Introduction

As social animals, many of our behavioral decisions, including travel-related ones, are made in coordination with members of the social networks we are embedded in. However, while it has been acknowledged that mobility patterns are strongly interwoven with social networks (Puhe et al. 2021), joint decision-making processes, particularly related to social activities remain poorly explained in traditional behavioral models. A key reason for this is the lack of empirical data, and the difficulties associated with collecting such data in the first place.

Axhausen (2005) was among the first in the transportation field to point out the importance of social networks in explaining joint travel and identify data requirements to do so. Building on his work, in recent years, several ego-centric network data-collection efforts have been conducted to get a better understanding of ego-centric social networks characteristics and social interactions (Parady et al. 2021). Such efforts include similar surveys in Canada (Carrasco and Miller 2006), Switzerland (Frei and Axhausen 2008; Kowald and Axhausen 2012; Guidon et al. 2018), the Netherlands (van den Berg et al. 2012), Chile (Carrasco and Cid-Aguayo 2012), Japan (Parady et al. 2018, 2019, 2020) and the U.K. (Calastri et al. 2020). These efforts have largely focused on the influence of social network characteristics on social activity frequency (Carrasco and Miller 2006; Frei and Axhausen 2008; van den Berg et al. 2012; Calastri et al. 2017; Parady et al. 2020), location type (van den Berg et al. 2010, 2014), travel distance (Moore et al. 2013), activity duration (Nurul Habib et al. 2008; Habib and Carrasco 2011), and travel mode (Moore et al. 2013; Sharmeen and Timmermans 2014). These studies have incorporated network level characteristics such as network size, relational attributes like relationship type, relationship strength, and homophily, and geographical characteristics such as ego-alter distance. More recently, Han et al. (2023) used data from a follow-up survey to an egocentric network survey conducted by Parady et al. (2020) where egos were asked about their most recent eating out activity for each clique identified in the original survey, including destination location, date and time, origin location (of ego), and other places usually visited with this clique. They showed that explicitly incorporating average travel time of all participating members of a clique increases the predictive ability of the destination choice model by up to 32% against a model considering only egos’ travel times, a considerable increase in performance.

A key limitation of these efforts is that since data is collected using an ego-centric approach, the data that can be collected on alters (other group members) is limited to what ego can observe and recall. This limitation is particularly critical for modeling group activity patterns beyond frequency, since spatio-temporal constraints are a key set of constraints defining travel (Hagerstrand 1970). For example, while ego might be able to recall roughly where alters live, they might not recall or even know more specific details of alters’ trips such as origin locations, travel modes, or critical spatiotemporal constraints.

Regarding the modeling of the joint-decision making process in social networks, Arentze and Timmermans (2008) proposed a microsimulation framework to model activity travel patterns within networks, explicitly considering the formation of social networks, the generation of social activities, and the influence of social networks on preferences and behavior of individuals. This model is formulated in such a way that is consistent with existing activity theory (such as needs-based theory of activity generation) and social network analysis theory such as reciprocity, transitivity and homophily of social ties, social influence, etc. Extending this framework, and based on agent interaction theory (e.g., Kraus 1997), Ronald et al. (2012) proposed an experimental model that considers group negotiations on type, purpose and location of activities. However, due to lack of empirical data, model parameter estimation and model validation remains a pending task for both models.

An early application using empirical data can be found in the work of Arentze (2015). Using data from an experimental activity travel-survey task with simulated group settings, he proposed an empirical model that assumes a negotiation process. He showed that the negotiation influences groups’ spatial choices and that fairness is a significant factor in individual social preferences for joint activities, especially when travel costs are negotiated. This study used completely hypothetical settings and unlabeled group members and activities. While this was done by design to avoid interference of actual preferences, this means that this approach cannot capture the actual idiosyncrasies of groups and their effects on the decision-making process of actual choices.

