Changing social attitudes with virtual reality: a systematic review and meta-analysis

ABSTRACT

Although it has been shown that Virtual Reality (VR) can positively impact political, civic and environmental views, the question of whether and how VR influences social attitudes more effectively than less immersive conditions has been a subject of debate. To address this question, this article provides a systematic review with meta-analysis of social attitude research in VR by examining 50 separate findings from 39 studies. We find that VR influences social attitudes significantly more than non-immersive interventions, while embodiment has no moderation effect. Further, outgroup VR perspective-taking fosters positive social attitudes more effectively than ingroup VR perspective-taking. Findings are categorized according to social attitude object and discussed in the light of emerging challenges and opportunities for VR attitude research.

KEYWORDS:

• Virtual reality
• attitudes
• meta-analysis
• embodiment
• perspective-taking

Introduction

The advent of Virtual Reality (VR), once predicted to challenge our definitions of communication and its effects (Biocca & Levy, 1995), has now led to a surge of scholarly interest in how and why VR interventions may change social attitudes toward social matters (Aitamurto et al., 2021). Virtual Reality pioneer Jaron Lanier noted that ‘the most important thing about a technology is how it changes people’ (Lanier, 2011, p. 4) and, in light of the current discourse on the Metaverse and the future of human existence in virtual spaces (e.g. Zuckerman, 2021), this question is timely for both scholars and practitioners. The purpose of this study is to investigate how VR changes the ways we think and behave toward social attitude objects by systematically reviewing and summarizing extant empirical evidence that addresses this question.

Over recent decades, researchers have been investigating how VR experiences evoke social attitude change, often posing similar questions and finding differing outcomes. Further, the variety of theoretical and methodological approaches make it difficult to identify common trends that could help accelerate progress in the field. A meta-analytic review of the research can contribute to providing a more accurate estimate of the effect investigated as well as a holistic picture of the literature and, thereby, point toward new research directions (Rosenthal & DiMatteo, 2001).

Given the above, the aims of this contribution are twofold. First, we systematically review relevant studies according to social attitude objects and identify current challenges and opportunities in their approaches. Second, we follow the quantitative procedures of a meta-analysis (Rosenthal, 1991; Rosenthal & DiMatteo, 2001) to test whether VR treatments are more effective for changing social attitudes than traditional media. To gain a better understanding of the effect of interest, we further examine the variables embodiment and perspective-taking in the literature.

Related work

Social attitudes

The topic of media-induced attitude change has a long research tradition in Communication Science and Social Psychology (Petty & Briñol, 2010); simply defining what an attitude actually is has occupied theorists for decades (cf. Oskamp & Schultz, 2005). Originally used to describe a person’s physical posture, the term ‘attitude’ now commonly refers to a ‘posture of the mind’ (Oskamp & Schultz, 2005, p. 8) toward an attitude object. The most prominent approaches to defining attitude include the ABC model, whereby attitudes consist of affective, behavioral, and cognitive components (e.g. Ellis, 1962), and Fishbein and Ajzen’s view (1975) of attitude as an exclusively affective construct that indicates favorability. We rely on the definition by Oskamp and Schultz (2005), who viewed attitudes as ‘a summary of all of a person’s evaluative beliefs/opinions about, affective reactions toward, and behavioral responses to an attitude object’ (p. 14). This definitional approach entails beliefs that are value judgments about an object; stereotypes, namely general beliefs about social groups; opinions, defined as evaluations with a narrower scope; affective reactions (e.g. liking or disliking); and behavior toward attitude objects.

Such objects may be as simple as consumer products or as complex as topics of broad societal relevance, for example, climate change, racial bias, or discrimination against minorities. In the latter case, objects are commonly referred to as ‘social attitudes’ (e.g. Campbell & Stanley, 1963; Haddock & Zanna, 1999). Social attitudes are vital for democratic processes as they serve as driving factors for voting propensity (Boonen et al., 2014), public service motivation (Taylor, 2010), and support for democracy (Stevens et al., 2006). Therefore, the processes by which they change and the reasons for such changes are central to the literature (Briñol & Petty, 2010). One line of this research investigates the influence that media exert on social attitudes, from mass television, radio (Petty et al., 2008), and social media (Bail et al., 2018) to emerging technologies like Virtual Reality (Kalyanaraman & Bailenson, 2020).

Virtual reality as a medium for communicating social issues

Virtual Reality has been defined as ‘a real or simulated environment in which a perceiver experiences telepresence’ (Steuer, 1992, p. 78). Telepresence, frequently referred to as spatial presence or presence, describes the feeling of being there within the mediated reality (Slater & Wilbur, 1997; Lee, 2004). Presence can be induced by specific hardware, namely head-mounted displays, computers, and headphones (Steuer, 1992), and enhanced through immersive features such as increased levels of user-tracking, stereoscopic visuals, and wide fields of view of display (Cummings & Bailenson, 2016). With immersive technology, users feel as though they are present within mediated realities that range from spherically filmed 360-degree videos to fully immersive virtual environments (IVEs). Experiencing the feeling of being present in VR has potential value for the communication of social and political concerns, given that these tend to be perceived as psychologically distant: first, their immediate consequences may lie in a temporal distance (e.g. climate change); secondly, they may lie in a spatial distance (e.g. migration); or thirdly, they may not directly affect one’s in-group (e.g. aging for younger populations). Scholars have argued that VR might decrease users’ psychological distance in relation to such abstract mental construals (Kalyanaraman & Bailenson, 2020), thus transforming distant attitude objects into near ones. This conversion is key for changing social attitudes because attitudes toward psychologically near objects are more likely to change than those toward distant ones (Trope & Liberman, 2010).

Other approaches have examined the role of social presence, that is the feeling of being there together with mediated others (Lee, 2004). Avatars or agents may exert social influence on users, specifically when the interaction is perceived as realistic and relevant (Blascovich et al., 2002). Furthermore, socially present characters are more likely to induce empathy (Schutte & Stilinović, 2017) and, in turn, drive attitude change toward vulnerable groups (Herrera et al., 2018). Another distinct subtype of presence is self-presence, which is the subjective experience of the virtual self as the actual self (Lee, 2004). Self-presence depends on the degree of connection users feel to the virtual self’s emotions, identity, and physical body (Ratan & Hasler, 2010). Because presence and its subtypes are often associated with persuasive outcomes in regard to commercial products (Lombard & Snyder-Dutch, 2001), researchers have shown a growing interest in its potential effects on social attitudes.

One of the main approaches in this line of research is to test how varying levels of immersive modality affect attitudinal measures. This body of literature is aligned with media richness theory (Daft & Lengel, 1986), the overarching hypothesis being that the richer the immersive modalities of the medium, the greater its persuasive effect on social attitudes. Such empirical investigations often compare VR to traditional media, for example, text or two-dimensional video, whereby some studies have identified significant effects (e.g. Ahn et al., 2014), while others have not (e.g. Ma, 2020).

A meta-analytic review of empirical studies may help explain such inconsistencies by drawing a more accurate and precise effect estimation based on the accumulation of separate findings (Rosenthal & DiMatteo, 2001). In addition, a synthesis of the literature can inform scholars about patterns in the theory and methodology of immersive modality effects research. Such an analysis would provide scholars and practitioners with a better understanding of the immersive technology needed to yield attitudinal change, and in turn, help generate new hypotheses. Moreover, in the event that challenges in extant designs are identified, a systematic review could help future researchers to address and overcome these shortcomings. For the purpose of our meta-analysis, we pose our first prediction:

H1: VR interventions will lead to more positive attitudes than less immersive media conditions.

Embodied experiences

Studies that compare social attitude effects on the basis of their immersive modality often use embodied experiences as VR treatment (Ahn et al., 2014). Embodiment requires the visual representation of the virtual self, manifested from virtual representations of hands (e.g. Heeter, 1992) to full-body avatars (Kilteni et al., 2012). When users experience a sense of ownership toward the virtual body, agency over its movement, and a feeling of intention over its location, they tend to experience a sense of embodiment (Kilteni et al., 2012). However, these additional variables make it difficult to compare embodied VR to print or 2D video because the treatments differ in regard to not only their immersiveness but also the sense of embodiment they evoke.

