Multiple Imputation Methodology for Missing Data, Non-Random Response and Panel Attrition

Modern travel-behavior surveys have become quite complex; they frequently include multiple telephone contacts, travel diaries, and customized stated preference experiments. The complexity and length of these surveys lead to pervasive problems with missing data and non-random response biases. Panel surveys, which are becoming common in transportation research, also suffer from non-random attrition biases. This paper shows how Rubin's (1987a) multiple imputation methodology provides a unified approach to alleviating these problems. Before discussing solutions to problems caused by missing data and selection, it is important to recognize that their presence causes fundamental problems with identifying models and even "simple" population estimates. Section 2 reviews this work and stresses the need to make generally untestable assumptions in order to carry out any inference with missing data