Most of the work on joint-decision-making has, however, focused on intra-household interactions. Several studies have used empirical data to model intra-household decisions, either considering individual preferences exogenous to the group choice such as Molin et al. (2002) and Zhang and Fujiwara (2009) in the context of residential location, and Srinivasan and Bhat (2005) in the context of activity participation, or endogenously, such as Beck and Rose (2019) in the context of household vehicle choice. The work of Beck and Rose is particularly relevant to this study, as it used an interactive agency choice experiment (IACE) to make individual preferences endogenous to the choice of the group, an approach that can be easily extended to out-of-household interactions. IACE experiments are based on a sequence of simultaneous choices by agents, followed by rounds of feedback, review and revision of initial preferences. More specifically, agents are asked to make initial choices independently. If there is an agreement, the choice experiment is terminated. If there is no agreement in the agents’ choices, then each agent receives feedback about other agents’ preferences (i.e., the choices made by other agents) and are then asked to revise or keep their choices. This process, which can be extended to as much rounds as the researcher desires, allows the researcher to track how agents’ preferences are modified from their initial preferences. However, the experiment design might not reflect the true negotiation and decision-making process among households or groups. Furthermore, IACE is also subject to the same key limitations of SP experiments, that is, the lack of consistency between stated and actual choices.

In the context of driving cessation behavior among the elderly in Japan, and taking into consideration that the decision on whether to cease driving or not is usually made in conjunction with other household members, Fukui et al. (2020) conducted in-depth interviews to observe each household member’s preference and intention on driving cessation of elderly drivers, and whether elderly drivers gave up or intended to give up their driver’s license after discussion with other household members. Such an approach allows for explicitly accounting for the effects of the content of the communication itself. The main limitation of this study is that questions related to the behavior of interest (i.e., driving cessation) were answered in preference space, which can be significantly different from actual decisions.

Against this background, in this article we propose x-GDP (Text-aided Group Decision-making Process Observation Method), a novel survey method to collect data on joint activities and their underlying joint decision-making processes. We implemented the method for joint leisure activities with a particular focus on destination choice, but this method can be easily generalized to other dimensions of travel behavior. Through this method we are able to observe not only the outcome (i.e., the final location chosen) but also the decision-making process itself, including the alternatives that compose the choice set, individual and clique characteristics that might affect the choice process, as well as the discussion behind the choice via texts. Observing such a process will allow us to first understand the decision-making process qualitatively, including how alternatives are weighted, how members interact with each other, and finally how the choice is made, and then move on towards a more quantitative understanding of these processes.

Text-aided Group Decision-making Process Observation Method (x-GDP): an overview

The main objective of this method is to collect real-time data on the joint decision-making process of travel-related activities of a given clique, a group where all members know each other. In a nutshell, the general idea of x-GDP is to ask participant cliques to coordinate an activity (or set of activities), using a group chat interface. In this paper, we use as a case study eating-out activities, because eating-out is the most frequently executed joint-leisure activity (Stauffacher et al. 2005). A key aspect of this method is that participants have to actually conduct the activity coordinated in the group discussion (proof of execution is one of the conditions to receive the participation monetary incentive), hence, there are incentives in place to guarantee a real discussion that takes into consideration the preferences and constraints of clique members, thus overcoming the SP limitations stated earlier. The experiment was approved by the Research Ethics Committee of the Graduate School of Engineering, the University of Tokyo.

Figure 1 illustrates the flow of an x-GDP survey. In particular, the crux of the method is Step 3, executed over a virtual meeting (via any market-available web-conference app). The widespread of tele-commuting and tele-conferencing resulting from the COVID-19 pandemic made such a virtual implementation much easier as people are now very familiar with such meeting tools.

Fig. 1
figure 1

Flow of an x-GDP survey

Full size image

Step 1: recruitment and pre-registration

x-GDP requires the participation of existing cliques. As such, registration of all members is necessary for schedule coordination. Since sampling frames of cliques do not exist, there were some challenges to sample recruitment. In the experiment presented in this study, we targeted cliques composed of at least one University of Tokyo (hereinafter UTokyo) student to simplify the sampling process. This was also done to limit to some extent the spatial distribution of participants to cliques with similar daily life activity spaces (i.e., given the expected sample size, we wanted to avoid sampling cliques with very sparse and spread-out geographical distributions that would make destination choice modeling difficult). Provided this condition was met, no constraints were imposed on the eligibility of other members.

Recruitment was done via social media (the Urban Transportation Research Unit Twitter account), and registration was done through a dedicated website on the Urban Transportation Research Unit’s homepage, where groups were asked to register the names and contact information (e-mail) of all members.

