Integration of predictions and afferent signals in body ownership

We aimed at investigating the contribution of sensory predictions triggered by the sight of an object moving towards the body for the sense of body ownership. We used a recently developed psychophysical discrimination task to assess body ownership in the rubber hand illusion. In this task, the participants had to choose which of the two right rubber hands in view felt most like their own, and the ownership discriminations were fitted to psychometric curves. In the current study, we occluded the visual impressions of the object moving towards one of the rubber hands (during the first two-thirds of the path) and only revealed the final third of the object's movement trajectory when it touched the rubber hand (approach-occluded condition). Alternatively, we occluded only the final part so that the main part of the movement towards the model hand was visible (touch-occluded). We compared these two conditions to an illusion baseline condition where the object was visible during the entire trajectory and contact (no-occlusion). The touch-occluded condition produced equally strong hand ownership as the baseline condition with no occlusion, while ownership perception was significantly reduced when vision of the object approaching the rubber hand was occluded (approach-occluded). Our results show that tactile predictions generated from seeing an object moving towards the body are temporally exact, and they contribute to the rubber hand illusion by integrating with temporally congruent afferent sensory signals. This finding highlights the importance of multisensory predictions in peripersonal space, object permanence, and the interplay between bottom-up sensory signals and top-down predictions in body ownership.


Introduction
"The body is our general medium for having a world". When Merleau-Ponty formulated this thought in 1962 in Phenomenology of perception, did he imagine the tremendous number of past and present studies that would prove him right by investigating the constant interplay between environmental stimulation, sensory processing, and selfbody perception? Our body influences how we perceive the world (Maselli, Kilteni, López-Moliner, & Slater, 2016;van der Hoort & Ehrsson, 2016;van der Hoort, Guterstam, & Ehrsson, 2011), and conversely, the sensory stimulation we receive can influence our bodily selfperception. For example, one specific aspect of bodily awareness is the sense of "body ownership,", i.e., the feeling that your body belongs to you, that strongly relies on the integration of visual, tactile, and proprioceptive inputs (Ehrsson, 2012;Ehrsson, Spence, & Passingham, 2004). From the integration of multiple sources of sensory signals related to the body emerges the phenomenological quality that a body part is part of one's body (Martin, 1995) and the subjective awareness of a limb or the whole body as one's own (Gallagher, 2000;Gallagher & Daly, 2018). As such, body ownership constitutes a fundamental component of self-awareness (de Vignemont, 2018). Being able to perceive which body parts constitute one's body is fundamental for an individual's ability to survive in their environment by supporting the preservation of the body's physical integrity through defensive actions (Graziano & Cooke, 2006). However, being able to perceive the boundary between the bodily self and the external environment is not enough to protect the body. Effective defensive action requires that the individual anticipates and predicts what will happen to the body during dynamic interactions with environmental objects that play out in realtime over brief episodes. Will the rapidly approaching tennis ball hit you in the face, or will you return it with a quick volley?
Moreover, the generation of visual predictions also influences other sensory modalities such that touch can be expected from the sight of a looming object that moves towards one's face. Such anticipatory responses are associated with changes in activity in multisensory neurons, i.e., neurons responding to at least one sensory stimulation but that respond differently when two or more signals from different sensory modalities are combined, found in several cerebral regions (Meredith & Stein, 1986;Stein & Meredith, 1993). Electrophysiological studies in monkeys have shown that neurons involved in visuotactile integration are found in the ventral premotor cortex (vPM; Fogassi et al., 1996;Graziano, Hu, & Gross, 1997a;Rizzolatti, Fadiga, Fogassi, & Gallese, 1997) and the ventral intraparietal area (VIP; Duhamel, Colby, & Goldberg, 1998). Moreover, neurons presenting matching receptive fields for the direction and speed of a stimulus have been identified in the VIP (Duhamel et al., 1998), and the neuronal responses to visuotactile stimulation observed in this region could explain the integrative behavior observed both in nonhuman and human primates (Avillac, Ben Hamed, & Duhamel, 2007;Avillac, Denève, Olivier, Pouget, & Duhamel, 2005). Importantly, visuotactile neurons in these multisensory areas show anticipatory responses when one's body is approached by a looming object (visual stimulation), which suggests that the time and location of the tactile consequences of the approaching visual object are predicted (Avillac et al., 2005;Graziano et al., 1997a;Graziano, Hu, & Gross, 1997b;Graziano & Cooke, 2006).
