An activity-episode generation model that captures interactions between household heads: development and empirical analysis

https://doi.org/10.1016/S0191-2615(01)00039-XGet rights and content

Abstract

In this paper, we develop an approach for modeling the daily number of non-work, out-of-home activity episodes for household heads that incorporates in its framework both interactions between such members and activity setting (i.e. independent and joint activities). Trivariate ordered probit models are estimated for the heads of three household types – couple, non-worker; couple, one-worker; and couple, two-worker households – using data from a trip diary survey that was conducted in the Greater Toronto Area (GTA) during 1987. Significant interactions between household heads are found. Moreover, the nature of these interactions is shown to vary by household type implying that decision-making structures and, more generally, household dynamics also vary by household type. In terms of predictive ability, the models incorporating interactions are found to predict more accurately than models excluding interactions. The empirical findings emphasize the importance of incorporating interactions between household members in activity-based forecasting models.

Introduction

Since the late 1970s, the rapidly expanding literature subsumed under the activity-based paradigm has increased significantly our understanding of urban travel behavior and provided insights into new approaches to replace current models of urban travel demand – namely, the Urban Transportation Modeling System (UTMS). Unlike trip-based approaches, the activity-based paradigm, more commonly known as activity analysis, recognizes explicitly that travel is a demand derived from the need to participate in out-of-home activities. In other words, discrete activities or patterns of activities are investigated, not trips. Jones et al. (1990) identify several features of the paradigm including recognition that decision-making occurs in a household context, taking into account interactions among household members. This differs, however, from the reality of activity-based research to date.

With few exceptions, the decision-making unit in both empirical studies and modeling efforts is the individual, not the household. Obviously, this implies that few studies have investigated interactions among household members in terms of their activity-travel behavior. Of those that have, the importance of such interactions is clearly demonstrated. For example, the findings of Kostyniuk and Kitamura (1982) indicate that participation in joint out-of-home activities during the evening is quite common for some couples. Specifically, young couples without children and dual-earner couples in which both husband and wife work on a given day are oriented toward such activities. In a more recent study that employs a structural equations model to capture interactions between male and female household heads in terms of their participation in work, maintenance and discretionary activities, Golob and McNally (1997) find that male work activity governs interactions. Their study shows that an increase in such activity leads to an increase in female maintenance activity and travel and to a decrease in female discretionary activity and travel. Furthermore, the number of young children in the household is related to the substitution of work and maintenance activities between male and female household heads. Finally, Gliebe and Koppelman (2001) estimate a share model of daily time use in two-adult households that distinguishes between independent and joint activity participation. Their findings demonstrate the importance of employment levels and the presence of children on time allocation to such activities.

Despite the paucity of empirical research to date, there is convincing evidence which suggests that interactions among household members can no longer be ignored in the development of behaviorally sound, activity-based forecasting models (e.g. Gliebe and Koppelman, 2001; Golob and McNally, 1997; Kostyniuk and Kitamura, 1982). Moreover, from a pragmatic point of view, such household-level models (i.e. models that capture interactions among household members) are superior to those developed for individuals in that they offer potentially more accurate and reliable forecasts. Two reasons are suggested for this. First, individual-level models are incapable of handling complex responses to travel demand management (TDM) strategies. For example, a person who performs an activity during the evening commute may forgo the activity when working a compressed workweek. This response obviously favors the alternative work-schedule strategy. However, the individual-level model does not consider that this activity may be reassigned to another household member who also undertakes it after work. In this case, the TDM strategy would prove ineffective in reducing travel demand. Second, such models do not account for joint out-of-home activities. This means that predictions of activity-travel behavior are likely to be inaccurate. For example, multiple out-of-home activities may be predicted for household members when, in fact, only one exists.

To date, only Wen and Koppelman's model is developed at the household level, accounting explicitly for interactions among household members (cf. Wen and Koppelman, 1999, Wen and Koppelman, 2000). This is surprising given the fact that the household is incorporated conceptually in other activity-based forecasting models (e.g. Kitamura et al., 1996; Stopher et al., 1996). Wen and Koppelman postulate a decision-making process whereby households first generate out-of-home maintenance activities and then assign them to specific members for execution. They incorporate the number of such activity episodes in their modeling framework, which is a necessary prerequisite for forecasting models. Furthermore, they extend the household decision-making process to include the allocation of automobiles to household members, which is defined in terms of the number of activity episodes for which an automobile is available. A three-tier nested logit model is used to implement the modeling framework, which, despite its advances over previous research, still has two shortcomings. First, it is limited to couples that do not engage in out-of-home maintenance activities together, thereby ignoring joint activities. Second, although the nested logit model allows the household decisions to be estimated simultaneously, it does not account for the ordinal nature of the first decision – that is, the number of out-of-home maintenance episodes.

