An a priori power analysis was calculated using the SPSS (28, IBM Corp., Armonk, NY) MANOVA procedure (54), which requires the number of participants for each group, means and standard deviations for the dependent variables, and correlations between the variables. Values were taken from the ordered pretest of the study by Schütz and Schack (2), who found hysteresis in a sequential drawer opening task. In their study, posture was measured as hand angle projected onto the drawer surface. Thus, the dependent variable was comparable to the projected shoulder, thorax, and pelvis angles used in the current study. Presuming a comparable, that is, large, effect size in the current study, 15 participants should be sufficient to achieve a power of over.95 for detecting a main effect of a within subject factor “order”, indicating hysteresis.

In Experiment 1 (real environment), 27 participants [15 female, 12 male, age 24.3 ± 2.5 (SD) years] were measured. Twenty-five participants were right handed [handedness score (HS): 0.98 ± 0.06] and two left handed (HS: −1.00 ± 0.00) according to the revised Edinburgh inventory (55). In Experiment 2 (virtual environment), 21 participants (12 female, 9 male, age 23.3 ± 3.0 years) were measured. Eighteen participants were right handed [handedness score (HS): 0.96 ± 0.12] and three left handed (HS: −1.00 ± 0.00). All participants were recruited at Bielefeld University and participated in exchange for either course credit or a participation fee (10 €). They reported no known neuromuscular disorders and were naïve to the purpose of the study. Before the experiment, each participant read a detailed description of the study, a data protection declaration, and provided written informed consent. The study was approved by the Bielefeld University ethics committee (proposal EUB-2022-200) and conformed to the 1964 Declaration of Helsinki (56).

In Experiment 1 (real environment), the setup consisted of two slalom poles (foot: 3 cm high, 12 cm diameter; pole: 180 cm high, 3 cm diameter) that were placed at the halfway point of a 5 m walkway (Figure 1A). Two blue rubber disks (30 cm diameter) on the floor marked the start and end position at each end of the walkway. The slalom poles created an aperture that could be varied in 5 cm steps from 35 cm (width 1) to 75 cm (width 9), thus creating nine different aperture widths through which participants had to pass.

The placement of the poles and the aperture width were controlled via two parallel strips of yellow adhesive tape that were attached 12 cm apart (Figure 1B). The slalom poles were placed between these strips and, for each pole, two white marks on either side of the foot had to be aligned with two black marks on the strips. For each pole, nine black marks on each side marked where the pole had to be positioned to create the correct aperture width.

In Experiment 2 (virtual environment), the setup was faithfully reproduced in the Unity game engine (2022.3.48f1, Unity Technologies, San Francisco, CA). Figure 1A depicts a render of the virtual scene used in the experiment. The virtual environment was presented with a Valve Index virtual reality headset (Valve Corporation, Bellevue, WA) at a resolution of 1,440 × 1,600 pixels with a 120 Hz refresh rate.

A rigid body with four retro reflective markers was attached to the headset and tracked with a Vicon Shogun motion capture system (v1.12.0, Vicon Motion Systems, Oxford, UK) with 16 Vicon Vero v2.2 cameras at a 100 Hz recording frequency. In the Shogun software, a rigid segment (or prop) was created for the headset, with its origin on the bridge of the nose and its x-axis pointing right, y-axis upward, and z-axis backward. Motion data of the headset prop was then streamed to the Vicon client (part of the Vicon Unity plugin, v1.3.1) in Unity and processed by the HMDScript component to fuse the position data from Shogun with the inertial measurement data from the Valve Index. Thus, the position and orientation of the virtual environment were perfectly aligned with the real environment.

Four character models of different sizes were created for both females (150, 160, 170, 180 cm) and males (160, 170, 180, 190 cm) in MakeHuman (v1.3.0, the MakeHuman team) and attached to a custom skeleton rig in Blender (v4.1, Blender foundation, Amsterdam, Netherlands) that matched the skeleton rig (i.e., the bone names and bone hierarchy) of Vicon Shogun. In the Shogun software, a skeleton (or subject) was created for each participant, based on the default FrontWaist3Fingers marker set with 59 retroreflective markers. Motion data of the subject (position and orientation of each bone) were streamed to the Vicon client in Unity and processed by a custom-written C# script to map the motion data onto the character model. The custom-written script mapped both the orientations and the positions (i.e., lengths) of the bones and, thus, faithfully reproduced the body dimensions of the participant in the virtual environment. The script further added a rigid body and a capsule collider to each bone.

