Capturing random utility maximization behavior in continuous choice data : application to work tour scheduling
Recent advances in travel demand modeling have concentrated on adding behavioral realism by focusing on an individual’s activity participation. And, to account for trip-chaining, tour-based methods are largely replacing trip-based methods. Alongside these advances and innovations in dynamic traffic assignment (DTA) techniques, however, time-of-day (TOD) modeling remains an Achilles’ heel. As congestion worsens and operators turn to variable road pricing, sensors are added to networks, cell phones are GPS-enabled, and DTA techniques become practical, accurate time-of-day forecasts become critical. In addition, most models highlight tradeoffs between travel time and cost, while neglecting variations in travel time. Research into stated and revealed choices suggests that travel time variability can be highly consequential. This dissertation introduces a method for imputing travel time variability information as a continuous function of time-of-day, while utilizing an existing method for imputing average travel times (by TOD). The methods employ ordinary least squares (OLS) regression techniques, and rely on reported travel time information from survey data (typically available to researchers), as well as travel time and distance estimates by origin-destination (OD) pair for free-flow and peak-period conditions from network data. This dissertation also develops two models of activity timing that recognize the imputed average travel times and travel time variability. Both models are based in random utility theory and both recognize potential correlations across time-of-day alternatives. In addition, both models are estimated in a Bayesian framework using Gibbs sampling and Metropolis-Hastings (MH) algorithms, and model estimation relies on San Francisco Bay Area data collected in 2000. The first model is the continuous cross-nested logit (CCNL) and represents tour outbound departure time choice in a continuous context (rather than discretizing time) over an entire day. The model is formulated as a generalization of the discrete cross-nested logit (CNL) for continuous choice and represents the first random utility maximization model to incorporate the ability to capture correlations across alternatives in a continuous choice context. The model is then compared to the continuous logit, which represents a generalization of the multinomial logit (MNL) for continuous choice. Empirical results suggest that the CCNL out-performs the continuous logit in terms of predictive accuracy and reasonableness of predictions for three tolling policy simulations. Moreover, while this dissertation focuses on time-of-day modeling, the CCNL could be used in a number of other continuous choice contexts (e.g., location/destination, vehicle usage, trip durations, and profit-maximizing production). The second model is a bivariate multinomial probit (BVMNP) model. While the model relies on discretization of time (into 30-minute intervals), it captures both key dimensions of a tour’s timing (rather than just one, as in this dissertation’s application of the CCNL model), which is important for tour- and activity-based models of travel demand. The BVMNP’s ability to capture correlations across scheduling alternatives is something no existing two-dimensional choice models of tour timing can claim. Both models represent substantial contributions for continuous choice modeling in transportation, business, biology, and various other fields. In addition, the empirical results of the models evaluated here enhance our understanding of individuals’ time-of-day decisions. For instance, average travel time and its variance are estimated to have a negative effect on workers’ utilities, as expected, but are not found to be that practically relevant here, probably because most workers are rather constrained in their activity scheduling and/or work hours. However, correlations are found to be rather strong in both models, particularly for home-to-work journeys, suggesting that if models fail to accommodate such correlations, biased application results may emerge.