Predictive Quality of a Time-Dependent Sequential Logit Evacuation Demand Model

Abstract

This study examines the predictive accuracy of a time-dependent sequential logit evacuation demand model. The model was estimated on revealed preference data and applied to stated choice data collected in the same survey. The survey was conducted in New Orleans, Louisiana, after Hurricane Gustav; 300 households participated. The stated choice scenarios of the survey consisted of nine hypothetical storms of varying strength, direction, forward speed, and time at which the storm made landfall, as well as decisions made by emergency managers on the type and timing of evacuation orders. The model was applied to each hypothetical storm, and the predicted evacuation demand from the model was compared with the response from the stated choice survey. The predictive performance of the model varied by hypothetical storm with a percentage root mean square error of 10.4 to 26.7 evacuations per 6-h period over 3 days. The model overpredicted total evacuation on all nine hypothetical storms by an average of 35.6%, with a range of 8.1% to 83.0%. The overprediction may be a result of the Katrina effect. After Hurricane Katrina, evacuation was more readily adopted. The model was estimated on the first major hurricane to hit New Orleans after Hurricane Katrina. Errors in predictions in individual periods are considered to be primarily a result of the limited number of stated choice scenarios used in the survey and the coarse representation of the storm these scenarios represent in the input to the model.