Discrete Choice Model For Resident Travel Willingness

Discrete choice model is a widely used model for the trip split model. Oneimportant study of this model is to obtain more accurate estimated coefficientsfrom surveyed data. This thesis studies the coefficient estimation of discretechoice model in transportation choice modeling when some the surveyed dataare missing. For the uncertainty of traffic survey, it is very common that some ofthe respondents refuse to answer all the questions, while these data areespecially important for the estimation of the coefficients. In this paper, westudy in detail that under this circumstance, how to use all the respondents’ datato finish the coefficient estimation of discrete choice model to improve themodel’s precision.We start from summarizing the common methods for coefficientsestimation of discrete choice model. Specifically, we focus on the calculation oflikelihood function in Logit model and Mixed Logit model. To meet thenumerical calculation in R, the variables in utility function are divided into three types: individual coefficient, common coefficient variables, individual-alternative variables and alternative-alternative variables. A detailed example isalso given to illustrate the coefficient estimation in R language.Based on the introduction to the likelihood, the solution methods fordiscrete choice model when missing data are at random are investigated. Here,we propose a two-stage model: First we use Multiple Imputation method to fillin the missing data and obtain several complete data tables; then all the completedata tables are use to estimate the coefficients respectively; finally, differentestimation results are combined together. A simple R simulation is alsoconducted, which indicates that this two stage model can get more preciseestimated coefficients.Lastly, the thesis discusses the situations that the missing data are non-ignorable, i.e., dependent of those missing data. We adopt the Monte CarloExpectation Maximization (MCEM) method to estimate the likelihood using allthe observed data. The computational issues are clarified in detail. Thesimulation results show that the likelihoods calculated by MCEM method satisfythe true likelihood much more than piecewise delete method.