Variable Demand Modelling - Scope of the Model
TAG Unit 3.10.2

June 2006

Contents

1. Variable Demand Modelling - Scope of the Model

1.1 Background
1.2 Model Area and Zone Size
1.3 Forms of Trip Matrices: Production/Attraction or Origin/Destination
1.4 Forms of Trip Matrices: Vehicle Occupancy
1.5 Aspects of Incremental Modelling
1.6 General Data Requirements and the Reference Case
1.7 Segmentation: Trip and Person Types
Trip frequency
Trip distribution
Mode choice
Time of day choice
Value of time
1.8 Assignment Modelling
Is Dynamic Assignment Required?
1.9 Division into Time Periods
Feeding back costs from assignment to demand model
1.10 Generalised Cost Formulation
Components of generalised cost
Composite costs

2. Further Information

3. References

4. Document Provenance

1. Variable Demand Modelling - Scope of the Model

1.1 Background

1.1.1 TAG units 2.9 and 3.10 explain why variable demand modelling needs to be considered and provide guidance on how to carry out such modelling for highway schemes. This unit forms the second stage of the process, TAG Units 3.10.3 and 3.10.4 detail subsequent stages, and explains how to establish the scope and coverage of the model for the particular application.

1.1.2 Important recommendations are shown highlighted and in bold. If those actions are not followed analysts will need to provide rigorous justification for the course of action taken.

1.1.3 A summary of the advice in this TAG unit is as follows:

• The demand and supply processes need to allow for trip redistribution when designing the zone system and provide a fine enough level of detail for the scheme and strategies being assessed. This is discussed in Section 1.2.
• Various stages of the demand modelling and forecasting process require travel movements to be described in terms of the factors that generate or attract trips - i.e. by productions and attractions (P/A). The guidance discusses the conversion between P/A and origin/destination (O/D) form for use in the multi-stage modelling process.
• Modelling variable demand uses reference case growth forecasts and socio-economic data that influence the travel of individuals. The need for these data and possible sources are discussed in Section 1.6.
• The impacts of different policy measures on particular groups of people can only be represented realistically and forecast satisfactorily if the demand modelling process is suitably segmented. Modelling should use groups of travellers (segments) which it is expected will continue to behave in similar fashion over time.
• A variable demand model will need to include a highway assignment stage to provide cost information to the demand model.
• Travel demand and traffic levels vary throughout the day and this usually requires the modelling of different time periods. The need to divide the day into different periods related to the daily profiles of road traffic, is no more onerous than is normally needed in assignment modelling itself, unless it is intended to model peak spreading, as described in VDM Key Processes VDM Key Processes (Unit 3.10.3).
• All transport models depend upon relating people's travel choices to estimates of their generalised cost of travel - a weighted sum of time and other costs of travel which can be measured in units of money or (preferably) time.

1.2 Model Area and Zone Size

1.2.1 Traffic assignment modelling connects the underlying zones which contain the demand to a network of links using "zone centroid connectors" serving the entire zone. For this reason it is generally desirable to have both zones and network at as fine a level of detail as possible within a 'fully modelled area' - the area in which both routeing and demand responses to the scheme being assessed are expected to occur. That may be larger than the area necessary for assignment modelling alone.

1.2.2 It is usually unnecessary to extend this level of detail beyond the 'fully modelled area' though there need to be 'intermediate' regions around the fully modelled area to provide realistic demand into the area. Zones just outside the fully modelled area should be of intermediate size enough to provide realistic demand. Anything outside the intermediate zones should be at the coarsest acceptable level for the schemes and policies being assessed. In general, the smaller the fully modelled area relative to the scheme, the more important is the accuracy of the external connectors. However, it is important to represent in a realistic fashion:

• travel to and from external zones beyond the boundary of the fully modelled area,
• congestion on centroid connectors within a zone, and
• mechanisms which allow trips to interchange between intrazonal and interzonal trips.

1.2.3 Other considerations in determining zone size relate to the need to forecast the numbers of intra-zonal trips, and the mix of modes being considered. Further advice on zone sizes for public transport modelling is given in TAG Units 2.10 and 3.11.

1.2.4 More detail is given below of how the fully modelled area and its hinterland should be defined and how the inclusion of demand modelling involves extra considerations about the realistic representation of competing destinations.

1.2.5 Guidance is available as to the level of network detail required for congested assignment models. For example, the IHT's Guidelines on Developing Urban Transport Strategies suggests that "all roads that carry significant volumes of traffic" should be included and, more generally, that the network "should be of sufficient extent to include all realistic choices of route available to drivers". When considering zone size, the integrating characteristic is that the entire zone is assumed to act as if concentrated at the centroid connectors. Consequently, if a zone contains distinctly different types of population or activity in different areas which are likely to access different points on the network, consideration should be given to sub-dividing it.

1.2.6 However, a balance has to be struck between detail and simplicity of representation, and the guiding principle should be to judge whether combining travellers and activities into a single zone is likely to make the conclusions of the assessment less reliable, or make it impossible to answer some questions of relevance to the assessment. Since the finest level of disaggregation of data is generally at the level of Census Output Areas for the 2001 Census, the model zones will generally be built up from aggregations of these or earlier equivalent areas.

1.2.7 If the fully modelled area is well-defined, then the boundary for both the demand and supply systems will coincide: i.e. inside the boundary both zones and network will be detailed, and outside they will be coarse (though there may also be an 'intermediate' region around the fully modelled area). In the assignment, the fully modelled area may use delay calculations from junction simulation, while the intermediate and external areas may be modelled in link-based representation only. Zone size is likely to increase from core to intermediate to external.

1.2.8 Movements between the internal and external areas need to be represented at an appropriate level of detail, for four reasons:

• On the demand side if only internal movements are properly represented as trips, then zones near the border will have (apparently) lower levels of trip-making.
• Secondly on the demand side, when modelling destination choice, travel opportunities to both internal and external zones need to be represented. Thus although the external area can be represented at a coarse geographical level, it is important that it should contain sufficient close destinations, and appropriately attractive ones, to take a realistic share of demand from within the modelled area.
• Similarly, zones just outside the fully modelled area need to provide a realistic demand into the area. Hence, the fully modelled area should be surrounded by a ring of zones with a dimension a little larger than the internal zones, and outside these will be very large zones representing the rest of the external area.
• On the supply side (network), movements from one external zone to another external zone may form 'through traffic' in the modelled area and this may respond to factors beyond the scope of the model. It is usually satisfactory to treat through-traffic as fixed in volume, but free to choose its routeing within the modelled area.

1.2.9 The size of internal zones will need to be carefully considered in relation to intrazonal trips: the larger they are, the larger will be the proportion of trips originating in them which remain intrazonal, i.e. their destination is also within the origin zone itself. It is important to represent the costs of such trips realistically, otherwise there may be biases in the demand model. At the distribution stage it is important to be able to redistribute intrazonals to become interzonals, and interzonals to become intrazonals, if relative costs change. If the zone sizes are small this is less of a problem, but for large zones it is important that the average intrazonal costs are as realistic as possible.

