UC Irvine Institute of Transportation Studies

Modern panel surveys frequently suffer from high and non-ignorable attrition, and transportation surveys suffer from poor travel time estimates. The initial sampling process for most transportation surveys is also non-ignorable since rare travel modes are oversampled (and mode choice is the key dependent variable). This paper examines new multiple imputation methods for adjusting forecasts and model estimates to account for these problems in a new panel survey of 1500 commuters in San Diego, California. These data are collected to evaluate charging solo commuters to use an existing 8-mile underutilized freeway carpool lane. We illustrate the impact of attrition and measurement error on a standard conditional logit model of commuters' mode choice (solo drive in free lanes, pay to solo drive in the carpool lanes, or carpool for free in carpool lanes). Although the attrition rate between waves is 40% and non-ignorable, the quantitative impact on the results is negligible. However, measurement error in travel time does have an important impact on the key results from our model. Finally, failure to account for the measurement error process using multiple imputations yields a downward bias of at least 50% in the standard errors of the logit coefficient estimates.