The literature shows a general consensus that embodiment can yield significant persuasive outcomes. Previous work has argued that, because of the illusion of control over the mediated environment, users tend to become susceptible to subliminal persuasive messages (Sundar, 2008). Further, embodied experiences can be considered more direct and personal than secondhand events and are, thereby, more likely to have a stronger influence on attitudes (Ahn, 2021). Moreover, significant effort has been made to investigate the effects of embodiment on body, identity, and attitudes. Yee and Bailenson (2007) coined the term ‘Proteus effect’ to describe how users internalize features of the avatar they embody. For instance, the embodiment of an avatar that is tall elicits more confident behavior than the embodiment of a short avatar (Yee & Bailenson, 2007).

Because literature offers evidence of a positive impact of embodied experiences on attitudes, we suspect that embodied treatments outperform non-embodied treatments when compared with traditional media. A moderation analysis may clarify this assumption by providing a more precise effect estimate. Therefore, we state our second hypothesis:

H2: Sense of embodiment will positively moderate the relation between VR intervention and attitude.

Perspective-taking

Embodiment in IVEs, as well as the spherical viewfield in 360-degree VR, can create the illusion of seeing the world through the eyes of another. VR technology thus eliminates the need to mentally construct a different viewpoint by programming it into the user’s visual field. According to traditional perspective-taking scholarship, the perception of the world through another’s eyes shifts the visual focus to the external environment; as a result, individuals begin to think about the other’s behavior based not on ascribed dispositional traits but on situational factors (Jones & Nisbett, 1987). Situational rather than dispositional information-processing, in turn, elicits feelings of sympathy (Batson et al., 1989) or congeniality toward the person whose perspective is taken (Davis, 1994).

Frequently, these approaches compare the effect of the first-person perspective (1PP) on in-group and out-group biases. For this, specific social markers such as race or gender are used to categorize subjects into an in-group and then test how an out-group 1PP in VR affects their attitudes toward the out-group. Mental out-group perspective-taking has been consistently linked to reductions in racial bias (Todd et al., 2011), a decrease in out-group stereotypes (Vescio et al., 2003), and improved social functioning in general (Davis, 1983). Naturally, a significant amount of scholarly effort has been made to examine how 1PP in VR affects attitudinal constructs (e.g. Aitamurto et al., 2018). However, it is unclear whether outcomes from mental 1PP research can be related to virtual experiences. The reason is that mental 1PP depends on important additional factors like mentalizing skills (Frith & Frith, 2006), which, in the case of VR 1PP, are not needed to put oneself in another’s shoes. To test whether current perspective-taking and attitude literature in VR aligns with mental perspective-taking research by means of a meta-analysis, we hypothesize:

H3: Out-group 1PP in VR will affect social attitudes more positively than in-group 1PP in VR.

Methodological considerations in VR attitude research

Our meta-analytic review addresses the hypotheses stated above while additionally accounting for factors relating to the experimental designs. This is because the findings from attitude change experiments can be context-sensitive (Holbrook, 2011).

One relevant factor is sampling. For example, student samples might be threatened by social desirability bias to a larger extent than other populations when the study is undertaken by familiar research staff. Further, sampling can determine the level of familiarity with VR and, thus, account for novelty effects that might influence levels of flow with the mediated contents (Tokunaga, 2013). Presence can also be affected by familiarity with (Lachlan & Krcmar, 2011) and trust in VR (Salanitri et al., 2016). Regarding other sampling demographics, there is no conclusive evidence about the influence of gender (Felnhofer et al., 2012; Felnhofer et al., 2014; Giannopoulos et al., 2008; Kothgassner et al., 2013) or age (Corriveau Lecavalier et al., 2020).

Closely related to sampling, monetary compensation for participation is another factor we seek to consider. High incentives for participation can increase perceived risk of participation (Cryder et al., 2010) or attract subjects with different personality traits (Hsieh & Kocielnik, 2016), thus resulting in a systematic sampling bias. For example, low income populations might be more attracted to a study that offers monetary compensation than high-income populations, yet the attitude measure might be income-dependent (e.g. willingness to donate). A common alternative to monetary compensation for student subjects can be course credit; however, there is mixed evidence about the effect of payment vs. course credit on task performance (Brase et al., 2006; Kvaløy et al., 2015).

Another methodological issue that warrants discussion is duration of treatment. On the one hand, the comparability of different immersive modalities might be hampered by differing exposure times. For instance, a VR condition can last 5 minutes because it is bound to the duration of the video; however, the text condition depends on the subject’s rate of reading, i.e. how quickly or slowly he or she reads. As a result, exposure time across treatments may vary, thus adding an extra differential factor to the independent–dependent variable relationship. On the other hand, long durations of exposure to VR treatments may produce feelings of cybersickness (Kennedy et al., 2000), which can, in turn, decrease levels of presence (Weech et al., 2019) and thus interfere with the outcomes. Further, it is also possible that longer exposure increases a sense of presence or engagement with the message and, thereby, the potential for attitudinal change.

We account for these three factors—sampling, incentives, and duration of treatment—in the review of the literature in order to provide a nuanced understanding of found effects. Given the lack of sufficient evidence to state hypotheses, we formulate a more general research question:

RQ: How are the experimental design factors of sampling, compensation, and duration of treatment considered in the extant literature?

Materials and method

The first step is the identification of relevant literature. The search for the review and meta-analysis was performed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. A collection of relevant literature was extracted by completing keyword searches of the databases that publish research related to Communication and Virtual Environments, namely Web of Science, Mass Media Complete, and PsycInfo; the IEEE conference proceedings; and the ACM Digital Library as well as Google Scholar. The search terms applied to the full text included a combination of ‘virtual’, ‘immersive’, ‘virtual reality’, ‘immersive virtual environments’, ‘HMD’, ‘360VR’, ‘experiment’, ‘attitude’, ‘bias’, ‘opinion’ and ‘belief’.

We included studies that (1) were published in a peer-reviewed journal, a conference proceeding, a dissertation or pre-print; (2) measured an attitude toward a social object as a dependent variable; (3) collected empirical data in an experimental research design with random treatment assignment; (4) included one experimental condition in which an HMD was used; and (5) reported an effect size adequate for the aims of the review and meta-analysis. After identifying relevant studies, we searched for additional work in their reference lists. Based on this process (Figure 1), we identified 39 studies with 50 separate research findings and a total of 4,210 subjects (Table 1).

Figure 1. PRISMA Flowchart of study selection for review and meta-analysis (Moher et al., 2015)

Table 1. Summary of studies.