1.2.10 Various approaches may be used to derive intrazonal costs:

• assume the average cost of an intrazonal trip is a fixed proportion of the costs of interzonal trips to the neighbouring zones, or
• assume the mean distance of an intrazonal trip is a proportion of distance to the neighbours and costed accordingly.

Intrazonal costs should reflect the prevailing level of congestion via the mean journey speeds, and preferably its response to changing demand: basing costs on those of trips to the neighbouring zones will generally be sufficient.

1.2.11 The key considerations are that:

• It is desirable to have both zones and network at as fine a level of detail as possible within the 'fully modelled area' - i.e. the area in which both routeing and demand responses are expected to occur. This requires a judgement about the general level of detail appropriate to the assessment, i.e. its likely consequences for the accuracy of assessment or for policy.
• Zone size outside this area can be much coarser, but the size of external zones, and the distance of the boundary between the internal and external areas from the scheme itself, has implications for the appropriate length of the external zonal connectors and external network links. Generally, the smaller the fully modelled area relative to the scheme, the more important are zones just outside the fully modelled area and the accuracy of their external connectors.
• Average intrazonal trip costs should be calculated as accurately as possible to remove bias in the distribution model.

1.3 Forms of Trip Matrices: Production/Attraction or Origin/Destination

1.3.1 There are two alternative ways of describing the travel pattern:

• When travel patterns are constructed from roadside surveys the trips are logically described by the place the trip started and the place the trip finished, and the trip purpose of each end. This is usually known as an Origin-Destination (O/D) based trip pattern. Assignment models use this definition of the trip matrix.
• An alternative way of looking at the trip pattern is from the viewpoint of the factors that produce or attract trips, i.e. on a Production-Attraction (P/A) basis, with home generally being treated as the "producing" end, and work, retail etc as the "attracting" end. To properly define trip production and attraction, it is important to understand what home based and non home based trips are. Home based trips are trips where the home of the trip maker is either the origin or the destination of the trip. Non home based trips on the other hand are trips where neither end of the trip is the home of the trip maker. Trip production is usually defined as the home end of a home based trip or the origin of a non home based trip. Trip attraction on the other hand is defined as the non-home based end of a home based trip or the destination of a non home based trip. Changes in these P/A trip end forecasts over time or by scenario will lead to changes in the trip pattern. This definition of the trip matrix has normally been used in modelling travel demand

1.3.2 The distinction between production/attraction and origin/destination matrices is most easily explained by the example of commuting trips from home. On an O/D basis a commuter from the suburbs with a workplace in the city-centre completes one trip from suburb to centre in the morning and one trip from centre to suburb in the evening, say. On a P/A basis, the suburb 'generates' two commuter trips^[1] (there and back) and the centre 'attracts' two commuter trips. This distinction is most important when forecasting travel patterns in the future since, for instance, changes in workplace distribution may well be different from those in employee's residences. This can, in most circumstances, lead to different forecasts of trips depending on which of the two trip matrix definitions is used.

1: In a number of transport models the modelling is not based on trips but on tours. A "tour" is applied to any round trip, starting and finishing at home, and may contain stops at several different destinations, but few models handle multi-destination tours. Journeys between non-home destinations are handled automatically in models which represent tours, but most demand models treat them as Non-Home Based trips.

1.3.3 In current modelling practice, trip end modelling is usually done on a P/A basis, as with the TEMPRO forecasts, but assignment is always done on an O/D basis since the actual direction of travel at a particular point in time is important. Somewhere during a multi-stage modelling process trip matrices must be converted from a P/A basis to an O/D basis.

1.3.4 The conversion is usually done after time of day choice, distribution and mode choice and before assignment. P/A based trips are converted into O/D based trips by using conversion factors disaggregated by time of day and trip purpose (distinguishing between inbound and outbound home-based trips). Although these factors may change over time in reality, it is usually acceptable practice to assume constancy before any time-period choice is applied. Such factors can be obtained locally or by using NTS data tables (this is how the DfT's TEMPRO produces O/D based forecasts from P/A forecasts).

1.3.5 Figure 1 below indicates what might be described as the "traditional" form of the model, in which an "absolute" demand model is used, in PA form, to estimate the matrices. It is then converted to O-D format, and assigned to the network, and the costs are then taken from the network, converted back to P/A format where required, and the demand model is updated. The dark black arrows apply in all cases, but the dashed arrows apply only when the "traditional" approach is being used.

1.3.6 In practice, as we discuss in section 1.5, it is often necessary to take an incremental modelling approach. This introduces further considerations into the conversion between PA and OD.

1.3.7 Where a non-uniform growth is forecast at either the production (home) end or the attraction end, forecasts produced using O/D matrices will be less accurate than those produced using P/A based matrices. For this reason P/A matrices should be used, even if no explicit trip distribution modelling is performed.

1.3.8 Although P/A based matrices are strongly recommended in demand modelling, there are a number of circumstances where it may be satisfactory to use O/D based matrices for forecasting:

1. Where forecasts are based on a simple overall growth rate, by purpose. In this case the forecast will not be biased.
2. For peak period modelling only. For this time-period it can reasonably be assumed that the vast majority of trips are starting from home so an O-D based trip matrix is the same as a P/A based matrix. For other time periods the only way to keep consistency would be to base the modelling on the basis of purpose and direction, for instance to model Home to Work separately from Work to Home.
3. Where the demand responses are estimated in conjunction with the assignment process in some simplified elasticity calculation, see VDM Key Processes (Unit 3.10.3), there must be common definitions, and these are perforce those of O/D since that is the basis of the assignment process. The preceding arguments suggest that using this method would be a better approximation for forecasting if undertaken in the morning-peak and/or with direction based purposes.
4. There will be many circumstances when the original roadside or household survey data has been lost, and the only trip data available are O/D matrices, by time-period, and perhaps purpose, for use in assignment. This should only be used where new surveys are impractical.

1.4 Forms of Trip Matrices: Vehicle Occupancy

1.4.1 Whilst assignment modelling is concerned with vehicle movement, demand modelling is concerned with individual traveller decisions. Before the assignment stage is reached car occupancy factors need to be applied to the private travel demand matrices to convert them to vehicles. Values of Time and Operating Costs (Unit 3.5.6) give default values by trip purpose and time period, as well as assumptions about how these factors are expected to change through time. Local factors can be calculated from RSI data to see if there are other local factors affecting car occupancy, such as direction of travel or type of flow. These local factors should be used if there are significant differences from the national ones and if there is confidence that the RSI-based factors are an unbiased estimate of all vehicle travel in the area.

1.5 Aspects of Incremental Modelling

1.5.1 As is discussed in (Unit 3.10.3), demand models may be either absolute or incremental: in the latter case, this means that they "pivot" off a base matrix, and only model the changes to the matrix brought about by changes in (generalised) cost. However, the implications of section 1.3 are that this base matrix needs to be in P/A format. In section 1.5 some guidance is given as to how this base matrix can be obtained.

1.5.2 Whether the demand model is absolute or incremental in form, there will be a need to validate the base matrix at the network level. In practice, this means that the conversion from P/A to O-D is carried out, and the resulting matrix assigned. Then the assignment process is validated according to the procedures given in (DMRB.12.2.1).

