Navigating the Virtual Gaze: Social Anxiety's Role in VR Proxemics

SA is characterized by an intense fear of being evaluated by others [89] and ranges from experiencing elevated anxiety symptoms in certain situations (e.g., giving a speech), to clinical levels [57, 65]. SA manifests through behavioral, physical, and cognitive symptoms [23, 89]. Behaviorally, individuals might exhibit withdrawal and avoidance tendencies [66]. Physical symptoms include blushing and trembling or even excessive sweating in social situations [5, 24, 89]. On the cognitive side, SA can lead to biased recollections of social events [80] and the feeling of being observed when they are not [32].

Those affected by SA often encounter difficulties in forming and sustaining relationships [15] and experience elevated risk of bullying [15]. Individuals with SA, thus, often dread scenarios where they are either directly interacting with others or under the perception of being observed [89]. Such fears are not limited to the moment and can amplify disproportionately, beginning days or weeks prior to the actual interaction [89].

In direct encounters, those with SA show distinct avoidance-related patterns: direct eye contact avoidance [99] or consciously keeping larger IPDs in conversations [34, 55, 57, 74, 105]. Prior work shows evidence that socially anxious individuals tend towards an enhanced self-directed perception of subtle gaze clues [82], particularly for negative and neutral facial expressions [83], which then may ampilfy experienced social stress [95]. These avoidance behaviors highlight the necessity of understanding SA’s nonverbal patterns.

Proxemics, the study of personal space in human interactions, delineates the surrounding area into different spaces [37, 102]: intimate space, for close relationships (0-45 cm); personal space, for friends (45-120 cm); social space, for interactions with unfamiliar people (120-365 cm); and beyond that, up to 762 cm, for addressing public space [37]. Note, however, that recent studies have found personal space to have a radius of about 1 m [39].

Personal space violations can create discomfort, often leading to defensive reactions [8, 38, 43, 58, 101]. However, personal space size is not static, being influenced by personal characteristics, such as gender [12, 37, 38, 39, 43, 44, 78, 103] and SA level [34, 55, 57, 74, 105]. There is a tendency for SA to increase IPD, which has neurological underpinnings: with increased SA, distance is perceived as closer, more salient, and as requiring more attentional resources [34, 74, 105].

One important situational determinant of IPD is gaze [11, 37, 67], which plays a role in personal space perception and regulation in virtual environments [11]. Some studies have explored gaze influence on proxemics, but results are mixed. Some reveal an enlarged IPD in mutual gaze conditions [11, 12, 67], while others show no significant effect [86]. It is important to note that minimal distance paradigms to measure proxemic behavior in these studies have been shown to be inferior to approach paradigms, as pointed out both by empirical HCI work [43] and meta-analysis [38].

Theoretical factors can also explain the mixed results. Equilibrium Theory (ET) offers a framework to understand the relationship between gaze and proxemics [3, 9, 11]. This theory posits that nonverbal cues help balance intimacy levels, by adopting avoidance and approach behaviors. Avoidance behaviors include gaze aversion, negative facial expressions, and increased IPD, whereas approach behaviors are marked by happy expressions, direct gaze, and reduced IPD [2, 102].

For VHA interactions we can posit different hypotheses based on different ET interpretations: one can argue that direct gaze intensifies intimacy and arousal, leading to increased IPD to reach an intimacy equilibrium [11, 12]. In contrast, traditional psychological theories suggest direct gaze is an affiliative behavior, leading to smaller IPD [9]. Previous SA research has shown that direct gaze cues inflict arousal and avoidance patterns among individuals with SA [89, 99]; it is possible to pit these interpretations of ET against each other when considering SA to resolve the discrepancies within the literature.

Social VR spaces, such as VRchat [28], have risen in popularity [22, 59]. These are characterized by full body movement and gestures in real-time, support of vivid spatial and temporal experiences, and mediation of both verbal and non-verbal communication [62, 63, 70]. Research on Social VR effects is inconclusive: some argue that Social VR may increase the risk of harassment, while others argue that Social VR spaces produce more satisfying social experiences due to the increased sense of co-presence [16]. Personality traits, and mental health conditions manifest virtually, increasing the risk of a low-quality experience. In SA, the design of the user’s self-representation [27] and agents’ social cues [21] can cause users to experience social stress in the game. With various hardware capabilities, this misunderstanding intensifies as users don’t know if the lack of social clues (e.g., gazing at them) is intentional by the user or a hardware issue. Given the important role of gaze and its interpretation on IPD in the physical world [82, 95] there is a need to better understand the relationship between SA, gaze, subtle social clues like facial expressions, and IPD in virtual environments.