In spite of the nonprobability sampling method, it is important to point out that given the student population of UTokyo was 27,233 students as of November 2022, we collected a sample equivalent to 2.28% of the student body. In addition, compared to the general population, the university population is rather homogeneous in terms of sociodemographics. As such, we were able to capture a set of participants that resemble the general characteristics of the student body of the university. For example, the male ratio among UTokyo students in our sample was 79.7% compared to the 76.5% reported in the official university statistics as of November 2022. If we narrow it down to undergraduate students, which account for 72.2% of the UTokyo students in our sample, the male ratio is 79.3% compared to 79.9% reported in the official university statistics. Given this fact, and the experimental nature of this study, the current sampling method was judged adequate. In total, data on 816 individuals, belonging to 217 cliques was collected. Out of the 816 participants, 76% were U Tokyo students, 20% were students from other universities and 4% were non-students (see the appendix for a summary of individual sociodemographic characteristics of the sample).

Step 2: virtual meeting schedule coordination

Once registration of all cliques was done, scheduling coordination was conducted via online forms. As shown in Fig. 2, the Schedule Coordinator matched Experiment Moderators (the person in charge of moderating and guiding the experiment over Zoom) with cliques, based on their time availability. In the experiment presented in this study, schedules were allocated in 2-week phases and constrained by the availability of Moderators. That is, registered groups were contacted via e-mail and asked to specify their time availability for a set of time slots spanning a two-week period. If a particular clique remained unmatched during a particular phase, they were reassigned to subsequent phases by the Schedule Coordinator. E-mail reminders were sent on need-basis to unresponsive groups.

Fig. 2
figure 2

Simplified diagram of the logistics of the x-GDP experiment after recruitment

Full size image

Once a clique was successfully matched, all members were informed via e-mail of the date and time of the experiment, as well as the Zoom link. In addition, detailed experiment explanations (including conditions for payment of participation reward) and informed consent forms were also sent. Scheduling was a rather challenging task given the need to coordinate time for all clique members and the Experiment Moderators (in total 4–7 persons). Early-morning (before school or work) and late evening (after school or work) were popular time slots. Although rare, in instances where some group members did not show up to the zoom meeting, they were contacted either by the Experiment Moderator or by other group members. In case of no-shows, the experiment was conducted if and only if the group size did not fall below the required minimum of three members. If the group size fell below the minimum, groups could decide to either re-schedule or withdraw. No-show information was documented in the data.

Step 3: zoom-moderated survey execution.

The Zoom-moderated experiment was executed as scheduled in Step 2. This was the crux of the experiment (see Fig. 2 for the logistics of this step). Although explanation of the experiment and informed consent forms were sent beforehand, at the beginning of the experiment the Experiment Moderator explained verbally the details of the survey as well as the conditions for the payment of the participation reward.

Guided by the Experiment Moderator, participants were first asked to respond to Survey 1 and Survey 2 via an online survey platform developed in-house. The Experiment Moderators shared via Zoom chat the link to the questionnaires as well as login information. Survey 1 collected data on individual socio-demographic characteristics (see the appendix for a summary of individual level characteristics of the sample). Survey 2 collected data on clique characteristic and was answered at the clique-level. To do so, the Experiment Moderator asked one member to share his/her screen while answering the survey. Participants could freely speak during Survey 2.

After Survey 2 was completed, the Experiment Moderator invited all members to a LINE group chat (LINE is the most popular instant communication freeware app in Japan, and we correctly anticipated that all participants would already have an account by the time of the experiment). Participants joined the LINE group chat on their personal accounts; however, the Experiment Moderator joined via a Line Works account (a cloud-based business chat tool that can link to LINE). This was done for privacy and ethical reasons as well as data management reasons. By connecting via a Line Works interface, data was not saved in the Experiment Moderators’ personal accounts. Instead, it was saved centrally in the Line Works database, which the Experiment Managers can control access to. This also protected the privacy of the Experiment Moderators who needed not share their personal account information.