1.5.3 Problems may be incurred when, after reasonable adjustments to the network, it is concluded that significant errors remain which are essentially attributable to the matrix. Ideally, further data should be introduced to the whole modelling procedure in such a way that the base P/A matrix is modified. Unfortunately, there is very little experience of how to do this, and conventional methods of "matrix estimation" (using, in particular, link counts as a source of information) only operate at an O-D level. If the O-D matrix is adjusted in this way, in order to improve the quality of the assignment, there is no direct way in which these adjustments can be conveyed to the P/A-based demand model. The result is that there will be a discrepancy between the demand model and the assignment model.

1.5.4 With the current state of knowledge, if this position is encountered, the best approach is to use an incremental version of the assignment model. Essentially, after converting the output of the demand model from P/A to O-D, the resulting matrix is not directly assigned, but is compared with a base case, and the implied changes are used to adjust an independently validated "assignment matrix". This adjustment could be done in a number of ways, proportionally, additively, or by a mixture of the two^[2].

2: Note, although for the greater part of the matrix, no problems will be incurred by either an additive or a proportional approach, both these methods can give rise to problems in specific cases. It is possible that the demand model could imply a decrease in demand for a particular ij cell which causes the adjusted assignment matrix to go negative. Alternatively, a large proportionate effect predicted by the demand matrix in the case of a low base demand could correspond with a much larger cell in the assignment matrix. Some care is therefore required in applying the method, and a small amount of re-allocation between cells may be necessary, with the aim of ensuring that the total change predicted by the demand procedure is maintained.

1.5.5 There are thus two types of incremental modelling, a) on the demand side, on a P/A basis, and b) on the assignment side, on an O-D basis. These two incremental variants are entirely independent of each other, so that there are four possible combinations, which can be written according to the following matrix:

1.5.6 These alternatives are illustrated in figure 2 below, where the red text and arrows indicate an incremental or marginal demand model where the model "pivots" off a base matrix (in PA form) and is driven by the change between the scheme costs and the base costs. The blue text and arrows indicate an incremental or adjusted assignment process, in which the demand matrix, after conversion to OD format, is compared with a base case, and the implied (OD) changes are used to adjust an independently validated "assignment matrix".

1.5.7 Type 2, with an absolute demand model and an adjusted assignment, is used by a number of models (for example, the LTS model). Type 3, with an incremental demand model and a direct P/A to O-D conversion, is essentially what is done in the START model (although it does not strictly carry out an assignment). Type 1 is the "conventional" approach illustrated in Figure 1 in Section 1.3.

1.6 General Data Requirements and the Reference Case

1.6.1 The detailed advice below lists the type and sources of data required for variable demand modelling, beyond the data used for assignment modelling. In most cases only a sub-set of these data sources will be needed for a given model; less detailed models will require less data, and may base some of their categorisation on assumptions or transfer of appropriate data from other models. As a minimum, however, any multi-stage demand model will require a database which provides, in the base year for each origin zone:

• the total number of "car-available" trips,
• to each destination zone,
• for each of several purposes,
• by each of the modes modelled, and
• within each modelled time period.

1.6.2 If a model is being built to describe and validate the base year as well as the forecast years, this information can be obtained from a variety of sources as summarised in the Table above. These sources are not independent. TEMPRO offers a valuable source of most of the data required to predict trip ends, both trip productions based on household characteristics and trip attractions based on employment etc, as well as car availability forecasts. For further information on TEMPRO and the data used see www.tempro.org.uk.

1.6.3 In many cases TEMPRO may be sufficient, and there will be no need to go back to the original sources. In other cases the TEMPRO zones may be too large, so that the TEMPRO data may have to be adjusted in the light of local knowledge or data specific to the study zones obtained from more original sources. Effort can be saved if there has been a suitable local transport survey, or if there is a local transport model from which the required data can be extracted. If the base data contains all the necessary detail but refers to a date which is considered too early for reliable prediction, then the data can sometimes be updated on the basis of more recent global indicators and good judgement.

1.6.4 There are two separate processes that need to be considered when developing forecast models:

• the production of a base year travel pattern, and
• the production of reference forecasts for future years.

1.6.5 The base P/A matrix can be constructed from observed travel movements based on road-side and passenger surveys, as well as household survey data. Although in theory it should also be possible to make use of traffic counts, the fact that these contain no directional information (ie, we do not know which the journey is from home, or to home, or, for that matter, non-home-based) means that there is no current methodology for doing this.

1.6.6 The detail available in these travel patterns is largely dictated by the richness of the survey data, and in reality the procedure to provide the "best" base matrices will always involve some synthesis as it is rare that all movements are surveyed. In carrying out a synthesis of a P/A matrix, it will generally be desirable to enforce the productions to agree with some reputable "trip end model", and TEMPRO should be treated as the default in this respect. TEMPRO provides forecasts of trip ends which have been derived by feeding TEMPRO planning data into the National Car Ownership and National Trip End Models at the TEMPRO or NTEM level of zoning. These models may also be applied at a local model level of zoning by feeding in appropriate planning data at the local model zone level. Alternatively, more detailed local data may be available to use with the model, but it will be necessary to ensure that at a broad level it is consistent with the assumptions of the National Trip End Model (NTEM) within the TEMPRO program.

1.6.7 If an absolute demand model is being created, then a theoretical structure representing mode and distribution will be chosen for each purpose, together with estimates of trip ends, and the mathematical functions will be calibrated to reproduce as much of the "observed" data as is possible (some guidance on the general calibration of such models is given in Unit 3.10.3).

1.6.8 However, for many reasons, this is not always the appropriate approach. This may be because of lack of calibration expertise, or it may be that even after considerable effort, the theoretical structure is still unable to accommodate the essential pattern of the observed data. In this case, the following approach may be helpful.

1.6.9 The essence of the approach is to begin with a wholly synthetic model which makes minimal, but reasonable, assumptions. In this way, initial P/A matrices at the required level of detail are developed. These are then treated as a "prior" which is then modified in the light of observed data, and the reliability thereof, leading to an improved version of the base matrix.

1.6.10 The initial synthetic model would start off with the all-day productions and attractions implied by TEMPRO for each purpose (or, better, make use of the underlying car ownership and trip end functions applied to local data on population, households and employment). The matrix cells would then be filled by means of a standard gravity model which should be constrained to reproduce (at least) average trip length for the purpose (taken either from local sources or national sources such as NTS). Next, factors giving modal choice and time of day (again, available as part of the NTEM database) can be applied. In this way the complete prior matrix is built up by mode and time period, distinguishing the outbound and return portions of home-based purposes.

1.6.11 The process of "introducing observed data" must then make allowance for the statistical accuracy of that data, based essentially on sampling theory (see guidance in DMRB 12.1). This could be done along the following lines. For each observed cell of the matrix (ie production zone, attraction zone, purpose, mode, time period as well as possibly some segmentation data relating to car availability etc), the "prior" value would be tested as to whether it lay outside of the confidence region of the observed data: if so, the prior data will need to be modified.