A stop-distance task was used in a within-participants design, where we measured IPD, our dependent variable, as the frontal distance (in cm) between participants and the VHA, logged by participants when approaching the VHA. We manipulated Gaze Dynamics as our independent variable in four conditions (dynamic averting gaze, dynamic centering gaze, static averted gaze, static centered gaze), creating seven gaze levels, given that in the first three levels gaze could be directed to left or right on a horizontal plane. Besides, we used four different VHAs (two male and two female), with seven gaze levels repeated three times for each VHA (3 repetitions x 7 gaze manipulations x 4 VHA), resulting in 84 trials for each participant. We decided on these conditions building on prior work, which emphasized that the gaze’s angle with neutral facial expression is key and found that interpretation of these angles is affected by SA, amplifying social stress when misinterpreted as staring [82].

We also measured SA using the Liebowitz Social Anxiety Scale (LSAS; [30, 57]). Since SA is a dimensional trait [52], and to avoid loss of power in the analyses due to binarising SA, we followed prior research [46, 100, 101, 103] assessing SA’s effect on proxemics continuously using bayesian linear mixed models.

Furthermore, we explored participants’ gaze behavior when approaching the VHA by dividing our environment into two areas of interest (environment vs. VHA). This study was approved by the research ethics committee of Aalto University (D/718/03.04/2023). All data and data analyses can be found online ².

Seventy-nine participants underwent the study. Three were excluded from further analyses due to experiencing motion sickness during the VR immersion, indicated by a score ≥ 14 on the Fast Motion Sickness (FMS) Scale [49] (remaining sample: M = 2.62, SD = 3.30). Two participants were excluded as they did not have normal or corrected-to-normal vision. We also tested for minimal visual acuity (all > LogMar ≤ 1) ³[10], confirmed by the Landolt C Visual Acuity Test [10]. Another six participants were excluded, due to data issues. Three participants were removed due to poor questionnaire data (e.g., most left-side options in nearly all questions), two due to one missing value on the LSAS and one because of lost IPD data.

The remaining sample comprises 68 participants (33 male, 34 female, one did not disclose gender, \(M_\text{age}\) = 26.15, \(SD_\text{age}\) = 6.31). The detailed demographics are reported in Table 1. Participants were recruited through flyers spread at Aalto University and the Helsinki region. Participants received a 20 EUR gift voucher as compensation for their participation.

The Liebowitz Social Anxiety Scale (LSAS) is a questionnaire designed to assess cognitive, behavioral, and somatic manifestations of social phobia and anxiety [30, 57]. It consists of 24 items, rated on a 4-point Likert scale, with each being answered twice: once rating how anxious or fearful you feel in the situation, ranging from 0 (none) to 3 (severe), and then how often the situation is avoided, ranging from 0 (never) to 3 (usually). The LSAS score was obtained by summing up all item values. Effectively, the score could range from 0 to 144. LSAS has a high test-retest reliability of r = 0.83 and an internal consistency reliability of Cronbach’s α = .79 - .92 [13], which aligns with our empirical data α = .93 [0.91, 0.95].

We used the Meta Quest Pro Head-Mounted Display⁴ to render the VR environment, an iPad to present the post-experiment survey, and a computer for data recording. The VR environment was implemented using Unity game engine 2021.3.16f1 [93], integrated with the Oculus XR Plugin [72], and the Ultimate XR Plugin [98]. Datalogging was integrated into the rendering pipeline for efficient performance data collection. While the rendering pipeline generated frames at a rate of 90 fps, data logging was limited to 35.9 Hz. The headset was calibrated by adjusting the interpupillary distance and calibrating the eye-tracking for each participant individually.

The virtual environment was designed to match the dimensions of the physical environment, to allow participants to walk up to the VHA in a natural manner without colliding with objects or walls in the real environment. To measure the sense of presence in the virtual environment, participants filled in the IGroup Presence Questionnaire (IPQ; [77]) after completing the experiment. The IPQ measures spatial presence, user involvement, and experienced realness of the virtual environment, using 7-point Likert scale items, ranging from 0 to 6. The IPQ showed participants felt relatively present in the virtual environment (M = 3.03, SD = 1.86).