In the LINE group chat, the Experiment Moderator asked the clique to first decide the date and time of the activity. In the study presented in this article, two scheduling constraints were imposed. First, for management reasons, the date of the activity must be within a maximum of two weeks from the day of the experiment (three weeks was allowed at the discretion of the Experiment Moderator if no consensus was reached). This time horizon was also set to allow for enough flexibility in the scheduling process, given that as group size increases so do scheduling difficulties. The second constraint was that the activity must be done from the evening on (from 17:00~). This was done to reduce the temporal variability of the activities and simplify the modeling process later on. Note that these constraints can be relaxed or modified depending on the activity or model of interest to the researcher.

Once the date and time for the activity were defined, participants were asked to elicit potential areas and restaurants to execute the activity. Note that we use the word restaurant in the broadest sense of the word to include establishments like bars, cafés, izakaya (Japanese pub), etc. There was no upper bound on how many candidates could be elicited but participants were asked to propose at least one location per person. Before moving on to the discussion phase to choose the activity location, participants were asked to respond to Survey 3, which asked them to rank the elicited candidate locations in order of their personal preference. This was done anonymously so that responses were not affected by the opinions of others.

After completing Survey 3, participants were asked to discuss and decide the location of the eating-out activity. No guidance was given regarding how to make this decision, so each clique was free to choose their own method. There was no time constraint imposed on the LINE group discussion. Once a decision was made, the LINE group discussion part of the experiment was completed. The average duration for the LINE discussion section including time decision, preference elicitation and location decision was 35 min (S.D. 16.42 min). The moderator then asked participants to respond to Survey 4 via a web-survey (at the clique level). This survey collected data on the chosen location as well other candidate locations considered. To avoid the issue of untraceable locations, participants were asked to use store hyperlinks from either Tabelog (a restaurant review site in Japan) or Google maps. Other links were allowed if and only if the restaurant did not have a Tabelog or a Google place link. Out of the 1,188 unique restaurants elicited during the course of the experiment, we were able to identify 1,182 (99.5%) of the restaurants via their public links and collect additional data on these restaurants.

Finally, once Survey 4 was completed, each participant was asked to report their expected schedule for the day of the activity in the form of an activity diary (Survey 5). This was done via a visual and interactive interface that greatly reduces the response burden (See the supplementary files for a screenshot of the survey interface).

Step 4: activity execution

On the morning of the day of the planned activity, participants were sent a reminder via LINE and were given explanations about proof-of-execution. More specifically, participants were asked to submit:

  1. 1.

    Their location during the activity via LINE

  2. 2.

    A picture in front of the restaurant along with a mobile phone showing date and time

  3. 3.

    A group picture inside the restaurant

  4. 4.

    Receipt showing the amount spent

Step 5: post-activity survey

Using the same interface as Survey 5, data was collected on the actual schedule executed on the day of the activity. To reduce the response burden, their responses for Survey 5 (i.e., the expected schedule) were presented first and respondents could edit these when changes in the scheduled had occurred.

Step 6: payment

A monetary incentive of JPY 4000 (approx. USD 29.80 as of Feb. 20, 2023) was provided to participants who responded to all surveys and provided proof-of-execution. Irrespective of reason, for participants who did not provide proof-of-execution (including non-participation) or did not complete Survey 6 after participation, the incentive was reduced to JPY 1080 (approx. US$8).

Summary

To summarize, Table 1 compares the type of data that can be collected with this approach vis-à-vis other approaches in the literature, assuming the target choice is a destination choice. In particular, x-GDP not only allows the researcher to observe both the expected behavior (at the time of scheduling) and actual behavior of all group members but allows for the observation of the decision-making process in a quasi-naturalistic manner.

Table 1 Comparison of methods to collect data on group decision making process and behavior for the case of a destination choice
Full size table