1.6.12 In the course of such modification, certain key features of the prior matrix may be lost: for example, the modifications may change the total productions, or the average trip length. There will therefore be a need for an iterative process which attempts to re-impose some features as "constraints". The modification can typically be done by means of a "proportionate fitting method". Note that, provided that purpose is available in all the observed data, this adjustment process can be carried out independently for each purpose.

1.6.13 When the adjustments have converged to a satisfactory degree, the resulting matrices can be taken as an acceptable P/A basis, and used as the pivot for an incremental demand model.

1.6.14 If a modified (O-D) matrix for assignment purposes has already been validated, then this matrix should be used as the basis for assignment, and predicted changes in demand (after conversion from P/A to O-D) will be used to adjust this matrix. Thus this corresponds to type 4 in figure 2 above.

1.6.15 If no assignment matrix is in existence, then the first step should be to see whether the derived base P/A matrix can, when converted to O-D form, be satisfactorily validated at the assignment level. If this turns out to be the case, then the form of the overall model will correspond to type 3 in figure 2. If the base P/A matrix cannot produce a satisfactory matrix for assignment, then it will be acceptable to carry out further "matrix estimation" work on the resulting O-D matrix, to produce an "assignment matrix" which is not compatible with the base P/A matrix: in this case the overall model will again be of type 4 form.

1.6.16 We now go on to consider the forecasting of the reference case, which is required when the demand model is of incremental form.

1.6.17 The reference case is a forecast at constant generalised cost and constant value of time. It is based on the assumption of no change in travel costs from the base year. It is not intended to be a realistic forecast, but is essentially a way of "locating" the demand curve for the future scenario.

1.6.18 The construction of the reference forecast requires the "reference case" growth factors and will involve the adjustment of the row and column of the base P/A matrix at an all-day all-modes level to reflect expected land-use and car ownership changes. Since by assumption the travel costs do not change, it is appropriate to assume that mode and time of day proportions, for each segment distinguished in the base matrices, are unchanged. These proportions therefore need to be derived from the base matrices, and re-applied after the rows and columns have been adjusted. As a default, the growth factors for both productions and attractions should be based on TEMPRO (or, as noted earlier, by making use of the underlying car ownership and trip end functions applied to local data on population, households and employment for both base and future scenarios).

1.6.19 The prediction of responses to different schemes can be obtained using an incremental demand or absolute modelling approach which pivots off the reference matrices (see VDM Key Processes (Unit 3.10.3)). When the demand model is non-incremental, the reference matrices are not strictly required, though the expected land-use and car ownership changes need to be reflected, again using TEMPRO as a default.

1.6.20 Modelling of incremental changes from the base matrix is likely to be adequate for most assessments. For very large schemes or situations where there will be substantial land-use and demographic changes within the timescale of the assessment, however, it may be necessary to make a detailed absolute forecast (i.e. with absolute values predicted by the demand model) of at least part of the future reference case.

1.6.21 The data required for the demand calculations depend upon the chosen level of segmentation (or disaggregation) of travellers and travel characteristics, as discussed in Section 1.3 of this Unit. The level and details of the segmentation should depend on the planned policy tests and on those differences which are considered to be important to people's travel behaviour. However, it is as well to be aware what data are available before deciding on the appropriate segmentation, since it may be necessary to modify the preferred segmentation according to the availability or lack of existing data.

1.6.22 For demand modelling the required matrices should be in production-attraction format as discussed in Section 1.3. This will then need to be converted to origin-destination format before the assignment stage of the modelling process. For trip ends (productions and attractions), much of the spadework has already been done for the National Transport Model, and the underlying data for this is available from the TEMPRO Trip End Model Program.

1.6.23 Trip distribution requires some measure of the attraction of each zone for different trip purposes, and the necessary data can be obtained from TEMPRO and more detailed data from the Census Special Workplace Statistics, Local Authority planning data, and local travel surveys. If there is a relevant local transport model then both data and appropriate parameter values may also be mined from the model. Such a local model could of course be used as the basis for building the new model. Modal split modelling requires the categorisation of travellers according to whether they have a car available for their journey. True car availability is difficult to model, and most models settle for categorisation of travellers according to whether they come from a household with no car, one car, or more then one car. This can be obtained from TEMPRO or local travel surveys if mode choice is to be explicitly modelled. It will also be necessary to obtain data on the alternative public transport options. The extent to which this is needed will depend upon how important competition between car and public transport is judged to be. If it is of fairly peripheral interest, limited ad hoc surveys may be sufficient, but if it is thought important for reasons such as the consideration of alternative public transport schemes then detailed modelling of public transport network will be necessary, and the reader is referred to the guidance on Public Transport Modelling and Forecasting for further advice. Again, a relevant local travel survey may provide relevant information on both public transport services and use. If slow modes are to be modelled, travel costs for walking and cycling can be derived satisfactorily on the basis of mean speeds by type of path with an allowance for street-crossing delays.

1.6.24 Good local data is preferable, but collection of detailed data from household surveys is expensive. As with all the modelling, the degree of effort to be applied to data collection depends upon the complexity and detail that can be justified by the particular application. Where a comprehensive demand model is to be built for general policy analysis and assessment work across a wide network it is obviously sensible to spend considerable resource in data collection for both calibration and validation.

1.6.25 Where, however, the intention is to test a smaller scheme against variable demand there may be insufficient data to calibrate a satisfactory mode choice or distribution model and illustrative parameter values should be used (see VDM Key Processes (Unit 3.10.3)). Inevitably, this approach will introduce uncertainties, and the modeller will have to judge their importance, but sensitivity testing against variation in the more important and uncertain parameters will demonstrate the robustness or otherwise of the conclusions.

1.6.26 In this more limited application of variable demand modelling, it will generally be sensible to use the demand model incrementally rather than absolutely, see VDM Key Processes (Unit 3.10.3). In that case the model merely predicts changes from the known base situation as travel speeds and costs change, rather than trying to predict the details of the base situation. When used incrementally, the variable demand model does not need to generate trips on the basis of detailed household socio-economic data: it needs a reference case incorporating those factors, but the variable demand model itself merely predicts how these reference case trips will be affected by changing costs. Similarly, it does not need to contain detailed zone attraction parameters for distribution, since these will be incorporated in the reference case, but merely needs to adjust scaled versions of the observed trip matrix as costs change. Such simplifications do not hold for those models which incorporate incremental modelling by forecasting the differences between two absolute models. See VDM Key Processes (Unit 3.10.3).

1.6.27 The validity of the base year matrix and the derivation of the parameters on which the demand model is based should be reported in an extended version of the validation report required by DMRB 12.2.1. That should report the assignment validation, identify which of the sensitivity parameters have been imported from elsewhere and report the standard errors and other confidence attributes of new data and new sources. Although some of those parameters may remain substantially untested, provided the model's responses to changing costs accord with what is known, see VDM Convergence, Realism and Sensitivity (Unit 3.10.4) it is likely to be adequate for its limited purposes.

