While virtual teams are not a new concept in the workplace (Warkentin & Bernak 1999), the rapid shift from in-person to virtual work during COVID-19 forced a rapid adaptation to new modalities of communication and collaboration (Klonek et al. 2021). Compared to physical teams, virtual teams often have unique considerations regarding how employees communicate and connect with each other, including information richness of communication available (Sprecher 2014), differences in spatial and temporal distribution of employees (Prasad, DeRosa & Beyerlein 2017), and cross-cultural interaction in the case of multinational teams (Collins et al. 2017). Adjusting to the unique challenges of virtual teams dramatically impacted team performance in the early period of the pandemic, and the continuation of the pandemic led to shifts in the action processes by which teams collectively made decisions or completed projects (Klonek et al. 2021). Furthermore, post-pandemic, several companies have opted to continue allowing telework for all employees (Conger 2020; Page 2020), while others have opted to gradually transition back to in-person work. This makes virtual teams during COVID-19 a fascinating case study, as they offer examples of communication among teams split across multiple collaboration modalities.

Chief among the challenges faced by virtual teams is the preservation of group coordination, allowing different branches within an overarching virtual team to communicate and coordinate effectively (Webster & Wong 2008) or allowing individuals within that team to maintain and build relational links to other members (Warkentin & Bernak 1999). On the technical front, specific capabilities of information communication technology (ICT) available to a virtual team can impact their communication, as different collaborative tasks require different methods and styles of communication (Garro-Abarca, Palos-Sanchez & Aguayo-Camacho 2021). ICT usage can impact teams both technically and psychologically, as sharing models of ICT use can lead to a shared work culture which positively impacts team coordination (Muller & Antoni 2020). Conversely, splitting modalities within a single context has been shown to negatively impact team performance (Incerti et al. 2020). On an interpersonal level, cohesion and performance in virtual teams are highly dependent on trust among individual team members, which in turn can be influenced by available technological capabilities. In a study on how communication modality impacts relationship formation, Sprecher (2014) found that forms of communication which allowed for more context or tonal cues, such as video-chat, led to greater trust between paired communicators after conversations than low-context (or less rich) communication. Unfortunately, maintaining high levels of richness is difficult in a virtual context.

All of the difficulties faced by virtual teams become more pronounced in hybrid team contexts, where some members operate face-to-face while others operate virtually. Because hybrid teams contain some members who can interact in more information-rich ways and others whose interactions cannot be as rich, schisms can result between the in-person and virtual divisions, leading to greater cohesion within the in-person team at the expense of the virtual team (Webster & Wong 2008). This creates a scenario wherein one part of a team is isolated from the other both spatially and technologically, and fragments teams into more divergent sites, both of which negatively impact team performance on the whole (Prasad, DeRosa & Beyerlein 2017; Muller & Antoni 2020). In addition, the asynchronicity of online communication such as e-mail can decrease the frequency with which virtual teams share critical information; in a hybrid context, these differences amplify, potentially creating information asymmetry between the in-person and virtual offshoots of an organization (Klonek et al. 2021). In essence, this means that hybrid teams operate as multidimensional networks, incorporating multiple agent and relationship types (Park et al. 2020). During COVID-19, most virtual teams were exclusively virtual, meaning that cohesion and shared practices could still form (Klonek et al. 2021). However, because of the gradual transition to in-person work, the difficulties of managing hybrid teams will become more problematic, requiring greater analysis of how communication and connection can occur within them.

One way to examine the spread of information within networks has been through the study of information cascades, or cases where social information within a network overtakes personal information obtained by individual members (Bikhchandani, Hirshleifer & Welch 1992). In a typical information cascade framework, individuals are faced with a binary choice, in which one outcome leads to a given reward. Individuals acquire two types of evidence pertaining to that choice: evidence they accumulate on their own, and the choices made by other people in their network. The first person in a group to make a decision can only base that decision on the evidence they observe directly, but subsequent decision-makers infer additional evidence from every prior decision made, gradually building a large sample of evidence (Anderson & Holt 1997). Given enough time and spread, the amount of evidence inferred socially will exceed the amount gained personally, leading to participants gravitating towards the socially preferred option more rapidly than the other option (Tump, Pleskac & Kurvers 2020). The principle of information cascades can be applied to collective behavior, such as in cases where a team must decide which of two strategies to adopt in a project or whether to engage in or refrain from a given task (Guarino, Harmgart & Huck 2011). Because information cascades are affected by the ability to observe social information, it stands to reason that they would operate differently in virtual team context, where that observation can be distorted through technological limitations or differences (Newman, Ford & Marshall 2019).