Reference & Location	Sample	Conditions & Duration	Presence Measure	Attitude Measure	Compensation	Attitudinal Outcome
Environment
Ahn et al., 2014 Study 1 US	N = 47 61,7% females (μ)age = 21,6 Student	Cutting a virtual tree in an IVE with embodied hands vs reading a print description; >2 min	none	Self-report: Ten items from the Environmental Action Internal Control Index (Smith-Sebasto & Fortner, 1994) Behavioral: Number of napkins used to clean up water	na	No significant differences of treatment effect on self-reported locus of control Participants in the IVE condition used significantly fewer napkins compared to participants in the print condition.
Ahn et al., 2014 Study 2 US	N = 60 75% females (μ)age = 20,32 Student	Cutting a virtual tree in an IVE with embodied hands vs reading a print description vs 2D video > 2 minutes; duration not reported	none	Self-report: Environmental behavioral intention; Environmental locus of control; Success in recycling-relevant behaviors at Time 2	na	Subjects in IVE reported significantly environmental locus of control, environmental behavior
Ahn et al., 2016 Study 1 US	N = 49 55,1% females (μ)age = 20,4 Student	IVE cow avatar embodiment vs 2D video; duration not reported	Five item interval scale adapted from Bailenson et al.'s Spatial Presence Scale (2005)	Self-report: Inclusion of nature in self (INS; Schultz, 2001)	na	Participants in the IVE condition perceived significantly higher inclusion of self in nature than participants in the video condition
Ahn et al., 2016 Study 2 US	N = 53 50,94% females (μ)age = 21,2 Student	IVE coral avatar embodiment with haptic feedback vs 2D video; duration not reported	Five item interval scale adapted from Bailenson et al.'s Spatial Presence Scale (2005)	Self-report: Inclusion of nature in self (INS; Schultz, 2001); Issue involvement was measured with six 7-point bipolar items asking participants to evaluate how important, of concern, relevant, meaningful, of matter, and involving the issue of ocean acidification was to them	na	Differences in inclusion of self in nature for participants in the IVE condition and those in the video group approaching significance, with the direct effect weakening over Time 2. No significant difference between groups was found for issue involvement
Ahn et al., 2016 Study 3 US	N = 126 65,08% females (μ)age = 20,1 Student	IVE coral avatar embodiment with haptic feedback vs 2D video; duration not reported	Five item interval scale adapted from Bailenson et al.'s Spatial Presence Scale (2005)	Self-report inclusion of nature in self (INS; Schultz, 2001); Issue involvement was measured with six 7-point bipolar items asking participants to rate how important, of concern, relevant, meaningful, of matter, and involving the issue of ocean acidification was to them	na	No significant differences in inclusion of self in nature between conditions
Bailey et al., 2015 US	N = 70 47,1% females (μ)age = 18,4 Student, Communication and Introduction to the Humanities class	IVE vividness (avatar-coal, coal) x personalization (personal sign, impersonal sign) during virtual shower; 6 min	na	Water amount and temperature during task	Course credit	Participants exposed to vivid messages used cooler water compared with not vivid messages. Personalization had no significant effect.
Breves & Heber, 2020 DE	N = 56 42,8% females (μ)age = 30,18 Half student, faculty members, distant acquaintances of researcher	Video about nature in 360VR vs 2D video; 11 min	Spatial Presence Experience Scale (Hartmann et al., 2016)	Self-report: commitment to environment (CTE) scale, Davis et al. (2009)	Course credit for students	The immersive nature video was more likely to elicit feelings of connection to nature than the regular nature video
Breves & Schramm, 2021 DE	N = 112 86,6% females (μ)age = 22,04 Half student, faculty members, distant acquaintances of researcher	Video about the negative environmental consequences of mono-cropping in agricultural practices in 360VR vs 2D; 4 min		Self-report: personal relevance of the environmental (Ahn et al., 2016); pro-environmental behavioral intentions (Toomey & Domroese, 2013)	Course credit or a small amount of money (10 EUR) for students	No significant main effect of condition
Filter et al., 2020 DE	N = 50 72% females (μ)age = 23,76 Student	Video about wolves in their natural habitat in 360VR vs 2D video; 3,36 min	Spatial Presence Experience Scale (Hartmann et al., 2016)	Self-report: Attitudes towards wolves, based on attitudes toward wildlife by Kaczensky et al. (2004)	none	No significant main effect of condition
Fonseca & Kraus, 2016 DK	N = 64 42,19% females age = 19-53 University environment	Video about meat consumption and its environmental effects in 360VR vs 360 video on tablet vs 360 control video	PQ (3 items)	Self-reported attitudes toward meat consumption Behavioral: Post-stimulus choice experiment, whereby subjects could choose between a snack with or without meat	na	Significant effect on self-reported attitudes, no effect on behavior.
Pimentel et al., 2018 US	N = 20 55% females (μ)age = 14,95 Highschool students	IVE interaction with small vs large penguin; 5 min	none	Self- report: environmental threat perception (Leiserowitz, 2006; O’Connor et al., 1999); likelihood of pro-environmental behavioral intentions (e.g. donating money, volunteering, following social media pages)	none	Small sized penguins had a significant effect on attitudes, and perceived threat than large sized ones; no effect on behavioral intentions
Pimentel & Kalyanaraman, 2021 Study 1 US	N = 60 na% females (μ)age = 20,97	IVE visit at virtual sea turtle hospital, one vs many victim	Spatial presence (Harms & Biocca, 2004) Spatial presence (Sundar et al., 2017)	Self-report: environmentalism (Gatersleben et al., 2002); behavioral intentions (e.g. likelihood of donating money)	5 USD	Participants in the single victim condition donated more towards sea turtle conservation than those in the many victim condition.
Pimentel & Kalyanaraman, 2021 Study 2 US	N = 98 75,5% females (μ)age = 20,63	IVE (visuomotor synchrony: yes/no) x 2 (visuotactile synchrony: yes/no); 2-5 min	Spatial presence (Harms & Biocca, 2004) Spatial presence (Sundar et al., 2017)	Self-report: environmentalism (Gatersleben et al., 2002); behavioral intentions (e.g. likelihood of donating money); threat perceptions; inclusion of self in nature	Course credit	Embodiment increases the salience of the threats to the group, which contributes to charitable giving intentions
Pimentel & Kalyanaraman, 2021 Study 3 US	N = 90 43,3% females (μ)age = 18,95 Museum visitors	IVE victims: self vs one other vs seven others (yes/no)	none	Self-report: single item measure of perceived extinction, imagining winning 5USD and then rating donation amount	na	No effect on dependent measures
Pimentel & Kalyanaraman, 2021 US	N = 25 88% females (μ)age = 18,95 Student	IVE victims: number (one vs many) x species (similar vs different); 2-5 min		Self-report: environmentalism (Gatersleben et al., 2002); threat perceptions; inclusion of self in nature; behavioral intentions (e.g. likelihood of donating money)	5USD	No effect on perceived extinction and threat perceptions. Donation intention was higher for the one-victim condition than the many-victim condition. The effect was reversed when victims were of different species
Markowitz et al., 2018 Study 2 US	N = 47 highschool students 61,7% females age = 18-28 (Median = 20) Prior experience with VR (over 90%) student, Department of Communication’s research participant pool	IVE coral embodiment vs scuba diver embodiment; 6,5 min	Nine items by the authors	Self-report: New Ecological Paradigm Scale (Dunlap et al., 2000)	Course credit	No significant effect of condition
Markowitz et al., 2018 Study 4 US	N = 43 55,81% females Student	IVE swimming by using physical arm movements as gestures vs swimming by using a joystick; >3 min	Five items about the degree to which participants felt surrounded by the virtual environment.	Self-report: New Ecological Paradigm Scale (Dunlap et al., 2000)	Course credit
Racial Bias
Banakou et al., 2016 Study 1 ES	N = 60 100% females (μ)age = 21,9 University environment	IVE White vs Black embodiment while performing Tai Chi movements with Asian avatar teacher; 10 min	none	Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	15–20 EUR	Black embodiment reduced the degree of bias but did not change bias to non-bias
Banakou et al., 2016 Study 2 ES	N = 30 100% females (μ)age = 21,9 University environment	IVE White while performing Tai Chi movements with European avatar teacher; 10 min	none	Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	15–20 EUR	No effect of teacher appearance
Banakou et al., 2020 ES	N = 92 100% females (μ)age = 21,8 Caucasian/White; University environment	IVE BodyType (White, Black) x Crowd (Negative, Neutral, Positive); 35 min	One item on a Likert Scale from −3 (strongly disagree) to 3 (strongly agree) (I felt I was in the middle of a crossroad with people walking by)	Self-report: attitude to Blacks scale (Brigham, 1993) Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	15 EUR	No significant effect of conditions on self-reported attitudes- Neutral crowd condition: IAT decrease with Black embodiment and increase with White embodiment Positive crowd condition: IAT decrease with Black embodiment Negative crowd condition: IAT increase with Black embodiment
Groom et al., 2009 US	N = 98 60,2% females (μ)age not reported 46,94% White 21,43% Asian 16,33% African American 7,14% Hispanic 8,16% Other; Student	IVE Black vs White avatar embodiment during job interview	none	Self-report: Racial Argument Scale (RAS) (Saucier & Miller, 2003) and portions of the Modern Racism (MRS) scale (McConahay, 1986) Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	Class credit or payment	Increase in IAT for Black avatar embodiment
Hasler et al., 2017 ES	N = 32 100% females Caucasian/White Student	IVE Black vs White avatar embodiment during interaction with White and Black virtual character (different sessions,counterbalanced order); 6 min	none	Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	na	No effect of condition on IAT
Maloney, 2018 US	N = 14 0% females Caucasian/White University and surrounding environment	IVE Black vs White avatar embodiment during shooting task	iGroup Questionnaire (Schubert, 2003)	Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	na	IAT decrease in White avatar condition
Patané et al., 2020 FR	N = 40 67,5% females (μ)age = 25,8 Caucasian/White	IVE Black avatar embodiment during cooperative vs non-cooperative interaction with Black virtual character; 9 min	none	Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	na	IAT decrease in cooperative condition
Peck et al., 2013 US	N = 60 100% females age = EL: 23 ± 5, ED: 25 ± 7, ND: 20 ± 3, EA: 21 ± 1 Student	IVE Embodied-Light-Skinned (EL) avatar vs Embodied-Dark-Skinned (ED) vs Non-Embodied Dark-Skinned (ND) vs Embodied-Alien-Skinned (EA); subjects first explored and viewed their virtual body infront of a mirror, then were approached by 6 light and 6 dark-skinned avatars; 11,5 min	none	Self-report: explicit racial bias (Avenanti et al., 2010) adapted for Spain Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	na	IAT decrease in Embodied-Dark-Skinned condition
Salmanowitz, 2018 US	N = 92 50% females (μ)age = 28 university environment, 51% from Psychology department participant's tool, 49% from the local community	IVE Embodied VR/Black Defendant vs Embodied VR/White Defendant vs Sham (not embodied) VR/Black Defendant vs Sham VR/White Defendant; 5 min	Custom (7 items)	Self-report: Symbolic Racism Scale (SRS) (Henry & Sears, 2002) Racial bias Implicit Association Test (IAT) (Greenwald et al., 1998)	na	No effect of condition on explicit attitudes; IAT decrease in embodied groups and favorable rating of mock legal case in comparison to Sham groups
