International Choice Modelling Conference, International Choice Modelling Conference 2017

Font Size: 
Assessing the Performance of Berkson-Theil Method on Multiple Choice Sets and Aggregated Choice Data
Weibo Li, Maria Kamargianni, Philip Krammer, Lynnette Dray, Andreas Schafer

Last modified: 28 March 2017

Abstract


One type of input data in choice modeling is that there are multiple or even a large number of choice sets and only an aggregated demand for each choice is available. Such data can usually be derived from official sources or big data platforms representing a high level of aggregation. An example could be when developing an air travel itinerary choice model at regional or national level, such that each different Origin-Destination pair will be a unique choice set with a number of itineraries in it as the choices. As an alternative approach to logistic regression based on maximum likelihood estimation (MLE), Berkson-Theil method that uses least squares is rarely being remembered. However, the method has significant practical advantages in terms of handling a large number of choice sets and software compatibility when dealing with this particular data type. As a result, this paper offers a leading research that assesses the performance of Berkson-Theil method in such a data case by comparing the model estimation results of Berkson-Theil method based on ordinary least squares (OLS) to a logistic regression based on MLE and also testing the predictive powers of the two methods. The comparisons reveal that the two methods can offer similar model estimation results; however, the results of logistic regression can lead to more accurate predictions. Heteroskedasticity is discovered in the end implying that the choice of OLS could be a cause for the lower predictive power of Berkson-Theil method. Overall, the findings suggest that Berkson-Theil method can be an effective approach in dealing with aggregated choice data with multiple choice sets and it may perform better if heteroskedasticity can be captured using weighted least squares (WLS) as the estimation technique instead. Choice modelers in air transportation and other domains that often deal with big data could therefore make use of this method given its significant practical advantages.


Conference registration is required in order to view papers.