Don’t worry, be active: how to facilitate the detection of errors in immersive virtual environments

Participants

Experimental sessions were conducted in a field setting during a science and technology exhibit at Trieste Next 2017 of a custom made IVR system used to support ship design review. Visitors and scholars served as participants according to a well-defined scheduling of the science and technology exhibit lasting two days (self-selected sampling). All had normal or corrected to normal vision and were naive as to the purpose of the experiment. None of them reported specific visual diseases relevant for our task, like ocular dominance, colour blindness, and severe forms of eyestrain. A total of 98 participants took part in the study (mean age = 14.42, SD = 2.76, range [12, 20]), subdivided in 10 small groups of variable sizes ranging from a minimum number of five to a maximum number of 12. The groups were recruited one after the other during the exhibit. Six groups were composed by middle-school students (n = 61), two groups by high-school students (n = 22) and two by undergraduate students (n = 15). A total of 10 active participants vs. 88 passive participants were thus recruited, being each group composed by only one active participant. The active observer was selected among those participants who were willing to wear HMD, reported some previous experience with 3D gaming and low levels of motion sickness. This was assessed by orally asking them about previous experience of motion sickness symptoms according to Kennedy et al. (1993): oculomotor (eyestrain, difficulty focusing, blurred vision, and headache), disorientation (dizziness, vertigo), and nausea (stomach awareness, increased salivation, and burping).

Apparatus and stimuli

The active participants’ head motions were tracked in real time using an Oculus Rift CV1 HMD (with a FOV of 93° × 104° and a frame rate of 90 fps) connected with a PC equipped with an Intel Core i7 7700K 4.20 GHz processor with 16 GB RAM and nVIDIA GeForce GTX 1080 Ti graphics card. The PC controlled the simultaneous active and passive visualization of our complex immersive virtual environment and sampled the tracker. The immersive virtual environment was rendered through an IVR system designed by Arsenal S.r.L and was updated in real time according to a combination of head and hand movements of the active observer (Jez et al., 2018). In particular, the translation/rotation of the point of view through the immersive virtual environment was supported by the combination of head movements and manual control through a 3D Connexion SpaceMouse Pro, with four Degrees of Freedom (DOF; i.e. transition in the directions of three directional axes—forwards/backwards, left/right, up/down—and rotation on the vertical axis). Accelerometers within the HMD together with two camera motion sensors converging with an angle of about 40° on the active observer’s position at rest (at a distance of about 1.5 m) were used to calculate the x, y, z coordinates of the observers viewpoint. These coordinates defined the head movements of the active observers and were used by our programme to update in real time the geometrical projection of the 3D graphic. A precise visualization of the virtual environment was achieved by carefully calibrating the 3D layout to the height and the inter-ocular distance of each active observer following the Oculus Rift procedure. The virtual reality experience generated by the HMD was paralleled on two large LCD screens (Samsung 50″ J5200 Full HD LED set at a screen resolution of 1,024 × 768 pixels) connected with the PC. The two LCDs simultaneously displayed a cyclopean view of the exact same dynamic optic information generated by the movements of the active observer within the IVR.

Figure 1 provides a bird eye view of the virtual environment (Fig. 1D) used for the hide and seek task together with the design errors implemented throughout the path (screenshots, in Figs. 1A and 1B, for AEs and PEs, respectively). Our virtual environment consisted of a custom-made digital mock-up of an L-shaped 3D ship corridor. The corridor was 2.8 m high, 2.4 m large, and was composed by two straight segments: a first 35 m long segment connecting the starting position to the corridor’s curve (Fig. 1D, horizontal part of the corridor), and a second 40 m segment connecting the corridor’s curve to the corridor’s end (Fig. 1D, vertical part of the corridor). The corridor included different features, like: floor texture (a green floor with lateral yellow stripes and horizontal arrows indicating the directions), long pipes on the ceiling (with different sizes and colours), doors (five along the wall and one louvered door in the middle of the corridor), security systems (such as fire extinguisher, warning lights), signals (e.g. ‘fire point’, exit), handrails, bollards, and several technical equipment. The screenshots of the eight design errors actually included in the digital mock-up of the corridor (as presented to the participants to the experiment during the response encoding phase) are shown in Figs. 1A and 1B, together with the screenshot used as a catch trial (Fig. 1C). The catch trial was a misplaced pipe holder that was not implemented as an actual error in the digital mock-up of the corridor though being visually consistent with the features of the corridor. Within our experimental design, valid screenshots (i.e. those depicting actual errors) were equally subdivided into the two types: AE and PE (Figs. 1A and 1B, respectively).

