This presentation describes the development of a discrete choice model for vehicle ownership in the Dallas-Fort Worth region. The process covers four stages of data analysis for variable selection, estimation of the model, calibration of the model, and validation of the model. The observed data sets are provided from NCTCOG-NHTS 2009 survey. Analysis of data and feedback from estimation provides a list of final variables with the most significant impacts on vehicle ownership for each household. They are household size, number of workers in each Household, household income, residential density for each Tract, percentage of households with income equal or lower than 30k in each Tract, and number of transit routes in each Tract.
The estimation process considers both multinomial logit and nested logit model structures. The multinomial logit is eventually chosen because of its simple structure, lower computational requirements, and lack of evidence of the nest structure from observed data.
The calibration step is done based on 2005-2009 5-year ACS data. To apply the model, Tract level zone structure is used. However, ACS data did not include breakdown of households by income, household size and number of workers and it only provides marginal totals in Tract level. Therefore, Iterative Proportional Fitting (IPF) process is applied to obtain market segment values for each Tract. CTPP 2000’s data is chosen as the seed cell value. Calibration process moves beyond traditional goal of matching the regional targets. It also considers sub-regional targets and 0 car ownership share for large counties. These considerations are done as conditions in the optimization program for calibration. Brute Force Attacking (BFA) method is first applied to find a global optimal solution which is the starting point to further find the local optimal solution. Generalized Reduced Gradient (GRG) Nonlinear Solving method is executed to start with this point to reach the Karush–Kuhn–Tucker (KKT) conditions for local optimum. Inclusion of the sub-regional values in calibration decreases the degree of freedom in the calibration program and it prevents over calibration of the model.
The validation process includes County and Tract level vehicle ownership comparisons. Specific analysis is done for 0 vehicle ownership since the performance of the model and the number of observations are far less than other choices. Special graphs are created to show Marginal of Error (MOE) of observed values for households with 0 vehicle at each Tract. Two other types of comparison are utilized to evaluate the performance of the model for 0 vehicle households based on number of households and income groups.