Measuring errors with latent variables in mode choice models
The Value of Travel Time (VTT) is fundamental in transport economics. In the last 30 years the best practice for VTT estimation has been to use Stated Choice (SC) data. However, there is now considerable evidence of reference dependence and gain-loss asymmetry in SC data, implying that such data do not reveal long-term preferences. This is a serious problem since the value of time is often applied in welfare analyses, where long-term stability of the preferences is a key assumption. A potential reason for the strong reference dependence found in SC data is the emphasis on a short-term reference point often used in SC data to reduce hypothetical bias. In the long-run there is no stable reference point. An alternative to SC data is to use revealed preference data and a mode choice model to estimate the VTT. Observed behaviour has adapted to travel conditions and should thus be ruled by long-term preferences. Many countries collect NTS (national travel survey) data and spend considerable resources on making them representative, which is an argument for using them for VTT estimation. However, a key problem in the use of NTS data for VTT estimation is the measurement errors in the travel time and travel cost variables. Time and cost in NTS data is either self-reported or derived from a network assignment model. In this project we explore the errors in the self-reported and network-computed time and cost variables by treating travel time and travel cost as latent variables in the estimation of a mode choice model. We also explore the errors in the time and cost variables in a descriptive analysis, for instance with regard to rounding errors and driving costs. We admit that we face a potential identification problem, i.e. that the random error in the choice model cannot easily be separated from error in the latent time and cost variables. In this case the assumption of the error structure in the choice model influences the estimated errors in the time and cost variables. We explore this issue by making sensitivity tests in the model formulation. However, given that we use a state-of-practice mode choice model, we argue that is it also relevant to explore the errors in the time and cost variables given this model. We use maximum likelihood measurement-error models, focussing on their application to mode choice models. To our knowledge, no previous study on large-scale transport models has explored the impacts of different model assumptions in error quantification in this way.