The VHAs were selected from the Microsoft Rocketbox Avatar library given its’ high-definition, fully rigged human-like avatars that are popular and well-used in AR/VR and HCI research [35, 61, 103]. We selected four white adult VHAs (two females; and two males) previously used in proxemics research in HCI [43]. Voice responses for each were prerecorded and implemented using the Amazon Text-to-Speech Software: Amazon Polly [4].

All VHAs were set to have a neutral facial expression (see Figure 1), as validated by participants’ emotionality ratings of each VHA (depicted in Table 1). To control for potential effects of gaze direction [9, 11, 79, 103], VHA’s height was dynamically adjusted to participants’ height. At the beginning of every trial, participants were positioned in front of the VHA, facing it directly.

For the static gaze condition, VHA’s gaze was set constant during the participants’ approach, fixed at the starting position either centered (0º; VHA established mutual eye contact during participants’ approach) or averted (+15º/left or -15º/right, VHAs eye gaze was deviated from participants). In the dynamic gaze condition, VHA’s gaze shifted in response to the participants’ approach. In the dynamic averted condition, the VHA’s gaze was centered at the beginning of the trial and gradually averted to the right (+15º) or left (-15º). In the dynamic centered condition, VHA’s gaze was averted at the beginning of the trial (+15º or -15º) and gradually centered (0º). The dynamic shift in gaze direction started when participants stood 2.5 m from the VHA, reaching its’ endpoint when participants stood at 1 m (horizontal body movement did not change gaze). Horizontal eye movement was modelled with an inverse Lerp Function using the following formula:Thus, VHA’s eyes changed linearly dependent on the IPD. Vertical gaze was not manipulated (e.g., see [31]) and kept constant at the participants’ eye level [100].

To validate gaze manipulation, participants were asked to report their subjective sense of feeling looked at on a visual analogue scale ranging from 0 ("I don’t feel looked at") to 1 ("I feel looked at."), in increments of 0.01. In the static gaze condition, participants felt more looked at when the VHA had a mutual gaze (0º; M = 0.64, SD = 0.31), compared to averted gaze (M = 0.18, SD = 0.24). In the dynamic gaze condition, participants felt more looked at when the gaze was centered (15º → 0 º; M = 0.66, SD = 0.31), compared to when the gaze was being averted (0º → 15º; M = 0.24, SD = 0.28). There was no effect of SA on the subjective feeling of feeling looked at (all p_b > 19.67%).

The social situation was standardized to minimize situational effects on IPD (e.g., [37, 100, 101]). Participants had to imagine a scenario in which they were in an unfamiliar location, asking a stranger for directions. In the VR stop-distance task, participants approached the VHA until a comfortable IPD was reached and pressed the controller’s trigger button. Then, on the participants’ left hand, a slider on which they assessed their subjective feeling of feeling looked at appeared. After, the VHA instructed on whether to press a white or a black button and disappeared. Then, participants would take a step forward, making the buttons to be pressed appear on the wall. After pressing, participants turned around to initiate the next trial. No time constraints were imposed.

According to the Declaration of Helsinki, participants gave written informed consent before starting the study. Participants were informed about the possibility of experiencing motion sickness during the VR immersion, explaining control with the FMS Scale [49]. This was followed by 10 practice trials (using a centered-gaze-female VHA), where participants could clarify doubts and adjust the headset volume to control for effects of sound on IPD [37]. Then, participants completed the stop-distance task. The post-experiment survey was completed on an iPad, provided to them digitally on Qualtrics⁵ [76]. The survey included the FMS Scale [49], general demographic information (i.e., age, gender, race, and occupation), previous VR experience, VHAs’ gender and emotionality ratings [102], the IGroup Presence Questionnaire (IPQ; [77]) measurement of experienced co-presence [14]; the LSAS [30, 57]) and the Triarchic Psychopathy Measure Screening (TriPM; [73] for separate replication purposes; not analyzed in this study). A summary of all descriptive statistics can be found in Table 1. Participants were debriefed after the experiment. Participation in the whole study took approximately 60 minutes.