1.7 Segmentation: Trip and Person Types

1.7.1 "Segmentation" is the dividing of travel, traveller and transport attributes into different categories so that all travellers in the same category can be treated in the same way. In the extreme, every link of the road and public transport network would be represented in the model, and zones would be small enough to allow identification of every trip. The trip matrix would be segmented by purpose and person/commodity type, and into time periods, so that every segment is completely homogeneous. All possible responses to changes in trip costs would be included. Such an all-embracing model is clearly impractical, and compromises have to be made. In general assignment and demand models require different forms of segmentation:

• In assignment models different categories of road and vehicle are identified because they require different parameter values in relation to traffic flow and speed. Some assignment models may categorise by broad trip purpose because these may grow at different rates over time, or be present in different proportions in different time periods. More segmentation may be required when testing policy options that may affect different groups of road users in different ways (for example charging for road use may require segmentation to reflect 'willingness-to-pay'). Otherwise, segmentation by time of day (see Section 1.8) is the most important issue for assignment modelling.
• Demand modelling generally requires more categorisation, both in order to estimate how much demand, and of what type, a particular zone may produce or attract, and because different types of traveller respond differently to changes in travel conditions and costs.

1.7.2 To be accepted by the policy-makers, forecasting and assessment must be seen to deal realistically with the variety of external factors which will contribute to changes in travel demand in the coming years. Moreover, policy makers may wish to know whether policies impact differently on different types of traveller, and if so, how. However, segmentation increases the size, complexity and run times of models, as does a more detailed spatial description using smaller zones and judgements have to be made about how much detail is necessary in a particular application. The same degree of segmentation may not be necessary at all stages of the model, and each of the stages of the demand model is considered in turn in the detailed discussion below.

1.7.3 Ultimately the segmentation adopted in the modelling process must depend on the nature of the study area, the objectives of the study, the data available, the outputs required and the intended model structure. The Table below suggests the minimum levels of segmentation for demand modelling. Note that these are guidelines on minimum segmentation, they are not necessarily adequate, and the degree of segmentation used should depend upon the particular application and the resources available.

Minimum segmentations for a multi-stage demand model

1.7.4 While it is undoubtedly useful to use a more elaborate segmentation of the population at the trip generation stage in order to facilitate forecasting, there is generally less requirement to carry such segmentation forward into subsequent stages of the model. A distinction between purposes is however essential; a suitable starting point would be - Work, employers business and others. Currently values of time used in appraisal are considered different for these purposes see Values of Time and Operating Costs (Unit 3.5.6). Where modal choice is modelled, it will also be important to make a distinction between travellers who have a car available for a trip and those who do not and are therefore limited in their choice of modes.

1.7.5 Not all stages of the demand model require the same degree of segmentation. The paragraphs below give more detail on what is needed for each stage, and consider value of time issues.

1.7.6 Currently illustrative values are only available split by trip purpose (see VDM Key Processes (Unit 3.10.3)). If the policies to be tested require more detailed segmentation, then the decision to use a simplified variable demand modelling approach should be reconsidered.

Trip frequency

1.7.7 For most purposes it will be satisfactory to take the observed trip pattern and modify this pattern incrementally by making it respond to changes in travel times and costs. Estimation of base year or reference case trip-rates often involves a relatively high level of disaggregation. Some of these categories can then be aggregated in later stages of the modelling process. The main socio-economic variables leading to household trip generation are now reasonably well understood. They are:

• size of the household
• age structure
• numbers of employed members, students, retired members
• level of car ownership/availability
• number of driving licence holders
• income or socio-economic group (SEG)
• other area attributes (for example, population density)
• accessibility/ level of service

DfT's TEMPRO program forecasts trip ends based on those factors.

1.7.8 Separate trip generation models based on that level of detail may be required where there are significant new development areas, to estimate the number of trips generated by a household (or person) with specified characteristics, for every purpose distinguished.

1.7.9 In other cases, the detailed local trip generation modelling is irrelevant and trip end growth commonly obtained through the application of TEMPRO is likely to be sufficient. TEMPRO aggregates over car availability categories and its outputs in Version 4 splits travel into:

• Home based (HB) shopping
• HB work
• HB visit friends and relatives
• HB recreation
• HB personal business
• HB education
• HB employers' business
• HB holiday

with a similar split, excluding visit friends and relatives, for non-home-based trips. Again, these categories are generally aggregated into a smaller set for later stages of the modelling process.

1.7.10 Whilst TEMPRO gives some details of how traffic veh-kms could change in the light of changes in the Value of Time over time, evidence of the impact of changing travel costs on traffic veh-kms is not available for highly segmented categories, and in reality is not that well understood even for simple categorisations. Thus a categorisation by trip purpose (where the Values of Time are assumed to differ) is more than sufficient. It is also possible to assume that only certain trip purposes will change their frequency in response to changing travel costs, for instance trip frequency changes may be modelled for leisure trips but not for commuting trips.

Trip distribution

1.7.11 The distribution model estimates the number of trips between each pair of zones, and ideally includes intrazonal trips which begin and end in the same zone, as well as the interzonal trips. It should be noted that it may be necessary to apply area-specific constants or movement specific deterrence functions within the distribution model to reflect the difference in the nature of travel to certain areas (e.g. longer distance trips to city centres). If this particular problem arises for the application being considered, some form of income or SEG segmentation may be appropriate to reflect how, for example, jobs in city-centres may have a high component of 'high social economic groups (SEGs)' such as professional and managerial posts in finance, banking and other business services, where workers may be drawn from further away producing higher average trip distances. A similar pattern may emerge for shopping trips to the city centre and both may require a 'white collar /blue collar' distinction, for example at a zonal trip attraction level. However, for most applications such a complication will be unnecessary.

Mode choice

1.7.12 Since the choice of mode depends on whether a traveller has a car available for the journey it is desirable to categorise travellers according to car availability for the trip, but since this is hard to identify in practice the segmentation is often merely by the level of household car ownership - 0, 1 or 2+ cars - rather then true availability. Variable demand models allocate trips to different modes according to the mean generalised cost of travel between the zonal origin and destination, in the relevant time period. There may be several modes to choose from: car, various forms of public transport, and possibly walk and cycle, though many models omit the 'slow' modes of walk and cycle. These issues are discussed in VDM Key Processes (Unit 3.10.3). It is standard practice to develop models with different parameter values for different purposes and different categories of car availability (car-owners and non-car-owners, or the distinction may be based on a more elaborate indicator of car availability), since this obviously influences mode choice. These values may be estimated during calibration of the model to the observed travel patterns, but Stated Preference surveys may also be helpful in distinguishing between the travel sensitivities of the different groups. VDM Key Processes (Unit 3.10.3) provides illustrative values for the sensitivity parameters of the various demand mechanisms, based on a review of existing models.

Time of day choice

1.7.13 Where time of day choice is modelled explicitly this choice mechanism can represent either Macro time period choice (the broad choice between time periods, e.g. 2 to 3 hours in length) or Micro time period choice (which uses fine periods e.g. 15 minutes). Obviously, the definition of the modelled time periods should be consistent with the choices to be made, as described in Section 1.8 in this Unit, and the necessary segmentation by trip purpose, since obligatory travel such as work and education is likely to have less flexibility in adjusting its time of travel than are more optional purposes such as shopping or leisure.