Gender Bias
Aitamurto et al., 2018 US	N = 67 58% females 67% of sample were aged 25–34 University	Video about gender inequality in 360VR vs 180VR vs 2D video; 10 min	Space presence and Social presence (Ahn et al., 2013; Oh et al., 2016)	Self-report: attitudes toward gender inequality; responsibility to address gender inequality (personal, community’s, employer’s and government’s); willingness to take action against gender equality	Voluntarily or 25 Amazon gift card	No effect of condition on attitudes or willingness to take action; Significant effect on personal responsibility
Fox & Bailenson, 2020 US	N = 83 48,19% females (μ)age = 20,82 38,6% White 24,1% Asian/Asian-American 13,3% Black/African/African-American 10.8% Latino/Latina/Hispanic 13,3% multiracial Convenience sample, university environment	IVE 2 Dress (suggestive vs conservative) x 2 Gaze (high vs. low); <1 min	Custom (12 items)	Self-report: Rape Myth Acceptance scale (RMAS; Burt, 1980); Two subscales from The Ambivalent Sexism Inventory (ASI; Glick & Fiske, 1996): Hostile Sexism (HS) and the Benevolent Sexism (BS)	Course credit or 10 USD	Increase in sexism and greater rape myth acceptance in the suggestive dressing x high gaze condition
Fox et al., 2013 US	N = 86 100% females (μ)age = 21,16 53,5% White/Caucasian 24,4% Asian/Asian–American 7% Black/African/African– American 7% Latina/o/Hispanic 2,3% Multiracial 5,8% Other Student	IVE embodiment of Sexualized Self vs Sexualized Other vs Nonsexualized Self vs Nonsexualized Other	Custom (12 items) based on Fox & Bailenson (2009), Price & Anderson (2007)	Rape Myth Acceptance scale (RMAS; Burt, 1980)	Course credit or 15USD	Increase in rape myth acceptance particularly in the Sexualized Self condition
Gonzalez-Liencres et al., 2020 ES	N = 32 0% females age = 18-64	IVE female first-person perspective vs observer during verbal abuse by male virtual character; 5,5 min	none	Gender bias Implicit Association Test (IAT) (Greenwald et al., 1998)	10 EUR	No effect of condition
Lopez et al., 2019 US	N = 24 0% females (μ)age = 29,8 White/Caucasian Students, University staff and members of the general public	IVE male vs female embodiment during Tai Chi exercise; 13 min	na	Gender bias Implicit Association Test (IAT) (Greenwald et al., 1998)	25 USD	Increase in IAT for female embodiment group
Steinfeld, 2020 IL	N = 73 64% females (μ)age = 24 students	Video about sexual harrasment at the workplace in vs 2D video vs print; 5 min	na	Self-report; Sexual Harassment Attitude Scale (Mazer & Percival, 1989) and the Ambivalent Sexism Inventory scale (Glick & Fiske, 2018)	Monetary compensation	Decrease in negative attitudes through interaction effect of gender (men), method of consumption (immersive), and assessment of the effectiveness and emotionality of the content
Homelessness
Herrera et al., 2018 Study 1 US	N = 117 64,1% females (μ)age = 22,94 28% White/Caucasian 6% Hispanic 6,8% Indian 41,8% Asian 6,8% African American 11.1% Multiracial University environment	IVE first-person perspective of being homeless vs text-based perspective-taking; 15 min	Adapted social presence scale, 6 items (Nowak & Biocca, 2003)	Self-report: attitudes toward the homeless (Batson et al., 1997) Behavioral: Agreement with proposition that supports affordable housing; signing petition for this proposition; donation question; letter writing to elected officials about the issue of homelessness; agreement with proposition about supportive initiatives	100USD Amazon gift card for completing the four parts of the study	No significant effect of condition on self-reported attitudes at T0, but significant effect of VR condition over the course of eight weeks No effect of condition on agreement with proposition or donation amount Increase in positive letter writing and petition signing for VR group
Herrera et al., 2018 Study 2 US	N = 439 56,9% females (μ)age = 29,2 48% White 20% Asian 9% Multiracial 6% Indian 5% African American 2% Middle Eastern 1% Native American 7% Hispanic 1% declined to answer; students and recruitment at schools, museums and senior citizen centers	IVE first-person perspective of being homeless vs interactive 2D perspective-taking vs text-based perspective-taking vs informative text; 15 min	Adapted social presence scale, 6 items (Nowak & Biocca, 2003)	Behavioral: Agreement with proposition that supports affordable housing; signing petition for this proposition; donation question	Course credit or a 10USD Amazon gift card	VR condition increased support for proposition and petition signing significantly more than other treatments No effect of condition on donation amount
Ageism
Banakou et al., 2018 ES	N = 30 0% females University environment	IVE Einstein embodiment vs normal adult embodiment during a numeric exercise; 10-12 min	na	Age bias Implicit Association Test (IAT) (Greenwald et al., 1998)	15 EUR	Decrease in age bias for Einstein embodiment group
Oh et al., 2016 Study 1 US	N = 148 64,19% females (μ)age = 21,03 43,2% White 31,8% Asian 10,8% Latino 7,4% Black 6,8% reported another ethnicity Student	Low threat article about the elderly vs high threat article x IVE elderly avatar embodiment with perspective-taking instructions vs mental perspective-taking vs control	non	Kogan's Attitudes towards Older People Scale (KOP) (Kogan, 1961) (sixteen items)	na	No significant effect of on ageism
Oh et al., 2016 Study 2 US	N = 84 54,76% females age < 35 53,57% White 11,90% Asian 8,33% Latino 13,10% Black 13,10% reported another ethnicity Student	Low threat interaction with virtual elderly avatars during a ball game vs. high threat interaction x IVE elderly avatar embodiment with perspective-taking instructions vs mental perspective-taking vs control	non	Future intentions to communicate with the elderly.	na	No significant effect of on ageism
Yee & Bailenson, 2007 US	N = 48 50% females Student	IVE embodied elderly avatar vs embodied young adult during interaction with confederate	non	Self-report: attitudes towards the elderly; Word association task; ambiguous story task (Levy, 1996)	Course credit	No effect of condition on attitudes or ambiguous story; Increase of positive words associated to the elderly for the elderly embodiment group.
Physical impairment
Ahn et al., 2013 Study 1 US	N = 44 54,55% females (μ)age = 22,57 Student	IVE being colorblind vs imagining being colorblind; 5 min	na	Self-report: Attitude Toward Disabled Persons Survey (Yuker et al., 1966) Behavioral: Number of completed tasks	na	Positive but non-significant effect of condition
Ahn et al., 2013 Study 2 US	N = 97 54,64% females (μ)age = 21,59 Student	IVE being colorblind vs imagining being colorblind	adapted from Bailenson & Yee, 2007; Nowak & Biocca, 2003; Witmer & Singer, 1998 (5 items)	Self-report: Attitude Toward Disabled Persons Survey (Yuker et al., 1966) Behavioral: Seconds invested in helping; number of words written to help identify the problematic areas of a website	na	Positive but non-significant effect of condition
Ahn et al., 2013 Study 3 US	N = 57 61,4% females (μ)age = 22,54 School	IVE being colorblind vs imagining being colorblind; 5 min	na	Behavioral: Seconds invested in helping; number of words written to help identify the problematic areas of a website	na	Significant effect of condition, IVE group spent significantly more time helping the colorblind person and also wrote significantly more words to help compared to participants in the perspective-taking condition
Chowdhury et al., 2021 US	N = 71 39,44% females (μ)age = 20,3 Student	Two (IVE vs desktop) x two interface (gamepad vs wheelchair); 30-35 min	PIFF ²	Implicit Association Test (IAT), (Greenwald et al., 1998)	none	Decrease in bias for the IVE treatment, no interaction effects
Migration & Human rights
Kim, 2009 US	N = 105 52,38% females (μ)age = 23,46 45,7% White 5,7% Black/African American 1,9% Hispanic/Latino 44,8% Asian/ Pacific Islander 1,9% Prefer not to disclose Student	Video about Syrian refugees crossing the Mediterranean sea in IVE vs 2D video; 7 min	TPI Social and Spatial Presence	Self-report: Scale by Telhami (2014) on ‘American attitudes on refugees from the middle east’ Behavioral intention; donation interest	na	IVE increased social and spatial presence, which in turn increased positive attitudes
Ma, 2020 US	N = 216 54,6% females (μ)age = 19,42 53,2% White 22,2% Asian 12,5% Black or African American 8,8% Hispanics 3,2% other racial backgrounds Students from participation pool	Video about the fight against malaria in Tanzania in 360VR vs 2D; 7 min	TPI Social and Spatial Presence	Self-report: Attitudes toward helping people who need protection from malaria and willingness to help (Peng et al., 2010)	na	No effect on attitudes or willingness to help.
Bujić et al., 2020 FI	N = 87 females 2:1 ratio (μ)age = 26,38 26,4% Finnish 10,3% Chinese 9,2% Spanish 9,2% Russian 5,7% Vietnamese	Video about Syrian refugees and war in 360VR vs 2D vs text; 7 min	na	Self-report: abridged version of the Human Rights Questionnaire (Diaz-Veizades et al. 1995)	Movie tickets	No significant effect difference between 360VR and 2D conditions
Bias against minorities
Chen et al., 2021 Study 1 SG	N = 77 42,5% females (μ)age = 24,28 University environment	IVE minority embodiment experiencing workplace aggression x perspective-taking instructions (affective vs cognitive)	na	Feeling thermometers measuring favorability	5 SGD	Increase in ingroup bias for both instruction groups
Chen et al., 2021 Study 2 SG	N = 73 53,42% females (μ)age = 24,51 University environment	Witnessing workplace aggression against minority x perspective-taking instructions (affective vs cognitive)	Self presence (Ratan & Hasler, 2010)	Feeling thermometers measuring favorability	5 SGD	Increase in ingroup bias for both instruction groups
Intergroup Conflict
Hasler et al., 2017 IL	N = 55 72,73% females (μ)age = 23,86 Israeli Student	Video from Palestinian couple’s perspective during interaction with Israeli soldier in 360VR vs in 2D	none	Self-report: moral evaluation of the soldiers’ actions	Course credit	Subjects in the 360VR condition judged the soldiers’ actions as significantly less justified and less moral than the 2D group
Bullying
McEvoy et al., 2016 US	N = 78 67,95% females (μ)age = 19,74 Student	Witnessing bullying in 360VR vs witnessing bullying in VR where avatars wear clothes with the university’s logo vs 2D	none	Self-report: atittudes toward bullies and bullying victims (Batson et al., 1997); attitudes toward bullying (Swearer et al., 2008); future anticipated bystander intervention behavior in bullying situations (Banyard et al., 2007)	none	No effects of media condition on measures
Mental health
Kalyanaraman et al., 2010 US	N = 112 60,71% females (μ)age = 22,25 University environment, psychology research	Virtual simulation of schizophrenia vs a written empathy-set induction of schizophrenia vs a combi-nation of both the simulation and written empathy conditions vs a control; 4,5 min	none	Self-report: adapted scale based on Attitudes towards people with AIDS (Batson et al., 1997)	Course credit or 20 USD	Increase in positive attitudes for the combined treatment group
Vaccine Skepticism
Nowak et al., 2020	N = 171 67,3% females age = 18-49	Story in IVE vs 2D vs electronic pamphlet; 5 min	Yes, not specified	Self-report: community immunity beliefs (Nowak et al., 2020)	30 USD	Significant effects of VR treatment on attitudes through higher presence scores