A looming object not only can induce such predictions at the cellular level but also can have a demonstrable influence on perception at the behavioral level. For example, a looming visual stimulus can influence the detection of a tactile stimulus, with the enhancement of tactile detection being maximal when the predicted impact of the visual stimulus is spatially and temporally congruent with the tactile stimulus, both in terms of correct detection (Cléry, Guipponi, Odouard, Wardak, & Hamed, 2015) and response time (Kandula, Hofman, & Dijkerman, 2015). In addition, this type of visually induced tactile prediction is accurate. For example, based on looming visual stimuli, participants in Neppi-Mòdona and colleagues' study Neppi-Mòdona, Auclair, Sirigu, and Duhamel (2004) were capable of distinguishing predicted impact locations with an accuracy of less than four millimeters. This capacity to accurately predict the tactile impact of an approaching object fundamentally depends on the representation of peripersonal space, i.e., the space immediately surrounding the body mapped by multimodal neurons (Graziano et al., 1997a;Rizzolatti et al., 1997). Close interactions between the peripersonal space and body ownership have been observed in several behavioral studies as well (Brozzoli, Ehrsson, & Farnè, 2014;Brozzoli, Gentile, & Ehrsson, 2012;Guterstam, Zeberg, Ö zçiftci, & Ehrsson, 2016;Lloyd, 2007;Makin, Holmes, & Ehrsson, 2008). These previous observations emphasize the need to further study the interaction between visually induced sensory predictions and body ownership in peripersonal space.
In the past twenty years, the relationship between perception of ownership towards a specific body part and multisensory integration has been studied in healthy people using body illusions, i.e., the experimentally induced embodiment of a fake body part (for a recent review, see Ehrsson, 2020). The most well-studied bodily illusion is the rubber hand illusion (RHI), in which the repeated synchronous stroking of a participant's unseen real hand, which is hidden behind a screen, and a rubber hand, which is placed in full view in an anatomically plausible position, elicit the illusory sensation of the rubber hand as one's own (Botvinick & Cohen, 1998). The incorporation of the rubber hand into one's own-body representation is considered to involve the integration of concurrent and spatiotemporally congruent visual and somatosensory feedback in multisensory brain areas (Ehrsson, 2020;Ehrsson et al., 2004). A multitude of factors have been shown to influence visuotactile integration leading to the emergence of the RHI: the synchronicity of the visuotactile stimulation (Botvinick & Cohen, 1998;Shimada, Fukuda, & Hiraki, 2009), the visual resemblance between the rubber hand and a human hand (Tsakiris, Carpenter, James, & Fotopoulou, 2010), the postural correspondence between the rubber hand and the real hand (Ehrsson et al., 2004;Pavani, Spence, & Driver, 2000), and the spatial proximity of the rubber hand to the real hand (Kalckert & Ehrsson, 2014;Lloyd, 2007;Preston, 2013; for a review: Kilteni et al., 2015). In the classic RHI setup, a participant sees the brush, or whatever object is used to touch the rubber hand, approaching the rubber hand in peripersonal space. The participant therefore receives visual information about an object approaching the model hand that will trigger multisensory predictions (Avillac et al., 2005;Cléry et al., 2015;Graziano & Cooke, 2006;Neppi-Mòdona et al., 2004) about the trajectory of the object motion and about where and exactly when the upcoming touch event will be experienced on the "owned" rubber hand. However, the relationship between such multisensory predictions and the RHI is unknown, and the role of multisensory predictions in body ownership remains unclear (Ferri et al., 2013;Guterstam, Larsson, Zeberg, & Ehrsson, 2019;see Discussion). Indeed, the current study aimed to investigate the role of visuotactile predictions in the RHI and specifically clarify the contributions from the sensory predictions generated by observing the object moving towards the rubber hand before the actual physical contact and somatosensory feedback. To quantify body ownership, we used a recently developed two-alternative forced choice (2-AFC) discrimination task (Chancel and Ehrsson, 2020) that allows the registration of subtle changes in body ownership perception during the RHI in a rigorous way. In this task, a participant is asked to choose which of two identical right rubber hands feels the most like their own hand. These rubber hands are touched by probes held by computer-controlled robotic arms, either synchronously with the touches applied to the participant's hidden real hand by an identical robot or asynchronously, with systematically varying degrees of temporal delays between the touches on the rubber hand and the real hand. The movements of the robotic arms holding the probes approaching the rubber hands are clearly visible. Thus, the participant can generate natural predictions about the precise location and time point of the touch event on the rubber hand because the robots always make the same movement with the same velocity and duration, with a full view of the probe's movement trajectory towards the rubber hands; this condition corresponds to the current study's baseline condition.