Value of time

1.7.15 The effect of distributing VOT (or willingness-to-pay) according to the various types of traveller may be picked up sufficiently by using a different mean VOT for different trip purposes. However, where some form of charging is central to the scenarios being tested, it will probably be important to include this type of variation explicitly, since the effect of having some drivers with high VOTs, who are more willing to pay a modelled toll, is more than offset by those with low VOTs who elect to divert. The net effect depends upon the number of willingness-to-pay (WTP) bands into which the travellers are divided, but three or four seem to be adequate. The available evidence suggests that VOT should increase with an elasticity of 1.0 relative to income for employer's business trips and 0.8 for other trip purposes (see Values of Time and Operating Costs (Unit 3.5.6).

1.7.16 A smaller number of user classes are likely to be required for Goods Vehicle and Employer Business trips. As a minimum, car travel for commuting, "other" purposes, and business, and ideally also light and heavy goods trips, should be distinguished.

1.8 Assignment Modelling

1.8.1 A variable demand model must include a highway assignment stage to provide travel cost information to the demand model. That assignment stage must include speed flow responses and must be adequately converged. Assignment can be considered separately from the other mechanisms, but it is essential that an equilibrium solution between demand and supply is obtained.

1.8.2 Assignment modelling that includes predicting traffic paths, loading traffic onto chosen paths, calculating speed changes, and iteration to convergence is a long established basis for estimating link flows and costs for input to economic appraisal procedures. Assignment has commonly been used on its own to appraise highway infrastructure improvements using one of a number of proprietary software packages. This advice assumes that the reader is already familiar with assignment, and has probably used it for fixed demand appraisals with COBA, but it is discussed here in relation to demand modelling.

1.8.3 If an adequate assignment model already exists, then this model can be used to supply the travel times and costs necessary for generalised costs, see Section 1.9, which are a central input to the demand calculations. Care should however be taken to ensure that the existing model is suitable for this purpose, particularly in the geographical (spatial) representation and the generalised cost formulations.

1.8.4 To ensure true converged solutions, unless there are good reasons for not doing so, the ratio of journey time to journey distance should reflect the generalised cost formulations set out in Section 1.9 below. This requirement does not preclude the use of time-only assignments where they have been clearly shown to validate better than time plus cost, but where this is true a check should be made to see why this is so. It may be due to other factors affecting route-choice or that demand modelling requires the consideration of other elements of travel costs, for instance parking costs and access times are more likely to be needed in such models' generalised cost formulations than if the transport model consisted of only an assignment model.

1.8.5 An important issue is that 'good' assignment convergence is necessary for 'good' convergence in the iterations between supply (assignment) and demand. Convergence criteria for assignment are discussed in DMRB 12.2.1. The issue of convergence between supply and demand is considered further in VDM Convergence Realism and Sensitivity (Unit 3.10.4).

1.8.6 There exists a substantial volume of comprehensive and satisfactory advice and information to guide the model development of an assignment model. Modelling (Unit 3.1) discusses some of the issues and further DfT advice can be found in:

• DMRB 12.1.1, especially Chapter 9.
• DMRB 12.2.1 ("Traffic Appraisal in Urban Areas"), especially Chapter 4.

Useful guidance can also be found in:

• "Guidelines on Developing Urban Transport Strategies" (IHT), Section 6.10.

1.8.7 The user is recommended to consult these documents and the user manuals of proprietary traffic assignment suites.

Is Dynamic Assignment Required?

1.8.8 Dynamic assignment can replicate the variation of demand within each time period modelled and this may be carried forward to micro time period choice models (see VDM Key Processes (Unit 3.10.3)), but other variable demand model mechanisms only need average costs. Both needs can be met using dynamic assignment packages that represent explicitly variation of demand with time.

1.8.9 In demand modelling, a judgement must be made as to how best to define the time periods so that, within each, travel conditions are sufficiently constant to provide a realistic mean cost for the modelling purposes, this is further discussed in Section 1.8 below. A balance needs to be struck between the level of detail in the assignment step of the multi-stage model and the need for detail elsewhere in the modelling process (the number of time periods, the level of detail in the various segmentations, the stages included in the demand model etc).

1.8.10 Assignment models typically use a matrix of demand flows for the whole modelled time period, though it is common practice to use separate matrices representing the morning peak and inter-peak periods, or the peak hours. But once the modelling is broadened to examine how demand might respond to changes in travel conditions there will be a need to re-examine the adequacy of the modelled time periods.

1.8.11 In general, demand modelling uses relatively broad time periods. When examining times and places of high congestion, it may be desirable to introduce a higher level of time-dependent responsiveness into assignment, either by modelling a series of short time-periods, or by dynamic assignment which represents explicitly the variation of demand over time, in order to obtain a better estimate of these average costs. This may be appropriate because:

• levels of demand change substantially during the assignment period (either across the whole matrix or in individual cells), so that travel times also vary substantially, or
• traffic demand exceeds capacity at some points on the network for some periods of time, so that not all demand can be satisfied

1.8.12 The use of short time periods in a dynamic assignment will be valuable in modelling micro time period choice as part of the assignment stage (see below). In other aspects of demand modelling, the short assignment periods should be aggregated to provide costs consistent with the longer time periods used in the demand model.

1.8.13 All congested assignment suites can allow for journeys taking longer as congestion increases. Several can also model the formation of over-capacity queues and flow metering effects. These effects may imply that, in a given time period, fewer trips arrive at their destination than depart. This broadens the peak in the trip profile and is known as passive peak spreading. A different form of peak spreading arises from travellers actively choosing a different departure time to allow for congestion or to avoid it. Passive peak spreading, and travellers' decisions to change their departure times by small amounts within the peak period (micro time period choice as discussed in VDM Key Processes (Unit 3.10.3)), sometimes referred to as active peak spreading, are best addressed by the use of dynamic assignment models, or by "quasi-dynamic" modelling in which equilibrium assignment models are used to represent short time periods and queues remaining at the end of one period are handed on correctly to the start of the next period.

1.8.14 As with all decisions about detail in a multi-stage model, the choice of whether to use dynamic assignment must be driven by the added value of the extra complexity, or the error or deficiency introduced into the model by ignoring these dynamics. If the system is strongly peaked, the assumption of a flat profile in the corresponding time period will underestimate average delays and also calculate an incorrect spread over routes. These effects can be quite serious.

1.8.15 If a case is made for dynamic assignment modelling, travel conditions are likely to differ significantly over the model period and modellers should generally also assume that travellers may change their departure times. This is discussed in VDM Key Processes (Unit 3.10.3). Fully-specified dynamic assignment offers an opportunity to model both passive peak spreading because of increasing journey times as congestion increases, and a behaviourally-based choice of travel departure times on a much finer timescale than is possible in the higher-level demand mechanisms.

1.8.16 Dynamic or quasi-dynamic assignment packages also have the potentially important benefit of providing properly flow-weighted travel costs to the variable demand model. This approach can provide a more representative average across the broader time period than can be obtained by simply assuming a constant level of traffic flow throughout the period.

1.8.17 Overall, the guiding principles are that your model should be:

• justified by the situation it is applied to and the policy requirements of the assessment;
• supported by the available data and by appropriate calibration or the plausible importation of the relevant model parameters;
• supported by the expertise available: if you are unsure of the understanding required, but see the need for elaboration, get expert advice.