Note: Following abbreviations were used in the ‘Presence Measurement’ column: Witmer & Singer’s Presence Questionnaire (PQ), Presence Involvement Flow Framework (PIFF2), Temple Presence Inventory developed by Lombard, Ditton and Weinstein (2009) (TPI).

The final data pool was coded by two authors trained in a specific coding scheme that included detailed guidelines for the definition of each variable. Coded variables included study details (participants’ demographics, sample type, participation incentive); aspects of the dependent variable (e.g. attitude type, measurement, effect sizes pre- and post-stimulus); presence measurements; aspects of the VR experience (e.g. sense of embodiment, perspective-taking); moderator variables; and additional conditions and technological affordances of the VR experience (duration of stay in minutes, HMD, tracking, audio). After specific coding intervals, the authors compared their results and discussed each coding step thoroughly to reach consensus on the ascribed values.

Dependent variable

Social attitude. This variable considers attitudinal measures defined as evaluative beliefs, opinions, affective reactions, and behavioral measures (Oskamp & Schultz, 2005) toward an attitude object of broader public relevance (Haddock & Zanna, 1999). Examples include abstract objects, such as the environment and migration, as well as narrower objects, such as policies or social groups. To ascertain internal validity, we restricted the search to studies that explicitly labeled their dependent variable as attitude and that employed at least one self-report or behavioral measure of that attitude. Thus, we excluded conceptually different constructs like empathy, which describes ‘the capacity to share and understand emotional states of others in reference to oneself’ (Decety & Moriguchi, 2007, p. 22). Attitude research related to marketing, advertising, and brand communication was excluded as the attitude objects did not match the research objectives.

Independent variables

VR interventions. For purposes of internal validity, we restricted the definition of VR interventions to experiences facilitated by the use of a Head Mounted Display, thus ensuring that immersive factors such as field of view, field of regard, display size and resolution, stereoscopy, rendering, realism, frame rate, and refresh rate (Bowman & McMahan, 2007) were within the same range. Interventions induced by less immersive devices, for example Google Cardboard, were excluded. In addition to fully immersive VR, we included studies that used 360VR as it realistically conveys an external environment that can be experienced on multiple axes (Huang et al., 2017).

