International Choice Modelling Conference, International Choice Modelling Conference 2017

Font Size: 
A Peek at the Future: Capturing the Anticipation Effects in Dynamic Discrete Choice Models
Charisma Choudhury, Stephane Hess

Last modified: 28 March 2017

Abstract


Many of the choices we make involve ‘planning ahead’ based on anticipation and apprehension. For instance, investment decisions may be based on anticipated trends of the market; residential location/re-location choices may be based on anticipated life-events; route choices may be based on anticipated travel times, etc. From the modelling perspective, this has been partly captured in dynamic discrete choice models where a decision maker chooses the alternative in the current period that maximizes his/her expected utility over the current and future periods (Train, 2003). These models have however primarily focused on anticipation or apprehension involved with the attributes of the alternatives and less on anticipation and apprehension involving personal circumstances (e.g. getting married/divorced, having a child, changing a job, etc.). Availability of disaggregate panel data, where the evolution of personal circumstances are available alongside the evolution of choices and attributes of alternatives, provides us with the opportunity to incorporate these effects in the modelling framework. For example, dynamic latent plan models for driving behaviour (Choudhury et al. 2011) provide evidence that the observed decisions are manifestations of unobserved and evolving plans; dynamic lifestyle models (El Zarwi et al. 2015) demonstrate that evolution of lifestyles (as a function of the built environment and individual socio-demographics) can be used to predict structural shifts in travel behaviour. Recent researches have also focused on joint modelling of life events and travel choices (e.g. Crastes et al. 2016 and Eherke et al. 2015) where the correlation between the error terms have been accounted for within the Multivariate Probit framework. The models account for state-dependence, but ignore the planning ahead aspect of decisions.In this paper, we focus on modelling the effect of anticipated future life events (e.g. change in job, birth of a child, marriage/divorce, etc.), on present medium term decisions (e.g. car ownership). While long panel datasets allow us to know the future life events and their timeline, there are several challenges in incorporating them in the choice framework. Firstly, though the panel data provides us with the opportunity to observe the time of occurrence of the events, for certain life events like birth of a child, there may be discrepancies with the planned time. Since the anticipation effect is based on planned timings rather than actual time of occurrence, this introduces measurement errors in this variable and the observed times are mere indicators of the planned times. Further, like other plans, plans regarding life events may change dynamically as well leading to further modelling complexities. Secondly, there is significant heterogeneity in the planning horizon of people. This makes it difficult to ascertain from when the anticipation effect starts to impact the decisions. Thirdly, the panel data provides us with information within a certain time window and what happens after the observed time periods, though unknown to the modeller, can still be a source of anticipatory behaviour within the observed time period.In order to account for these issues, we extend the dynamic latent plan models in several directions (Figure 1). First, we assume a two level dynamic latent plan model where the first level involves planning related to life events (e.g. change in family composition), denoted by L in Figure 1. These plans are latent or unobserved as they may change over time. The observed life events (E) serve as indicators of the life plans (rather than directly observed variables) given the potential difference between their planned and actual timing. Further, the heterogeneity inthe plan-ahead time of the respondents is accounted for by probabilistic latent classes whereeach class has a different plan-ahead time. The second level captures the evolution of mediumlevel plans (e.g. plan to buy a car), denoted by P. These plans are influenced by the plans in thefirst level. This level is also unobserved since due to external constraints, it may or may not bepossible to materialize the plan to an action (e.g. actual purchase of a car). Hidden Markovformulations are used to account for the effects of state dependence and interactions betweenthe plans and actions in a tractable manner.Figure 1: Model StructureThe model structure is estimated using Puget Sounds Transport Panel (PSTP) which spans over1989 to 2002 and includes information about life events as well as travel decisions including carownership. In a previously completed analysis using nine waves of PSTP and age as the timeprogression variable, Dalal and Goulias (2013) found significant differences in the initial statusand growth trajectories for households that anticipate a life cycle change (e.g., a new child in thehousehold). The findings of our study quantify the anticipation effect and provide valuableinsights regarding the relationship between life events and travel decisions. The findings canlead to better transport planning modelsReferences:Train, K.E., 2009. Discrete choice methods with simulation. Cambridge university press.Choudhury, C.F., Ben-Akiva, M. and Abou-Zeid, M., 2010. Dynamic latent plan models. Journal ofChoice Modelling, 3(2), pp.50-70.El Zarwi F., Vij A., Walker J. (2015) Integrated Hidden Markov and Discrete Choice Models:Developing a forecasting framework for the transition matrix model, 14th InternationalConference on Travel Behaviour Research, Windsor, UK.Crastes dit Sourd, R., Calastri, C., Hess, S., Scheiner, J. , and Holz-Rau, C. (2016) Modellinganticipation and interrelations between long-term mobility decisions over the life course: arecursive multivariate probit approach, 5th Annual Conference of the European Association ofResearch in Transportation, the NetherlandsEhreke, I., Crastes dit Sourd, R., Beck, M., Hess, S., Axhausen, K.W., Holz-Rau, C. and Scheiner, J.,(2016) A Dynamic Approach to Long Term Mobility Decisions in the Life Course, TransportationResearch Board 95th Annual Meeting, USA.Dalal P. and K.G. Goulias (2013) Preparation and Adaptation Processes in Travel BehaviorDynamics: A Latent Growth Modeling Approach. Geotrans Report 2013-01-01. Santa Barbara,CA.

Conference registration is required in order to view papers.