Research on information cascades has been conducted both at the level of individual evidence processing and the influence of network structures. On the individual level, Anderson & Holt (1997) modeled evidence accumulation through a simple Bayesian updating system, where individuals made decisions in sequence and incorporated evidence from others’ decisions into their priors. Later iterations of cascade research expanded upon this by incorporating the Diffusion Drift Theory of decision-making, where individuals begin with a bias towards one decision or the other, then gradually accumulate evidence until said evidence reaches a threshold needed to make a decision (Ratcliff et al. 2016). In the Diffusion Drift view of cascade behavior, social evidence is specifically incorporated in the rate at which individuals move towards one threshold or the other, parameterized as the drift rate (Tump, Pleskac & Kurvers 2020). Typically, information cascade studies assume relatively unified networks with identical rates of information transfer through linked agents; few account for cases wherein uncertainty, distortion or delay exist in communication among nodes of different types, as is the case with hybrid or displaced virtual teams (Webster & Wong 2008). Therefore, the current study will examine the formation and spread of cascades within a network system wherein two distinct groups exist, and the rates of communication within groups are different from those across groups.

Individual agents within the NetLogo network are classified as one of two agent types, representing In-Person or Virtual members. In-person team members are modeled as circles, while virtual team members are modeled as squares. Agents from both groups occupy a shared network space in the model. At setup, agents are randomly distributed through the space, and each agent will link to one other agent in the network, following the sequential model established by lab-based cascade studies (Anderson & Holt 1997; Tump, Pleskac & Kurvers 2020). One agent, representing the project initiator or leader in a virtual team, is selected to choose first and highlighted in grey at the start of each setup. For ease of user interaction, all agents are arranged in a circle on the display.

Agents are modeled through the parameters of a simplified Diffusion Drift model (Wagenmakers et al. 2007), specifically threshold separation θ and mean drift rate. In NetLogo, these are referred to as “threshold” and “drift-mean” in the code. Threshold separation represents the amount of evidence needed to reach a given option, while mean drift rate represents the average rate at which an agent accumulate evidence. Threshold separation will be held constant across agents, but can be set within the code by the experimenter. Meanwhile, mean drift rate will begin at 0, representing an agent that is completely neutral – that is to say, they accumulate evidence at random – but will update at the start of each agent’s turn based on the current state of the collective evidence displayed.

Agents within the simulation each must choose between one of two options, a correct option A and an incorrect option B. In the virtual team context, these represent potential approaches to a given task that will yield favorable or unfavorable outcomes, respectively. At the beginning of the simulation, agents start with a pool of prior evidence, expressed through the variables “evidence-a” and “evidence-b”. These represent the evidence agents have for each option. The experimenter can set these priors using the “a-prior” and “b-prior” sliders in the NetLogo interface, but for the purpose of the forthcoming study they will remain fixed at 1 each, indicating that agents believe A and B are equally likely. In addition, agents have a probability of drawing one of two signals from the environment when they make their choice, with the with the probability of receiving a signal for the correct option being set by the experimenter through the “prob-a” parameter.

Each turn, the active agent will receive an environmental evidence signal from a random draw and a social evidence signal from the agent who chose prior to them. Based on the signals each agent receives, they will update their mean drift rate based on the difference between signals for each option. Changes due to personal evidence contribute linearly to mean drift rate, while changes due to social evidence contribute via a power-law function; this is consistent with current literature on social diffusion drift (Tump, Pleskac & Kurvers 2020). Here, q is a constant set outside the function parameters, and is kept fixed at 1 for the current study, meaning that personal and social evidence contribute equally to an agent’s decision. Once their mean drift rate has been updated, agents calculate their probability of choosing option A via a logistic choice function of both drift mean and threshold separation.

After the active agent has calculated their probability of choosing option A, NetLogo will randomly draw a float between 0 and 1. If the float exceeds the probability calculated, the agent will choose option A; otherwise, they will choose option B. The active agent will then change color to a shade of blue if they chose A and red if they chose A, as well as encode their choice into their out-link through that link’s “info-type” parameter. They will then signal to the node receiving a link from them to begin their own decision process under the same rules. In accordance with Anderson & Holt (1997), each agent sees the total evidence for A and B at the start of their turn, including the initial priors plus the choices of all agents prior to them in the sequence. Users will also be able to track the frequency and length of cascades within the simulation, indicated by the fadedness of agents’ colors; agents with more faded colors did not cascade. The model converges when every agent has chosen.

The principal difference among agent groups is their interpretation of evidence from the other group. Uncertainty in cross-group communication will be modeled through a beta distribution, with parameters determined by a customizable “uncertainty” parameter; α = (6 – uncertainty) and β (uncertainty). This means that the higher the user-set uncertainty, the more skewed the beta distribution is to the left, and the more likely information is to be lost through communication. In the context of virtual teams, this uncertainty represents difficulties in communication due to discrepancies in technology, leading to distorted messages. For In-Person and Virtual agents, an agent’s choice will add 1 to the evidence for A and B if it is communicating with a member of its own group, but a random beta draw if it is communicating with the opposite group.

