Emotional contagion: From unidirectional to bidirectional

Emotional contagion is the “tendency to automatically mimic and synchronize facial expressions, vocalizations, postures and movements with those of another person and, consequently, to converge emotionally” [7; p. 5]. Emotional contagion processes have been widely studied in organizational contexts, with previous studies finding that leaders’ moods can influence the mood of followers, impacting the performance of the latter [8, 9]. The transfer of emotional states from leaders to followers can also influence the perception of the latter towards the leader’s style, honesty, and overall effectiveness [10–13]. However, all of these studies have examined emotional contagion between leaders and followers as a unidirectional phenomenon, ignoring that the social interaction process is a process of mutual influence. This study builds on this argument and examined whether followers can influence leaders via emotional contagion in a particular setting, namely computer-mediated communication.

Bidirectional anxiety contagion in computer-mediated communications

Within the realm of CMC, some studies have found that emotional contagion is not only possible in CMC but also that interactions on social media platforms allow the transfer of discrete emotions, including anger or happiness [14] as well as anxiety [15]. Social media platforms, especially public ones such as Twitter, allow the investigation of the effects of emotion transfer from followers to leaders and vice versa [16]. And indeed, less than a handful of studies have suggested that the transfer of emotions via social media platforms, in principle, can be a bi- or even multidirectional effect [14, 17, 18].

In the leadership literature, most attention is given to leaders’ effects on followers, though there is growing evidence that followers affect leaders in numerous ways and that the leadership process is indeed bi-directional. In the context of emotions, a similar pattern can be observed—with most attention given to the ways that leaders produce emotional effects on followers. Yet, this lopsided attention is incongruent with the fact that emotions are highly contagious, regardless of who is the ‘producer’ and observer of the emotions. Indeed, Kramer and colleagues [19] demonstrated that slight manipulations of the emotions ‘displayed’ by one’s friends resulted in the observers displaying similar emotions. Hence, if reading about another person’s emotions (and indeed, these may be people the observer hasn’t spoken to in years) on a screen creates evidence of emotional contagion, it is reasonable to believe that leaders and followers, who typically interact frequently, may be prone to similar emotional contagion in both leader-follower and follower-leader directions.” However, conclusive empirical evidence to that effect remains elusive. In this respect, this study both builds on existing literature and expands prior empirical evidence by investigating the transfer of a particular emotion, namely anxiety, within a CMC context.

Anxiety is present in both state and trait forms [20]. And while state anxiety is short-term and event-specific (e.g. anxiety due to pandemic effects, etc.), trait anxiety refers is a stable trait-like form of anxiety, similar to other personality traits [20, 21]. In terms of behavior, individuals who score high on trait anxiety tend to be both more anxious than others (i.e., intensity) and experience anxiety more often (i.e., frequency).

We argue that state anxiety is a particularly important emotion to study given the increased likelihood of negative emotions spreading more easily on social media than positive emotions. In addition, while we recognize that individual responses in the context of the SARS-CoV-2 crisis individuals vary [22], anxiety is considered the default crisis emotion [23] and is experienced by all individuals during a crisis before they experience any of the other crisis-related emotions (i.e., fear, anger, sadness). Therefore, we argue that the study of state anxiety is particularly important [24] when assessing communications between followers and leaders in a social media context. And while several previous studies have found that leaders’ state anxiety influences followers’ state anxiety [25, 26], including in social media contexts (e.g., [20]), to the best of our knowledge, there are no previous studies to examine anxiety contagion from followers and leaders on social media. Yet, based on previous evidence in face-to-face teams [6] we expect that follower state anxiety in CMC influences leader state anxiety over time.

H1: In CMC, follower state anxiety positively predicts leader state anxiety.

In addition, trait anxiety, in combination with state anxiety, can influence the way that a person experiences anxiety in a specific context [24]. Previous studies have also shown that the role of leader trait anxiety in particular is an important moderator to consider in the transfer of anxiety from leaders to followers [15]. We argue that leader trait anxiety also forms an important variable in the presented study. That is because highly trait anxious leaders (and individuals more broadly) are more anxious in general. Individuals who are pre-disposed to experience trait anxiety tend to perceive given situations as more hostile, and therefore give more attention to negative information [27]. And because such individuals are likely to be constantly on the alert and the lookout for possible threats, we would expect highly trait anxious leaders to be less influenced by their followers’ emotions, because they are predisposed to being anxious already. Put differently, we would expect the effect of followers’ state anxiety to be most impactful on less trait anxious leaders’ state anxiety. Hence, we would expect leaders with higher trait anxiety to be less influenced by their followers than less trait anxious leaders.

