Realism and Field of View Affect Presence in VR but Not the Way You Think

2.1 Visual Realism

It has been argued that the realism of a VE is the most important factor ultimately driving user presence [91]. The visual realism of a VE has itself been described as composed of two main components. First, geometric realism refers to how realistic objects within the VE look, or how close they are to their real world counterparts [93]. The second aspect of realism is illumination realism, which refers to the fidelity of lighting and shadows cast by objects in the VE [93]. These can be further divided into quality of objects and terrain [110], texture and lighting [118], and shadow quality [60, 93, 95]. It has been shown that when presented with a VE, users will invariably compare the look of virtual objects with real life ones, in order to judge the level of congruence [100]. Indeed, Weber and colleagues [106] state that from the perceptual and conceptual point of view of the user, realism can be sub-divided into separate components such as coherence [89], fidelity [3], judgement of reality [5] and perceived realism [14, 81, 90]. Thus, most factors that determine the level of realism of a VE are heavily dependent on the way a VE looks.

However, due to the multitude of factors manipulated and hardware utilised across studies, there is currently no consensus as to the exact effect of visual realism on presence. The strong effects of visual realism on user experience are hard to contest. For example, there is evidence that perceived VE realism can even affect user behaviour in VR [88]. Some studies have found that visual realism can be beneficial for presence [49, 93, 110], while others found no such effect [23, 57, 60, 118]. Still, a portion of this work did not account for affective content, which is a characteristic of most VR games. One more recent study aimed to address this limitation and presented users with two versions of a VR game that elicited fear [38]. The authors manipulated polygon count and texture resolution and found that a higher level of realism enhanced presence [38]).

Moreover, not only the software technical component contributes to visual realism. Even in a hypothetical scenario where a VR VE could be perceptually indistinguishable from reality, the way in which this VE is perceived is still mediated by the hardware of the HMD, in particular the display. In other words, presence may not be solely determined by what users perceive in a VE but also how users perceive it.

2.2 Field of View

The impact of field of view on presence has long been studied with 2D screens, e.g. [12, 37, 47], with results suggesting that wider screens enhance the immersive features of an application. VR HMD displays are still evolving, with many features such as pixel density and colour accuracy still in need of improvements [18]; the afforded FoV of HMDs has been constantly improving since their appearance on the market. It is important to acknowledge, however, that the human’s average binocular field of view (FoV) reaches up to 190° [2], whereas the most popular consumer HMDs currently stand at around half that. The benefits of wider FoVs for VR presence have been highlighted by early studies [53, 73, 98] but the findings were not unanimous [43]. Again, these studies are dated, employing setups with far lower FoVs than are prevalent today. A more recent meta-analysis by Cummings and Bailenson [19] showed that the FoV of an HMD can play an even more important role in presence formation than visual realism. However, this meta-analysis was published before the recent wave of modern consumer HMDs [86]. A recent study from 2021 showed that participants using a variety of modern HMDs experienced lower presence with a reduced FoV [102]. This study, however, was conducted remotely and due to differences in the native FoVs of various HMDs, it is difficult to draw conclusions as to what precisely the wide and narrow FoVs were. A reduced FoV not only affects presence directly but can lead to skewed distance estimation within VR [62], which influences users’ ability to co-locate themselves among other landmarks within the VE, which is a prerequisite for presence [97].

Despite the apparent advantages of wider FoVs for presence, deploying wider screens to HMDs presents several challenges. First, a wider FoV implies that a larger portion of the VE is displayed at one time, which can be extremely taxing on limited computational power. At the hardware level, as remarked in a recent review by Angelov et al. [4], HMDs often have to compromise other qualities, such as pixel density, in exchange for a wider FoV. Lower pixel density can reduce presence and potentially contribute to VR motion sickness [61]. Simultaneously increasing the FoV and VE verisimilitude meets the same challenge of limited computational resources. It is clear that hardware and software technical features need to be balanced to maximise presence while staying within the bounds of available processing power. Before beginning to understand how such technical factors impact presence, however, there is a need also to take into account factors originating from the user.