Virtual 1PP. Virtual 1PP refers to the experience of the external environment from the first-person perspective of a person other than oneself. Subcodes included in-group 1PP, when the assigned experimental treatment aligned with a subject’s in-group (e.g. a male subject assigned to male 1PP), and out-group 1PP, when the treatment provided the subject with the perspective of an out-group member (e.g. a male subject assigned to female 1PP). In-group and out-group had to be clearly defined by the study to be included.

Moderator variable

Sense of Embodiment. Studies in this code category included an embodied experience as experimental treatment whereby users experienced (1) a sense of self-location (first-person perspective overlaps with position of artificial body’s eyes), (2) a sense of agency (motoric and visual correlation, being in control of global motor control), and (3) a sense of body ownership (visibility of a virtual body, enhanced synchrony of bodily movements), hence, taken together, a Sense of Embodiment (Kilteni et al., 2012). For this category, in-text descriptions, screenshots, and videos were used for the coding. When no information except for the treatment contents (e.g. title of video) was provided, online searches helped gather the material.

Statistical analysis and procedure for meta-analysis

For the quantitative synthesis and graphical presentation, we used the R packages metafor (Viechtbauer, 2010) and metaviz (Kossmeier et al., 2019). To ensure comparability of the studies included in the meta-analyses, we included only the effect sizes regarding post-treatment measurements the first time they were collected (therefore, no effects were measured over time), self-reported and implicit attitudes (therefore, no intent or behavioral data), and when participants received the same content (therefore, no unrelated control conditions). In cases in which more than one effect size within a sample was eligible to be included in the meta-analysis, we summarized those effects by calculating the mean (Borenstein et al., 2009).

We performed two separate meta-analyses for our independent variables, IVE interventions and virtual perspective-taking. Due to the diversity of the studies, we used a random-effects model with the Hedges’ estimator to conduct the meta-analyses because the Hedges estimator is mainly unbiased in the case of meta-analyses with few effect sizes and a heterogeneous sampling variance (Viechtbauer, 2010). To investigate the effect size of heterogeneity, we calculated I², τ², and Q and used a Baujat plot to identify studies that contribute disproportionately to the overall heterogeneity and the summary effect size. We excluded two samples (Banakou et al., 2020, positive condition; Lopez et al., 2019) from the second meta-analysis due to irregularities in the Baujat plot. Additionally, we conducted a leave-one-out sensitivity analysis to examine whether any study would change the summary effect size if not included within the meta-analyses. Further, we checked for indications of a publication bias using Egger’s regression test. To analyze the effect of the sense of embodiment, we calculated a moderator analysis (mixed-effects model) with ‘sense of embodiment’ as the moderator on the effect of VR interventions.

Review of the literature

We discuss the identified studies using the attitude object categories that emerged in our review of the literature.

Environment. A considerable number of empirical investigations (n = 17) explored VR exposure effects on environmental attitudes and attitude-expressive behavior. In terms of VR contents, small-sized animals and single-victim narratives (e.g. one turtle) enhanced feelings of environmental threat, conservation intentions, and small donations significantly more than large-sized animals and multiple-victim narratives did (Pimentel et al., 2018; Pimentel & Kalyanaram, 2021-Studies 1 and 4). Embodiment studies yielded mixed results. The visual representation mattered in some studies, wherein human and large-animal embodiments resulted in larger effects than other species embodiments (Ahn et al., 2016). Further, one study found that the vividness of the VR contents can lead to more water-saving behaviors (Bailey et al., 2015). Studies that compared the effects of varying immersive modality often found that VR outperformed less immersive treatments in eliciting attitudinal change, for example, use of paper and recycling intentions (Ahn et al., 2014), self-reported commitment to the environment (Ahn et al., 2016; Breves & Heber, 2020), and likelihood of engaging in gardening and talking about environmental issues with friends (Breves & Schramm, 2021).

Overall, it merits discussion that the investigated effects align with the low-cost hypothesis, according to which environmental behavior related to low costs and minor inconveniences (e.g. recycling, water-saving) is more likely to change than high-cost objects (e.g. no airplane flights, no car in the household) (Diekmann & Preisendörfer, 2003). Only one study investigated VR effects on a more high-cost environmental attitudinal object, namely not eating meat for environmental reasons (Fonseca & Kraus, 2016), and found no significant effect. The authors showed that, although subjects in the VR condition reported higher intentions of reducing meat consumption for environmental reasons than subjects in the 2D condition, both groups chose a meat option in a post-treatment behavioral measurement. Considering that most studies found VR effects on low-cost environmental behavior, it is possible that the nature of the measured object moderates found effect sizes. It may be theoretically relevant to further explore the extent to which low- and high-cost environmental behavior is influenced differently by VR treatments.

Racial bias. Several studies have reported that the embodiment of a Black avatar reduces racial bias in White subjects (e.g. Salmanowitz, 2018). These effects are most consistently observed when the avatar is placed in a neutral-to-positive social context, for example, athletic exercises (Peck et al., 2013) and cooperative tasks (Patané et al., 2020). In contrast, Black embodiment in a negative social context can increase racial bias for White subjects, for example, when placed in a negative crowd (Banakou et al., 2020) or in a job interview (Groom et al., 2009). Scholars have argued that hostile situations interfere with the ownership of the virtual body and, thus, prevent the formation of a positive relation to the avatar (Banakou et al., 2020). Moreover, Maloney (2018) found that White avatar embodiment during a gun shooting exercise can decrease IAT scores more than black avatar embodiment, and linked this finding to potential social desirability effects as the researcher conducting the study was Black. Further, we reviewed the avatar images, which were available in all the identified studies. Here, it is critical to note that the ‘Black’ avatars resembled the ‘White’ avatars but with a tanned or darker complexion. Specific characteristics, for example, Black hair textures, were not addressed in the studies except for one (Banakou et al., 2020). The visual representation of avatar-based racial identity should be taken into consideration in future work.

Gender bias. Several studies examined how VR experiences affect gender bias. One group of studies found that the embodiment of sexualized female avatars can activate stereotypes (Fox & Bailenson, 2009; Fox et al., 2013). Stereotype activation was also observed by Lopez et al. (2019), who found that female embodiment during a sports exercise increased implicit gender attitudes in comparison to male embodiment. The other group of studies examined how VR exposure can influence attitudes toward sexual harassment. Steinfeld (2020) showed that men who experienced workplace harassment from a woman’s viewpoint in VR changed their views on the matter, particularly when the issue felt emotional to them. Feeling for the male character also moderated VR effects on gender inequality attitudes in another study (Aitamurto et al., 2018). These findings point to the importance of individual and contextual factors that need to be further explored.

Homelessness. One study (Herrera et al., 2018) found positive effects on attitudes toward homelessness in two consecutive experiments. In their first empirical investigation, VR fostered more positive attitudes and behaviors compared to text-based narrative perspective-taking. Most significantly, the effect declined to a smaller degree for the VR group over the course of eight weeks. In a second experiment, VR elicited greater support for affordable housing than perspective-taking via a text-based narrative and a 2D video. Considering that a long research tradition has been discussing attitudes toward homelessness as a multifaceted construct, it is worth exploring how VR treatments affect its various components such as indifference, hostility, annoyance, revulsion, and curiosity (Bahr, 1973).

Ageism. A small number of studies explored how interaction with or the embodiment of elderly avatars affects ageism among users. Similar to research on racial bias, we found that positive social context yielded larger effects. For example, elderly avatar embodiment during a friendly interaction with a confederate increased positive associations with the elderly in contrast to young avatar embodiment, even though there was no direct effect on self-reported attitudes. (Yee & Bailenson, 2007). It might also be possible that the embodiment of an elderly figure associated with positive social characteristics such as wisdom and success can contribute to reducing ageism; one study found that an Einstein embodiment scored significantly lower levels of implicit ageism in comparison to the embodiment of an unknown adult (Banakou et al., 2018). Further, positive feedback during task execution suppressed ageism in subjects who embodied an elderly avatar, while negative or neutral reactions reinforced hostile ageism (Tsao et al., 2020). Likewise, Oh et al. (2016) demonstrated that subjects who felt excluded by elderly avatars during a virtual ball game showed increased ageism, regardless of the perspective-taking exercise that followed. However, playing games with the elderly might not be an ecologically valid behavioral measure for ageism reduction. For this, more research is necessary to test potential VR effects on forms of ageism occurring in real-world settings, for example, age discrimination in the workplace (Rupp et al., 2006).