Affordance errors included (Fig. 1A) a door handle placed too high and in the same side of the hinges (AE1), a fire extinguisher placed too high (AE2), a door handle placed too low and in the same side of the hinges (AE3), and a door handle in the same side of the hinges (AE4). PE included (Fig. 1B) a misaligned bollard (PE1), blue handrail embedded within yellow handrails (PE2), a louvered door missing an half (PE3), and a misaligned handrail (PE4). Three independent judges categorized the eight errors into the two categories, with an inter-rater agreement of 94%. By combining our two types of valid screenshots (AE/PE) with the two viewing conditions (active/passive) we obtained four experimental conditions of a mixed factorial design.

As shown in Fig. 1D, the design errors depicted within the screenshots were distributed all along the length of the 3D ship corridor, at a distance from the starting point of 16 m (AE1), 25 m (PE1), 31.5 m (AE2), 40 m (PE2), 49.5 m (PE3), 52 m (AE3), 53 m (PE4), 63 m (AE4).

The spatial arrangement of the setting is schematized in Fig. 2A, with Fig. 2B showing a photograph of the actual exposition context taken during a training session on a digital mock-up different from the one used during the experiment. The active observer comfortably sat on a stool right in between the two LCD screens, whereas the passive observers were arranged in front of the LCD screens comfortably sat on stools at a viewing distance of about 100 cm (Fig. 2A). A 90 cm wide desk was interposed in between the passive groups of participants and the screens in order to control at the best the viewing distance (observers were required to sit on the stools posing the elbows on the margin of the desk). At the viewing distance of 100 cm the retinal size subtended by the 3D dynamic layout when presented in full screen was of about 57.9° horizontally and 34.5° vertically.

Procedure

The experimental procedure included six phases lasting overall 10 min.

First, participants were informed about the experimental setting and the instructions of the modified hide and seek task using a Power Point presentation. The task was described as a game, in which the best-performing participant would be awarded with a gadget. This procedure was also aimed to motivate each single participant to perform at his/her best. To this purpose, an experimenter first explained to the entire group of participants the role they were about to endorse in the game. Participants were then introduced with their spatial disposition depending on the group they belonged: passive participants (one in front of the LCD to the left of the active observer, the other in front of the LCD to the right of the active observer, as by Fig. 2A) or active participant (right in between the two groups of passive participants, Fig. 2A). They were then informed about the aim of the game: ‘to search, find and memorize for design errors they would have encountered during the exploration of an immersive virtual environment’. Participants were also informed that at the end of the exploration they would have been administered a brief serial presentation of screenshots. After each screenshot’s presentation they would have been asked to raise their hand if the screenshot reported an error they saw during the exploration phase of the game.

Second, participants were then assigned to an active (n = 10) or passive (n = 88) role.

In the third phase, the group of participants was familiarized with the immersive virtual environment they were going to view during the game. During this familiarization phase, all participants watched on the LCD screen a 1 min video clip showing the same 3D ship corridor they would have been exposed during the exploration phase of the game but without including design errors. Participants were instructed to carefully watch the video clip. In order to encourage subjects to be as accurate as possible in seeking design errors, the experimenter stressed that in the video clip design errors were absent. This information also provided the subjects with a hint on the design errors they would have successively encountered during the experimental phase.

