Aggregated Transport Forecasting Model

Development of an Aggregated Transport Forecasting Model (ATFM)

Stage 1 - Final report including extension

Joyce Dargay, Phil Goodwin and Mark Hanly
Centre for Transport Studies
University College London

27 September 2002

ESRC Transport Studies Unit
Centre for Transport Studies
University College London
Gower Street
London WC1E 6BT

Tel: +44 20 7679 1586
Fax: +44 20 7679 1586

http://www.cts.ucl.ac.uk/tsu/

Executive summary

This is the final report of Stage 1 of the development of an Aggregated Transport Forecasting Model, ATFM, carried out for the Department for Transport. ATFM is intended to augment the Department's existing, more detailed, forecasting models, by adding statistical transparency in calculation and understanding, a coherent revisiting of input and output elasticities, and an ability to look at some aggregate properties that are too complex to deal with in detail. Stage 1 focuses on surface passenger transport, at the national level. The report includes discussion of the characteristics and specification of the model, as defined in the Project Brief, Proposal and contract documents, and also the detailed results of econometric estimations carried out following an amended work-plan agreed in January 2002, and additional econometric work as agreed in July 2002.

Model Specification

The proposed model will forecast passenger kilometres, vehicle kilometres, fuel consumption and CO2 emissions, for car, bus and rail travel. These forecasts would enable tests to be carried out of the effects of changes in assumptions, such as income growth, or different policies affecting transport prices and service levels for each mode.

Demand responses are modelled on economic principles, and distinguish between short-and long-run effects, connected by an explicit time scale of approach towards equilibrium. Responses include both the direct effects on use of each mode, and also substitution among them, as determined by a matrix of direct elasticities and cross elasticities, which are explicitly specified in the model (though provision will be made to substitute user-specified values). It is proposed that price and service effects will be estimated by distinct elasticities, rather than by combining them into a generalised cost for which historical data is lacking. Car costs include the costs of both ownership and use, and allow for endogenous changes in fuel efficiency. The model includes some simplified aggregate demand-supply feedback effects, notably the effects of increasing supply on demand, but does not treat supply itself as endogenous.

Considerations of logic, theory and operational requirements are used to derive a checklist of all the elasticities and other parameters needed for the model to work. It is established that in principle it would be possible to find values for all these necessary parameters, by identifying a contingent set of prior values, by reference to literature reviews, assumptions in other models, logical inference and previous research carried out by the authors and colleagues. These values are intended to act as hypotheses to be replaced by directly estimated parameters wherever feasible, and default values otherwise.

Data and Estimation

In collaboration with DfT officials, a time series database for the period 1974-2000 has been collated for the variables of interest for the model, or in some cases the closest available proxies for them.

All the parameters specified in the passenger demand models are then explored by econometric estimation, using a variety of different model forms, mathematical specifications and estimation techniques, as devised by the research team and also as suggested by members of the project Steering Group. All results were judged by consideration of formal statistical diagnostics, internal consistency, a priori logical expectations on sign and in some cases size, consistency with other evidence where available, and implications for performance. It was found that statistically acceptable estimates could be made, using the available data, for the full set of income and own-price elasticities for all three modes, albeit with best-fit results that were in some cases somewhat different, though not always significantly, from current assumptions about the values of these elasticities. A number of other coefficients could not be estimated, with a satisfactory degree of statistical confidence, while at the same time meeting some strong logical constraints on sign, or in some cases simple common-sense tests on credible size. This especially related to cross-elasticities and service level effects, where there are known to be problems in the reliability of the data, some of which are expected to be small and therefore unlikely to be significantly different from zero, where nevertheless a small non-zero value may have material implications for forecasting and appraisal.

Forms of the model were developed which made use of the directly estimated parameter values for those variables where estimation had been successful, in combination with prior values (in preference to zero) for those parameters which could not be reliably estimated. These hybrid forms of model were tested using statistical criteria, especially focussing on their ability to reproduce the base data as compared with the preferred forms of model using only directly estimated parameters. Inevitably, the hybrid forms perform worse statistically, but only to a very small extent - barely distinguishable in terms of tracking the data over time. They do however perform materially better when judged by logic, comprehensiveness, consistency with previous research, and ability to take account of effects which are small but may be material in policy appraisal. Therefore it is the view of the authors that such a hybrid form will be more suitable for further development than a form in which no variables are included other than those which can be estimated directly from the data set to hand.

Next Steps

Although it would be possible to extend statistical estimation indefinitely, by collecting better data and experimenting with an even wider variety of model forms, we judge the combined equations to be sufficiently robust to act as the parameters for the next stage of model development, testing and validation. This would be carried out by constructing an operational version of the model sufficient to produce back casts and other output for diagnostic tests on performance and properties.

We suggest that it would be useful to redesign the phasing of model development, as some of the work which was originally expected to be in series would more sensibly be carried out in parallel. If this is agreed in principle, a new schedule of deliverables and milestones will be drafted, for discussion and agreement.