very difficult to recruit at high passenger flows	From very few at peaks to 90% off-peak	Not during morning peak hours at busy stations
On-board self completion survey	Highly productive	Self-selection biases. Some response bias against short-hop passengers	Poor quality and completeness of data, limited scope of data	Up to 95% of survey forms returned but as low as 40% of questions answered	On highly crowded services, where only limited data required
At station self-completion survey	Productive	Self-selection biases, fewer time-critical journeys	Difficult to cover all bus stops on network, Poor quality and completeness of data, limited scope of data	Between 15% and 25% returned, with about 70% of questions answered	Where no permission for on-board survey, only limited data required, where it is acceptable that some stops are not surveyed

11.8.2 In summary, the best approach is for an adequate sample of face-to-face interviews to be conducted on-board a sample of public transport vehicles on each service in the modelled area. However, operators may not allow interviews to be conducted on their vehicles.

11.8.3 The next best approach is to conduct face-to-face interviews at stops and stations. While comprehensive coverage of Network Rail stations may be affordable, it is likely that, to be cost-effective, surveys at bus stops would have to be confined to particular areas. This means that some other source of data will be required to supplement the at-stop surveys.

11.9 Other sources of data

11.9.1 Three further sources of information on passenger movements should be mentioned here:

Census journey to work data;
household interview surveys; and
CAPRI data.

11.9.2 Use of the 1991 Census data would not now be appropriate. The 2001 data should be used only with caution, for two main reasons. First, this data source only includes trips between home and work and does not provide any information about trips for other purposes. Secondly, the correspondence between the trips recorded in the Census and those that actually take place is not direct. Nevertheless, the Census journey to work data may be of some use in developing a public transport model, providing that it can be supplemented by data from other sources.

11.9.3 Household interview survey sample sizes are rarely sufficiently large to provide acceptably accurate estimates of trips between pairs of public transport model zones. Therefore, data from this source should not generally be used directly in the creation of trip matrices at a zonal level. However, household interview surveys are a rich data source, in the sense that actual trip-making behaviour by all modes can be linked to the characteristics of the household and travellers. This data source is useful, therefore, for demand model estimation. It can also provide large area controls on the numbers of trips derived from other sources.

11.9.4 The Department has developed a set of national rail passenger trip matrices from the CAPRI (Computer Analysis of Passenger Rail Income) data. However, these matrices may not give a true picture of rail travel particularly where travel card use and free concessionary travel is prevalent. Further information is available in Data Sources (Unit 3.1.5).

11.10 Counts

11.10.1 Numbers of passengers using the public transport system are required for:

expanding interview samples;
use as constraints in matrix estimation; and
validation of trip matrices and assignments.

11.10.2 The counts required for expanding interview samples will depend on the method used to survey passenger movements.

11.10.3 Depending on passenger flows, passenger counts can be combined with on-board surveys. Often, on board face-to-face surveyors would be able to count the numbers of boarders and alighters at each stop before selecting respondents for interview. However, they would not be able to do so on crowded vehicles nor where each individual interview is likely to over-run the time between stops.

11.10.4 Where self-completion questionnaires are distributed, counts of the number of survey forms distributed and received at each station can be recorded along with the number of refusals. Surveyors would have an easier task of recording these details compared with surveyors who need also to conduct on-board interviews.

11.10.5 Even though tasks of distributing survey forms and conducting counts can be combined, where passenger counts need to be highly accurate then it may be necessary to conduct dedicated counts with on-board fieldworkers. Where dedicated counts surveys are conducted, electronic recording of numbers of boarders and alighters using palm-top computers will provide information not only of numbers but also allow an accurate record to be kept of times.

11.10.6 For multiple door vehicles such as trains, there may be substantial differences in passenger numbers counted at different sets of doors. Doors near station entrances will have a higher number of boarders than more remote doors. Because of this, an accurate level of count detail will only be available if counts are conducted on each set of doors. Experience shows that one surveyor can accurately conduct counts at only one set of doors during peak hours.

11.10.7 Electronic ticket machine data provide the numbers of people (a) using each section of each service, and (b) boarding and alighting at each stop or station. Thus, where this information is available, the need for counts by other means may not be very great.

11.10.8 Estimates of the numbers of passengers on public transport vehicles made by observers standing at the roadside will not, generally, be sufficiently accurate for any of the purposes listed in paragraph 11.10.1.

12. Public Transport Passenger Trip Matrices

12.1 Introduction

12.1.1 Matrices of public transport trips are required for development of a public transport passenger assignment model. For this purpose, they will be required for each time period which is to be modelled separately. They may, in certain circumstances, be required by ticket type (for example, travel card and others) or person type (for example, concessions and others). Matrices for assignment will be required in origin-destination format.