Data characteristics and preliminary findings

Clique characteristics

Table 2 summarizes the clique-level characteristics of the sample. Clique size was limited by design so it ranges from three to five persons. Relationship length was measured for all the possible dyads within the clique (1191 dyads). For the majority of the dyads, the relationship length is longer than three years. As expected, eating out constitutes the most common joint activity of all cliques (44.7%) followed by hobbies (21.66%) and university club and circle activities (17.05%) such as sports, arts, etc. In terms of within-clique hierarchy, 72.4% of cliques had no hierarchy. Two-level hierarchy cliques accounted for 18.4% of cliques. In such cliques there are only two degrees of hierarchy (for example, all members are either undergraduate 1st year or undergraduate 2nd year). Hierarchy was self-reported by cliques, but it is important to note that in the context of Japanese social networks, hierarchy is usually very clearly defined as age or grade differences. While in student groups in many countries, “hierarchy” might be defined as a function of other attributes such as academic or sport prowess, popularity, etc., in the Japanese context, age seniority is the strongest criteria defining hierarchy among groups. In such a context, yielding to, or attempting to accommodate the preferences of the senior members of the group is likely to be observed. This phenomenon is well documented in Chie Nakane’s Japanese Society (Nakane 1970). Although hierarchical structures manifest this way in the Japanese context, more generally, it has been hypothesized that joint decisions need not be fully consensual and might be often driven by power differentials with one or more persons having a larger weight on decisions (Neutens et al. 2008), which merits further exploration.

Table 2 Clique level characteristics of sample (n = 217)
Full size table

Contact frequency was measured as meeting or ICT-mediated frequency of at least three members of the clique (contact between dyads are not considered). 12% of cliques meet once to three times per week, but the majority meets roughly once a month or less often.

Event characteristics

Figure 3 illustrates both the chosen restaurant location as well as other considered candidates. The first thing to point out is the agglomeration of locations around Tokyo sub-centers such as Shibuya, Shinjuku, Ikebukuro, Ueno, Tokyo and Ginza connected via the Yamanote loop line, in addition to areas around the University of Tokyo’s Komaba and Hongo Campuses. Historically, the Tokyo sub-centers have exhibited high degrees of agglomeration of commercial and other facilities due to their high levels of access both from the railway-connected suburbs as well as other central areas. In addition, smaller agglomerations can be seen around the intersection of railway lines even though they are not central.

Fig. 3
figure 3

Location of chosen restaurants and alternatives considered during the experiment

Full size image

Table 3 summarizes the characteristics of the scheduled events. As mentioned above, the start time of the joint eating-out (hereinafter, the main activity) was constrained by design to after 17:00. 80.6% of all activities were conducted on weekdays.

Table 3 Characteristics of scheduled joint eating-out activity (n = 217)
Full size table

Figure 4 shows the origin–destination (OD) distance distribution for the trips towards the main activity location as reported by the clique on Survey 4, and the trips towards the individually top-ranked location as reported in Survey 3. We hypothesized that individuals would highly rank locations that are closer to them because the preference elicitation survey (Survey 3) was anonymous. However, contrary to our expectations, there were no large differences at the aggregate level between OD distances for the chosen restaurant by the clique and for the individually top-ranked restaurants. The median difference was 450 m. However, since individual preferences were measured after alternative elicitation, members might have also considered the alternatives proposed by other members, which can be one factor explaining this result.

Fig. 4
figure 4

OD distance distribution of trips towards restaurant chosen by clique (a) and its respective box plot by hierarchy (c). OD distance distribution of trips towards the individually top-ranked location (b) and its respective box plot by hierarchy (d)

Full size image

Regarding differences by hierarchy level, we hypothesized that more junior members of the clique would defer to the more senior members’ preferences in terms of trip distance, however, although a weak association between hierarchy and trip distance was observed, it was not as strong as we expected. Note that four-level hierarchy was omitted from the figure, because there were only 2 cases out of 217 that fell in this category.