In the next condition, however, the approaching movement of the robotic arm towards one of the rubber hands was occluded for the first two-thirds of the movement trajectory, and the participant could only see the last third of the movement, i.e., when the robot's probe touched the rubber hand. Consequently, the participant's capacity to generate visually induced tactile predictions based on visual information about the object moving in peripersonal space is reduced. We hypothesized that this condition would lead to a weaker illusory ownership over the rubber hand compared to the original condition without any visual occlusion. Finally, in the third condition, the last third of the robot arm's movement towards one rubber hand was occluded, i.e., the moment when the robot's probe touched this rubber hand was not visible by the participant, while the initial two-thirds of the approaching movement of the robotic arm was visible. In this condition, the participant can thus form predictions about the unseen contact event on the rubber hand behind the occluding screen on the basis of the visual information about the movement trajectory of the object approaching the model hand. Thus, we expected the RHI to work in this condition because the participants should be able to continue to represent the movement of the object as it disappears behind the occluding screen and use information about the expected touch event to integrate with the tactile information from the real hand. Indeed, it has been well known since Piaget's classic studies Piaget (1952) that we are able to mentally represent an object and follow its position even when it has disappeared from view, a concept known as "object permanence". This concept was also identified in different animals (Anderson, Hunt, Stoep, & Pribram, 1976;Thinus-Blanc, Poucet, & Chapuis, 1982) and was later confirmed by electrophysiological recording of visuotactile neurons in the vPM encoding the presence of an object in arm-centered coordinates that was no longer visible (Graziano et al., 1997b). Nevertheless, compared to our first condition with full visibility we expected a relative decrease in the RHI strength in both conditions with visual occlusion in line with the principle that increasing the availability of sensory information generally leads to a more vivid multisensory perceptual experience (Stein & Meredith, 1993). Finally, by directly comparing the two different occlusion conditions, we could explore the relative importance of the visual feedback from the two phases of the touch episodes: seeing the approaching movement of the object towards the rubber hand and the actual touch event.

Participants
We conducted an a priori power analysis using G*Power3 (Faul, Erdfelder, Lang, & Buchner, 2007) on previous data using the same discrimination task (Chancel and Ehrsson, 2020). The results showed that a total sample of 20 participants was required to achieve a power of 0.80. Of the 35 recruited individuals, only those who could experience the RHI took part in the study (see Inclusion test). As a result, we included 28 healthy naive participants (17 females) between 18 and 50 years of age (M = 26.30, SD = 6.25) in the main study. From our previous study, we know that participants who do not experience the RHI produce more or less random forced ownership decisions in the discrimination (Chancel and Ehrsson, 2020), consistent with their subjective report, but this made them unsuitable as participants in the current study. All participants had corrected-to-normal or normal vision. We made no distinction between left-and right-handed participants, since the dominant and nondominant hand seems to produce similarly strong RHI (Smit, Kooistra, van der Ham, & Dijkerman, 2017). All participants provided written consent at the start of the experiments and received 300 SEK for participating. The Regional Ethical Review Board of Stockholm approved the study (since 2019, the Swedish Ethical Review Authority).