Fourth, the active observer was briefly trained in the usage of the SpaceMouse to control the translation of his/her point of view. The training was performed keeping static the HMD and visualizing its view on the LCD screen. A digital mock-up different from the one used during the experimental phase was used for such a phase (see an example in Fig. 2B). The observer following the instruction of the experimenter was trained for about 1 min on the main commands of the SpaceMouse (rotation, lateral and back/forth viewpoint translation). During this phase, in order to similarly train the passive observers with the 3D complex structure of the environment, the passive observers were required to look at the same scene the active observer was looking at.

Fifth, the experimental phase took place lasting about 5 min (1 min of instruction + 4 min of game). Active and passive observers were first required to keep their positions as by instruction (Fig. 2). The experimenter then focused the observers’ attention on the experimental task and instructed them to memorize the errors they found throughout the exploration phase of the game without disclosing their identification. Participants were thus explicitly discouraged to use neither verbal (i.e. claims) nor non-verbal behaviour (i.e. gestures) as soon as they detected a design error. The exploration phase of the game then started: the active observer was required to begin moving freely along the 3D ship corridor (now including the eight design errors) in search for design violations. This phase was terminated after 4 min of exploration. On average, this duration led to travel along the corridor back and forth for about one time.

Finally, the sixth response-encoding phase took place. Participants were serially presented with nine screenshots: the eight design errors presented serially in the order they appeared along the pathway from the start position to the corridor’s end plus the catch trial presented in a random serial position. As by initial instructions, participants were then required to raise their hands after each screenshot if the screenshot reported an error that they found during the 4 min exploration phase of the game: the experimenter registered the number of raised hands as well as the participants who raised their hand. At last, participants were thanked, debriefed and awarded with a gadget. Importantly, participants, throughout the game, were never aware of the number and types of design errors included in the 3D ship corridor.

Results

We analysed the individual likelihood of detecting a design error as an index of performance in our modified hide and seek task following Knoblauch & Maloney (2012) by applying a generalized linear-mixed effect model (glmm) with a probit link function to the whole set of individual binary responses on valid screenshots (1 = error detected; 0 = error not detected). In particular, we used a glmm with the type of design error (AE vs. PE) and the viewing condition (Active vs. Passive) as fixed factors, with by subject and by error (i.e. the eight valid screenshots) random intercepts and by subject random slope for type of design error (Baayen, Davidson & Bates, 2008). Importantly, this type of analysis has been proved to be optimal for a research in which the distribution of responses and participants across conditions is inevitably unbalanced as in the case of our design in which the proportion of individual responses resulting from the active observers was of about 0.105. Two-tailed p-values were obtained from type 3 F-statistics with the denominator’s DOF calculated by subtracting from the total number of observations the model’s DOF minus one. As indices of effect size we reported the partial eta-square (η_p²), and the concordance correlation coefficient, rc. Following Vonesh, Chinchilli & Pu (1996; but see also Rigutti, Fantoni & Gerbino, 2015) such a latter index provides a reliable measure, in the −1 to 1 range, of the degree of agreement between observed values and values predicted by generalized linear-mixed effect models. Finally, we reported Cohen’s d as a standardized measure of significant difference between means.

Disregarding catch trials, the data analysis was based on the 100% of active participants’ individual responses to valid trials (n = 80 resulting from the combination of 10 active observers and eight errors), and 97% of passive participants’ responses (n = 680 out of the total of 704 responses, resulting from 85—three excluded from the total of 88—passive observers and eight errors).

We removed three passive observers from the analysis because of the application of two exclusion criteria both aimed at minimizing any possible biasing effect of social desirability intrinsic in our field setting:

1. relative the sample mean deviation, one passive participant achieved an individual error-detection proportion that deviates from the one of the active observer of the group he/she belonged more than ± 2.5 SD;

2. two passive participants provided a positive response (i.e. raised their hand) after the catch trial presentation.