Decision-making process characteristics

Table 4 summarizes the reasons for choosing the chosen restaurant for the main activity. These data were categorized from free-answers in Survey 4, where the clique was asked to briefly summarize the main reasons behind their choice. As expected, restaurant quality and accessibility were the most frequently mentioned factors (78.8% and 57.1% of cliques mentioning them, respectively). This is also consistent with the attitudinal responses collected in the individual survey (Survey 1) where respondents were asked to rate on a 7-point Likert scale (1 being not important at all, 7 being extremely important), the importance they place on different factors when eating out with a group (Fig. 5). Group evaluation of restaurants and group transit access were rated six or seven by 71.2% and 76.7% of the individuals, respectively. In contrast, individual evaluation of restaurant and individual transit access were rated six or seven by 65.7% and 59.3% of the individuals. This might explain why no large difference were observed in Fig. 4 between OD distances for the chosen restaurant by the clique and for the individually top-ranked restaurants. That being said, what Table 3 does not capture is whose accessibility is being prioritized, or whose preference. As shown in Table 5, in less than 12% of cases, all members’ individually top-ranked locations were actually chosen, with this percentage reducing as clique size increases. Furthermore, irrespective of clique size, in around 17–20% of cases, no one’s top-ranked location was chosen by the clique. This might be a result of compromise among members and/or changes in preferences during the process as a result of new information. To illustrate this point, let’s briefly consider two instances. In the first instance, a group of four members elicited eleven alternatives, out of which, two members chose a Japanese hot pot restaurant chain as their first choice. However, one of these members (who chose the hot pot place) said that while he wanted to go to this restaurant, he was OK with going somewhere else if others wanted to, suggesting that his preference was rather flexible and that he was willing to accommodate the rest of the group’s preferences. Consequently, alternatives were reevaluated, and this time, an Oden (Japanese fishcake and vegetable stew) restaurant was considered a good option because it was winter. Ultimately, this restaurant was chosen, at a location where the group could go for drinks to an Izakaya restaurant afterwards.

Table 4 Reasons for choosing restaurant categorized from free answers (multiple criteria allowed)
Full size table
Fig. 5
figure 5

Factors considered important for group-level restaurant choice by individuals (n = 816)

Full size image
Table 5 Degree of matching between individually top-ranked locations and clique choice
Full size table

The second instance is a three-member group, whose meeting time and place was determined by a previous event that was expected to have ended by 20:00. This group elicited five alternatives, and the first preference for two members was a café restaurant near the meeting area, the other first preference was a Ramen (Japanese noodles) shop. During the discussion, the group decided that ramen would not be a good option because “the noodles would get soggy while they talk”. Here, a previously elicited Okonomiyaki (sometimes referred to as a Japanese pancake) restaurant alternative was re-evaluated. This option was finally chosen, when members noticed that while the café restaurant was open at night, it closed at 21:00, so would not be a good choice given the group’s previous constraint.

As shown above, the fluidity of the choice process and individual preferences underscores the importance of observing the actual decision-making process to gain a better understanding of within-group dynamics, but also highlights the challenges of modeling such processes.

Two case studies

To further elucidate the properties of the data collected we will briefly introduce the decision-making process in two particular cases, as summarized in Figs. 6 and 7 using information from Surveys 1 to 5 as well as the LINE group discussion text record (a). In particular, the plots of members’ schedules (b) and activity places (c) were created using data from the individual preference elicitation survey (Survey 3) and the expected activity diary of the meeting day (Survey 5).

Fig. 6
figure 6

Extract of collected data for a clique (example 1). 1a. LINE chat excerpt. 1b. Schedule of all members on the day of the activity (Survey 5). 1c. OD lines to individually top-ranked location

Full size image
Fig. 7
figure 7

Extract of collected data for a clique (example 2). 2a. LINE chat excerpt. 2b. Schedule of all members on the day of the activity (Survey 5). 2c. OD lines to individually top-ranked location

Full size image

The first clique (Fig. 6) is composed of five same-year students (no hierarchy). Two of the members had previous commitments on the suburbs of Tokyo on the day of the activity (1b and 1c). In this particular case several features of the decision-making process can be highlighted (1a). For instance, Mr. A pushed from early in the discussion for his preference, eating French food at Ginza, an upscale district in central Tokyo. Other members, like Mr. C, had a personal preference but showed high degree of agreeableness and willingness to compromise for the group, stating: “My preference is for meat, but if everyone is in for French at Ginza, I don’t mind.” While other alternatives were raised during the discussion such as Japanese BBQ or and oyster bar in Shibuya, Mr. A kept insisting on his preference by posting a link to the restaurant’s online site and menu: “Let me give you an idea of what French at Ginza will be like.” It should be noted that most members’ individually top-ranked locations were close to their expected origin locations on the day of the activity. Another constraint in the process was that some students were under 20 years old, hence could not drink alcohol, which tilted the choices towards restaurants rather than bars or Japanese izakaya. In the end the group agreed on Mr. A’s preference. In this particular case, Mr. A’s strong opinion clearly influenced the final decision, given the other members’ agreeableness and willingness to compromise. In other words, the weight of Mr. A’s opinion was larger than other members'. At the same time, we can speculate that had other members had similarly strong opinions, the resulting outcome might have been different. Such information cannot be observed from the outcome alone, but we were able to capture it with the proposed x-GDP method.