12.1.2 Matrices may also be required for the calibration of mode and destination choice models. For this purpose, the matrices will be required by trip purpose and by car availability. Depending on the structure of the demand model, they may be required for a 24-hour period or for peak and inter-peak periods. Matrices for demand model calibration will be required in production-attraction format. Advice on the development of trip matrices for demand model calibration can be found Variable Demand Modelling - Scope of the Model (TAG Unit 3.10.2). The remainder of this sub-section concentrates on matrices for the assignment model.

12.2 Choosing the general approach to matrix building

12.2.1 Typically, information for the construction of trip matrices will come from up to six sources:

new face-to-face interview (or, undesirably, self-completion questionnaire) travel demand surveys;
electronic ticket machine data;
pre-existing models if suitably up-to-date;
infilling of movements not available from the previous three sources using a gravity model;
controls on the synthesised trip ends using data from a household interview survey; and
matrix estimation with the aim of ensuring an adequate validation of the final matrix against passenger counts.

12.2.2 The following considerations relate to each of these sources.

New interview or questionnaire travel demand data should enable the maximum level of segmentation required.
Electronic ticket machine data includes no information on which to base segmentation and the data will need to be factored using information from another source, such as interview data.
The segmentation of matrices from a pre-existing model may or may not be sufficient for the purposes of the new model.
Infilling partial trip matrices using a gravity model may be more difficult with sparser prior matrices, in the sense that sensible model parameters may be harder to achieve. Thus, it may be preferable to carry out this process at a level more aggregate than the full level of segmentation required.
Controls from household interview data should be available at whatever level of segmentation is required, except perhaps ticket type (travel card or other) and person type (concession or other) unless the survey has been conducted specially for the development of the public transport model.
Matrix estimation has to be done at the most aggregate level, although matrix estimation software may allow input proportions of trips by type to be preserved in the adjusted or output matrix.

12.2.3 Where a data source cannot support the level of segmentation required, information from another source may be used to split the data from the first source in the required manner.

12.2.4 At the outset of the matrix building process, it is important to consider:

whether the trip matrices are required for the demand model calibration and, if so, what level of segmentation is required - matrices for this purpose will be required in production-attraction format;
whether the trip matrices are required for the assignment model alone, in which case segmentation by time of day and, possibly, ticket type (travel card or other) and person type (concession or other) - matrices for this purpose will be required in origin-destination format;
the level of segmentation that is to be used in processing each data source; and
the need for subsequent splitting or aggregation of trip matrices to yield the final matrices that are required.

12.3 Combined public transport mode trip matrices

12.3.1 In creating, from a variety of sources, a matrix of trips which use all modes of public transport, it is quite possible that estimates of some movements will be available from more than one source. Consideration should therefore be given, at the outset, to the order in which the data sources will be used. In general, data from direct interviews should be used in preference to other sources. The choice between ETM data and data from a pre-existing model is less clear cut and will depend on the age and source of the data in the pre-existing model.

12.3.2 In assembling the partial trip matrices from sources of observed data, the following principles should be followed:

only movements between an origin and a destination which have been fully sampled should be included in the matrix building process;
corrections should be applied to account for movements which have been fully sampled more than once, either in the same data set or in more than one data source, giving greater weight to the more accurate data; and
corrections should be applied to ensure that all the data relate to the same year and the same day, or average day in the same week, of that year.

12.3.3 If the matrices derived from the sources of observed data are complete, they should then be validated by assigning them to the combined public transport network. Matrix level validation should involve comparisons of assigned and counted passengers across complete screenlines and cordons (as opposed to individual services, which are dealt with in the validation of the assignment model - see Sub-section 10). At this level of aggregation, the differences between assigned and counted flows should in 95% of the cases be less than 15%.

12.3.4 If the matrices do not validate satisfactorily, matrix estimation may be used to adjust the trip matrices to accord more closely with the validation counts. The changes brought about by the matrix estimation process should be examined to check for particular distortions. If distortions have been introduced, the count data being used as constraints should be checked for consistency. Use of matrix estimation is most satisfactory when the adjustments required are small refinements - that is, when the process is almost not required. Use of the technique is likely to be most suspect when it produces large changes from the prior trip matrix, especially if the changes are, in proportional terms, uneven across the matrix.

12.3.5 If the matrices derived from the sources of observed data are not complete, the missing trips should be either derived from a pre-existing model or synthesised using the partial matrix gravity model technique. Advice on the partial matrix technique can be found in DMRB 12.2.1 paragraphs 4.3.17 to 4.3.27. Matrix estimation is not a suitable technique for infilling a partial trip matrix.