According to the result of the a priori power analysis, the total sample size of 760 observations, resulting from the application of the exclusion criteria and entered onto the glmm model in order to estimate the global active vs. multi-passive advantage, resulted to be large enough to rely on reliable statistical conclusions. We performed the power analysis following Faul et al. (2009; see the section 5.1) and using the generalized linear regression tool of G* Power (3.1.9.2 version) for testing logistic regression models based on the z-score distributions for the proportion as those implied in our glmm model (with the probit as a link function). In particular, the required sample size resulting from the analysis was equal to 695, which was indeed smaller than our total sample size of 760 observations obtained from the total number of valid binary responses of our 95 participants in our repeated measures design (with eight trials—one valid screenshot each—for each participant). The required sample size was obtained as follows. Based on a pre-experimental planning of the size of our group sessions we estimated an average proportion of active participants within each group of about 0.11 (eight passive observers for every single active observer). We entered this ratio as the best estimate of the expected value of our binomial distribution of the binary encoding of the type of viewing condition variable (1 = active; 0 = passive). We assumed an odds ratio of about 1.86, resulting from an expected gain in the detection of error proportion of the active over the passive condition of about 0.15 with p(H0) = 0.5, a power of 0.80, and no correlation between our two predictors (i.e. the type of viewing and the type of error). Notably, the required sample size resulted to be still smaller (728) than our total sample size of observations when calculated entering into the analysis (as the average of the binomial distribution associated to our binary set of type of viewings) the actual ratio (0.105) that we observed from the number of active observations (80) and the total number of observations in our design (760).

The pattern of average error detection proportion as a function of the type of design error (x-axis) and viewing condition (colour coding as by the legend) shown in Fig. 3A is only in part consistent with our expectations. Expectation 2 (not 2b) indeed holds for AE but not for PE.

This was confirmed by the results of the glmm analysis (rc = 0.52, 95% CI [0.48, 0.55]) revealing an unexpected trend towards significance of the interaction between type of design error and viewing condition (F_{1, 751} = 3.90, p = 0.048, η_p² = 0.005, 95% CI [0.000, 0.017]). Such a trend was due to the fact that the likelihood of detecting AEs was larger for active (M = 0.70 ± 0.07, with M indicating average proportion of error detection plus/minus one standard error), rather than passive (M = 0.53 ± 0.02) observers (consistent with Expectation 2, not with Expectation 2b). This difference produced a marginally significant glmm estimated active vs. multi-passive error detection advantage of about (0.22 ± 0.23, z = 1.61, one-tailed p = 0.053, d = 0.34). Differently, the likelihood of detecting PEs was similar for active (M = 0.77 ± 0.07) vs. passive (M = 0.80 ± 0.02) conditions thus producing a non-significant active vs. multi-passive advantage (−0.03 ± 0.15; z = 0.43, one-tailed p = 0.332).

The glmm analysis further revealed that Expectation 1 was only in part validated by our pattern of data as revealing a main effect of the type of error (F_{1, 751} = 5.04, p = 0.025, η_p² = 0.007, 95% CI [0.000, 0.020]) with an overall glmm estimated facilitation for PE over AE of about 0.31 ± 0.17. However, this facilitation was moderated by the type of viewing, as qualified by the type of viewing × type of error interaction. The facilitation for PE over AE was indeed present only for passive, but not for active viewing conditions. For the passive viewing condition the likelihood of detecting an error was indeed smaller for AEs (M = 0.53 ± 0.02), rather than PEs (M = 0.80 ± 0.02) type of errors (z = −2.38, one-tailed p = 0.009, d = 0.60). No such a difference was found for the active viewing condition in which the likelihood of detecting an error was instead similar for AEs (M = 0.70 ± 0.07) and PEs (M = 0.77 ± 0.07) type of errors (z = 0.47, one-tailed p = 0.318).