Physical impairment. Two studies measured effects on attitudes toward physical impairments. Ahn et al. (2013) compared mentalizing to being virtually color-blind, finding mixed results in three consecutive experiments. In two investigations, no conclusive attitudinal and behavioral effect was found; however, a third experiment with improved believability showed that subjects who experienced color-blind vision in an IVE demonstrated twice as much helping behavior than subjects who imagined being color-blind. In another study, experiencing how it feels to be in a wheelchair in VR decreased implicit bias toward people with disabilities more effectively than the gamepad and desktop condition (Chowdhury et al., 2021).

Migration, human rights, and humanitarian aid. In comparison to video, Kim (2019) found that VR can induce levels of telepresence and social presence, which in turn factored into more positive attitudes toward and more donations for Syrian refugees. Another study found that both the VR and the 2D treatment outperformed text in eliciting positive attitudes toward human rights (Bujić et al., 2020). More research is needed on factors that might moderate these effects, for instance the role of geographical location (Czaika & Di Lillo, 2018). Further, Ma (2020) did not find a significant main effect of 360VR on self-reported attitudes or willingness to help Tanzanian populations in their fight against malaria; however, it indirectly affected spatial presence, transportation, counterarguing, and, eventually, favorable attitudes. Given that the stimulus was released by the United Nations Foundation to promote support in the fight against malaria, future studies should explore the effectiveness of VR use by non-governmental and non-profit organizations.

Biases against minorities. In two empirical investigations, Chen et al. (2021) compared whether cognitive or affective perspective-taking instructions would affect biases against minorities. In the VR video, subjects either embodied a migrant worker who experiences aggression at the workplace in 1PP (experiment one), or they observed the interaction in 3PP (experiment two). Regardless of instructions and perspective, both studies found that in-group biases increased after VR exposure. In other words, subjects perceived their in-group as more negative rather than their out-group as positive.

Intergroup conflict. One study showed that, when Israeli subjects adopted a Palestinian’s perspective in 360VR, feelings of dehumanization and threat toward Palestinians could decrease (Hasler et al., 2017). Similar effects were found in a second measurement five months after the VR treatment. While one study does not suffice to draw conclusions, these early findings point to the practical implications of VR experiences on real-world intergroup conflict perception.

Bullying. McEvoy et al. (2016) studied VR as a tool for bullying prevention and found no significant effect of media condition on the perception of the school’s responsibility to address bullying or on attitudes toward bullying and future bystander behavior intention. On the contrary, more subjects who watched the video in 2D reported perceiving school bullying as a problem than participants in the VR condition. Consistent with literature on the relation of immersive affordances and presence (Cummings & Bailenson, 2016), subjects in a subsequent focus group reported that a higher degree of realism, interactivity, and prominent customization would have affected their responses.

Mental health. Kalyanaraman et al. (2010) conducted research to determine whether experiencing how it feels to suffer from schizophrenia affects users’ attitudes toward the condition differently than a written empathy condition or both treatments combined. The combined condition outperformed the others; however, in contrast to the desired effect, the VR-only condition increased social distance. Thus the authors suggested that such simulations should be used with caution (Kalyanaraman et al., 2010).

Vaccine skepticism. One study found that feelings of presence in VR mediated the belief that the flu vaccination can help protect other members of the community, particularly in contrast to video, tablet, and text (Nowak et al., 2020). However, considering the COVID-19 pandemic and the vaccine debate it sparked, such empirical investigations need to be replicated in order to reach conclusive results.

Summary of review sample

Regarding our final research question about the role of sampling, compensation, and duration of treatment, we identified several challenges and opportunities in the reviewed studies. To begin with, samples were commonly not controlled for familiarity with VR technology. Although a few investigations tested whether subjects were familiar with VR (e.g. Markowitz et al., 2018-Study 2; Tsao et al., 2020; Breves & Schramm, 2021), precise numeric estimates or scales were not included. This observation is surprising, particularly given that studies often reported that novelty effects might have affected the measured outcomes (e.g. Pimentel et al., 2018; Bujić et al., 2020; Filter et al., 2020). The lack of novelty control is especially critical in studies that compared the effects of VR to less immersive yet familiar media. In other words, it is likely that, in some of the reviewed studies, one experimental group was assigned a medium they have known for decades (e.g. video), whereas the other group was assigned a medium they had potentially never encountered before. For this reason, it is crucial to pretest familiarity with VR and to account for it in the analyses, particularly because novel experiences can increase interest in the object and excitement about its applications (Izard, 1977).

Concerning studies that compared out-group 1PP to in-group 1PP, only one or two social demographic characteristics were used as sampling criteria; for example, Banakou and colleagues (2020) included Caucasian female subjects to test for the effects of Black female avatar embodiment on racial bias. Other information that might factor into feelings of social identity, for example, age or socioeconomic status, were not used as inclusion criteria. Thus, subjects shared one or two in-groups but potentially differed in others. This trend emerges across the reviewed in-group–out-group attitude literature and makes it difficult to draw definitive conclusions. A more precise approach could be to determine a certain level of homogeneity of potential in-groups, for example, age, race, and socioeconomic status, in order to isolate the causal effect of the manipulated out-group variable (e.g. gender).

Further, because studies mostly used university-environment samples, it is difficult to determine whether social desirability had an impact on effect sizes. More empirical research with non-student samples is needed to reach conclusive answers. This also applies to the level of education and political orientation. Only two studies controlled for political inclination. Here, one study noted that only 10.3% reported being conservative (Bujić et al., 2020), while in the other study, subjects with a left-wing orientation scored higher levels of empathy and perspective-taking tendencies than participants with a right-wing orientation. These two early cases indicate that the analysis of prior political attitudes might be a valuable source of information for the causal isolation of found effects.

Concerning compensation, studies commonly offered either course credit (for student samples) or monetary compensation in the form of gift cards or direct financial incentives. These ranged from 5 SGD ( = 3, 72 US dollars) to 25 dollars, regardless of the duration of treatment. Subjects assigned to the VR condition received the same compensation as subjects in other groups. We could not identify any specific reasoning that determined the amount of compensation given to participants. Given that most of the identified studies were conducted in overwhelmingly WEIRD nations (i.e. Western Industrialized Rich Democratic; Henrich et al., 2010) with predominantly White, young, well-educated samples, it would be valuable if future studies were to examine how compensation levels vary across time, location, and population.

Finally, we found that, when a time specification was given, VR treatments lasted 7 minutes on average (min: 0.82 min; max: 35 min). It must be noted, however, that some studies reported the total amount of time subjects were immersed in the virtual world, while others reported the duration of the specific treatment task excluding preparation and instruction time. VR interventions naturally did not differ in duration from 2D interventions because the same stimulus video was used (e.g. Fonseca & Kraus, 2016). However, when compared to print treatments, the VR experience was often longer than the reading of the text. For example, one study reported that ‘the entire experiment lasted between 20 and 30 min, depending on the treatment, as the text article did not necessarily require the allocated 7 min to read’ (Bujić et al., 2020, p. 1413). To avoid interfering effects resulting from varying exposure time, future work should consider keeping treatment times consistent. Further, some studies referred to cybersickness as reason to exclude subjects from the study (e.g. Breves & Schramm, 2020). We could not identify a relevant role of cybersickness in regards to outcomes.

Taken together, the roles of sampling, compensation, and duration of treatment need to be explored further to gain a more precise causal isolation of main effects.

