Modelling non-ignorable attrition and measurement error in panel surveys: an application to travel demand modeling

Modern panel surveys frequently suffer from high and likely non-ignorable attrition, and transportation surveys suffer from poor travel time estimates. This paper examines new methods for adjusting forecasts and model estimates to account for these problems. The methods we describe are illustrated using a new panel survey of 1500 commuters in San Diego, California. These data are being collected to evaluate a federally-funded "Congestion Pricing" experiment investigating the impacts of allowing solo drivers to pay use freeway carpool lanes. The panel survey, begun in Fall 1997, collects data on travel behavior and attitudes at six-month intervals through telephone interviews. The panel sample is refreshed with new respondents at each wave to counteract the attrition between waves. Both the original and refreshment samples are stratified on commuters' mode choices (solo drive in free lanes, pay to solo drive in the carpool lanes, or carpool for free in carpool lanes) to insure sufficient sample size for estimating our models.