A further post hoc glmm analysis was carried out on the distribution of the likelihood of errors detection amongst our two types of errors (Figs. 3B and 3C for affordance and perceptual, respectively). This analysis revealed the active vs. multi-passive advantage for AE type of error was mainly due to two out of the four AEs: AE3 and AE4 (Fig. 3B). We indeed found a significant active vs. multi-passive error detection advantage for AE3 (z = −2.10, one-tailed p = 0.018, d = 0.71) and AE4 (z = −1.84, one-tailed p = 0.033, d = 0.64), but not for AE1(z = −1.18, one-tailed p = 0.120) and AE2 (z = 0.92, one-tailed p = 0.178). No reliable differences were instead found for any one of the four PEs (PE1: z = −0.07, one-tailed p = 0.470; PE2: z = 1.24, one-tailed p = 0.107; PE3: z = 0.37, one-tailed p = 0.354; PE4: z = −0.21, one-tailed p = 0.415; Fig. 3C).

Discussion

In the current Experiment, we found a trend towards an active vs. multi-passive viewing advantage with AEs but not PEs. As we hypothesized, the failure to find a difference between active and multi-passive observers in PEs can be the result of an overall ceiling effect. According to Expectation 1, such a ceiling was more likely to occur on PEs rather than AEs, given that PEs was expected to be easier to be detected than AEs. Such a response ceiling is likely to be induced by the type of field setting used in the current experiment favouring positive (hand raised) rather than negative (hand not raised) responses. Positive responses were indeed more socially desirable than negative responses in our task, thus producing a social desirability bias. Given these mixed results, the field experimental setting and the heterogeneous sample that we employed in Experiment 1, we ran a second experiment using a more controlled setting possibly reducing social desirability bias.

Participants

A total of 100 undergraduates of the University of Trieste (mean age = 20.60, SD = 3.29, [18, 40] range, 77% female), all with normal or corrected to normal vision, participated in the experiment in return for course credits. According to our pre-experimental questioning (same as in Experiment 1) our participants did not suffer of any specific visual diseases relevant for our task (ocular dominance, colour blindness, and severe forms of eyestrain). Participants were subdivided in 17 groups with variable size, each of which ranged from three to eight. Groups were tested in individual sessions and were formed following the same procedure used in Experiment 1. A total of 14 groups were composed by first year bachelor psychology students (n = 91), and three groups each composed by three participants included second year master psychology students (n = 9). Participants were assigned to active (n = 17) and passive (n = 83) conditions following the same selection criteria used in Experiment 1.

Apparatus and stimuli

The experiment was conducted in a VR laboratory equipped with the same experimental apparatus of Experiment 1, with three major differences:

1. The stereo projection system was driven by an MSI laptop instead of a PC, equipped with an Intel Core i7 7820HK 2.90 GHz processor with 32 GB RAM and nVIDIA GeForce GTX 1070 graphics card;

2. the LCD screens used to simultaneously display the dynamic cyclopean view of the 3D ship corridor explored by the active observer, were smaller (22″ not 50″, Samsung S22E450M LED set at a screen resolution of 1,024 × 768);

3. the overall viewing distance of the passive observers was settled at about 43 cm (not 100 cm) so to equate the displays in term of retinal size.

In order to provide a method that corroborates the reliability of the errors reported by the active observers, the IVR system was aided by a functionality that allows to store in real time the screenshots of the cyclopean views of the active observer. A screenshot was stored during the exploration as soon as the active participant pressed the R button of the SpaceMouse.