The primary questions for this study are:

(1) How do choice environment and individual evidence thresholds affect cascade behavior?

(2) How does uncertain communication impact the formation of cascdes?

To address the first question, default model settings were used. Each simulation included thirty randomly-distributed agents, all communicating without uncertainty. Thirty iterations of the first simulation were run at each of six probabilities of A, from 0.1 to 1, to determine whether agents in the model could perceive changes in their environmental state. Threshold separation was also measured from 0.2, its default setting, to 2, in order to determine whether higher threshold separations affected the likelihood of agents to choose. As a secondary check of the model, response time for each agent was calculated through the equation . If the model ran correctly, agent response time should have decreased as social evidence in favor of an option increased. Finally, the distributions of maximum cascade lengths at each level of threshold separation were calculated.

To address the effect of uncertainty, the above outcomes were measured again across different environmental conditions. However, rather than threshold separation (here fixed at 0.2), uncertainty was varied alongside probability of receiving A signals. Each simulation contained thirty agents, approximately evenly distributed between in-person and virtual team members. Again, response times per agent were examined as a quality control check. The effect of overall environmental probability on agent choice was measured across uncertainty levels from 0.1 (no information lost in intergroup interactions) to 5.9 (almost all information lost in intergroup interactions).

Based on the first parameter sweep, agents were responsive to changes in the environmental distribution of signals, with the average number of agents answering A increasing as the probability of drawing an A signal increased; at default threshold separation, the average amount of A choices when no A signals were received was 13, while the average amount of A choices when only A signals were received was 21. The average maximum cascade length at each probability was also fairly consistent, with an average maximum length of 11.53 agents out of 30 and an average standard deviation of 6.8 agents.

Increasing threshold separation had an extremizing effect on participant choice distributions; when the probability of A was low, increasing agent threshold separation decreased the amount of agents choosing A, while when the probability of A was high, more agents on average chose A at higher threshold separations. Diffusion drift literature indicates that threshold separation indicates that higher threshold separation leads to more accurate choices, that is to say more choices favoring the correct option (Wagenmakers et al. 2007). In this cascade framework, however, this translates to more accurate choices related to environmental signals received, or greater attunement to one’s choice environment. In accordance with prior diffusion drift literature, response times decreased over the course of a cascade, and increasing threshold separation increased overall response times at early cascade stages.

From the second parameter sweep, the effects of uncertainty proved less clear-cut than those of probability or threshold separation. At minimum uncertainty, the model behaved similarly to its previous iteration; however, the relationship between probability of receiving an A signal and proportion of agents choosing option A was slightly less linear. As uncertainty increased, the relationship between environment and outcome remained consistently positive but became more erratic, with more fluctuations at signal probabilities closer to 0.5. This suggests that agents were able to extrapolate from their more limited social evidence and reach analogous conclusions to agents under certain communication conditions overall, but that the addition of uncertainty led to even more variance within individual runs.

The effect of uncertainty on maximum cascade lengths was similarly obscure, with average maximum cascade length across uncertainties decreasing from 11.53 to 8.00; the average standard deviation across uncertainties decreased only slightly to 5.09. However, compared to runs where communication was always certain, there were fewer cascades of lengths higher than 10 within the uncertainty model, with the distributions of maximum cascade lengths either showing truncated range (uncertainties of 1.0 and 5.5) or greater right-skewedness (other levels of uncertainty). It remains unclear what this effect implies about the effects of uncertainty on cascades, as the effect was consistent across different levels of uncertainty; based on the algorithms used to define diffusion drift here, it is possible that agents relied more on personal information to compensate for their reduced social information.

Future modifications to the model could bring the simulation closer to a realistic virtual team scenario. Under current assumptions all agents can see the decisions of every previous agent. However, obtaining such a record is more difficult in reality, especially in denser networks. Studies of cascades in larger networks examined the effects of local vs. global nodes, with global nodes observing every other agent and local nodes observing only the agents in their general vicinity (Kobayashi et al. 2015). Introducing this dynamic into our model, either as a separate agent type or as a substitute for our current uncertainty distribution, could allow more generalization with it.

To expand our model in a different direction, rather than having agents make decisions in a linear chain, the model could assume the form of a preferential attachment network, wherein agents connect to other agents on a power-law distribution (Barabasi & Albert 1999). Network centrality, through network position or number of connections, has been shown to affect the degree to which agents rely on personal information over social information and to which cascades spread (Jalili & Perc 2017; Feng 2016). Modeling this effect within our network would add more robustness to the model, rather than the uniform uncertainty parameter the model currently uses. In addition, although this network was multidimensional in the sense that multiple relationship and agent types existed, all agents were still treated as part of the same continuous sequence. A greater effect might be observed if agents were split across multiple semi-connected networks which would then interact at a group level, as within- and between-network structures can also affect group performance, particularly in virtual teams (Warkentin & Bernak 1999).