The second clique is composed of three futsal club friends, one of them being one year more senior than the other two (two-level hierarchy). The joint activity time was set based on two time constraints. First, Mr. A had a part-time job at Shinjuku until 19:00 and second, all members wanted to watch the FIFA World Cup (Qatar 2022) after dinner. Once the time slot was defined, several candidate locations were proposed, but they were all in Shinjuku. A possible reason for this is that Mr. A had a non-flexible activity schedule during that day. It is also worth noting that Mr. A was the more senior member of the group. The rest of the discussion focused on the restaurant type, such as hotpot, Brazilian BBQ and gibier. In this case, economic constraints were taken into consideration and Brazilian BBQ was selected.

Potential avenues of research

In the above sections we have summarized the details of the proposed x-GDP survey method and introduced the results of a survey implementation focusing on joint eating-out activities in the Greater Tokyo Area, giving a detailed overview of the survey components, execution logistics and initial insights on the data. To conclude we want to discuss potential avenues of research that can be pursued with this kind of data.

Identification of patterns of decision-making

We have illustrated with a few examples that clique-level decision making is rather heterogeneous. As such, a first step forward is the qualitative evaluation of the discussion logs in search for patterns of decision-making. An applicable methodology is Grounded Theory developed by Glaser and Strauss (1967), a qualitative method used to uncover behavior of groups and other social processes (Noble and Mitchel 2016). Such processes can form the basis to quantitatively model joint decisions. Of particular interest is the understanding of the conditions that give way to a specific decision-making pattern or decision rule, such as relational characteristics of group members (e.g., do decision rules between groups that exhibit hierarchical relationships differ from groups that do not?), or spatio-temporal constraints (e.g., do the weight of a particular member's opinion changes given specific spatio-temporal constraints?). Such analyses become feasible due to the quasi-naturalistic nature of x-GDP, and although as discussed in the introduction, several theoretical frameworks have been proposed to model joint decision-making, the nature and quality of the collected data might help discover processes not yet identified.

Modeling group decision explicitly considering the decision-making process

The high resolution of the data collected via x-GDP can be used to model joint choices explicitly considering the decision-making process, as illustrated in Fig. 8, including the effect of relational characteristics among members, members’ contextual attributes such as spatio-temporal constraints as well as individual preferences. Differing from existing methodologies, since x-GDP allows for the observation of the discussion process itself in a quasi-naturalistic manner, models can be tailored to specific decision-making patterns observed in the data with few a-priori assumptions. This also means that x-GDP allows not only the consideration of members’ and groups’ attributes, but also the content of the bargaining process itself. As such, we would be able to empirically confirm, for example, (1) how various constraints, which would vary across group members, are embedded into their group decisions, and (2) to what extent a group decision-making model produces different results compared to the conventional individual decision-making models. These analyses would allow us to identify potential biases caused by ignoring the group decision-making process in travel demand models.

Fig. 8
figure 8

Conceptual diagram of joint decision making within group members

Full size image

Estimation of joint accessibility

Another potential avenue of research is the empirical estimation of measures of joint accessibility. Departing from Miller (1999)’s work on space–time accessibility, theoretical models to estimate joint accessibility based on time geography and random utility theory have been proposed by Neutens et al. (2008). This model takes into account network-based travel times, individual activity schedules, the attractiveness of facilities and their temporal availability, and can explicitly account for the relative influence of participants. While Neutens et al.’s model focuses on the maximization of the group’s joint accessibility, it can be generalized to accommodate other decision-making rules (e.g., rules identified during the qualitative analysis of the discussion logs). However, estimation of model parameters has been identified as a major challenge that may limit the applicability of joint accessibility models (Neutens et al. 2008), with lack of data being a key limitation. Here too, empirical data collected through x-GDP is adequate for such a task, where model parameters could be estimated with a random utility model of joint destination choice.