Experimental setup
The setup was adapted from Chancel and Ehrsson (2020) (see Fig. 1). The participant was seated in front of a small box. On top of the box, two identical right rubber hands were placed side-by-side in parallel and anatomically plausible orientations. The box was slightly tilted with the front end upwards (30 • ) so that the participant could fully see the rubber hands while sitting down. To ensure that all participants were seated in the same way with their head and body with respect to the box, a chinrest with height restrictions was used and adjusted to a height of 38-43 cm above the table based on each participant's height. The participant placed their right hand inside the box in an identical orientation to the rubber hands and equidistant from the two fake hands. The vertical distance between the rubber hands and the participant's hand was 15 cm, and in the horizontal plane, each rubber hand was located 5 cm to the left or the right from the participant's real hand. Thus, the rubber hands were located within the peripersonal space (Holmes & Spence, 2004) of the participants' real right hand, a distance within which the RHI can be elicited (Brozzoli et al., 2012;Kalckert & Ehrsson, 2014;Lloyd, 2007), including ownership of two fake hands (Fan & Ehrsson, 2020). A white fixation dot was placed halfway between the two rubber hands to prevent the participant from looking more at one rubber hand than the other. During the experiment, the index fingers of the participant's hand and the rubber hands were touched by three robots. Each robot consisted of two metal plates (17 × 3 cm), at the end of which a small plastic tube (1.7 cm) was attached, and was controlled by two motors (Hitec Multiplex®, USA; see Chancel and Ehrsson, 2020 for details). The robots could thus make a smooth approaching movement towards the rubber hands and the participant's Fig. 1. A. Basic experimental setup with the two identical rubber hands above (lRH: left rubber hand; rRH: right rubber hand; both in dark gray) and the participant's hand below (pH in light gray). The black rectangles represent the robot arms touching the rubber hands (on top) or the participant's hand (in the lower box). B. Noocclusion condition. The participant could see the robots approach and touch both the lRH and the rRH. C. Approach-occluded condition. The screen occluded the approaching movements of the robot touching the rRH. The participant could therefore not see the rRH being approached but was able to see this rubber hand being touched. D. Touch-occluded condition. The screen occluded the index finger of the rRH and 3 cm of the surrounding space. The participant could therefore see the robot approach the rRH but did not receive the visual confirmation of touch on this rubber hand. NB: the outlines of the screens in Fig. 1B and C are highlighted with a blue line for the purpose of clarity. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) hidden real hand, touch the hands, and then make a receding movement back to the original starting position. A black piece of fabric covered the part of the participant's right arm and shoulder that was not hidden inside the box; thus, the whole real upper right hand was occluded, and the participant could only see the two model right hands. During the experiment, the participants wore earphones playing white noise to block auditory information from the robots' movements. The volume of the noise was adapted to each participant so that they would not hear any sounds from the robots' motors or mechanical parts while still being comfortable to listen to for the duration of the experiment (sounds synchronized with tactile feedback can otherwise influence the RHI; Radziun & Ehrsson, 2018). A frame of firm cardboard (8.5 × 24 cm) was placed 5 cm to the left of the rRH. The frame itself did not obstruct the participant's view of the robot or the rRH in any of the conditions, but two different sized screens were attached to this frame in the two conditions with partial occlusion of the visual stimulus (see Conditions below).