2.3 Human Factors

Earlier models of presence placed a heavy emphasis on the technical factors of VR which enabled VEs to feel more realistic e.g. [112]. However, with significant advances in those areas, more recent research has started to investigate human factors too. More recent models acknowledge that ultimately the user determines whether presence is formed, and this also depends on how they feel within the VE [21, 39, 42, 85].

Agency, or the perception of acting within a VE, has been at times neglected in accounts of user presence. However, with advances in hardware, in particular tracking technology, agency has received more attention within the presence literature. Sanchez-Vives et al. [82] describe presence as grounded in the feeling not only of “being there” but also “doing there”. The user perceived verisimilitude of the interaction is also amongst the elements that contribute to overall perceived realism and thus user presence [106]. This is supported by Magnenat-Thalmann [58] who points out that amongst the crucial aspects that drive user presence are the presentation of the environment and the interaction that is afforded to the user within it, i.e. agency. Jang and Park [39] aimed to create a SEM model to explain user retention and adoption of HMD VR. They showed that both technical features and agency contributed to presence. However, their study did not immerse users in an actual controlled VE, but merely asked them which factors they considered important. Such reporting could be dependent on any variety of VR applications that the users had engaged with, and although informative does not allow for clear design recommendations. Moreover, the authors did not test for interactions between the design factors, nor for affective content, which can have a strong impact on presence [41, 42].

Despite this evidence that both technical factors (realism of the VE or the HMD’s FoV) and human factors (such as emotion and agency) contribute to presence, to our knowledge no previous research has investigated how they interact in the formation of presence. For example, Hvass et al. [38] immersed users in two VEs with different levels of realism, both of which afforded users agency and elicited fear. It was found that fear levels were lower in the condition presenting poorer realism. This suggests that technical factors are indeed able to moderate the intensity of felt emotion, which could in turn affect presence. However, as shown by Jicol et al.’s [42] PEA model of presence, agency also moderates the effect of fear on presence. This makes it problematic to expand the findings of Hvass et al. [38] to VEs where users do not have agency. Agency could in fact interact with technical factors in their effect on fear and perhaps presence, but such an effect has not been tested with modern hardware.

Hence, a clear understanding of whether and how technical and human factors may interact to create presence is missing. Past literature has clarified the relevance of technical characteristics such as VE realism and FoV as well as fear and agency [42]. The current study expands this knowledge by systematically investigating not only their contributions to presence in isolation but also their interactions with each other. To answer RQ1 and RQ2, we collected a new large data set and used it to create a novel SEM model including the technical factors realism and FoV, and the human factors agency and emotion. When considering emotion, our particular focus is on fear, as it holds significant importance to the VR industry and has shown intricate relationships with technical factors [38], agency and presence [42].

3.1 Apparatus

A Valve Index HMD was used to display the VEs. This was chosen for its high FoV of 130°, which is considerably higher than the most popular commercial HMDs at the moment. Other HMDs such as the 8K Priax can offer higher FoVs but present other limitations, such as limited frame rate. Moreover, past research has shown no differences in presence between 140° and 180° FOVs [53], suggesting that the Valve Index maximises the effects of FoV as far as impact on presence is concerned. The Valve Index also provides a high maximum frame rate of 144Hz, which allows us to discount frame rate as a confounding variable, and high resolution (two 1440×1600 LCD IPS Fast Switching Type Displays). Additionally, because we tested every combination of our four factors and levels of each, we used the same headset for all conditions to control for confounding variables and subtle differences between hardware, such as frame rate and pixel density. This also avoids issues that the advertised FoV of an HMD is not necessarily perceived as such by the human eye, due to a variety of factors such as optics and positioning of the user’s eyes relative to the screen. As such, the actual perceived FoV of the Valve Index can be closer to 110°. However, this would affect both the high and low FoV conditions equally given that we used the same HMD in all conditions. The HMD was powered by a desktop computer running Windows 10 with an Intel i7-9900k processor, an RTX 2080Ti GPU and 64GB of RAM. These specifications align with recent studies using similar VR stimuli and Unity recommendations [42, 64].