The spatial disposition of participants within the setting, and the immersive virtual environment explored during the task (the 3D L-shaped ship corridor, Fig. 1D) together with the design errors (Figs. 1A and 1B) were the same as in Experiment 1. The screenshots used during the response-encoding phase were the same as in Experiment 1 (8 valid screenshots, 4 AEs + 4 PEs, plus one catch trial screenshot), with one additional screenshot used as catch trial. This additional screenshot displayed a missing holder in a ship thruster (Fig. 1C, screenshot 2). This catch trial screenshot was apparently different from all of the other screenshots for colour properties, being reddish not greenish. Such a difference in colour purposely magnifies the un-relatedness of some trials with the 3D ship corridor thus reducing the tendency to favour positive over negative responses in our task. In Experiment 2, the same 2 × 2 mixed factorial design of Experiment 1 applied, with four experimental conditions resulting by the full factorial combination of two types of design errors × 2 viewing conditions.

Procedure

The procedure was similar to the one used in Experiment 1 with the major difference that the ludic dimension was now intentionally removed from the task so to reduce the social desirability bias. The procedure included the same six phases described in subsection 2.1.3 and lasted almost the same time (10 min). Major differences regarded:

1. the first phase, in which task was described as a game, was omitted from the current instruction;

2. in the second phase, in which, although following the same selection criteria used in Experiment 1, albeit a relatively larger number of participants were assigned to the active (n = 17) and the passive (n = 83) conditions;

3. the fifth experimental phase, that although requiring the participants to perform the same modified hide and seek task of Experiment 1 (lasting 4 min), it included an additional task for the active participant (to push the R button of the SpaceMouse as soon as he/she detects a design error during the exploration phase);

4. the sixth response encoding phase that was conducted following the exact same procedure as in Experiment 1, but that included the serial presentation of 10 (not 9) screenshots.

At last, we collected demographic information (age and gender), and participants were thanked and debriefed following the same procedure used in Experiment 1.

The debriefing in Experiment 2 also served the purpose of gathering information from those active observers that during the exploration phase did take a number of screenshots that differs from those signalled by the raising of their hand. A total of 10 observers out of the total of 17 took a screenshot of all of the errors they reported during the response-encoding phase. The remaining seven observers either took no screenshot at all (n = 1), or took one (n = 5) to three (n = 1) screenshot less than the number of errors they reported during the response encoding phase. The average number of correspondence between screenshots and errors reported by the raising of hands was of about 82%. This showed that our response method was reliable. In the debriefing phase, all seven participants with a smaller correspondence between the screenshots they took and the errors they reported claimed that they might have: (1) forgotten about pushing the SpaceMouse button during the exploration phase; (2) pushed the SpaceMouse button accidentally or to signal relevant parts of the environment in addition to mere errors. In line with this latest claim, the analysis of the screenshots reveals that observers in general used the SpaceMouse button not only to signal a detection error but also to signal a salient part of the 3D environment. The average individual proportion of screenshots relative to the total number of actual errors (n = 8) was indeed reliably larger than the average individual error detection proportion (0.72 ± 0.09 vs. 0.48 ± 0.05, t = 3.91, df = 16, p = 0.001, d = 1.95). Consequently, we decided not to use the screenshots as an additional exclusion criterion for our participants, given that our result suggested that they were used for a mixture of purposes (accidental button press, to signal a detection error, or to signal a relevant part of the environment).

Results

As in Experiment 1, we analysed the individual likelihood of detecting a design error as an index of performance in our modified hide and seek task by applying the same glmm model (with probit as a link function) to the whole set of individual binary responses on valid screenshots (1 = error detected; 0 = error not detected), with type of design error (AE vs. PE) and viewing condition (Active vs. Passive) as fixed factors, and with by subject and by error random intercepts and by subject random slope for type of design error. We used the same indices of effect size as in Experiment 1 in order to support the reliability of our statistical effects. Disregarding catch trials, the data analysis was now based on the 100% of active participant’s responses to valid trials (n = 136 resulting from the combination of 17 active observers and eight errors), and on the 96% of passive participant’s responses (n = 640 out of the total of 664 responses, resulting from 80—three excluded from the total of 83—passive observers and eight errors). After the application of the same exclusion criteria used in Experiment 1, we removed from the analysis three passive observers who raised their hand when one of the two catch trials was presented.