We used Bayesian parameter estimation (for benefits on the Bayesian approach in HCI, see [1, 36, 48, 96]) to quantify effects, using brms [17], a STAN-sampler wrapper [18] in R [92]. We computed 4 Hamilton-Monte-Carlo chains, each with 40000 iterations and a 20% warm-up. We use the metric p_b for inference decisions. It is similar to the traditional p-value [42, 64, 85], which denotes the proportion of probability that the effect is negligible or reversed. A p_b value less than or equal to 2.5% was deemed significant. We also calculated the 95% High-Density Interval (HDI) for all parameters and conducted mean comparisons on standardized outcome variables. We utilized δ_t as an effect size estimate, comparable to Cohen’s d [40, 47].

To analyze eye-tracking data, we defined one area of interest, a cylindric shape centered at the middle of the body (with a radius of body = 37.5 cm, head = 10 cm) around the VHA. We compared the number of samples in relative amount of samples in a given trial until participants logged their preferred IPD, where participants looked at the VHA as compared to the environment.

Running a linear mixed model with Gaze, SA and their interaction as a predictor (with a normal prior of M = 0, SD = 10%), we find a contrasting pattern of VHA gaze and participant gaze. With static centered gaze at \(\tilde{b}_\text{static centered}\) = 18.84 % [17.22, 20.41], we see that averted static gaze increased the percentage of looking at the VHA by about 2%, \(\tilde{b}_\text{static averted}\) = 1.88% [1.40, 2.37], p_b = 0%, \(\tilde{\delta }_\text{b}\) = 0.22 [0.16, 0.28], this also holds for the difference to dynamic centering, \(\tilde{b}_\text{dynamic centering}\) = -0.46% [-0.95, 0.02], p_b = 3.08%, \(\tilde{\delta }_\text{b}\) = -0.05 [-0.11, 0.00], and dynamic averting, \(\tilde{b}_\text{dynamic averting}\) = 1.54% [1.05, 2.03], p_b = 0%, \(\tilde{\delta }_\text{b}\) = 0.18 [0.12, 0.24] (see also Figure 7). Comparing the other conditions directly by inspecting the posterior predictive distribution, we find all conditions to differ from each other, dynamic averting vs. dynamic centering, \(\tilde{b}_\text{diff}\) = 2.00% [1.61, 2.40], p_b = 0%, dynamic centering vs. static averted, \(\tilde{b}_\text{diff}\) = -2.34% [-2.74, -1.95], p_b = 0%, but dynamic averting vs. static averted, \(\tilde{b}_\text{diff}\) = -0.34% [-0.74, 0.06], p_b = 4.61%, and dynamic centering vs. static centered, \(\tilde{b}_\text{diff}\) = -0.46% [-0.95, 0.02], p_b = 3.08%. Therefore, participants looked the least at the VHA when it had a dynamic centering gaze and most at the VHA when the gaze was averted (see again Figure 7). We also found an effect of SA for centered static gaze. Each standard deviation in SA scores increased the relative time looking at the VHA for centered static gaze by about 2.7%, \(\tilde{b}_\text{LSAS total}\) = 2.73% [1.13, 4.31], p_b = 0.05%, \(\tilde{\delta }_\text{b}\) = 0.32 [0.13, 0.50[0.12, 0.60]. There was no interaction effect of SA and Gaze condition, all p_b > 3.49%.

Our study illuminates the interplay between gaze, proxemics, and SA. We used Bayesian parameter estimation to account for the uncertainty in estimating the size of the effects. We found, contrary to Bailenson et al. [11, 12] and H1.1., that participants prefer shorter IPD to VHAs with a centered gaze as compared to averted or dynamically averting gaze. The difference between static averted and centered gaze was diminished in SA up to the point where participants with higher levels of SA preferred closer IPD when the gaze was averted compared to when it was centered.

Aligning with H2.1., participants preferred larger IPD when the gaze was averted compared to when it was centered. Notably, dynamically centering the gaze produced the smallest IPD in our study, highlighting that dynamic gaze is less ambiguous and can enrich our interactions with VHAs. However, no support was found for H2.2. We did not find any indication that SA increased IPD (H3); however, our Bayesian approach to analyzing the data could show that we did not have enough data to be conclusive about the effect of SA on IPD.

Conversely, by exploring participants’ gaze using eye-tracking, we found that VHA gaze centered/centering diminished the amount of time participants looked at the VHA and that SA increased the overall time spent looking at the VHAs.