There are important applications of joint accessibility measures. For example, Farber et al.(2013) have hypothesized that cities or regions with low levels of joint accessibility might be conducive to segregation, social isolation, and reduced social capital, while higher joint accessibility levels might be conducive to artistic and intellectual innovation, as well as greater levels of social capital. The relationship between activity participation and social capital has also been quantitatively analyzed by Tahlyan et al. (2022).

Choice set generation analysis

Finally, choice set generation is a persistent problem in location choice models, and the need for proper specification of choice sets to avoid biased parameters is well acknowledged (Pagliara and Timmermans 2009). At the same time, true choice sets are rarely, if ever observed in traditional surveys, so choice set generation, while based on theory, usually cannot be validated against empirical data. Furthermore, as opposed to other choice dimensions like travel mode, the issue of choice set generation for spatial choices makes dealing with spatial aggregation and alternative sampling unavoidable. Against this background, x-GDP allows researchers the possibility of observing the actual choice set against which actual choices (as opposed to stated choices) are made, at the elemental alternative level. Such information can be used to test the effectiveness of different choice set generation regimes considering both spatial aggregation and alternative sampling.

Transferability to other contexts

Although in this study we have focused on eating-out destination choice, x-GDP can be applied to other dimensions of travel. This method is particularly well suited for frequently occurring joint activities in daily life given difficulties associated with sampling groups and the requirement to actually execute the target activity. While we have targeted groups of 3–5 members to be able to observe more complex processes, there is no particular reason why x-GDP cannot be applied to dyads.

For less frequent activities such as vacations, vehicle purchase, or home relocation, the method would require some tailoring. For example, the researcher could target specifically groups or households that are considering vacations, vehicle purchase, or home relocation, and inquire about the actual behavior (actual place visited, actual vehicle purchased, or actual relocation place) via a follow-up survey. This, however, would result in increases in sampling burden and execution costs associated with keeping track and contacting respondent at different time periods.

In addition, the proposed methodology is general enough to be transferable to a different socio-cultural context. For example, while we discussed that in Japanese society, hierarchical structures are largely defined by age differences between members, this is not necessarily a universally transferable trait. However, with minor modifications to the survey instrument contextual idiosyncrasies can be easily accounted for.

Limitations

In terms of limitations, while qualitative analysis of the data suggests a relation between members' personality traits and the decision-making process (as shown in the first case study) direct measurements of personality traits were not conducted in this study. The main reason for this is that existing instruments to measure personality traits are long and would considerably increase the response burden of an already long experiment. For example, the Tokyo University Egogram (TEG) which measures personality traits is composed of 53 questions and requires 5–10 min to answer (Oshima et al. 1996). That being said, abridged versions of the Big-5 personality test have been validated for Japan (Namikawa et al. 2012), and even though this version still includes 27 questions, simple modifications to the x-GDP survey could be made to accommodate it.

Another limitation of the study is that the provision of the monetary incentive of JPY 4000 might affect the choice set formation for some groups. That is, participants might consider alternatives that they would not have consider otherwise given the incentive at hand. While this might be true to some extent, the discussion logs suggest that in many cases considered alternatives were alternatives that the groups usually frequented, some members had been to in the past, or were already interested in going to in the first place. The incentive provision also means that cost effects estimated in a model will be conditional on this incentive. The average spent amount per capita was JPY 3846 (min. JPY 444; 1st quartile: JPY 2348; median JPY 3610; 3rd quartile JPY 4953; max. JPY 12773) suggesting that on average, groups aimed to spend the incentive amount in the event (See Table 8 in appendix). The provision of the incentive was necessary to (i) incentivize participation in a high response burden experiment and (ii) so that participation in the study did not result in a monetary loss, as execution of the activity was required. That being said, while this might affect the choice set formation for some groups, it should not affect the decision-making process itself, which is what we are interested in observing.

Finally, the generalizability of the results presented here is limited by the fact that the majority of participants are students at a specific university in Tokyo. As stated earlier, due to the experimental nature of the study, this was done first to simplify the recruitment and execution of the experiment, and to limit to some extent the spatial distribution of participants to cliques with similar daily life activity spaces. By doing so, we avoided, given the expected sample size, sampling cliques with very sparse and spread-out geographical distributions that would make destination choice modeling difficult. However, this should not affect the methodological contribution of this study.