Psychophysical discrimination task
The psychophysical discrimination task is a 2-AFC task in which the participant had to decide whether the rubber hand on the left (left rubber hand: lRH) or the rubber hand on the right (right rubber hand: rRH) felt most like his or her own hand (Chancel and Ehrsson, 2020). Each trial started with a 12-s stimulation phase, during which six touches were applied to the index fingers of the participant's hand and the rubber hands by the three robots. The index fingers were touched on three locations, the second knuckle, the third knuckle, and between these two knuckles, in a pseudorandom order. All touches were applied as short taps, with regular intervals (0.5 Hz) and a constant speed (9 cm/ s). The taps were applied by the plastic tube ending (7 mm diameter) of the robots. After this stimulation phase, the participant heard a beep, which served as a prompt for them to verbally indicate which rubber hand ("left" or "right") felt most like their own. The (a)synchronicity between the seen and felt touches was manipulated in seven steps (i.e., hereafter referred to as the level of visual-tactile asynchrony). At least one rubber hand was synchronously stroked with the participant's hand in each of these conditions. The other rubber hand was either stroked synchronously with the participant's hand or after a delay of 50, 100 or 200 ms. For the remainder of this paper, negative values refer to delays applied to the rRH (− 50, − 100, − 200), and positive values refer to delays applied to the lRH (+50, +100, +200). Asynchronies in this range can lead to reductions in the illusory hand ownership in favor of one or the other rubber hand, and the wider the asynchrony was, the more significant the reduction (Chancel and Ehrsson, 2020). Thus, when asked to choose which rubber hand they most perceived as their own, the participants tended to favor the synchronously touched model hand, and the greater the asynchrony for a particular rubber hand was, the more often the subject chose the other synchronously stimulated hand (Chancel and Ehrsson, 2020).

Conditions
The left robot arm's movement approaching the lRH always remained fully visible, but the three different conditions differed in terms of the visibility of the right robotic arm that approached and touched the rRH. (i) In the first condition (no-occlusion), the participant could see the right robot approach and touch the rRH under full viewing conditions without any occlusion (Fig. 1B). The approaching movement lasted 1.00 s (with estimated error of +/− 8 ms), so did the receding movement. In this condition, visually induced tactile predictions about when and where the robot's probe will touch the rRH are automatically generated as part of the natural viewing experience, and the subject receives visual feedback about the touch event. This no-occlusion condition, therefore, was used as a baseline for the two other conditions. (ii) In the second condition (approach-occluded), the robot's approaching movements towards the rRH were occluded with a large screen which blocked the participant's view of the first two-thirds of the movement path (19.5 × 24 cm; Fig. 1C). Thus, the initial approximate 0.66 s of the robot's movement trajectory, were occluded from view; therefore, tactile predictions could not be formed based on this dynamic visual information. However, the screen in the approach-occluded condition did not occlude the rRH and approximately 3 cm of the surrounding space, so the participant could see the final one-third of the robots' movement path, i.e., the last 0.33 s, when it was moving close towards and then touching the rubber hand. (iii) In the third condition (touch-occluded), the participants could only see the first two-thirds of the robot's movement path towards the rRH but not the final part when the rubber hand was touched (Fig. 1D). In this condition, the view of the rubber hand's index finger and 3 cm of the surrounding space was obstructed with a smaller screen (17 × 5.5 cm; only the rRH index finger was occluded, but the rest of the rubber hand was visible). This arrangement allowed the participant to see the robot moving for approximately 0.66 s before the tip holding the probe disappeared behind the screen (plus 0.66 s of the receding movement when the robot moved back to its starting position after the touch event). Thus, in the touch-occluded condition, only visually induced predictions about when and where the robot's probe will touch the rubber hand could be used to infer single touch events and lead to integration with the somatosensory feedback and the triggering of the RHI.