Rigid bodies and capsule colliders were also added to the virtual slalom poles to ensure that the participants could actually knock them over if they did not rotate their bodies sufficiently when passing through the aperture. The start and end position were updated to change their color from red to green when the participants (i.e., the headset) were directly above them, to provide additional visual feedback in the virtual environment.

In both experiments, participants first read the study description and data protection declaration and provided written informed consent. Next, they changed into tight-fitting clothing and sneakers. Their shoulder width (between the left/right great tubercle) and hip width (between the left/right greater trochanter) were measured with a body caliper (WZ101, Uniprox GmbH, Zeulenroda, Germany).

In Experiment 1, ten retroreflective markers were attached to the thorax and pelvis of the participants: left and right Articulatio acromioclaviculare (L/R AC), Processus spinosus of the 7th cervical and 8th thoracic vertebra (C7 and T8), Incisuria jugularis (IJ), Processus xiphoideus (PX), left and right Spina iliaca anterior superior (L/R ASI) and Spina iliaca posterior superior (L/R PSI). A static calibration of the participants standing in an A-pose (palms down) was done in Vicon Nexus (v2.16.0, Vicon Motion Systems, Oxford, UK). Participants read the task instructions and could clarify open questions with the experimenter before they assumed the start position, facing away from the aperture. The preparation phase in Experiment 1 took approximately 20 min.

In Experiment 2, participants additionally put a motion capture suit over their tight-fitting clothing. Fifty-nine retroreflective markers were attached to the participants in accordance with the default FrontWaist3Fingers marker set of Vicon Shogun. A static A-pose-calibration and a range of motion of the participants was recorded in Vicon Shogun. The character model most closely matching the size and gender of the participants was activated in Unity. Participants read the task instructions and could clarify open questions with the experimenter. Next, the headset was adapted to the participants and they were given a 2 min familiarization phase in the virtual environment. During the familiarization phase, participants were encouraged to walk back and forth between the start and end position until walking in the virtual environment felt natural. Then, they assumed the start position, facing away from the aperture. The preparation phase in Experiment 2 took approximately 35 min.

Experiment 1 consisted of 2 tasks. Each task contained 5 or 10 sequences of 9 trials each. A trial was defined as a single passing through the aperture. The participants started each trial on the start position, facing away from the aperture. The participants then had to 1) turn 180° on the start position, 2) walk to the end position at normal speed, starting with their right foot, 3) pass through the aperture without knocking over the slalom poles, and 4) stop at the end position, facing away from the aperture. Two research assistants then aligned the slalom poles with the correct marks for the next trial. The end position of the current trial then became the start position of the next trial. If participants knocked over a slalom pole during a sequence, the sequence was aborted and restarted.

In Task 1, participants performed 5 sequences of the 9 aperture widths [5 repetitions × 9 aperture widths = 45 trials]. In each sequence, the 9 aperture widths were presented in a pseudo-randomized order, that is, each width was presented once (e.g., 2,7,6,8,4,3,1,5,9). To this end, a list of 5 pseudo-random permutations [Mersenne twister algorithm (57)] was created before the experiment for each participant. Referring to this list, the research assistants updated the positions of the slalom poles once the participants had reached the end position.

In Task 2, participants performed 10 ordered sequences of the 9 aperture widths: 5 increasing (1,2,3,4,5,6,7,8,9) and 5 decreasing (9,8,7,6,5,4,3,2,1) sequences [2 orders × 5 repetitions × 9 aperture widths = 90 trials]. The sequence order was pseudo-randomized (see above) with no more than two repetitions of the same order in a row (e.g., d,i,d,i,i,d,i,i,d,d). Again, the research assistants updated the positions of the slalom poles after each trial in accordance to an individual sequence/trial list for each participant.

All participants conducted the randomized task (Task 1) first, to get accustomed to the experiment. Posture selection in randomized tasks is unaffected by hysteresis (3, 10) and, thus, should not affect posture in the subsequent ordered task (Task 2). Participants had a resting period of 30 s between each sequence and of 2 min after each set of 5 sequences.

