|Regimes in Social-Cultural Events-Driven Activity Sequences
Theo Arentze and Harry Timmermans
Urban Planning Group
Eindhoven University of Technology
PO Box 513
5600 MB Eindhoven
Abstract. This paper reports the results of an empirical analysis of factors influencing activity sequences that are triggered by social events. The study is motivated by the intention to examine the wider context in which activity-travel decisions are made and to model such decisions under different time horizons. We assume that social events trigger a series of interrelated activities. A Bayesian network is applied to data collected in the Eindhoven region, The Netherlands and shows that indeed many travel choices are influenced by particular events and that these influences vary by socio-demographic variables.
Keywords: Activity-based models, events, Bayesian networks, network learning, regimes.
The field of activity-based analysis and modelling has truly matured over the last decade. In addition to the hundreds of descriptive and analytical studies, focusing on particular aspects of activity-travel patterns, recently the first fully operational, comprehensive activity-based models have become available and are currently tested and applied in daily transportation planning practice. For example, Vovsha et al. (2005) reporting the development and application of discrete choice based models in the United States, mention Portland, New York City, San Francisco County and Columbus, Ohio. Similarly, CEMDEP developed by Bhat and his co-workers (2004) as a set of loosely integrated advanced econometric models has found application in Texas. Likewise, Famos (Pendyala et al., 2005) and its predecessor PCATSs (Kitamura and Fujii, 1998) have been applied in Japan and Florida. Finally, Albatross (Arentze and Timmermans, 2000, 2003) has received a stamp of approval by the Dutch Ministry of Transport, while TASHA (Miller, 2005) is getting close to application.
While certainly a lot of effort is still required to move activity-based models to practice, the academic community can start addressing a wider set of questions. One of the limitations of current activity-based models is their focus on a typical daily pattern. Most models are cross-sectional in nature and are not really based on a behavioural process representation. There is a need to describe, analyze and model activity-travel decision-making for different time horizons. To contribute to this line of research, the Urban Planning Group has started various projects. First, a stream of research has been developed in an attempt to model activity rescheduling behaviour, ultimately leading to the Aurora model (Timmermans, et al., 200; Joh et al. 2004. 2005, 2006; Arentze et al., 2005). The main application of this model concerns activity-travel rescheduling during the implementation of a planned daily schedule in case of unexpected events. Second, realizing that activity-travel patterns emerge in response to a changing (transportation) environment and institutional context, with habitual behaviour being a special case, we have developed a conceptual framework for learning and adaptation to simulate the dynamics in activity-travel behaviour (Arentze and Timmermans, 2005, 2006a). Even in stationary environments, individuals learn about their environment when conducting activities and their changing cognitions in turn may lead to changes in activity-travel patterns. This will be reflected in gradually changing patterns: short term dynamics. Third, at the other end of the spectrum, we have argued that sudden changes in activity-travel patterns may occur due to critical incidents and key (lifecycle) events. Events such as moving house, a change of job, a divorce, birth etc may trigger major changes in certain facets of activity-travel patterns. Based on some early descriptive analysis, a model linking lifecycle events to changes in travel resources and travel behaviour is under development (Waerden et al., 2002 a,b; Verhoeven et al., 2006).
In an attempt to ultimately link all of this into a consistent theoretical framework, Arentze and Timmermans (2006b) have suggested a needs-based theory which not only allows incorporating these dynamics into a single operational framework, but also addresses the issue of complementarity and substitution of activities. Activities are conducted to satisfy certain needs. Using needs as a conceptual basis is similar to the theoretical orientation underlying Tasha (Miller and Roorda, 2003; Miller et al., 2005; Roorda, et al., 2005). The main difference is that our theoretical framework has some explicit mechanisms that endogenously account for the dynamics of activity generation due to changing needs.
In this context, it should be realized that one of these basic needs is social interaction. Moreover, we should realize that individuals and households are part of a social system with its institutions, regulations and culture. The social system has a major impact on the organisation of society and therefore also on how particular needs are organised over time. Belonging to a social system means that individuals and households organize their life in accordance with social norms, rules and expectations, conducting a set of interrelated activities to maintain the social system. Examples are religious events such as Christmas, celebrations of anniversaries and holidays. We have therefore argued (Arentze et al., 2006) that activity-based models should gradually be incorporated into event-based models in an attempt to incorporate the larger decision context and various time horizons into the modelling effort, thereby enriching the modelling of transport demand.