1.9 Division into Time Periods

1.9.1 Travel conditions vary considerably across the day, and across the days of the week and time of year. Models usually represent a weekday during a 'neutral' or representative month. In order to capture the variation in conditions within the modelled day, and especially the fact that many schemes are aimed primarily at times of maximum travel demand and highway congestion, it is conventional practice to divide the day into different periods for modelling purposes.

1.9.2 Demand modelling depends upon the time-divisions of the traffic assignment because the relevant travel costs and journey times which are extracted from the assignment are averages across the assignment periods. Hence, it is important to ensure consistency between the time-periods used in the calculation of these averages and the key time periods for the main demand segments.

1.9.3 However, the demand modelling can be assumed to take place over different time-periods, such as 24 hour weekday, 16 Hours or just a peak period. In theory, different demand responses can be modelled over different time-scales. A good deal of the survey data will be collected over a 12 hour or 16 hour period, and the background changes in trips estimated from TEMPRO data is on a 24 hour basis. Many of the current large regional models estimate trip frequency, mode choice and distribution over a 24 hour time period. However, procedures need to be adopted to convert such 24 hour trip patterns to be compatible with the shorter time-scales generally required for assignment modelling (a peak hour or an average inter-peak period, for instance). Paragraph 1.8.9 below explains how this can be done.

1.9.4 Guidance on division of the modelled period into time periods and time slices is given in Traffic Appraisal in Urban Areas DMRB 12.2.1. It recommends that automated traffic counts should be used to establish a daily profile across at least a week's traffic data, and that subdivision should only be made where there is a clear difference in traffic congestion and/or travel patterns, or where there is an intention to model time of day choice. However, if modal transfer between private and public transport is important, and public transport offers different fares or frequencies at different times of day, it is advisable to choose time periods to reflect their different costs.

1.9.5 In general, it will be necessary to distinguish at least between peak and non-peak travel. Where the traffic profile is very variable leading to significantly different conditions at different times of the day, it may be desirable to sub-divide the peak period into a fairly narrow "peak of the peak", say the peak hour, with shoulder periods on either side. This requires a judgement of the extent to which some form of average flow and speed across the modelled period can be taken as representative of the whole period and whether any off-peak model should represent a "typical" inter-peak hour, or a more extended "off-peak" encompassing early morning and evening travel in addition to the inter-peak.

1.9.6 It is unlikely that inclusion of variable demand modelling will require any greater segmentation of time periods than is satisfactory for assignment, except where there is an interest in modelling time of day choice, as discussed in VDM Key Processes (Unit 3.10.3).

Feeding back costs from assignment to demand model

1.9.7 The assignment model provides the travel times and costs required by the demand model, and generally both assignment and demand models will use costs averaged across the defined time periods into which the model has been divided. Many models relate all demand in a given purpose category to the costs in the period where most of these trips are made. Depending on which trip purposes are categorised in the model, the cost bases may be approximated as follows:

1.9.8 Obviously, this is not strictly correct, and more detailed models should ideally calculate the costs as weighted means across the periods according to the proportions of trips of each type in each period. Note that in the demand model these average costs for particular types of trips may be averaged again across different modes, or different destinations, to obtain the "composite" costs on which the demand mechanisms operate, as described in VDM Key Processes (Unit 3.10.3).

1.9.9 Few models have as detailed a disaggregation of trip purposes as listed in paragraph 1.8.8 and it is usual to aggregate them into fewer categories (home-based shopping, leisure and social are often combined into "other" category, for example). For each category, the mean generalised costs of travel in the appropriate period, or in a combination of periods (or sub-periods) weighted according to the proportions of trips in each, are calculated and fed back to the demand model. Where the assignment model used is dynamic, or quasi-dynamic, (see below) mean costs can be obtained on a flow-weighted basis taking proper account of the variation over time, but the mean across the broader time periods should still form the basis of the demand modelling.

1.9.8 As Section 1.7 of this Unit noted, each segment considered will have, in principle, different parameters in the generalised cost function. Central to this is the concept of value of time, whereby money costs are converted into time units or vice-versa. Different values of time are appropriate to different segments of the travel market, particularly according to different journey purposes. Further information on values of time can be found in Values of Time and Operating Costs (Unit 3.5.6).

1.9.9 Generalised cost normally includes elements relating to, for private car:

1.10 Generalised Cost Formulation

1.10.1 All transport modelling should recognise that people's travel choices depend upon the cost, in both time and money. It is important to combine time and money into a single disincentive to travel, so that demand can be assumed to rise or fall with reductions or increases in either. To do so it is necessary to apply appropriate weights to the time and money components of this combined cost so that travellers can trade money for time, as in choosing between a faster but more expensive mode or a slower but cheaper mode. This combination of time and money costs is termed the "generalised cost" of a trip. Generalised costs are intrinsic to assignment modelling too, but for most applications car travel times and operating costs are highly correlated with journey distance and the modelling sometimes uses travel time alone as the cost basis, rather than the combination of time and money used in demand response modelling (see below and VDM Key Processes (Unit 3.10.3)).

1.10.2 The term "disutility" of travel is sometimes used to refer to this combination of time and money, and for many purposes the distinction between the terms has been lost. In principle, however, disutility encompasses more than the travel times and costs, including aspects of travel which cannot be quantified in the generalised cost, such as general convenience or unknown local factors. Conversely, the term "utility" of travel may be used to denote the value of activities at the destination less the cost of getting there. The "disutility" is then the negative component (i.e. disincentive to travel) of the larger positive "utility", i.e. the overall value of travelling. In this advice we will use "generalised cost" to refer to the weighted sum of time and money costs, and "disutility" to refer to the sum of generalised cost with any additional components, such as mode or area specific constants, which stand proxy for other local and unknown aspects of travel.

1.10.3 Note that, although it is referred to as generalised "cost", it can be measured in terms of either money or time, and the parameter values suggested in VDM Key Processes (Unit 3.10.3) use units of time based implicitly on the value of in-vehicle car time^[3].

3. Units of disutility. In this TAG Unit it has been assumed that generalised cost and disutility are measured in units of generalised time based on 1 minute of in-vehicle car time being valued as 1 minute. In some demand model formulations the relative value of in-vehicle time is not assumed to be unity, but it will normally be possible to translate the formulations so that the equivalent assumptions to those in this Unit holds.,/p>

Components of generalised cost

1.10.4 Two kinds of variable can enter into the function of generalised cost:

• variables which relate to the trip under consideration, and
• variables which relate to the individual making the choice.

1.10.5 Taking mode choice as an example, the cost function developed for the choice of, say, rail by an individual can be influenced both by variables relating to rail (e.g. travel time, fare) and by variables relating to the individual (e.g. income, gender, journey purpose). In principle the generalised cost structure permits a considerable level of variation in behaviour to be examined and allowed for in the forecasting process.

1.10.6 Different groups of people will trade off time and money in different ways: for example, company car owners may be less affected by rises in fuel prices, and holders of certain kinds of public transport tickets may receive free marginal travel. There is likely to be further variation by trip purpose and time of day, which can be modelled using segmentation or disaggregation.