Experiment 2 consisted of 4 tasks. Each task contained 4 or 8 sequences of 9 trials each. A single trial in Experiment 2 was identical to Experiment 1. After each trial, the positions of the slalom poles in the virtual environment were updated by the experimenter with a keypress. If participants knocked over a slalom pole during a sequence, the trial was automatically marked as failed and removed from the later data analysis. The order of the tasks and the sequences, the positions of the slalom poles, the start and end of each trial, and the marking of failed trials was controlled by the UXF toolbox (43).

In Task 1 (no visible body) and Task 3 (visible body), participants performed 4 sequences of the 9 aperture widths each [4 repetitions × 9 aperture widths = 36 trials per task]. In each sequence, the 9 aperture widths were presented in a pseudo-randomized order. To this end, 4 pseudo-random permutations of the 9 aperture widths were created before the experiment for each participant and stored as csv-files. Based on these files, UXF updated the positions of the slalom poles after each trial. In Task 2 (no visible body) and Task 4 (visible body), participants performed 8 ordered sequences of the 9 aperture widths each: 4 increasing and 4 decreasing sequences [2 orders×4 repetitions × 9 aperture widths = 72 trials per task]. Sequence and repetition rules were identical to Experiment 1, and sequence lists for each participant were stored as csv-files.

To test whether behavior in the virtual environment depended on the visibility of the own (virtual) body, we further varied the factor “body visible” in Experiment 2. In Tasks 1 and 2, the virtual body and the colliders were deactivated, thus replicating a simple VR setup with headset only. In Tasks 3 and 4, the virtual body and the colliders were activated, thus providing real-time visual feedback and physical interactions with the environment, which requires a considerably more expensive VR setup with real-time full-body motion capture. In these 2 tasks, participants could actually knock over the virtual slalom poles, similar to the real environment. The order of Tasks 1 and 2 and Tasks 3 and 4 was counterbalanced across participants; the order of the randomized and ordered condition was fixed (as in Experiment 1).

Participants had a resting period of 30 s between each sequence and of 2 min after each set of 4 sequences. During the 2 min breaks, participants were asked to remove the headset and rest their eyes.

The experimental phase took approximately the same time in the real (Experiment 1) and virtual (Experiment 2) environment, that is, 50 min. This was due to the fact that the number of trials was higher in Experiment 2, but the updating of the aperture widths between trials took less time.

In Experiment 1, marker trajectories were reconstructed in Vicon Nexus v2.16.0, labeled manually, and exported to MATLAB (2021b, The MathWorks, Natick, MA) for data analysis. In Experiment 2, reconstructed and prelabeled marker trajectories in Vicon Shogun v1.12.0 could be directly exported to MATLAB. Marker names of the FrontWaist3Fingers marker set were adapted to the marker names used in Experiment 1, to use the same subsequent computations. The laboratory's coordinate system in both experiments was defined with the x-axis pointing to the right, the y-axis pointing to the front, and the z-axis pointing upwards while standing on the start position facing the aperture (Figure 1B).

To identify the moment when participants were passing through the aperture, we first calculated the segment center trajectories of the shoulder girdle (SC), thorax (TC), and pelvis (PC; Table 1) To identify the moment of aperture passing for each segment, the distance in the y-direction (i.e., the walking direction; Figure 1B) between the aperture center and the segment center was analyzed. The distance graph exhibited 9 local minima for each sequence. The corresponding frames of these minima were used to calculate the segment rotation angles α.

To calculate the rotation angle α, a segment direction vector v was defined for each segment. For the shoulder girdle, v was defined as the connection vector from the left to the right acromion (Table 1), thus pointing along the x-axis if participants passed the gap without upper body rotation (Figure 1B). To define comparable direction vectors for the thorax and pelvis, we first defined support vectors s that pointed forward and then calculated the cross product of s with the global z-axis (Table 1).

From the vector components v_x and v_y, the segment rotation angle α was calculated with the four-quadrant inverse tangent function of MATLAB. This was equivalent to a projection of the segment direction vector onto the floor (x-y-plane, Figure 1B). The absolute value of α was used as the dependent variable.

The segment rotation angle α was zero when participants passed straight through the aperture and increased if the segment was rotated to the left or right. Participants tended to rotate their body away from the leading foot when stepping through the aperture. Whether they rotated to the left or right was largely determined by the participants’ step lengths. Within each participant, rotation direction was consistent, as we asked them to always start the movement with their right foot. To render the data of left- and right-rotating participants comparable, the absolute value of α was used.