H2: In CMC, leader trait anxiety moderates the relationship between follower state anxiety and leader state anxiety, so that increased follower state anxiety is more likely to predict increased leader state anxiety in the case of less trait anxious leaders.

Sample and procedure

The final sample consisted of both the leaders and followers. Leaders were defined as top managers, namely C-suite executives, i.e., individuals with a job title of e.g., Chief Executive Officer (or CEO), Chief Financial Officer (or CFO), etc. However, unlike Gruda and Ojo [15], we did not classify all remaining individuals in the provided dataset as followers. We argue that it is unlikely that lower-level followers (e.g., analysts) can influence top-management leaders in the same manner that top leaders can influence their employees. That is because C-suite executives are considered senior organizational leaders [5] who wield great signaling power due to their access to privileged information, access to resources, and resource distribution power [28]. This power is not available to all other followers. However, we would expect that direct followers of such leaders or followers who are close to top leaders in the organizational hierarchy (e.g., Directors, Senior Vice Presidents, Head of Departments, etc.) do have the power to influence their leaders and affect their emotions. Hence, we specified such individuals as followers. Follower status scores were obtained from the original authors [20]. All employees were from the United States of America. This is because the ground truth training data was obtained exclusively from individuals in the United States. This led to a final dataset of 108 leaders and 178 followers across 50 organizations and a total of 6,988 dyadic daily interaction observations.

Measures

Big-five personality traits.

We also included the Big Five personality traits as controls in our analysis. Previous work [e.g., 32–35] has demonstrated that social media data provides a continuous stream of data that can be used to accurately and reliably measure personality traits. The Big Five personality traits for all individuals in our sample were obtained using the IBM Watson Personality Insights API [36], an open-vocabulary machine-learning approach trained using a reference sample of 1,000,000 individuals. The IBM Watson algorithm also provides six individual facets for each Big Five personality dimension [21, 37], which were combined into higher-order Big Five personality dimensions (all Cronbach α ≥ .70, except for Openness to Experience: α = .68, see Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Pairwise correlations of main variables.

https://doi.org/10.1371/journal.pone.0279164.t001

Analytical strategy

As followers are nested within leader-follower dyads (i.e., several followers per organizational leader) and within companies, we use a multilevel cross-classified mixed-effects model for repeated measures to test our hypotheses [38]. We define a three-level model with random intercepts at the company level, using the Aikake Information Criterion (AIC) and the Bayesian Information Criterion (BIC) to compare models. All analyses were conducted using Stata 17.0.

Robustness check

Anxiety is closely related to neuroticism, as is also evident in this study (see Table 1). To check whether results diverge when considering alternative models which do not include neuroticism as a control, we rerun a separate model without this control. Results were comparable and remained robust (b = -.002, t = -2.01, p = 0.045).

Limitations and future research

Our study has some limitations. Firstly, because our data relies on follower and leader tweets, it could be that social media users do not present their own selves but rather a positively enhanced and promotable version of themselves when communicating online. While this argument makes intuitive sense, previous research has found strong support for negating this belief [34, 43]. In addition, we would argue that the applied methodology is particularly suited to detect anxiety in an unobtrusive manner [31]. Put differently, due to behavioral residue found in tweets, the applied approach allows us to detect changes in anxiety more accurately [21].

Secondly, our data was solely restricted to the U.S., due to the training dataset of the applied algorithm. However, because of the country’s size, the infection rate of SARS-CoV-2 differed in time from state to state and indeed even from county to county. Hence, a more detailed assessment of COVID-19 infections over time could have provided different results concerning the impact of the crisis on follower anxiety. We encourage future research to examine this more closely.

Finally, another important limitation is the issue of endogeneity and lack of causality in the presented model. To prove causality, a randomized control trial would be needed, in which individuals (leader and followers) would be randomly assigned into two groups, and followers would be asked to tweet highly anxious or less anxious tweets. Leaders, themselves divided into two groups (i.e., highly trait anxious vs low trait anxious) in turn, would be asked to read these tweets and provide self-report ratings of their respective state anxiety at that moment. In addition, we also were interested in testing the lagged influence of follower state anxiety and leader state anxiety. Therefore, after the experiment takes place, we would also need to test leader state anxiety several days later. However, such an experimental setup—albeit it would allow researchers to visualize the direction and measurement of the contagion effect—would prove difficult to implement and would lack generalizability, due to the artificial conditions of the aforementioned laboratory experiment. We argue that the presented approach using machine learning to infer (state and trait) anxiety of both leaders and followers, allowed us to test the contagion effect in a naturalistic and unobtrusive setting. Finally, the applied multi-day lagged design allowed us for associations between follower and leader anxiety over time.