Despite the fact that our existing models are not up to the job of predicting travel behavior in today’s rapidly changing landscape, and despite considerable evidence that attitudes help us explain behavior more completely and more meaningfully, attitudes are nowhere to be found in practice-oriented travel demand forecasting models. Two main objections have been raised to their inclusion: they are too cumbersome to measure, and difficult-if-not-impossible to forecast. This paper reports on the considerable progress that has been made toward overcoming the first objection, through the use of machine learning methods to train a prediction function on smaller-scale research-oriented survey datasets, and then applying that function to impute attitudes into large-scale household travel survey datasets. Internal evaluations show that we can estimate attitudinal factor scores with moderate fidelity when using socioeconomic/demographic, land use, and targeted marketing variables, and with high fidelity when using just a few attitudinal marker variables. External evaluations demonstrate that the imputed attitudes lead to improved behavioral insight and predictive ability for forecasting-oriented models. With respect to the second objection I have only sketched some ideas for moving forward, but there are clearly some practical steps that could be taken at very little marginal cost, such as including as few as 10 attitudinal marker statements in future household travel surveys.
Read full abstract