We assume that these events trigger a series of interrelated activity sequences. Individuals engage in activities to make preparations (before the event), implement the event itself and to take care of the event aftermath (after the event). This notion is akin to Axhausen’s concept of a project as a set of coordinated activities tied together by a common goal or outcome (Axhausen, 1998), albeit positioned differently in the sense that our focus is on socially and culturally triggered events. To test that assumption, in this paper we will report the results of an empirical analysis, which was conducted to find regimes in activity sequences, triggered by social events. To our knowledge, it is the first empirical study examining such sequences. Before discussing the results, we will however first elaborate some key concepts and the approach taken. This will be followed by a summary of the data collection and a motivation of the modelling approach. The paper is completed with some conclusions and a discussion of potential avenues of future research.
CONCEPTS AND APPROACH
We define events as those occasions or activities in the life of an individual that are special in that they fall outside the domain of regularly recurring activities conducted by the individual to fulfil his/her basic needs at person or household level. Even when they occur on a regular basis, events are salient in the perception of the individual and they tend to break the routine of every day life. An event does not only disturb the daily routine on the day it happens. Typically, an event requires also that certain activities are implemented before and after the event to make necessary preparations and take care of the aftermath of the event. An example of an event is the celebration of a person’s birthday. Such events typically entail activities before and after the event; before, to make sure that food and drinks are available for the guests and the house is tidy and, after, to dispose of the litter and bring the house back in an orderly state after the guests have left. Operationally, we define an event as any activity or occasion that: (i) occurs once or maximally several times a year, (ii) has an impact on the daily activity schedule and (iii) may entail preparing activities and/or aftermath activities before and after the event. As implied by this definition, what counts as an event for one person does not need to be an event for another person. Table 1 shows a classification of events which illustrates the concept and which we used in the survey (explained below) to assist respondents in recalling relevant events. As the list indicates, most classes of events are social in nature, i.e. involve gatherings of people to celebrate a special occasion related to a person (e.g., birthday) or a community (e.g., thanksgiving). Other events may have another goal as their primary purpose such as for example recreation (e.g., a day-out), health (e.g., a therapy) or maintenance (e.g., a house job).
A question that interests us here is: what is the impact of such events on travel demands of individuals? Events often involve groups of people coming together at someone’s home or at some central location to participate in a joint activity such as a celebration, a sports event, a religious event, etc. But also events that do not involve gatherings of people may still generate trips, namely when they necessarily (e.g., a hospital) or preferably (e.g., a beach) take place out-of-home. Apart from the event itself, also the preparation and aftermath activities, such as for example dropping off or picking up persons, going to a barber, getting money, etc., may induce travel. In sum, an event possibly can generate travel related to the event itself and through the activities preceding and succeeding the event.
In an earlier paper (Arentze et al., 2006), we described results of a first exploratory analysis, based on the same data collection as we use in the present study. There we focused on the frequency and timing of events. We found that events are quite diverse with respect to the extent that their temporal regimes are flexible and influenced by socio-economic variables or relatively fixed by existing institutions. In the present study, we focus the attention on the activity and trip generation characteristics of events.
In particular, we develop and test a model to explain and predict activity-travel chains associated with particular events. Figure 1 schematically shows the conceptual framework that we propose for the model. The unit of observation is an event. That is to say, we assume that an event is given and we are interested in predicting the event type and associated activities and travel. We consider variables of the individual, household, the spatial environment (e.g., such as accessibility of locations) and temporal setting (e.g., time of year) as explanatory variables. The variable Event defines the type of event in terms of the 8-way classification mentioned above (see Table 1). An event may involve a trip, namely if the event takes place somewhere away from the person’s home. Furthermore, an event may require one or more preparation activities, which precede the event, and one or more aftermath activities, which succeed the event. We define preparation and aftermath activities in terms of a general classification of activities which include in-home as well as out-of-home activities. An out-of-home location will generally involve a trip. A trip has several facets such as transport mode, travel time and possibly others. Therefore, the node labeled ‘trip’ refers to a set of variables instead of just one variable.
In the scheme, solid arrows represent the causal relationships that we expect to find between these (sets of) variables. In words, we expect the following relationships. The event type influences the choice of activities preceding and succeeding the event. Furthermore, it has an influence on characteristics of the trip, if any, required for the event itself. In turn, the choice of activity type before or after the event has an influence on travel time and transport mode of the trip, if any, needed to implement the activity. In sum, according to this network, events generate and influence the choice of trips for the event itself and through activities required before and after the event. In turn, event type, given that an event occurs, is not a fully random variable in the model. Rather, we expect (and indeed as we found in the previous study) that probabilities of events co-vary with socio-economic and situational (spatial and temporal) variables. For example, a birthday party may have a higher occurrence probability in households with children compared to households without children. We also expect that these variables have an impact on trip choices, given activity characteristics. For example, whether or not the person has a driving license will influence the choice of a transport mode even if the activity is known.