Data were aggregated over participant, task, and experiment. In Experiment 1, 4 of 27 participants knocked over a slalom pole and, thus, had to abort and restart one sequence. Trials from aborted sequences (1.0% of all trials) were excluded from the analysis. Similarly, in Experiment 2, trials that were marked as failed (6.5%) were removed. Therefore, only successful passes were included in the analysis of both experiments. To render the data more comparable across participants of different body dimensions, the aperture width was divided by the shoulder width of each participant (A/S ratio) for the analysis of the shoulder and thorax rotation, and by the hip width (A/H ratio) for the pelvis rotation. Thus, participants of different shoulder widths should initiate their shoulder rotation at a similar A/S ratio. The same applies to thorax and pelvis rotation. A/S ratio and A/H ratio (from now on termed A/X ratio for simplicity) were then centered around the mean to be used as fixed effects in the models.

Five additional independent variables were used as fixed effects in the models: “environment” (real or virtual), “condition” (randomized or ordered), “order” (increasing or decreasing A/X ratio), “fatigue” (number of sequences carried out since the start of the experiment), and “body visible” (false or true, only defined in the virtual environment). These variables were scaled to values between −1 and +1.

In the randomized conditions, “order” was determined based on the A/X ratio of trial n − 1: if the A/X ratio of trial n − 1 was lower, “order” of trial n was set to increasing; if it was higher, “order” was set to decreasing. Thus, “order” in the randomized conditions was comparable to “order” in the ordered conditions. As no trial n − 1 was available for the first trial of each sequence, these trials were removed from the data sets of both the randomized and the ordered conditions.

Segment rotation angles were analyzed with a generalized linear mixed-effects model (GLMM) with a logit link function. The logit link function was used to better capture the clear bend in the rotation angle data found in an aperture task (38), resulting from a straight posture for larger A/X ratios and an increasingly rotated posture below a critical A/X ratio. To make the dependent variables (shoulder, thorax, and pelvis angle) compatible to the logit link function, the rotation angles for all participants were divided by their respective maximum rotation angle. Thus, the dependent variables were scaled between 0 (no rotation) and 1 (maximum rotation).

To test (1) if participants persisted in their previous postures and (2) if participants' postures differed in the real and virtual environment, we calculated GLMMs (logit link function) on the scaled rotation angles of shoulder, thorax, and pelvis, with “A/X ratio”, “environment” (real or virtual), “condition” (randomized or ordered), “order” (increasing or decreasing), and “fatigue” as fixed effects, and “participant ID” as a random effect. We added two interactions to the model: one between “environment” and “order” to test whether a potential hysteresis effect differed between the real and virtual environment, and one between “condition” and “order” to account for the fact that hysteresis effects have been found to be significantly smaller in randomized than in ordered sequences of trials (10, 21). For this model, all trials from Experiment 1 (real environment) and only trials with a visible body and active colliders from Experiment 2 (virtual environment) were used, based on the assumption that behavior in VR would be most similar to real-world behavior with real-time visual feedback of the own body and physical interactions with the environment.

To test (3) whether real-time visual feedback of the own body and physical interactions with the virtual environment were necessary to reproduce real-world behavior or if simple head-tracking was sufficient, we calculated GLMMs on the scaled rotation angles of shoulder, thorax, and pelvis, with “A/X ratio”, “body visible” (false or true), “condition” (randomized or ordered), “order” (increasing or decreasing), and “fatigue” as fixed effects, and “participant ID” as a random effect. Two interactions were added to the model: one between “body visible” and “order” to test whether a potential hysteresis effect differed depending on the visual feedback, and one between “condition” and “order” to account for potential differences in hysteresis effects between randomized and ordered sequences of trials. For this model, only the trials from Experiment 2 were used.