Obviously, as highlighted by the discussion above, there remains a need to incorporate interactions among household members in the development of behaviorally sound, activity-based forecasting models. Our paper presents one such attempt at addressing this issue. Specifically, the daily number of non-work, out-of-home activity episodes is modeled for the heads of three common household types: couple, non-worker, one-worker and two-worker households. To capture interactions between such household members in terms of their participation in independent and joint out-of-home activity episodes, we develop the trivariate ordered probit model. To our knowledge, this is the first attempt to extend Mckelvey and Zavoina's pioneering work on the ordered probit model to account for three ordered decisions whose outcomes are determined by a joint decision-making process (cf. McKelvey and Zavoina, 1975).

Our focus on modeling the daily number of non-work, out-of-home activity episodes for household heads is based on a consensus among researchers that the decision-making process underlying observed activity-travel behavior can be split into two distinct, yet interrelated, components: activity generation and activity scheduling (e.g. Bhat and Koppelman, 1993, Bhat and Koppelman, 1994; Doherty and Axhausen, 1999; Ettema et al., 1993, Ettema et al., 1996; Gärling et al., 1989; Recker et al., 1986a, Recker et al., 1986b; Wen and Koppelman, 1999, Wen and Koppelman, 2000). While activity scheduling has been the focus of much research activity in recent years, the same is not true for activity generation. Despite this situation, however, a common assumption underlying the very few studies that have addressed conceptually or empirically the subject of activity generation is that household members generate activities collectively and then assign them to individual members for execution (e.g. Bhat and Koppelman, 1993; Wen and Koppelman, 1999, Wen and Koppelman, 2000). Obviously, this process is based on interactions among household members. For this reason, activity generation is a natural starting point for developing behaviorally sound, activity-based forecasting models.

Our approach to modeling activity generation captures interactions between household heads in terms of all non-work, out-of-home activity episodes that they participate in on a daily basis. Although our approach does not distinguish between specific activity types, it has several strengths, in addition to capturing interactions, which define its usefulness for practical forecasting purposes. These strengths include explicit recognition of membership identity within a household, explicit recognition of activity setting (i.e. independent and joint activities) and explicit recognition of activity episodes as units of analysis. Three alternatives exist that are also capable of capturing interactions between household heads – namely, nested logit models (e.g. Wen and Koppelman, 1999, Wen and Koppelman, 2000), structural equations models (e.g. Golob and McNally, 1997) and share models (e.g. Gliebe and Koppelman, 2001). Moreover, unlike our approach, these models are capable of distinguishing between activity types. Despite this strength, however, they appear to be of limited value for activity generation because they lack one or more of the strengths identified above with reference to our modeling framework. For example, structural equations models and share models are used to model continuous variables, not ordinal ones. Although our approach to modeling activity generation within households is not necessarily a panacea, it does provide a robust alternative to other methods that are capable of capturing interactions between household heads.

The remainder of this paper is organized as follows. Section 2 presents the modeling framework for the trivariate ordered probit model. Section 3 describes the data and sample used in the empirical analysis. Section 4 presents the empirical findings. The contributions of this study to activity-based research and suggestions for further research are summarized in Section 5.

Section snippets

Overview

As mentioned, we develop the trivariate ordered probit model to model the daily number of non-work, out-of-home activity episodes for household heads. The strengths of the model, in the present application, are threefold. First, the model accounts for two activity settings: independent and joint activities. These settings are based on the number of household heads participating in activity episodes. In other words, the presence of other household members, such as children, is not used to define

Data and sample

The data for this research are derived from a trip diary survey that was conducted for the Ontario Ministry of Transportation during February and March 1987. The sample of households for this survey was selected from households that responded to a much larger survey conducted in the Greater Toronto Area (GTA) in 1986 – that is, the 1986 Transportation Tomorrow Survey. The trip diary survey used a mail questionnaire to obtain socio-demographic information on each household surveyed, including

Model specification and variable selection

To ascertain whether or not household heads determine jointly the number of independent and joint, non-work, out-of-home activity episodes that they undertake on a daily basis, two models were specified for each of the above household types. The first specification, which we denote as a joint model, captured any unobserved interactions between household heads by correlating the error terms in (1). For the second specification, these error terms were left uncorrelated by setting the ρs in (4) to

Conclusions

At the outset of this paper, we argued that interactions among household members must be incorporated in activity-based forecasting models if they are to be behaviorally sound. As was mentioned, however, very little research has been undertaken in this important area. Our study, therefore, represents one effort to redress this issue. Specifically, we extend McKelvey and Zavoina's seminal work on the ordered probit model to account for three ordered decisions whose outcomes are determined by a

Acknowledgements

The Social Sciences and Humanities Research Council of Canada provided financial support for this research through grant 752-96-1505, which was awarded to the first author.

References (19)

There are more references available in the full text version of this article.

Cited by (0)

View full text