Dashed arrows in the scheme represent causal relationships that we consider less likely, but that nevertheless may exist in reality. More specifically, we expect that event type does not have a direct influence on choice of a trip. The possible influence is supposed to run through the activity (i.e., event type influences the choice of activity and activity has an impact on the choice of transport mode and possibly other trip facets). Furthermore, we expect external variables not to have a direct influence on activities that precede or succeed an event. Rather external variables influence activity choices indirectly through events (i.e., attributes of a person influences probabilities of certain event types and event type influences probabilities of certain activity types). Thus, we expect that external variables and event-related activities are conditionally independent of each other, given that the event is known. We should emphasize, however, that the existence of conditional independencies in this model is dependent on the classifications used for activities and events. For example, if a relatively course classification of activities is used, event type may give additional information regarding activity type so that knowing the event type helps to predict trip choices even if in reality the causation runs through the activity variable. So, we must keep in mind that findings of an estimation of this model cannot be generalized beyond the classifications used.
We use a Bayesian Belief Network (or in short a Bayesian Network or BN) as a framework for the model. A BN is a directed acyclic graph (a DAG) where the nodes represent variables and the arcs causal or temporal relationships between variables (Pearl 1988, Heckerman et al. 1995, Spiegelhalter et al. 1993). The state of variables may be uncertain under any given configuration of evidence (i.e., data) about the variables. Note that the structure of the network defines for each node a set of parent nodes (node X is said to be the parent of node Y if an arc runs from X to Y). Associated with each node is a conditional probability table (a CP table or in short CPT) which defines the probability distribution across possible states of the node under each state configuration of the parent nodes (if any). Collectively, the CPTs of nodes are referred to as the parameters of a network. The parameters allow the system to calculate the probability distribution at each node and to update these probabilities each time when evidence about variables is entered to the network. Existing probability propagation algorithms are consistent with Bayes’ perception updating rule.
The BN is a well-suited formalism for the present modeling purpose for several reasons. First, we wish to predict the multiple variables involved in event-activity-travel chains simultaneously and a network is the most flexible structure conceivable in that case. Second, the core variables we are dealing with, namely events, activities and transport modes, are discrete variables and a BN, in contrast to structural equation models and artificial neural networks, is designed to deal with discrete variables. Most importantly, however, methods exist to empirically derive a BN from data. Estimating a BN involves two subtasks, namely first learning the network structure and then finding the parameters for that structure that best fit the data (i.e., estimating the CPTs). Having established the network structure, the estimation of CPTs is relatively straight-forward. EM-learning (Lauritzen 1995) is the commonly used method for CPT learning and is implemented in software for BN learning such as Hugin (Hugin Expert A/S, 1995-2005, Anderson et al. 1989), PowerConstructor (Cheng et al. 2002) and other packages. Therefore, most research in this field has been and still is devoted to the first subtask, network learning. Two general approaches have emerged from this work (see Cheng et al. 2002). Dependency-based methods develop a network based on tests of conditional independencies between pairs of nodes. These methods are also known as constraint-based learning methods as they try to find pairs of nodes that are conditionally independent of each other and then add this information as constraints in constructing a graph and determining the directions of relationships. On the other hand, scoring-based methods seek a structure that maximizes some measure of goodness-of-fit of the network on the data in terms of the joint distribution of all variables involved. In the present study, we use a dependency-based method developed by Cheng et al. (2001) which is well-tested and available through the software PowerConstructor.
To collect the data for estimation, we conducted a survey among a large and a-select sample of households in the Eindhoven region, the Netherlands, which involved two rounds of data collection. In the first round, respondents were asked to specify the events they anticipated in the year starting from the beginning of the next month at the moment of the survey (which was September). On a form, which had the lay-out of a calendar, respondents could indicate which events they anticipated would occur in each month of the year. For each event they then could indicate the date (as exact as possible), type of event and which other persons besides the respondent him/herself were involved. In the second round of the survey, the same respondents were asked to keep a diary of event-related activities in a month that we had designated to them. Respondents were asked to indicate for each day of the designated month which event-related activities, including the event itself, if any, they participated in and to describe these activities in terms of a number of attributes. These included the begin time, the end time, with whom the activity was conducted, the location, transport mode, travel time, travel party, the event with which it was related, the nature of the relationship with the event and the impact the activity had on the normal activity schedule. Respondents could specify the activity type by selecting an item from a pre-coded list of activities. This list was based on a classification of activities that we standard use in activity diary surveys (including 33 labeled out-of-home activities and 14 labeled home-based activities).