12.3.6 Data from a pre-existing model should only be used in the area of prime interest if they are up-to-date. For parts of the matrix away from the centre of interest, data from an older pre-existing model may be used, providing that the data, after factoring to allow for the changes in demand levels over time, are considered to provide a reasonably accurate picture of movements in the areas concerned. It is likely that data from a pre-existing model will be coded to a coarser zoning system than that used for the public transport scheme appraisal. The necessarily approximate nature of the disaggregation process should be borne in mind when considering data from this source.

12.3.7 The partial matrix gravity model works best when the proportion of cells with missing non-zero numbers of trips is low, typically less than 30%. If the partial matrices are sparse, with data missing in more cells than present, consideration should be given to either using data from a pre-existing model or conducting more surveys.

12.4 Separate public transport mode trip matrices

12.4.1 The principles set out in the previous section apply generally also to the creation of matrices for each public transport mode separately.

12.4.2 Where the modes are quite distinct, that is, mixed mode trips are few in number, matrices can be derived straightforwardly from data collected on each mode.

12.4.3 Where mixed mode trips are prevalent, it is not appropriate, in principle, to use a model in which the allocation between the public transport modes is done at the mode choice stage and separate assignment models are used for each mode. However, if such an approach is chosen, the appropriate leg of the mixed mode trips would need to be included in the matrices for all the modes used if the assignments are to match the counts. Thus, in these cases, corrections for multiple sampling between data sources will not be required.

12.4.4 The alternative approach is to identify, for each mixed mode trip, the main mode. This can be done according to a rule such as: the mode on which the highest proportion of total in-vehicle time is spent; or according to a hierarchy. For example, if heavy rail is used the trip would be classified as a rail trip; then for the remaining trips, if light rail is used the trip would be classified as light rail and otherwise as bus. The drawback with this approach, however, is that, apart from mode at the top of the hierarchy, the assignments would show underestimates of patronage compared to counts.

13. Modelling the Responses of Public Transport Operators to Changes in Demand

13.1 Introduction

13.1.1 In principle, public transport operators can respond to changes in demand by either changing the capacity they provide or by changing the fares they charge. They can change capacity by either providing more frequent services or more spaces on existing services. Some operators have more freedom to respond than others. For example, it is relatively simple for a bus operator to provide more buses, either on existing routes or new routes, at relatively short notice. At the other extreme, in congested parts of the rail network, it may be impossible to provide more capacity without costly new infrastructure which would take a long time to provide. All operators can, in principle, increase their fares in order manage demand to fit within the capacity they are prepared to provide.

14. Commercial Software

14.1 Introduction

14.1.1 The public transport modeller should identify, using the guidance provided in the previous sections, the most appropriate modelling approach for the system in question. Then, he or she should decide on a commercial package that can support that approach.

14.1.2 In table 2 the key characteristics of the main commercial public transport assignment packages are presented, based on documentation and consultation with model developers. This was the status in November 2004, and offered as guidance. Modellers making final decisions on the most appropriate package for their problem are advised to check up-to-date details with developers.

Table 2: Summary of commercial software

	CUBE VOYAGER	EMME/2	VISUM	VIPS	PT SATURN	OMNITRANS	QPT	VISUAL TRANSPORT MODELLER
	CUBE incorporates functionality of TRIPS. It has a comprehensive range of PT modelling facilities within a much larger traffic modelling package.	Comprehensive PT modelling functionality within a larger suite of tools for multi-modal traffic modelling.	The latest version VISUM 9 incorporates much of the functionality of VIPS within a larger set of modelling functions. Multi-modal model.	PT modelling only.	This is an add-on to SATURN suite of programs and has limited range of PT modelling functionality.	A new package in the UK with PT modelling capability within a larger package for multi-modal modelling.	Part of the 'Q-Software' transport planning software. This is no longer supported.	A new package; multimodal.
(Y = Included, N = Not Included)
Walk Access Time	Y	Y	Y	Y	Y Same coefficient as egress and transfer	Y Same coefficient as egress and transfer	Y	Y Same coefficient as egress and transfer
Walk Egress Time	Y	Y	Y	Y	Y Same coefficient as access and transfer	Y Same coefficient as access and transfer		Y Same coefficient as access and transfer
Walk Transfer Time	Y	Y	Y	Y	Y Same coefficient as access and egress	Y Same coefficient as access and egress		Y Same coefficient as access and egress
Origin Wait	Y Derived from headways; default wait curve provided; node-specific	Y Derived from headways and statistical distribution of arrivals	Y Headways and statistical distribution of arrivals; Separate coefficient	Y Headways and statistical distribution of arrival times	Y Headways only; Separate coefficient	Y User-specified function of frequencies, by mode		Y Same coefficient as transfer wait
Transfer Wai