Robust Estimation Techniques for Trip Generation in Tennessee
Corresponding Author: Sumit Bindra, RSG
Presented By: Jason Chen, RSG
This presentation describes the robust estimation process using discrete generalized linear models (GLMs) for trip generation in the Tennessee statewide travel model.
Traditionally cross-classification method or, in some cases, linear regression is used for trip generation in trip-based travel demand models. Cross-classification models generally offer sensitivity to only two or occasionally three variables. A two- or three-dimensional cross classification can at times also suffer from low statistical efficiency in categories with too few observations (e.g. large HHsize and low number of vehicles).
Generalized linear models have distinct advantages for trip generation. They allow the incorporation of multiple factors affecting trip generation such as life cycle variables (e.g., number of seniors, children, workers, etc.). They provide maximum statistical efficiency with parameters supported by the fullest possible range of observations, minimizing issues common in cross-classification with too few observations for some rates. They do not require the response variable to have a normal error distribution, but rather can reflect the discrete nature of trip generation through the use of Poisson or Negative-Binomial distributions, and they can take advantage of robust estimation techniques.
The presentation will illustrate how k-fold cross-validation was used in model estimation and review the advanced variable selection methodology which tested various forms of the same or similar variables without including all of the forms in the model selection process because of their correlation. For example, household size can be represented by a continuous or discrete factor variable. The household size variable is also highly correlated to number of children or number of non-workers in a household and including all of these in the variable selection process is not advisable since together they are almost perfectly collinear. Ultimately, the presentation will illustrate how the use of these techniques produced robust trip generation models for Tennessee with sensitivity to a variety of factors that could not all be included in a cross-classification approach.