Performing the same a priori power analysis of Experiment 1 (assuming the same power, and odds ratio), our total sample size of 776 observations entered onto the glmm model resulted to be large enough to rely on reliable statistical conclusions. The sample size was indeed larger than the one used in Experiment 1, and more balanced across viewing conditions, with the planned proportion of active participants within each group raised to 0.2, which resulted from planned groups that on average were formed by one active participant out of four passive participants. The increased balance over the two viewing conditions delivered a required sample size of 427, that was much smaller than the observed one. Notably, the required sample size was still smaller (473) than the one we used also entering into the analysis (as the average of the binomial distribution associated to our binary set of type of viewings) the actual ratio (0.175) that we observed from the number of active observations (136) and the total number of observations in our design.

Figure 4A depicts the pattern of average error detection proportion as a function of the type of design error (x-axis) and viewing condition (colour coding as by the legend). The pattern is apparently different from the one of Fig. 3A (Experiment 1), being now consistent with both Expectation 1 (main effect of the type of design error), and Expectation 2 (main effect of viewing condition). Furthermore, differently from Experiment 1 the type of design error × viewing condition is now absent.

These observations were confirmed by the statistical glmm analyses (rc = 0.58, 95% CI [0.54, 0.61]). Indeed, as consistent with Expectation 2 (not 2b), glmm results showed a significant main effect of the viewing condition (F_{1, 767} = 5.02, p = 0.025, η_p² = 0.007, 95% CI [0.000, 0.019]), with a larger likelihood of detecting an error in active (M = 0.48 ± 0.04), rather than in passive viewing condition (M = 0.38 ± 0.02). This effect was quantified by an active vs. multi-passive error detection advantage of about 0.15 ± 0.13 (z = 2.15, one-tailed p = 0.016, d = 0.20). Expectation 1 was supported by the significant main effect of type of design error (F_{1, 767} = 8.23, p = 0.004, η_p² = 0.011, 95% CI [0.002, 0.026]), with a larger likelihood of detecting an error for PE (M = 0.56 ± 0.03), rather than for AE (M = 0.23 ± 0.02). Unlike Experiment 1, the interaction between type of design error and viewing condition was not statistically significant (F_{1, 767} = 0.12, p = 0.727, η_p² = 0.000, 95% CI [0.000, 0.005]).

As in Experiment 1 we further performed a post hoc glmm analyses on the way in which the likelihood of errors detection was distributed amongst our two types of errors (Figs. 4B and 4C for affordance and perceptual, respectively). This analysis revealed that the origin of the active vs. multi-passive advantage was likely to be qualified by two out of four AEs (Fig. 4B, AE1, z = −1.52, one-tailed p = 0.064, d = 0.38; and AE2, z = −1,43, one-tailed p = 0.077, d = 0.38), and two out of four PEs (Fig. 4C, PE1, z = −1.45, one-tailed p = 0.073, d = 0.37; and PE3, z = −1.39, one-tailed p = 0.083, d = 0.39), all of which showed a trend towards significance concerning the active vs. passive error detection advantage.

Discussion

Experiment 2 revealed that both the active vs. multi-passive viewing advantage predicted by Expectation 2 (not 2b), and the effect of facilitation due to the type of error predicted by Expectation 1, did occur under a laboratory setting. Joining data from both Experiments revealed an overall reduction of the tendency to provide a positive response in Experiment 2 relative to Experiment 1. The likelihood of detecting an error strongly decreased in the laboratory setting of Experiment 2 (M = 0.40 ± 0.02), relative to the field setting of Experiment 1 (M = 0.67 ± 0.02). Such a difference was confirmed by the significant main effect of the Experiment (F_{1, 1523} = 92.53, p < 0.001, η_p² = 0.057, 95% CI [0.040, 0.077]), when it was included as an additional fixed factor in the glmm analysis. The direction of this effect might be accounted for by social desirability, which more likely biases responses towards positive detection in field rather than in laboratory settings.