As a reviewer suggested, the posture chosen by the participants may be influenced not only by the previous postures or the perceptual judgement of aperture width, but also by intrinsic variance in movement execution (44, 47, 49) that could affect task success (33). Therefore, to test (4) whether posture variance may have influenced participant behavior in our task, we calculated the standard deviation (SD) as a measure of variance. For the first dataset (comparison of real and virtual environment), we calculated the SDs of the scaled rotation angles of shoulder, thorax, and pelvis for each unique combination of “A/X ratio”, “environment”, “condition”, “order”, and “participant ID”. In order to have multiple repetitions of each combination available for the SD calculation, the factor “fatigue” had to be aggregated. We then calculated a linear mixed-effects model (LMM) that was identical to the GLMM used for the posture analysis, except for the fixed effect of “fatigue”. As a linear function for “A/X ratio” did not match the variance pattern well, we added a logit derivative function for “A/X ratio” to the model to align it with the logit function used for the posture. For the second dataset (effect of visual feedback), we calculated the SDs for each unique combination of “A/X ratio”, “body visible”, “condition”, “order”, and “participant ID” and then again calculated an LMM identical to the GLMM used for the posture analysis (with a logit derivative function for “A/X ratio” and without a fixed effect of “fatigue”).

The LMMs and GLMMs were calculated in R (v4.3.2; R Core Team, 2023) with the lme4 package (58). Effect sizes for the LMMs were calculated with the r2glmm package (59), using the method of Nakagawa and Schielzeth (60). Effect sizes for the GLMMs were calculated as odds ratios.

For the shoulder rotation, the GLMM comparing the real and virtual environment showed a significant main effect of “A/X ratio” (Table 2). Participants passed straight through the aperture for the higher A/X ratios. Below a critical A/X ratio, they had to use an increasingly rotated posture to pass through the aperture (Figure 2A). There were significant main effects of “environment” and “order”, and no interaction between the two (Table 2). Thus, both main effects could be interpreted. The point-of-change (i.e., 50% of the individual maximum rotation) in the real environment was at a significantly lower A/X ratio (102.4% of shoulder width) than in the virtual environment (124.5%; Figure 2A).

Since it is customary in these types of studies to report not the point-of-change (50% rotation), but the critical value below which a rotation is initiated (5% rotation), only critical values are reported for all subsequent results. However, both measures are depicted in the figures (Figures 2,3) and, due to the negligible interaction terms in the models (resulting in parallel curves), the offset is constant across the whole range of A/X ratios. The main effect of “environment” thus also indicated that the critical value in the real environment (120.7%) was significantly lower than in the virtual one (142.8%). Participants in the virtual environment initiated their shoulder rotation at a larger aperture width.

The main effect of “order” indicated that the critical value in increasing sequences of A/X ratios was significantly lower (130.3%) than in decreasing sequences (133.2%; Figure 2B) of trials. Participants switched from a rotated to a straight posture at a smaller aperture width in increasing sequences, and from a straight to a rotated posture at a larger aperture width in decreasing sequences. Thus, participants did not persist in their previous posture (which would indicate hysteresis) but instead adopted their next posture earlier. The non-significant interaction between “environment” and “order” (Table 2; Supplementary Figures S1A,B) suggested that this inverse hysteresis effect did not differ between the real and the virtual environment, that is, in two independent participant groups. Presuming sufficient power, as indicated by the a priori power analysis, it can be concluded that behavior was similar in both environments. The non-significant interaction between “condition” and “order” (Table 2; Supplementary Figures S1G,H) indicated that no difference in the size of the inverse hysteresis effect was found between the randomized and ordered condition. In the randomized condition, “order” of each trial n was based on the A/X ratio of trial n − 1 (if the A/X ratio of trial n − 1 was lower, “order” was set to increasing; if it was higher, “order” was set to decreasing), and was thus comparable to the encoding in the ordered condition.

We did not find significant main effects of “condition” or “fatigue” (Table 2; Supplementary Figures S1C–F), suggesting that the critical values (i.e., the behavior of the participants) did not differ between the randomized and the ordered condition and also did not change over the course of the experiment. A full visualization of all the main effects and interactions in the model can be found in the Supplementary Figure S1.

All effect sizes were “very small”, except for the main effects of “A/X ratio” and “environment”, which were “large” (61). For the thorax and pelvis rotation, the GLMMs comparing the real and virtual environment qualitatively reproduced the results for the shoulder rotation, with significant main effects of “A/X ratio”, “environment”, and “order” and same signs for the corresponding estimates (Table 2). Thus, all significant shifts of the critical values were comparable to the reported results for the shoulder rotation.

For the shoulder rotation, the GLMM comparing the effects of different visual feedback in the virtual environment showed a significant main effect of “A/X ratio” (Table 3). Participants used a straight posture for the higher and a rotated posture for the lower A/X ratios (Figure 3A). There were significant main effects of “body visible” and “order”, and no interaction between the two (Table 3). Thus, both main effects could be interpreted. The main effect of “body visible” indicated that the critical value without visual feedback of the own body (154.0%) was significantly lower than with visual feedback (162.0%; Figure 3A). Participants with feedback initiated their shoulder rotation at a larger aperture width.