The month assigned to a respondent was determined in a weighted random way, as follows. The probability of assigning a certain month to a certain respondent was determined based on the distribution of events across the months included in his/her event calendar. The more events occurred in a month, the higher the probability of selecting that month and vice versa. The allocation procedure was constrained by a requirement that at sample level a sufficient number of cases (i.e., respondents) should be obtained for each month. In this way, we could enlarge the number of observations of event-related activities and, at the same time, obtain sufficient coverage of a year. However, July and August had to be excluded because of the high frequency of summer holidays in that period.
In both rounds, all persons older than 12 years in households that participated were invited to fill out the calendar (first round), event diary (second round) and a pre-questionnaire (preceding the first round). 815 Respondents completed the event calendar. Together, they reported 25,555 anticipated events, which comes down to a mean of 31.4 events per respondent per year or a mean of 2.6 events per respondent per month. 415 Respondents completed the event diary for a month. As a consequence of the way months were designated, the distribution of respondents across months is not completely uniform. December and January had the highest numbers (of 76 and 66 persons, respectively) and February and October had the lowest numbers (of 21 and 24 persons, respectively). For a more detailed description of the survey and sample characteristics readers are referred to Arentze et al. (2006).
The diary data allows us to determine which activities are associated with which events, the interval times between activities and events and characteristics of trips made to implement the activities. In terms of the relationship with the event, we distinguish three categories of activities, namely activities to prepare an event (before the event), activities to implement the event (during the event) and activities to take care of the aftermath of an event (after the event). In the following, as in the schematic representation of the conceptual model (Figure 1), we will not distinguish between the activity conducted to implement an event and the event itself. This means that we have three categories of activities: preparation activities, the event itself and aftermath activities. Before turning to the model-based analysis in the next section, we will first report some statistics describing the occurrence and timing of event-related activities.
A total of 772 preparation and 318 aftermath activities related to 3,200 events were reported over all months. Thus, on average a preparation activity occurs 0.24 times and an aftermath activity 0.10 times when an event takes place. Events show a considerable variation regarding the number of preparation activities. 84.8 % of the events have no preparation activity, 15.1 % have a single preparation activity and 1.2 % two preparation activities. For the aftermath activities, these numbers are 95.2 % (no aftermath activity), 4.6 % (a single aftermath activity) and .2 % (two aftermath activities). We should note, however, that the window of observation of an event is limited to the period of a month over which respondents could report the event-related activities. Especially for events taking place at the beginning or the end of the month preparation and aftermath activities could have occurred that were not reported since they occurred before or after the recording period. Probably, the under recording caused by the limited recording period is not that serious. In terms of timing of activities relative to the event, it turns out that 30.9% of preparation activities take place on the same day, 33.0% of these activities on the day before and 10.1% two days before the event. The percentages decline rapidly as interval time further increases. More than 87.0% of the preparation activities fall within a period of 6 days prior to the event. For aftermath activities, we find that 35.7% of these activities take place on the same day, 22.6% one day later and 4.8% 2 days after the event. 84.8 % of the aftermath activities take place within a period of 6 days after the event.
Table 4 shows for each event type the probabilities that certain activities occur as a preparation or aftermath activity for that event. In this analysis, we classified the activities into 6 categories: 2 home-based activities and 4 out-of-home activities. As the totals in the last row reveal, the probability of observing a preparation or aftermath activity at all varies considerably between event types. The probability is largest for Special-day events (32.6%) and Person/family/relatives events (32.0 %) and second-largest for Maintenance events. Note that the events mentioned are not only social events, that motivated our study. In case of other events, probabilities are considerably lower and vary between 9 and 13%. We also see that there is considerable variation in activity type. For Person/family/relatives and Special-day events a shopping activity has the highest probability of occurring as a preparation or aftermath activity (19.4% and 18.3% respectively). Leisure and social activities overall have the lowest occurrence probabilities suggesting that, as expected, preparation and aftermath activities mostly have a non-leisure character.