3.2 Stimuli

When selecting different levels for the technical factors we focused on realistic design decisions to maximise applicability to consumer-grade VR HMDs and PCs. As we were considering the same human factors identified by the recent PEA model proposed by Jicol et al. [42], we chose to modify their VEs to suit our design. These VEs have been validated for their effects on emotion and agency, and are freely available³(see Figure 1). Using these VEs allowed us to control the human factors we were interested in while considering changes to the VEs’ realism and FoV, allowing us to compare and contrast our findings with the previously reported PEA model.

Visual realism was adjusted as follows: for ‘high’ realism, ’Texture Quality’ was set to ’Full Res’ and ’Maximum LoD Level’ was set to 0, forcing the environment to use the highest LoD level for all objects, and for all textures to render at the highest quality natively available. By comparison, for ‘low’ realism, textures and shadows of a lower resolution were used, and polygon counts of 3D models were reduced compared to Jicol et al.’s original VEs. To achieve this, the quality settings within Unity were changed as follows: ‘Texture Quality’ was reduced from ‘Full Res’ to ‘Eighth Res’, decreasing the resolution of all textures to one-eighth; and the ‘Maximum LoD Level’ was increased from 0 to 3, forcing all objects to use the lowest level-of-detail (LoD) regardless of distance from the camera. All assets have several LoD profiles, ranging from the full quality high LoD object (stated by LoD level 0) to a low-quality low LoD object (stated by LoD level 3). These were either designed by the original asset creator or created within Unity. The ‘high’ FoV was 130°, which is the maximum afforded by the Valve Index. In order to maximise the study’s relevance to current popular HMD technology, the low FoV was set to 90°. This FoV is particularly relevant to investigate as it is the maximum FoV afforded by the Oculus Quest 2, a very popular HMD at the moment. To create the low FoV condition, a shader in the shape of a vertical rectangle was applied to the camera, blocking the user’s FoV outside of the rectangular area. This shape was created by altering the original circular 3D model using Blender’s 3D modelling tools. This was done as it more accurately represents the Oculus Quest 2’s FoV, while also ensuring that participants didn’t induce a higher risk of motion sickness due to restricting vertical FoV [115].

The VEs were designed by Jicol et al. [42] based on mood induction research [25, 111]. Two of the four VEs induce fear (one with agency and one without agency), while the other two happiness (again one with agency and one without agency). Participants were afforded six degrees of freedom, thus being able to look around or tilt. Users were instructed to remain roughly in the same position during the virtual experience so that they would not hit the VE boundaries and break presence. No virtual body of the user was implemented within the VEs to avoid confounding variables due to perceptions of body ownership [99]. The duration was also unchanged from Jicol et al. [42] as this duration had been suggested as optimal for presence before [117].

3.3 Measures

In order to facilitate comparability, we used the same questionnaires for assessing emotions, Agency and Presence as Jicol et al. [42], based on rating scales ranging from 1 (lowest) to 10 (highest). Emotions were measured using two items (“For the following emotions (happiness, fear), how intensely did you feel them during the VR experience?”), administered immediately after completing the VR experience. We will refer to reported levels of felt happiness and fear as Happiness and Fear , respectively. Moreover, Intensity in each VE will refer to the reported intensity of the specific emotion a VE is meant to elicit: it will be the reported Happiness scores in the VEs that were designed to induce happiness and the Fear scores in the VEs designed to induce fear. Perceived agency was measured using three items which were based on the “User Experience in Immersive Virtual Environments” questionnaire developed by [101].

We measured Presence with Tcha-Tokey et al.’s version [101] of the revised Presence Questionnaire (PQ) [113] initially developed by Witmer and Singer [114]. It should be noted that Witmer and Singer defined presence in a VE as not just the feeling of ‘being there’, but also as experiencing the VE as a real environment. This has led to several items in their questionnaire (and thus also in Tcha-Tokey et al.’s revised version) to be phrased around how realistic the perception of the VE is (e.g. “I could examine objects from multiple viewpoints.” and “The visual display quality distracted me from performing assigned tasks.”). This is different from other common presence measures such as the iGroup Presence Questionnaire (IPQ) [84], which solely focuses on the concept of ‘being there’. The PQ has been widely used for measuring presence in VR [31, 34] and has been correlated with other accepted measures such as the Slater–Usoh–Steed (SUS) [96] and the Multimodal Presence Scale (MPS) [59]. The revised version by Tcha-Tokey et al. is almost identical in its operationalisation of presence, but separates engagement into a separate measure.

3.4 Procedure

Participants were greeted by the experimenter and presented with an information sheet describing the study, before being asked for their consent to take part. They then completed a pre-task demographics questionnaire and were given a description of the type of VE they would experience, including whether or not they would have agency and whether the VE was designed to induce fear or happiness. This briefing was especially relevant for the fear condition as, in addition to the VE itself, fear can be elicited through a conceptual pathway by being told that there will be something scary [25, 36]. This may introduce a potentially uneven effect of priming on happiness and fear, however we opted for this method to create expectations about emotions and potentially further increase their intensity [85]. As such we did not aim for a similar intensities of emotion, and our later SEM analysis takes into account that intensity of emotions may not be felt equally. Participants were also told they would either have agency or not for purely instructional purposes. Those in the agency conditions were told about the interactive method they had at their disposal and the others were just told that they would observe the VE. We opted for such a text-based description so as to standardise the prior information given to each participant.

Participants were not told about manipulations in realism or FoV, and were not aware of conditions other than their own. Participants were assisted to put on the Valve Index HMD. They were then presented with a blank scene with text of different sizes and underwent a calibration phase during which the HMD was adjusted until they were able to read the smallest text presented. If assigned to one of the agency conditions, participants were shown how to use the hand controller to direct the laser pointer or the flashlight respectively.

At the end of the VR experience, a screen within the VE prompted participants to remove their HMD. Next, the emotion, presence and agency questionnaires were filled in on a normal screen in this order. While within-VR measures of presence have their merits, assessing presence with a traditional questionnaire after the VR experience has been shown to not affect obtained scores [74]. An open-ended question was also administered which was answered in text: “What elements/characteristics of the virtual environment made you feel present?”. The entire session lasted approximately 20-25 minutes. At the end, a debrief was given to participants and they were paid for their time.

3.5 Hypotheses

For clarity, we divide our hypotheses into sets that each focus on a particular aspect of our investigations. Each set of hypotheses provides a building block for the novel TAP-Fear model, which is assembled at the end. Most of our hypotheses are confirmatory, i.e. they are unambiguous a priori predictions that are based on the findings of previous work. Some of our hypotheses are speculative in the sense that they consider multiple possibilities, without a specific a priori prediction, as related work has not yet explored how technical and human factors work together. As discussed later, we test the speculative hypotheses using post hoc correction, which ensures that confirmatory and speculative results have similar validity.

Figure 2:

View Figure

3.6 Participants

A total of 360 participants (130 males, 230 females) were recruited amongst university students and staff members. Recruitment was done via online posts on the university noticeboard, posters and word of mouth. Participants were randomly assigned across the 16 conditions. Participants’ age ranged from 16 to 60, and had a mean of 23.847 and standard deviation of 10.09 years. We ensured that all participants had normal or corrected to normal vision and normal hearing. Due to the intensity of emotions elicited by the VEs all participants were also screened for neurological diseases, use of medication, psychological or emotional issues, epilepsy or use of medical devices before taking part in the study. However, no participant was excluded as all passed the screening questionnaires. Participants’ level of VR experience was assessed via a single item ("Please rate the amount of experience you have with virtual reality"), which was rated from "not at all" to "a great deal", on a scale from 1 to 10. Additionally, we ensured no participant suffered from cynophobia (fear of dogs), which was assessed through the same three items as used by Jicol et al. [42]. All participants were paid £5 in cash for their time. The study received ethical approval from the Department of Psychology Ethics Committee at the University of Bath (Ethics code: 21-233).

To ensure that we had sufficient statistical power, we conducted an a priori power analysis to calculate the necessary sample size per participant group. This was done for a between-factors ANOVA with main factors and interactions, through the widely used G*Power software (version 3.1) [24]. To estimate the sample size we used a partial eta-squared \(\eta _p^2\) of 0.06 (for a medium effect size), with a level of power of 0.80 for 16 groups, 1 numerator df (degree of freedom; for main factor 4 − 1 = 3, for interaction (4 − 1) × (4 − 1) = 9), and an α-level of 0.05. The analysis parameters were chosen so as to be similar to those of Jicol et al. [42]. The G*Power analysis indicated a minimum necessary overall sample size of 254, or approximately 16 participants per condition.

3.7 Statistical Methodology

Statistical analyses were performed using JASP 0.16.4.0 [56]. First we ensured the data satisfied the assumptions of ANOVA, using Levene’s test to check equality of variances and Q-Q plots to verify that distributions were close enough to normal. Then we conducted four-way ANOVAs using factors Emotion\(_\text{VE}\), Agency\(_\text{VE}\), Realism\(_\text{VE}\) and FoV\(_\text{VE}\) to test H1-H6 and H10-H15. For directed hypotheses, one-tailed tests were used, and otherwise two-tailed tests. Irrespective of our hypotheses, we always report all effects tested by an ANOVA to provide a complete picture. In tables, we report Bonferroni-Holm corrected p-values p\(_\text{BH}\) for speculative sets of hypotheses, highlight significant hypotheses and their p-values, and report effect sizes (η² for ANOVAs and standardised coefficients for regressions) for all effects. The error bars in the graphs show the 95% confidence intervals of the means.

Power analyses using GPower 3.1 show that the ANOVAs were able to detect medium, main and interaction effects (Cohen’s d = 0.5) at α = .05 with a power of 0.999. In other words, thanks to our sample size it is very likely that we will detect any effects that are of at least medium size, which is a common threshold for such analyses. For the four sets of speculative hypotheses H6A-D, H9A-D, H14A-D and H15A-D, we used the robust procedure discussed in [17]. We performed four related comparisons each time (A-D); so in order to control the Type 1 error rate α, we applied the Bonferroni-Holm post hoc correction. This correction ensures α = .05 despite the multiple comparisons by increasing p-values but it also decreases the power of some of the comparisons slightly to 0.998. This is still very high, given that a power of 0.80 or higher is generally considered good for a study. Thus, our sample size enabled us to effectively control both Type 1 and Type 2 errors even when testing multiple speculative hypotheses.

H7-H9 involve continuous predictor variables, therefore we applied linear regressions to test them, using the same robust procedure as described above for H9A-D to test multiple regression coefficients. In order to assemble our findings in a cohesive model, we constructed a structural equation model (SEM) as described in [68], using the SEM maximum likelihood estimator provided by the R package lavaan [80]. We then used accepted measures such as the root mean squared error of approximation (RMSEA) and the Comparative Fit Index (CFI) to evaluate model fit as discussed in [83].

4.1 Verifying and Extending the PEA Model

Table 1 shows the results of the four-way ANOVA of Emotion\(_\text{VE}\), Agency\(_\text{VE}\), Realism\(_\text{VE}\) and FoV\(_\text{VE}\) on Presence. The main effects of Emotion\(_\text{VE}\) and Agency\(_\text{VE}\) are both significant, and so are their interaction; therefore we accept H1-H3. The main effects of Realism\(_\text{VE}\) and FoV\(_\text{VE}\) are both not significant; therefore we reject H4&5. Of the set of speculative hypotheses investigating the interactions of Realism\(_\text{VE}\) and FoV\(_\text{VE}\) with Emotion\(_\text{VE}\) and Agency\(_\text{VE}\), which signify moderation effects on Presence, only Agency\(_\text{VE}\) × FoV\(_\text{VE}\) is significant; therefore we reject H6A-C and accept H6D.

4.2 Predicting Presence in the TAP-Fear Model

Table 2 shows the results of the linear regression analysis on Presence testing H7-H9. Perceived Fear and Agency are significant positive predictors of Presence, therefore we accept H7&8. By comparison, Intensity (r(358) = 0.027, p = .614) and Happiness (r(358) = −0.067, p = .202) do not correlate significantly with Presence. In line with our ANOVA results, Realism\(_\text{VE}\) and FoV\(_\text{VE}\) do not significantly predict Presence, which provides further support for rejecting H4&5. Of the set of speculative hypotheses investigating the interactions of Realism\(_\text{VE}\) and FoV\(_\text{VE}\) with perceived Fear and Agency, which signify moderation effects on Presence, only FoV\(_\text{VE}\) × Agency is significant; therefore we reject H9A-C and accept H9D. This is in line with the rejection of H6A-C and the acceptance of H6D, which consider similar interactions with design variables Emotion\(_\text{VE}\) and Agency\(_\text{VE}\) rather than perceived Fear and Agency.

4.3 Predicting Fear in the TAP-Fear Model

Table 3 shows the results of the four-way ANOVA of Emotion\(_\text{VE}\), Agency\(_\text{VE}\), Realism\(_\text{VE}\) and FoV\(_\text{VE}\) on perceived Fear. The main effect of Emotion\(_\text{VE}\) is significant, therefore we accept H10. The interaction between Emotion\(_\text{VE}\) and Agency\(_\text{VE}\) is not significant, so we reject H11. Of the set of speculative hypotheses H12A-D investigating the main effects of Realism\(_\text{VE}\) and FoV\(_\text{VE}\), and their interactions with Emotion\(_\text{VE}\), only the interaction of Realism\(_\text{VE}\) with Emotion\(_\text{VE}\) is significant, so we reject H12A,B&D and accept H12C.

4.4 Predicting Agency in the TAP-Fear Model

Table 4 shows the results of the four-way ANOVA of Emotion\(_\text{VE}\), Agency\(_\text{VE}\), Realism\(_\text{VE}\) and FoV\(_\text{VE}\) on perceived Agency. The main effect of Agency\(_\text{VE}\) is significant, therefore we accept H13. The interaction between Emotion\(_\text{VE}\) and Agency\(_\text{VE}\) is significant, so we accept H14. Of the set of speculative hypotheses H15A-D investigating the main effects of Realism\(_\text{VE}\) and FoV\(_\text{VE}\), and their interactions with Agency\(_\text{VE}\), none of the effects is significant, so we reject H15A-D.

4.5 The TAP-Fear Structural Equation Model

Figure 3:

View Figure

Figure 3 shows the Technical Agency-Presence-Fear (TAP-Fear) structural equation model (SEM), which was constructed based on our accepted hypotheses. Boxes are variables and arrows are regressions, so that the diagram illustrates the flow of effects from technical and design variables (marked with the subscripts VE) at the top and left, to perceived Fear and Agency in the middle, and finally to Presence. In other words, similar to the PEA model, the TAP-Fear model can be used to predict perceived Fear and Agency from technical and design variables, and finally predict Presence. As the TAP-Fear model focuses on fear, we encode the emotion the VE is designed to induce using a dummy variable called Fear\(_\text{VE}\), with value 1 meaning the VE is designed to induce fear and 0 meaning the VE is designed to induce happiness. By contrast, the PEA model encodes Emotion\(_\text{VE}\) as 1 for happiness and -1 for fear. Our encoding of Agency\(_\text{VE}\) is the same as in the PEA model, with 1 meaning the user is afforded agency and 0 meaning she is not. Fear_VE predicts perceived Fear (H10), and Agency_VE predicts perceived Agency (H13). Additional effects of Fear_VE on Fear_VE are moderated by Realism _VE (H12C). Additional effects of Agency_VE on Agency are moderated by Fear_VE (H14). Presence is formed through perceived Fear (H7) and Agency (H8), as well as effects of perceived Agency that are moderated by FoV_VE (H9D).

We evaluate the TAP-Fear model by considering several fit measures, as shown in Table 5. We compare TAP-Fear to the PEA model from Jicol et al. [42], as well as three model variations called Fear, TAP-Fear2 and TAP-Fear3, which have fewer variables. Considering these reduced models makes sense because many fit measures reward models that are parsimonious, i.e. are able to describe data accurately while avoiding complexity, which is a desirable model property. Different fit measures penalise the complexity of a model in different ways, so several fit measures should be taken into account when comparing models [83].

We first compare the PEA model to the Fear model, which aims to do the same as the PEA model: predict emotions (Intensity for PEA and Fear for Fear), Agency and Presence based on the design variables Fear_VE, Agency_VE and their interaction. The Fear model is the TAP-Fear model shown in Figure 3 without the technical factors, i.e. without the grey boxes at the top. Table 5 shows that Fear is better at predicting emotions (Fear) and then Presence than PEA, as indicated by markedly higher R² values. Fear also has a lower RMSEA (lower is better) and higher CFI (higher is better), with RMSEA and CFI measures indicating that PEA has a ‘mediocre’ fit and Fear an ‘adequate’ fit. This suggests that Fear is an improvement on PEA. The AIC and BIC values can only be compared when two models predict the same variables, which is not the case as PEA predicts Intensity where Fear predicts Fear. If we included also Happiness as a predictor of Presence into the Fear model, then the R² value would increase marginally (0.450) while the CFI would drop drastically (0.824) due to increased model complexity, which is an indication that including Happiness would lead to model overfitting.

By including the technical factor Realism_VE × Fear_VE, the TAP-Fear model improves its prediction of Fear, as seen by the higher R². However, the prediction of Presence gets slightly worse. The added complexity (more variables) cause some fit measures (RMSEA and CFI) to get worse, while some get better (AIC and BIC, lower is better). All coefficients of Fear and TAP-Fear are significant (p ≤ .023) except for the one for Agency_VE on perceived Agency (p = .529), as suggested by the very small standardised coefficient β = −0.04. We therefore removed this effect (the dotted lines at the bottom left in the diagram), resulting in a model variation TAP-Fear2 with a prediction performance similar to TAP-Fear but better parsimony (lower AIC and BIC). We further reduce TAP-Fear2 by removing the smallest effect FoV_VE × Agency (β = 0.08, p = 0.023) shown in dashed lines at the top right. This results in a new model TAP-Fear3 with slightly improved prediction performance and RMSEA and CFI values that put it into the ‘adequate’ to ‘good’ category.

6.1 Limitations and Future Work

Despite the significant undertaking of systematically investigating all possible combinations of four factors with two meaningful levels each, the TAP-Fear model only offers a restricted view over the multitude of design factors and levels that are at play in VR experiences. Examples of elements not considered are VEs designed to induce emotions other than fear or happiness, and HMDs with more extreme technical capabilities. While caution needs to be employed when generalising the TAP-Fear model, incorporating other design factors and levels can be addressed in future work because our between-participants experimental design allows for extensions of the model without retesting the conditions employed in this study. This data is made publicly available for other researchers to use and provides the foundation for future models on presence.

Our VEs only depict one scenario in a single environment (park). This is a location that all users will have a level of familiarity with. It has been shown that perceived realism of a VE is determined in part by the extent to which said VE meets users’ expectations; these expectations are in turn grounded in their prior knowledge about the setting depicted in the VE [90]. Arguably, perceived realism may not have been affected as much if we had used a less-familiar VE. For example, a fantasy world could have been used, which would still have been able in principle to elicit sensations of realism [30]. Other VEs depicting a variety of settings, real and abstract, should thus be tested in the future.

In this study, we adopted a binary approach to affording agency, mainly to limit the number of levels for our experimental design and create clear design guidelines. However, agency is a continuous variable, and being able to manipulate it continuously would be useful for future work, e.g. to assess individual differences in how affordances of agency are perceived. This is increasingly relevant with the introduction of richer interaction techniques to consumer HMDs, such as controller-less hand tracking and foot trackers [13, 104]. In addition, our VE did not afford participants any form of locomotion, which was done to control for sensory input. This means that our model may not be entirely applicable to cases where users are able to navigate the VE. One can infer, however, that in such cases the level of perceived agency would be higher, and so could be presence. It is hard to predict whether the same interactions with emotions and FoV would be found with such increased agency.

Based on the technical factors we chose, the applicability of the TAP-Fear model is currently limited to consumer-grade VR hardware. The two FoV levels we used were chosen to maximise the applicability of our findings to current consumer-grade popular VR HMDs, and the lower FoV was chosen to match that of the Meta’s Quest 2, which is the most popular HMD. However, we acknowledge that there may be floor or ceiling effects present in our results, and that testing significantly worse or better HMDs may lead to new insights.In future, testing lower FoV may provide interesting insights for the design of low-cost HMDs, however HMD technology is fast evolving. We note that the high FoV we used is lower than some special-purpose HMDs, such as the Pimax Vision 8K X which benefits from up to 200°. As the presented models are extendable, future work should aim to understand how a wider range of FoV and realism affect presence and the quality of experience, which is especially relevant to more specialised hardware.

Our two VEs are two points on a continuum of visual realism that extends beyond the low and high levels tested here. This was exemplified by the qualitative data where some users remarked that the low graphical quality lowered their fear levels, even in the high realism conditions. The applicability of the TAP-Fear model is thus limited by the parameters it was tested with. Still, the level of graphical fidelity in our study went beyond what is currently computationally possible on current untethered VR HMDs, as it was designed to make full use of the relatively high-end consumer-grade PC that was used. In addition, these models raise questions about how presence is affected in other technologies, such as MR and AR, which are still lagging behind VR in terms of technical factors and where the concept of presence is different. Furthermore, other technical factors can be incorporated into the models, such as the frame rate of the HMD.

Finally, Tcha-Tokey’s presence measure [101] was chosen for its reliability as well as to allow for validation of and comparison with the PEA model by Jicol et al. [42]. However, this also meant that presence was opeartionalised as a unidimensional construct, as it is very commonly done in literature. Still, it has been argued that presence can be divided into three separate dimensions, namely spatial, social and self presence [51]. This is relevant to the TAP-Fear model as emotion and agency may disproportionately affect the three sub-components of presence. Thus, further work should refine our model to separate between them.

6.2 Impact

Our results clarify the intricate ways in which human and technical factors can interact in the formation of VR presence. From our TAP-Fear model it becomes apparent that designers cannot ignore the influence of human factors when developing VR experiences. More precisely, it is the technical factors that should be adapted according to the specific emotion and level of agency afforded to the user, given the importance of the latter. Ultimately, technical factors need to be optimised due to limitations in computational power and high component cost that still characterise VR HMDs. Our model provides a framework for doing such an optimisation while prioritising the user experience, when designing VEs meant to elicit fear in particular. TAP-Fear can be interpreted as a structured decision tree, whereby the purpose of a VR application determines its properties and those of the HMD delivering it. In cases where the dominant intended emotion is fear, game designers should prioritise the enhancement of visual realism. In such cases, HMDs with an FoV above the 90° threshold should be required only for experiences that afford users agency, such as interactive games or training applications.