Research On Logit Model Estimation Accuracy Based On D-optimal Design

Discrete choice model, such as multinomial logit model, nested logit model, mixed logit model are widely used in the transport field, mainly used to forecast traffic demand, policy and traffic service benefit evaluate. The data used to estimate model is the key to guarantee model accuracy, efficient questionnaire experiment design can maximize the collected information, yielding more reliable parameter estimates. D-optimal design aims to minimize the covariance matrix determinant of model, essentially minimize the standard errors. In order to confirm the efficiency of the new method, mainly compared with widely used orthogonal design.A traffic corridor between Chengdu and Longquan district is selected as analyze target, and assumes that the traffic corridor opened the subway, the stated preference questionnaires adopted D-optimal design and orthogonal design, then collect the travel preference data of travelers. Cost, accessing time, waiting time, time in car and egress time etc were selected as the utility variables to construct the multinomial logit (MNL) model, the stated preference data was used to calibrate the MNL model, and then compare the model results based on two experiment designs. The results show that compare to orthogonal design, D-optimal design is more flexibility in aspect of the attribute level combination number, and it can set combination number on the basis of demand. No matter what the number of sample is, the parameters estimation standard errors based on D-optimal design are all less than that are based on orthogonal design, this illustrate D-optimal design is superior to orthogonal design in aspect of model parameters estimation accuracy, especially obvious under the circumstance of small sample, this state D-optimal design can well guarantee model parameters estimation accuracy under the circumstance of small sample. D-optimal design also can avoid dominated alternatives, which can avoid parameters become biased.Finally, the author uses the better model result to analysis choice behavior of travelers. The model results show that cost, accessing time and time in car significantly impact on traffic mode choice of travelers. The author uses elasticity analysis to quantify significant factors influence on choice behavior of travelers, and then forecast the alternatives choice probability on the different levels of three significant factors separately. This can provide some suggestions for the formulation of traffic policies and service measures.