The main effect of “order” (Table 3) indicated that the critical value in increasing sequences of A/X ratios was significantly lower (156.6%) than in decreasing sequences (159.4%). In the virtual environment, participants in increasing sequences switched postures at a smaller aperture width than in decreasing sequences, demonstrating an inverse hysteresis effect. There was no interaction between “body visible” and “order”, or between “condition” and “order” (Table 3), indicating that the size of the inverse hysteresis effect did not differ significantly depending on the visual feedback, or between the randomized and ordered condition. There also was no main effect of “condition”, indicating that the critical values did not differ significantly between the randomized and the ordered condition.

We found a significant main effect for “fatigue” (Table 3) in the virtual environment. Participants adjusted their behavior over the course of the experiment: The critical value to initiate a shoulder rotation significantly increased from the first (150.3%) to the last (165.7%; Figure 3B) sequence, that is, by 0.67% per sequence. With increasing fatigue, participants initiated their shoulder rotation at larger aperture widths.

Effect sizes were “very small”, except for the main effect of “A/X ratio”, which was “large”, and for the main effects of “body visible” and “fatigue”, which were “small” (61). For the thorax and pelvis rotation, again, the GLMMs comparing the effects of different visual feedback in the virtual environment reproduced the results for the shoulder rotation, with significant main effects of “A/X ratio”, “body visible”, “order”, and “fatigue” and same signs for the corresponding estimates (Table 3). Thus, all significant shifts of the critical values were comparable to the reported results for the shoulder rotation.

For the shoulder SDs, an LMM with only a linear function for “A/X ratio” did not accurately reflect the pattern of variance, as variance was higher at the central A/X ratios and lower at the borders (Figure 4A). We therefore added a logit derivative function for “A/X ratio” to the variance model, to align it with the logit function used for the posture analysis. The expanded model achieved a considerably better fit (BIC: −3,286) than the linear model (BIC: −3,241) when comparing the real and virtual environment. The linear effect of “A/X ratio” was significant (Table 4) with a negative slope (−3.0% per A/X ratio). This indicated that SDs decreased as A/X ratios increased (Figure 4A). The logit derivative effect of “A/X ratio” was also significant (Table 4), demonstrating that SDs were higher for central A/X ratios, that is, where the shift between the rotated and the straight posture occurred (cf. Figure 2A).

We further found a significant main effect of “environment” (Table 4), indicating that the posture variance was lower in the real environment (6.4%) than in the virtual environment (9.0%; Figure 4A). A significant main effect of “condition” (Table 4) further indicated that the mean SD was lower in the randomized condition (7.2%) than in the ordered condition (8.1%; Figure 4B). All remaining fixed effects and interactions were not significant (Table 4), indicating that no differences in posture variance were found between increasing and decreasing orders. While the effect size for “environment” was “large” (61), the effect size for “condition” was “very small”. For the thorax and pelvis SDs, the LMMs comparing the real and virtual environment qualitatively reproduced the results for the shoulder SDs. However, for the pelvis, the main effect of “condition” was only close to significance (Table 4). Again, the models expanded by a logit derivative function performed better than the linear models.

When comparing the effects of visual feedback for the shoulder SDs, again, the expanded model achieved a considerably better fit (BIC: −2,100) than the linear model (BIC: −2,007). Both the linear and the logit derivative effect of “A/X ratio” were significant (Table 5). There was a significant interaction between “body visible” and “order” (Table 5), with a “very small” effect size (61). In the increasing sequences, posture variance was higher with visual feedback (11.6%) than without visual feedback (9.9%). In the decreasing sequences, posture variance was more similar (10.7 vs. 11.1%). For thorax and pelvis SDs, the expanded logit derivative models performed better as well. The LMM for the thorax SDs qualitatively reproduced the results for the shoulder SDs (Table 5). The LMM for the pelvis SDs, in contrast, showed no significant interaction of “body visible” and “order”, but a significant main effect of “body visible” (Table 5), with a “very small” effect size (61). Posture variance was higher with visual feedback (11.6%) than without (10.6%; Figure 4C).