• Anderson, L.R., & Holt, C.A. (1997). Information Cascades in the Laboratory. The American Economic Review, 87(5), 847-862.

• Bikhchandani, S., Hirshleifer, D., and Welch, I. (1992). A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades. Journal of Political Economy, 100(5), 992-1026.

• Collins, N., Chou, Y.-M., Warner, M., & Rowley, C. (2017). Human factors in East Asian virtual teamwork: A comparative study of Indonesia, Taiwan and Vietnam. The International Journal of Human Resource Management, 28(10), 1475-1498.

• Conger, K. (2020, May 2021). Facebook Starts Planning for Permanent Remote Workers. The New York Times. https://www.nytimes.com/2020/05/21/technology/facebook-remote-work-coronavirus.html

• Feng, Y. (2016). Are you connected? Evaluating information cascades in an online discussion about the #RaceTogether campaign. Computers in Human Behavior, 54, 43-53.

• Garro-Abarca, V., Palos-Sanchez, P., & Aguayo-Camacho, M. (2021). Virtual teams in times of pandemic: Factors that influence performance. Frontiers in Psychology, 12, 232.

• Guarino, A., Harmgart, H., & Huck, S. (2011). Aggregate Information Cascades. Games and Human Behavior, 73, 167-185.

• Incerti, V., Bellesia, F., Bertolotti, F., Chudoba, K., Fadel, K.J., Mattarelli, E., & Ungureanu, P. (2020). Working in the Era of Multiple Virtual Team Membership. A Study on the Effects of Variety of Communication Rules on Individual Management of Knowledge Resources. ICIS Conference 2020, talk.

• Jalili, M., & Perc, M. (2017). Information cascades in complex networks. Journal of Complex Networks, 5(5), 665-693.

• Klonek, F.E., Kanse, L., Wee, S., Runneboom, C., & Parker, S.K. (2021). Did the COVID-19 Lockdown Make us Better at Working in Virtual Teams? Small Group Research, in press.

• Kobayashi, T. (2015). Trend-driven Information Cascades on Random Networks. Physical Review E, 92(6), 1-9.

• Müller, R., & Antoni, C. H. (2020). Individual perceptions of shared mental models of information and communication technology (ICT) and virtual team coordination and performance—The moderating role of flexibility in ICT use. Group Dynamics: Theory, Research, and Practice, 24(3), 186–200.

• Newman, S.A., Ford, R.C., & Marshall, G.W. (2019). Virtual team leader communication: Employee perception and organizational reality. International Journal of Business Communication, 57(4).

• Page, C. (2020, November 25). Square-Enix Expects 80% of Employees to Work from Home Permanently. Forbes. https://www.forbes.com/sites/carlypage/2020/11/25/square-enix-expects-80-of-employees-to-work-from-home-permanently/?sh=12f7baf0294c

• Park, S., Grosser, T., Roebuck, A.A., & Mathieu, J.E. (2020). Understanding work teams from a network perspective: A review and future research directions. Journal of Management, 46(6), 1002-1028.

• Prasad, A., DeRosa, D., & Beyerlein, M. (2017). Dispersion beyond miles: Configuration and performance in virtual teams. Team Performance Management, 23(3/4).

• Ratcliff, R., Smith, P.L., Brown, S.D., & McKoon, G. (2016). Diffusion Decision Model: Current issues and history. Trends in Cognitive Science, 20(4), 260-281.

• Sprecher, S. (2014). Initial interactions online-text, online-audio, online-video, or face-to-face: Effects of modality on liking, closeness, and other interpersonal outcomes. Computers in Human Behavior, 31, 190-197.

• Tump, A.N., Pleskac, T., & Kurvers, R.H.J.M. (2020). Wise or Mad Crowds? The Cognitive Mechanisms Underlying Information Cascades. Science Advances, 6(29), 1-11.

• Wagenmakers, E.-J., van der Maas, H.L.J., & Grasman, R.P.P.P. (2007). An EZ-diffusion model for response time and accuracy. Psychonomic Bulletin and Review, 14(1), 3-22.

• Warkentin, M., & Beranak, P.M. (1999). Training to improve virtual team communication. Information Systems Journal, 19(4), 271-289.

• Webster, J., & Wong, W.K.P. (2008). Comparing traditional and virtual group forms: Identity, communication and trust in naturally occurring project teams. The International Journal of Human Resource Management, 19.1, 41-62.

...Okay, making it this far deserves an Easter egg. Here:

https://www.youtube.com/watch?v=UnIhRpIT7nc