Chintakayala P. Investigation spatial transferability of models estimated on stated choice data collected on four different continents. Presented at the European Transport Conference 2010; October 11, 2010. Glasglow, Scotland.

Given increasing survey costs, transferring model estimates obtained from one location and survey and applying them to another location is becoming increasing appealing. The transfer of previously estimated model outputs to new application contexts has the potential to reduce the need for new large scale data collection in the new application context as well as reduce the effort required to develop new models. As such, significant savings in cost and time can be achieved. Nevertheless, advantages in time and cost savings may be outweighed due to biases introduced if the transferred model does not adequately represent the behaviour of individuals in the new application context. The majority of applications dealing with model transfer involve data collection within the same country or continent, or with relatively homogenous cultures. In this study, we compare the ability to transfer model outputs based on the same survey collected across four different continents, each with their own unique cultures and identities. The data which we use here were collected in Australia, Chile, Taiwan and the United Kingdom. By using the same survey and survey processes across each location, we are able to discard effects related to the implementation of different survey instruments and/or interviewing process, and truly test whether model transfer can be achieved across such a diverse group of countries. The main objective of this paper is to compare alternative methods of spatial transfer of discrete choice models across the four heterogeneous regions and cultures with particular interest being what outputs, if any, can be transferred. As previously identified within the literature, there are two main factors that impact upon the ability to transfer a model estimated from one set of data collected in one region and time, to another region and time. The first factor involves transfer bias, represented as differences between parameter estimates from the original model to corresponding coefficients associated with the application context data. The second factor is that of statistical reliability. Although the second factor is of interest, it is with the first issue that we are primarily interested in here. That is, given data from the four study areas, we are able to test which estimates may be transferred from one region to another. Of particular interest are whether estimated models, parameters, values of time, and elasticity measures, may be transferred from one area to another.

Share on: