The sample included 37 university student participants who all made eight eyewitness identifications each. All were recruited through email or via social media. From the 296 identifications, we excluded 28 observations because the participant recognized one or more of the images used in the line-ups. In other words, we had included a post line-up question where we asked participants if they recognized any of the faces prior to the experiment and we used this question to exclude line-up responses where participants had recognized one of the line-up images used (i.e., a familiarity control question).

Thus, the final sample included 268 observations from 37 participants. Nineteen of the participants reported their nationality as Finnish and 18 reported that they were from other countries. There were 21 male participants (mean age = 25.6, SD = 4.7) and 16 female participants (mean age = 23.5, SD = 8.6).

All aspects of the proposed research project were conducted in accordance with the Declaration of Helsinki. That is, participants were informed of the voluntary nature of the studies, and of their right to terminate their participation in the research at any stage, without giving a reason. Written and full informed consent was obtained from all participants. Research permits were sought and approved from IRB of the Department of Psychology at Åbo Akademi University.

We filmed eight separate targets (i.e., perpetrators) that acted in all eight conditions of our factorial design. The total duration of each of the scenes was approximately 10–15 seconds. The threatening condition included a target that stole a money-collection jar while threatening the salesclerk with a knife over a table/counter. In the non-threatening condition, the target placed money in the money-collection jar while holding a booklet in the same hand as the knife was in the threatening scene. In the first-person perspective, the point of view was that of the victim, and in the third person perspective, the point of view was that of an onlooker of the two people by the counter (i.e., the victim and the perpetrator). For still images from the video scenes used in the threatening (non-threatening) and first-person (third-person) scenarios, see S1 Appendix. The conditions were filmed using a Canon camera (for the 2D videos) and a Ricoh Theta Z 360-degree camera (for the 360-degree VR videos). During the experiment, the 2D videos were presented on a desktop computer monitor (Dell U3014, screen: 75.6 centimeters; cm (30 inches; in), resolution: 2560 x 1600 pixels; px) and the 360-degree videos were displayed in virtual reality using an HMD. The HMD was an Open Source Virtual Reality (OSVR) HDK 1 (Screen size: 1920x1080 px; height 12.2 cm (4.8 in.), width: 6.8 cm (2.7 in., per eye 1080 x 960 px, width 6.8 cm (2.7 in.), height 61mm (2.4 in.), with a field of view of 100 degrees and a refresh rate of 60 Hz).

The position of the 360-degree camera when recording the first-person perspective was 1 meter from the perpetrator and the camera lens was at a height of 1.5 meters from the floor. Since the 360-degree lens is a fisheye it warps distance to a certain extent (see for example the still images in S1 Appendix) so we had to move the Canon camera to a position of 4 meters from the perpetrator (same height) in order to create 2D recordings where the perspective was as similar as possible to the 360-degree recordings. This was done through trial and error prior to recording all the video scenes, we tested the perspective that most closely matches one another when viewing the scene on a computer screen and when viewing the scene in VR. When we filmed the third-person perspective, we placed the 360-degree camera at a diagonal angle 1.2 meters relative to the target/perpetrator, and we placed the Canon camera at 4.8 meters to the side of the target/perpetrator. The static images of the 360-degree video (see S1 Appendix) do not do justice to the point of view when using the HMD and experiencing the scene in VR; with the HMD on, the 360-degree perspective is much more similar (i.e., the perspective is closer) to the 2D perspective than what it appears from the still images. Furthermore, although the video quality is slightly lower in the 360-video compared with the 2D video, the distance to the face is small and, in all videos, it was easy to discern details of the face of the target/perpetrator, so we considered the difference to be negligible in relation to the line-up task.

The targets that were included in the video scenes were native Finnish university students. There were four female targets and four male targets, whose average age was 23.5 years (SD = 4.7), average was height 173.9 cm (SD = 8.2), and average was weight 68.3 kg (SD = 9.5).

We used a Lenovo Tab 2 A10-70 10.1" Android tablet in order to present line-ups and collect participant responses. At the start of each session, participants were asked to fill in demographic information (i.e., nationality, age, and own height) and prior to each line-up they were asked to answer five pre-programmed questions regarding the target’s appearance (not relevant to or analyzed in the present paper), followed by a line-up where identification or rejection of the line-up was achieved by selecting an image or selecting the option to reject the line-up. The tablet also recorded post-line-up questions regarding the participant’s certainty regarding their line-up choice and questions regarding their familiarity with any of the images used in the line-ups (for the sake of exclusion).

The line-up task was a computerized six-person photograph line-up that was randomized (per trial) to be either target present or target absent. The target present line-up included one target image and five filler images, and the target absent line-up included one foil image (i.e., target substitute) and the same five filler images as in the target present line-up. Identifications and rejections were coded as binary outcomes (0/1) by the software application. Identification accuracy was coded as “1” for target present (TP) target identifications, “0” for TP foil identifications, and “0” for TP rejections. Rejection accuracy was coded as “1” for target absent (TA) rejections and “0” for foil/innocent identifications. Confidence was registered on a continuous scale between 0–100% for confidence and response time in milliseconds. Descriptive estimates of the targets were registered as follows: Height (0–220 cm), Target Weight (0–150 kg), Target Age (5–99), Target Gender (Man or Woman), and Distance to target (0–200 meters).

The experimental procedure entailed 1) reading and signing an informed consent, 2) watching eight pre-recorded videos (each video was approximately 10–15 seconds), 3) answering questions and conducting a line-up tasks on an Android tablet. We pseudo-randomized the order of the conditions in our design, so that there were eight possible sequences of our eight conditions. Each participant was allocated to one of the eight sequences, which included eight separate videos and each of the videos included one of the eight targets. This meant that every participant saw eight scenes and conducted eight line-up tasks where one of the eight targets was presented (no one saw the same target twice). The participants either first saw four VR scenes and then four 2D scenes or vice versa (depending on the sequence that they had been placed into). In VR, the participants saw two threatening and two non-threatening scenes, where two were from a first-person perspective and two were from a third-person perspective (the same applied to the 2D scenes). The threatening and non-threatening and the first- and third-person perspective scenes were counterbalanced within the four VR and 2D scenes. After watching each video scene, participants were asked to estimate, by answering questions on a tablet, the target’s age, gender, height, and weight, as well as the distance to the target. This was then followed by a line-up identification task. Lastly participants were asked to make a confidence judgment (i.e., the self-perceived certainty with which decisions were made) and to state if any of the images used in the line-up were familiar from prior to the experiment.

The present experiment was a within-subject factorial design with three factors: two (2D or VR) by two (threatening or non-threatening) by two (first- or third-person).

The main analyses were conducted using the R platform [24] and by fitting data as a multi-level logistic regression in the lme4 package [25]. The general formula used was: Outcome∼Predictor(s)+(1|subject)+(1|Target) In lme4 the categorical variables are by default contrast coded and the outcome of the binary logistic regression analysis is a value of log odds [26]. This value is used to assess the likelihood of a response being 1, using the probit function (exp(x)/(1+exp(x)), where x is the log odds estimate of interest. The multilevel model (also known as random effects model), is an appropriate approach of analyzing the current data as it takes into account the hierarchical nature of our repeated measures design and includes the within and between subject variance as part of the statistical model. Due to the limited number of observations per cell (in the experimental design), we were not able to run full factorial analyses, using all three predictors and their interactions in the same model. For this reason, we analyzed the results sequentially, by first looking at main effects and then their interactions (including planned comparisons). This resulted in six separate analyses for TP identification accuracy and six separate analyses for TA rejection accuracy.

Based on the log odds from the logistic regression we also report the odds ratios (OR) and their confidence intervals for the analysis of main effects. Odds ratios are a common way of reporting effect sizes [27]. We also ran six separate post hoc power analyses for each of the analysis of main effects (i.e., three for the TP and three for the TA analyses). For each of these analyses, we used OR = 1.68 as an effect size of interest (comparable to Cohen’s d = .2) [27]. We did this in order to estimate the power to detect a small effect in either TP or TA line-ups when investigating main effects. Here, we used the powerSIM function that is part of the simr package in r [28] in order to calculate the power (1-β) and confidence intervals (CI), using 200 iterations, to detect a small effect (i.e., OR = 1.68). As can be seen in the results section, the power for each main effects analysis was low, which increases the risk of a type II error in each case, but as this was a proof of concept design, we argue that the results are still of interest for future studies.

In order to conduct an exploratory analysis addressing also the issue of confidence, we created two new confidence-accuracy variables for TP and TA line-ups. Here, we recoded the binary outcome variable 0/1 as -1/1 and multiplied by the post line-up confidence rating and divided by 10. This gave us a continuous variable where 0 to -10 represented the low to high confidence ratings for incorrect decisions in TP line-ups (i.e., filler selections or rejections) and TA lineups (i.e., filler selections). Correspondingly, the values 0 to 10 represented low to high confidence ratings for correct decisions in TP line-ups (i.e., target selections) and TA lineups (i.e., line-up rejections). This provided us with a more sensitive measure combining confidence with accuracy.

The data and R markdown script used to analyze that data can be found at the Open Science Framework at the following address: https://osf.io/gz3wy/

Out of the total 268 eyewitness responses, there were 138 TP decisions (VR: 71, 2D: 67), 130 TA decisions (VR: 61, 2D: 69); see Table 1. See also Table 2 for the average post line-up self-reported confidence and the average line-up response times.

We found an effect of our VR and 2D manipulation on TP identification accuracy (see Fig 1A), where 2D resulted in more correct identifications compared to VR (B = 1.63, SE = 0.51, p = .001, OR = 5.12 (95% CI: 1.88–13.91), 1-β = 26.00% (95% CI: 20.07–32.66). We found no effect (see Fig 1B) of our perspective manipulation on TP identification accuracy, where a third-person perspective did not differ significantly from a first-person perspective (B = -0.78, SE = 0.43, p = .072, OR = 0.46 (95% CI: 0.19–1.07), 1-β = 19.00% (95% CI: 13.81–25.13). We also found an effect of threatening vs. non-threatening scenes on TP identification accuracy (see Fig 1C), with non-threatening scenes resulting in less correct identifications compared to threatening scenes (B = -0.96, SE = 0.45, p = .032, OR = 0.38 (95% CI: 0.16–0.92), 1-β = 16.50% (95% CI: 11.64–22.38). The results do not support our hypothesis that VR (compared to 2D) results in a higher degree of identification accuracy; rather the results suggest that the reverse is true. We also did not find support for our hypothesis that a first-person (vs. third-person) perspective gives rise to a higher degree of identification accuracy. However, we found support for our hypothesis that threatening (vs. non-threatening) scenes leads to higher identification accuracy.

Investigating interactive effects, we found no significant interaction (see Fig 1D) between the 2D vs. VR and first- vs. third-person perspective on TP identification accuracy (χ² [5] = 0.91, p = .339). Planned comparisons revealed no significant differences between the perspectives within 2D or VR (see Fig 1D for exploratory difference between conditions). The results go against our hypotheses that the difference between the first- and third-person perspectives would be more pronounced in VR compared with 2D. We found no significant interaction (see Fig 1E) between 2D and VR and threatening and non-threatening on TP identification accuracy (χ² [5] = 2.56, p = .110). Planned comparison, however, revealed a significant difference between the threatening and non-threatening conditions in VR, where the threatening conditions resulted in higher accuracy compared with the non-threatening condition (see also Fig 1E for the exploratory differences between conditions). This falls in line with our hypotheses that threats in VR are more realistic and that threatening scenes enhance accuracy. However, as can be seen in Fig 1E, there appears to have been a ceiling effect of accuracy in both the threatening and non-threatening condition in 2D. It is, therefore, unclear if the lack of difference in 2D was due to the ceiling effect or the actual condition. When we investigated the interaction between first- and third-person perspectives and threatening and non-threatening manipulations on TP identification accuracy. We found no significant interaction between the manipulations (χ² [5] = 0.00 p = 1.000) and planned comparisons revealed no significant differences (see Fig 1F). The results indicate that there is no difference between threatening and non-threatening scenes seen from a first- or third person perspective, which goes against our hypotheses. However, this may also be due to overall ceiling effects in the study; indicating that the task was perhaps too easy. Lastly, using the variable combining accuracy and confidence information, we found that the results exactly mirrored the analyses using the dichotomous outcome variable in terms of significant effects.

We found no effect (see Fig 2A) of our VR and 2D manipulation on TA rejection accuracy, where 2D did not significantly differ from VR (B = -0.49, SE = 0.42, p = .247, OR = 0.61 (95% CI: 0.27–1.40), 1-β = 18.00% (95% CI:12.94–24.04). Nor was there an effect (see Fig 2B) of our first- and third-person perspective manipulation on TA rejection accuracy, so that the third-person perspective did not differ significantly from the first-person perspective (B = -0.50, SE = 0.43, p = .248, OR = 0.60 (95% CI: 0.26–1.42), 1-β = 23.00% (95% CI: 17.36–29.46). We found no effect (see Fig 2C) of our threatening and non-threatening manipulation on TA rejection accuracy, where non-threatening scenes did not differ significantly from threatening scenes (B = -0.28, SE = 0.42, p = .506, OR = 0.76 (95% CI: 0.34–1.71), 1-β = 19.50% (95% CI: 14.25–25.68). None of the results support our hypotheses.

Investigating the interactive effects, we found no significant interaction between 2D and VR and first- and third-person perspective on rejection accuracy (χ² [5] = 0.49 p = 0.486) and no significant interaction between the 2D and VR and threatening and non-threatening manipulations on rejection accuracy (χ² [5] = 0.00 p = 0.992). There was also no significant interaction between first- and third-person perspectives and threatening and non-threatening manipulations on rejection accuracy (χ² [5] = 0.15 p = 0.696). As can be seen in (Fig 2D–2F), planned comparisons did not reveal any significant differences within groups. The results illustrate that we were unable to find support for any of our hypotheses in relation to TA rejection accuracy and this is in contrast to the findings of TP identification accuracy. Lastly, using the variable combining accuracy and confidence information, we again found that the results exactly mirrored the analyses using the dichotomous outcomes in that there were no significant effects.

Due to the complexity of our design and the small sample size there were a relatively low number of observations per cell (in the factorial design), which did not allow for an analysis of the complete design, which might have revealed interesting interactive effects. Clearly there is need for further investigation with and increased sample size because low statistical power means that it is difficult to interpret both the effects found and the lack of effects. It also appears that the task was too easy, which resulted in a ceiling effects and this also limits the interpretation of the results. However, accuracy rates vary greatly between studies depending on the stimulus material used, such as, live, video, photo material, and the line-up type used, such as, simultaneous or sequential line-ups [29]. In other words, although accuracy was high it was not outside the range of what has been seen in earlier experimental eyewitness studies. A more difficult task would most likely be more adequate when investigating a similar design. Moreover, we did not include objective measures of stress and that limits the interpretation of the current design, which future studies should aim to include.

The practical implication of this research is that it was an attempt to investigate the potential of using 360-degree videos and VR as a more ecologically valid method compared to using standard 2D monitors. The generalizability of our results is limited by the parameters set by this small proof of concept study; however, the results indicate that VR might be better suited to investigate first- vs. third-person perspectives. There are also many more questions that can be combined with the first- and third-person perspective that are crucial to eyewitness research. For example, in eyewitness research it is generally accepted that stress negatively impacts eyewitness memory and identification accuracy [7], yet this is contrary to basic memory research that states that encoding is improved by stressful compared to neutral events [14]. The connection between stress and memory has been further elaborated on in a recent meta-analysis by Shields and colleagues (2017), where it was found that stress during encoding improves memory when a stimulus is closely associated with a stressor. Moreover, a recent eyewitness study by Sauerland and colleagues (2016) concluded that one of the possible reasons for the differences in findings regarding memory and stress between research paradigms is that eyewitness research has relied so heavily on laboratory-based 2D video stimuli that give rise to only small elevations in stress. The implications are that 2D video is lacking in both realness and immersion and this means that 2D video fails to elicit emotional reactions that are comparable to real life, which in turn leads to small effects. Sauerland and colleagues [16] also state that very few eyewitness studies (in contrast to neurological studies) have investigated stress with objective measures, relying instead on subjective assessments. This further complicates the comparison between 2D and real life, leaving a gap in the field that has not been sufficiently investigated. Future research should both include objective physiological measure of stress in threatening scenarios and investigate the potential of using 360-degree videos in VR to achieve more realistic events that can elicit reactions that may be more comparable to real life. Future research should also compare 2D, VR, and real-life using objective physiological measures to estimate possible differences between methods. The inclusion of a manipulation check of the sense of presence and perceived realism would also be of use when comparing 2D, VR, and real-life.