1.10.7 As Section 1.7 of this Unit noted, each segment considered will have, in principle, different parameters in the generalised cost function. Central to this is the concept of value of time, whereby money costs are converted into time units or vice-versa. Different values of time are appropriate to different segments of the travel market, particularly according to different journey purposes. Further information on values of time can be found in Values of Time and Operating Costs (Unit 3.5.6).

1.10.8 Generalised cost normally includes elements relating to, for private car:

• fuel cost
• in-vehicle time
• parking costs
• access time
• tolls or user charges

so that, for example, measured in units of time:

G_car = V_wk*A + T + D*VOC/(occ*VOT) + PC/(occ*VOT)

where A is the total walk time to and from the car, T is the journey time spent in the car, VOC is the vehicle operating cost per km, (note the advice in Values of Time and Operating Costs (Unit 3.5.6) is to assume that travellers in course of work (Employer's Business) take into account fuel cost and other operating costs of travel, whilst private travel only takes into account the cost of fuel) for a journey of D km, occ is the number of people in the car (who are assumed to share the cost), VOT is the appropriate Value of Time, and PC is the parking cost. V_wk is the weight to be applied to walking time (see below), and although in this formulation the generalised cost is measured in time, it can just as easily be expressed in monetary units by multiplying the whole equation by VOT. Similarly, out-of-pocket monetary costs such as parking charges and tolls may needed to be added. These would be converted into generalised cost units by dividing by the relevant value of time.

1.10.9 For public transport modes generalised cost will include:

• fares
• in-vehicle time
• walking time to and from the service
• waiting times
• interchange penalty
• non-walked access, e.g. park and ride

so that, for example, in time units

G_PT = V_wk*A + V_wt*W + T +F/VOT + I

where A is the total walking time to and from the service, W is the total waiting time for all services used on the journey, T is the total in-vehicle time, and I is the interchange penalty if the journey involves transferring from one service to another (I is normally calculated as a time penalty multiplied by the number of interchanges). V_wk and V_wt are the weights to be applied to time spent walking and waiting.

1.10.10 Values of walk and wait times and interchange penalties are usually related to the value of in-vehicle time by applying weights such as V_wk or V_wt above. IHT's Guidelines on Developing Urban Transport Strategies (May 1996) and ITS and John Bates's review of value of time savings in the UK in 2003 suggest:

Value of walk time = 1.5 to 2.0 times in-vehicle time

Value of wait time = 1.5 to 2.5 times in-vehicle time; and

Interchange penalty = 5 to 10 minutes of in-vehicle time per interchange

1.10.11 Equivalent weights are likely to be equally applicable to the walk from and to parking locations for car journeys. Equivalent information is available in Transport Models (Unit 3.1.2) and DMRB 12.1.1.

1.10.12 It should be noted that there are other factors that affect travel choices. Probably the most important omission is that of reliability. This has been subject to a considerable amount of research, and mechanisms whereby reliability can be included in the generalised cost formulation are currently under development. While recent research has suggested that this factor should include an explicit estimate of travellers' scheduling costs (i.e. the cost they place on not being able to travel at their preferred time), in practice the approach is likely to be based on including the standard deviation of travel or waiting time in the utility/generalised cost function, representing the uncertainty of arrival time. These effects are potentially important, and useful advice can be found in research reports on the DfT website (see www.dft.gov.uk/pgr/economics/). An interim approach is given in The Reliability Sub-Objective (Unit 3.5.7) as a post-model calculation.

1.10.13 For public transport schemes, the effects of comfort may need to be represented. Stated Preference exercises have produced plausible results whereby time spent in crowded or standing conditions incurs a higher cost than time spent seated in relative comfort. In these circumstances the estimation of the generalised costs of using public transport has an additional cost related to the degree of overcrowding, which in turn depends upon the number of passengers and capacity of the service, in terms of seating and standing capacity. This is most likely to be relevant only in peak-hour travel conditions. To be effective, models including an overcrowding feature need to be embedded in a feedback procedure so that they are demand-sensitive. In principle this is necessary if overcrowding changes significantly in either the base or forecast situations.

1.10.14 The example below, from a rail model, shows how the impact of seating and standing capacity can be modelled as influencing the perceived journey time by using a Crowding Factor F_c.

where : V = Volume

C_s = Seating Capacity

C_t = Total capacity seating and standing

In this model, the Crowding Factor increases the cost of in-vehicle time by a factor which is zero when 60% of the seats are occupied, rising to 1.12 when all the seats are occupied and to 2 when all the standing room is full.

1.10.15 In general, because the generalised cost methodology is relatively robust, the inclusion of additional elements does not present major modelling problems for demand forecasting. If required, it should be possible to build models that extend the standard definition of generalised cost, and also allow for greater behavioural variation between person-types and purposes.

1.10.16 All the above discussion has related to a (dis)utility function where the generalised cost is made up of a weighted linear combination of quantities such as time, distance toll etc. It is however possible that the (dis)utility function may include these quantities in a non-linear form e.g. costs may be expressed logarithmically. In these situations the concept of generalised cost, measured in time units, with a constant relationship between time and cost quantities does not hold. Such a modelling system is currently rare in the UK (the 2004 PRISM model of the West Midlands currently being the only example), but is more common elsewhere in Europe.

Composite costs

1.10.17 Where more than one demand response is incorporated in the demand model, the transfer of costs from one demand response to another is expressed through the concept of composite costs. The estimation of these composite costs takes account of the generalised costs and shares of the alternatives from the demand response considered lower in the modelling hierarchy. This concept is explained in more detail in VDM Key Processes (Unit 3.10.3).

2. Further Information

The following documents provide information that follows on directly from the key topics covered in this TAG Unit.

3. References

DETR (July 1998) A New Deal for Transport: Better for Everyone (available at www.dft.gov.uk/about/strategy/whitepapers/previous/anewdealfortransportbetterfo5695).

DfT (2000) Land-use indicators and trip-end models - Final Report, January 2000. DfT (available at www.dft.gov.uk/pgr/economics/rdg/).

TEMPRO Trip End Model Program (available at www.dft.gov.uk/pgr/economics/software/tempro/).

P.J. Mackie, A.S. Fowkes, M. Wardman, G. Whelan and J. Nellthorp (Institute for Transport Studies, University of Leeds) and J. Bates (John Bates Services) (2003) Value of travel time savings in the UK: summary report. ITS Leeds, January 2003 (also available on DfT web site).

National Travel Survey 1998-2000 (DfT) update, July 2001, DfT.

IHT (1996). Guidelines for Developing Urban Transport Strategies. Institution of Highways and transportation.

4. Document Provenance

This Transport Analysis Guidance (TAG) Unit reflects the consultation comments received on the Model Scoping Stage of the draft Variable Demand Modelling Advice produced by TRL in June 2003.

Technical queries and comments on this TAG Unit should be referred to:

Integrated Transport Economics and Appraisal (ITEA) Division
Department for Transport
Zone 3/06, Great Minster House
76 Marsham Street
London SW1P 4DR

E-mail: itea@dft.gsi.gov.uk
Tel: 020 7944 6176
Fax: 020 7944 2198