To reiterate, ET posits that individuals engage in a dynamic balance between intimacy and personal space [12, 51]. When confronted with cues that increase perceived intimacy (e.g., direct gaze), IPD may be increased to maintain a comfortable equilibrium, thus adopting compensation behaviors [51]. Argyle and Dean [9] have previously highlighted gaze’s significance as an affiliative signal, suggesting that it can serve as an invitation for closer interaction, opposite to what Bailenson et al. [12] proposed.

In our study, for average SA levels, the pattern was evident: centered gaze was preferred over averted gaze, possibly indicating a feeling of comfort and affiliation, thus supporting Argyle and Dean [9] proposal. Regarding SA’s effect on IPD, no main effect was found, contrasting with previous findings [34, 46], however, more data is needed to estimate this effect.

Despite this, a tendency was found within participants with increased SA: a centered gaze seemed to evoke heightened intimacy and arousal, leading to bigger IPD. When looking at the VHA, increased SA possibly led to higher arousal, caused by biased social cue interpretation. According to ET, individuals try to lower arousal promoted by the increase in perceived intimacy by increasing their IPD. Arguably, SA possibly moderates gaze-promoted intimacy: direct mutual gaze is not intrinsically positive or negative since it can either induce feelings of intimacy and signal attention in those with average levels of SA, or promote uncomfortable levels of arousal that lead to compensation behaviors in individuals with increased levels of SA.

This latter interpretation aligns with Bailenson et al. [11, 12]. Essentially, while Argyle and Dean [9] proposal holds in the broader context, the specific direction of the balance (approach or avoidance) can vary based on, for example, personality traits. Therefore, designers of inclusive social VR experiences have to know their audience and design VHA interactions with ET in mind.

Our study, while shedding light on several nuances of VHA interactions, is not one without limitations.

First, we must acknowledge the influence of cultural backgrounds on the experience of SA characteristics [41]. For instance, what might be deemed as an intimate distance in one culture might be perceived as too distant in another [37, 86, 88]. Future research should delve deeper into the role of cultural background affecting the user’s behavior and perception in-virtuo.

Secondly, physiological arousal, which could moderate the relationship between gaze, proxemics and SA, was not measured (see e.g., [19, 58]). Future studies may simulate personal space violations for SA and enhance measurements with physiological sensing (e.g., skin-conductance response;[43]).

Third, one could argue that the initial gaze and not the final gaze is critical for IPD and its interplay with SA. Be reminded that in the static conditions, gaze was either averted or centered, while in the dynamic centering condition, gaze was initially averted and then gradually centered and the opposite for the dynamic averting condition. IPD for static and dynamic conditions resemble each other regarding the end of the animation and not the beginning (averting/averted > centering/centered), mirrored in the ratings of gaze. Thus, the effect of gaze on IPD cannot be explained by initial gaze alone. Could this explain the differences in gaze conditions regarding SA, e.g., entertaining that with SA, one did not look at the avatars when approaching? This is also unlikely. We find an SA effect only for static conditions and not for dynamic conditions. Supporting this, we find no interaction between SA and Gaze condition in our eye-tracking data. Nevertheless, future researchers interested in dynamic gaze patterns should add conditions with fully averting gaze (i.e., looking away from the left to looking away on the right as one approaches.)

In line with van Berkel and Hornbæk [97], we will highlight theoretical and HCI-oriented implications for the design of social VR as well as the implications for social anxiety research:

The prospect of detecting personality traits in users, especially within virtual environments, raises several ethical concerns. One primary concern is that of consent [25]. Users might not be aware that their interactions, behaviors, and responses can be indicative of their personality traits. Extracting such information without explicit consent infringes on individual privacy rights [25]. Miller et al. [68] could identify people from 5 minutes of motion-data with high accuracy. They propose that such data should be regarded as personal data. While we were interested in the correlation pattern of IPD and personality to improve the design of virtual environments, given that personality traits can be distinctly linked to stimuli and behavior in VR [100, 102, 103], we encourage research into privacy-preserving techniques which can be adapted to users personality. However, if virtual environments are tailored to cater to identified personality traits, users might end up in echo chambers reinforcing existing beliefs and behaviors instead of deconstructing harmful behavior of users with SA [29, 57].