International Choice Modelling Conference, International Choice Modelling Conference 2015

Font Size: 
The Temporal Transferability of Mixed Logit Mode-Destination Choice Models
James Fox, Stephane Hess, Andrew Daly

Last modified: 18 May 2015

Abstract


Key to appraising the impact of transport policy is to predict travellers’ choices of mode and destination, which is best done through the use of models that predict these choices simultaneously. Increasingly, especially in North America, mode and destination choice are being incorporated in activity-based model systems that predict travel and other behaviour over the whole day. Typically, these models are of the multinomial or nested logit form. In some cases, however, the models incorporate distributed parameters, for example some models contain a log-normally distributed cost parameter. However, while many studies have demonstrated that adding distributed parameters can significantly improve the fit of models to the estimation data in the base year, to the authors’ knowledge no study has investigated whether adding such parameters improves the temporal transferability of the models, i.e. their success in forecasting. This is a wider issue in the development of more advanced choice models, where, across different fields, the key emphasis has been on showing improved model fit in estimation and greater insights into behavioural diversity, but little or no discussion of forecasting accuracy. Since transport infrastructure frequently has a pay-back period of several decades, it is important that models developed now are reasonably applicable quite far into the future. Similar implications arise in other fields, for example policy interventions in health, environment and finance have impacts that can last many decades.

 

This paper seeks to address this evidence gap with a specific focus on transport, and the context of mode-destination choice models. Data from the Greater Toronto data has been used to develop mode-destination models from 1986 and 2006 data, allowing the temporal transferability of the models with and without distributed parameters to be assessed over a 20 year transfer period. The tests have been undertaken in a step-wise fashion. First, different model specifications have been investigated using multinomial logit models, and the transferability of these different specifications has been assessed. Then, distributed parameters have been introduced into the model specifications, allowing the transferability of models incorporating these terms to be compared to the results from the multinomial logit model tests.

 

To assess the transferability of the models, three groups of measures have been used:

 

· measures based on overall model fit which compare the fit achieved by the transferred model to the identical model specification re-estimated on the transfer dataset (e.g. comparing the fit achieved by transferring a 1986 model to predict the 2006 choices to the fit achieved when the same model is estimated from the 2006 choices)
· comparison of changes in the model parameters between the two years, and in particular changes in both the mean value and shape of the distributed parameters

· tests of the ability of the transferred models to predict observed changes in mode share and trip length

 

In each case the impact of allowing distributed parameters relative to fixed parameters is a particular focus.

 

The application potential of the research is that it will help developers of models incorporating mode-destination models, such as activity-based model systems, to assess the value of incorporating distributed parameters in terms of model transferability rather than on the basis of base year model fit alone.

 

The policy relevance of the research is that it will help model developers to more effectively deploy model development funds by re-focussing the discussion on the ability of the models to predict future behaviour, rather than the current focus on best fitting the base year choices without full consideration of the impact of model specification choice on forecasting. More generally, better forecasting models will help policy makers in reaching quite important decisions concerning future transport investments and management policies.

 

For the wider choice modelling community, the research gives valuable insight into the relationship between how well a choice model fits the estimation dataset and how well it predicts future behaviour. This is only really possible in the case where data for a forecast period is also available, as has been the case here, and is very different from the traditional approach of using hold-out samples from the same point in time as the estimation sample. This will be informative to choice modellers across a range of different contexts in considering on the appropriate specifications for their models.

 


Conference registration is required in order to view papers.