Meta-Analysis findings

Vr vs. less immersive conditions

We conducted a meta-analysis for which we summarized the effects of 24 samples with a total of 1,878 subjects. The main meta-analytic results suggest that across studies, VR interventions had a significantly stronger positive effect on attitudes than traditional media interventions (Hedges’ g = 0.21, p <.01, 95% CI [0.08; 0.34], I² = 45.00%, τ² (SE) = 0.05 (0.04), Q(23) = 34.35, p = .06), confirming the assumption stated in the literature. Even though the heterogeneity was moderate, the Baujat plot did not indicate a single study with a strong influence on the overall result and overall heterogeneity, and the leave-one-out sensitivity analysis did not flag any study that would change the summary effect size a lot if it were left out. Further, Egger’s regression test did not indicate a publication bias (t(22) = 0.69, p = .50). H1 was supported. Effect sizes and the summary effect size are shown in the rainforest plot (Figure 2).

Figure 2. Rainforest plot depicting results of meta-analysis comparing effect sizes by VR interventions to effect sizes caused by less immersive experimental conditions

Sense of embodiment as moderator

A moderation analysis was employed to provide information about whether a sense of embodiment influences the overall effect of VR interventions on attitudes. Contrary to our expectations, a sense of embodiment did not impact the effectiveness of VR on the dependent variable across studies (see Figure 3). The moderator analysis showed that the average true effect does not significantly differ over the two conditions (QM(1) = 0.06, p = .80). H2 was not supported.

Figure 3. Rainforest plot depicting moderation analysis results of comparison between effect sizes related to VR interventions providing sense of embodiment and not providing sense of embodiment

Outgroup 1PP vs ingroup 1PP

Finally, we compared the effect sizes of in-group vs. out-group 1PP in VR. We used nine independent samples with a total of 411 participants. The results indicate a positive effect of out-group perspective-taking on prejudicial attitudes (Hedges’ g = 0.54, p < .001, 95% CI [0.28; 0.81], I² = 42.83%, τ² (SE) = 0.07 (0.09), Q(8) = 12.72, p = .12) (Figure 4). Neither sensitivity analysis nor the Baujat plot revealed any other than the earlier described irregularities. Most of the included studies used an implicit instrument to measure attitude (IAT, Greenwald et al., 1998) and an interactive IVE. Only one study (Hasson et al., 2019) used a questionnaire to measure attitude and a 360-degree video as IVE. However, the sensitivity analysis did not reveal a significant difference in the summary effect had this study been excluded. Therefore, we decided to include this study in the analysis. Further, Egger’s regression test was not significant (t(7) = 1.98, p = .08). Overall, consistent with prior literature, seeing the virtual world through the eyes of an out-group member did foster more positive attitudes toward the out-group than adapting the in-group’s perspective. H3 was supported.

Figure 4. Rainforest plot depicting results of a meta-analysis comparing effect sizes related to virtual ingroup-perspective taking and virtual outgroup-perspective taking. A positive effect size equals a more positive attitude after outgroup 1PP than ingroup 1PP.

Discussion

In this contribution, we provide an overview and meta-analysis of experimental research on the effect of VR interventions on social attitudes. First, we find that an increased immersive modality tends to yield greater attitudinal effects as VR outperformed print, 2D video, and oral narration treatments in the reviewed studies. VR elicited larger degrees of attitudinal change regardless of whether these were measured with self-reports, behavioral intention, or actual behavior in the literature. This finding can inform both media-richness researchers and practitioners about VR’s potential to influence certain types of attitudes, namely those directed toward social objects like climate change, racial biases, or intergroup conflicts.

However, this conclusion must be read with caution. From a theoretical perspective, most studies identified presence as the phenomenon that explains why VR outcomes are greater than traditional media effects while neglecting other possible theoretical explanations. This applies both to studies that measured actual levels of presence (e.g. Kim, 2019) and those that made a mere theoretical reference to its concept (e.g. Kalyanaraman et al., 2010). The assumption that telepresence, social presence or self-presence have a direct effect on attitudinal constructs requires further empirical investigation, ideally in experimental settings that maintain consistent levels of immersion, embodiment, or duration of treatment while manipulating exclusively the presence subtype of interest. Moreover, increased attention should be given to moderators and mediators, for instance, novelty effects (Tokunaga, 2013), feelings of interest or excitement (Izards, 1977), or political inclinations (e.g. Patané et al., 2020), before a causal relation between presence and social attitudes can be drawn. Further, homogeneity of age, race, and level of education in the pooled studies may have suppressed critical information that might have been obtained with a more diverse sample. Taken together, further efforts in theory and methodology are needed to reach conclusive answers about VR’s persuasive effect on social attitudes.

Secondly, we find that embodied VR experiences did not outperform non-embodied VR experiences in evoking larger effects compared to less immersive media. This finding stands in contrast to the literature, wherein embodiment is linked to persuasive outcomes because of its capacity to provide direct experiences (Ahn, 2021) and facilitate embodied cognition (Yee & Bailenson, 2009). We suspect that future analyses with a larger pool can confirm a moderating effect, as single studies (e.g. Banakou et al., 2018) reported successful embodied cognition, which indicates that a sense of embodiment can alter users’ self-identity and their behavior. Another explanation may lie in the novelty of embodiment, which potentially shifts the user’s attention away from the message contents, thus suppressing the desired effect. It is also possible that 360VR suffices to elicit attitudinal effects toward social issues. Moreover, 360VR videos can feature objects, humans, and events with photographic and behavioral realism, both of which contribute to visual realism and the overall media experience (Harris et al., 2009). For instance, viewing a spherical film of a refugee camp in 360VR might evoke a similar attitudinal change toward migration as a fully immersive, embodied experience. This assumption might be relevant for scholars and practitioners interested in the efficacy of 360VR experiences, especially given the high costs that come with IVEs.

Thirdly, our analysis indicates that, across studies, out-group 1PP was more effective in suppressing bias and negative attitudes than in-group 1PP. This finding aligns with prior literature on mental perspective-taking and corresponds with the notion that fundamental cognitive processes observable in real life also occur in VR (Sheridan, 1999). We suggest that scholars further explore the nuanced effects of virtual out-group perspective-taking, for instance, meta-stereotype activation (Vorauer & Sasaki, 2009). This is particularly important, given the backfire effects found in some of the studies (Groom et al., 2009; Lopez et al., 2019). Beyond their theoretical implications, it should be stressed that studies like that of Hasson et al. (2019) illustrate the potential of virtual out-group 1PP to impact deeply ingrained patterns of social attitudes. By enabling citizens to witness the consequences of violent intergroup conflict from the other’s perspective, VR experiences can potentially contribute to a shift in the public perception of complex historical and political disputes.

While the aforementioned findings contribute to understanding the current state of the art, our study has several limitations. Due to internal validity purposes and the need for a clear selection criterion, we included only studies that explicitly defined an attitude as a dependent variable. Thereby, we excluded work that tackled related constructs but did not refer to them as attitudinal (e.g. Kandaurova & Lee, 2019). Doing so limited the final pool of relevant literature, yet it permitted a fruitful comparison across findings. Further, the effect sizes that powered our meta-analysis are comparable on the basis of treatment medium but not measure. The aim of the analysis was to detect whether the effect sizes generated by VR interventions overall differed from those generated by less immersive interventions, regardless of the type of attitude. The same applies to the moderating role of sense of embodiment and the difference between virtual in-group vs. out-group 1PP. It is, therefore, important to stress that the identified studies explored different types of attitudes and assessed them with varying measures. Heterogeneity tests, however, allowed for a robust ground for comparison, and therefore, we hope that the trends we observe in our analysis are of a general nature.

In sum, our analysis of extant literature finds that VR affects attitudes toward social issues significantly more than less immersive interventions. With some areas of human experience and connection potentially migrating to virtual spaces over the next decades (Li, 2022), it is critical for scholarship to explore beyond main effects and further investigate how, and why, VR can influence the ways we think and act in the physical world.

Disclosure statement

No potential conflict of interest was reported by the author(s).