Inclusion test (susceptibility screening for RHI)
The inclusion test consisted of two parts. In the first part, we assessed whether the participant was sensitive to the classic version of the RHI involving one model hand. The participants needed to experience this illusion to make ownership decisions in the psychophysical discrimination task, but not everyone perceives the RHI (Kalckert & Ehrsson, 2014). Therefore, in the present study, we tested illusion susceptibility in three synchronous trials of a classic RHI paradigm. The participant placed their right hand inside a small box and was instructed to look at a right rubber hand (a life-like cosmetic male hand prosthesis; model 30,916-R, Fillauer®, filled with plaster, identical to the ones used in the main task) that was placed on top of this box. The experimenter subsequently used two firm plastic probes (firm plastic tubes, diameter: 7 mm) to tap the rubber hand synchronously with the participant's hand for 24 s. These tubes were similar to the robots' probes in the main experiment. We applied the taps to three different places along the real and rubber index fingers, at a frequency of 0.5 Hz. The characteristics of the taps and the duration of the stimulation were designed to resemble the stimulation later applied by the robot during the discrimination task (see Section 2.3 and further below). After each trial, the participants were instructed to fill out an RHI questionnaire (Botvinick & Cohen, 1998; see Supplementary material, Fig. S1). This questionnaire includes three items assessing the RHI, including sensing touch on the rubber hand and feeling the rubber hand as one's own, and four control items to be rated between − 3 ("I completely disagree with this item") and 3 ("I completely agree with this item"). Our inclusion criteria for a strong enough RHI to participate in the main psychophysics experiment were as follows: i) the mean score on the ownership statements (Q1, Q2, Q3) was greater than one, and ii) the difference between the mean score on the ownership items (Q1, Q2, Q3) and the mean score on the control items (Q4 to Q9) was greater than one (Abdulkarim & Ehrsson, 2016). Five out of the initial 35 volunteers did not meet this criterion and were excluded from further testing.
In the second part of the inclusion test, the participant received six practice trials of the psychophysical discrimination task. For three trials, the lRH was synchronously stroked with the participant's hand, and the rRH was touched with a delay of 200 ms. For the three other trials, the rRH rubber hand was synchronously stroked with the participant's hand, and the lRH was touched with a delay of 200 ms. Participants did not perform the main experiment if they always selected the same rubber hand as their own during these six practice trials. Two volunteers had a strong bias for the lRH and always indicated feeling ownership for this rubber hand even when this rubber hand was touched 200 ms after the participants' real hand (maximal asynchrony condition with the rRH being synchronously touched). These individuals were not included in the main experiment since it would be impossible to fit their discrimination data with the Gaussian cumulative models (see below), and their responses were judged unreliable (the 200-ms delay should lead to a reduction in hand ownership in participants who are susceptible to the RHI; Chancel and Ehrsson, 2020). At the end of the inclusion phase, the 28 remaining volunteers took part in the main experiment described below.

Main experiment
The main experiment consisted of six blocks of the psychophysical discrimination task, with two blocks for each condition. The order of the blocks was semirandomized for each participant to ensure that the same condition was not repeated in two consecutive blocks. At the start of each block, the experimenter checked whether the screens that occluded the touch probe's movement trajectory were correctly placed to exclude view from the participant's perspective, depending on the particular condition in the upcoming trial. We also asked the participant to confirm that they could see the rRH and/or the robot's approaching movement trajectory according to the particular occlusion manipulations used in the different conditions. The experimenter first explained which parts of the robot and rRH the participant was supposed to see and subsequently used a test trial to see if the screens needed to be adjusted. Each block contained 42 trials of the discrimination task, with each (a)synchronicity condition repeated six times in a randomized order. A six-second interval separated the trials, after which a beep indicated the start of the next trial.

Psychometric curve fitting
For each condition, we calculated the percentage of trials in which the rRH was chosen per level of visual-tactile asynchrony (i.e., the seven (a)synchronicity levels). The Palamedes Toolbox for MATLAB (Math-Works) was used to fit a cumulative Gaussian function to these data with a maximum likelihood criterion approach (for a detailed explanation of this procedure, see Prins & Kingdom, 2018). An example of these psychometric curves is displayed in Fig. 2A. The Gaussian function assumes that the probability of choosing the rRH at one level of visual-tactile asynchrony is constrained by the probability that this hand is chosen at the other levels. By comparing the likelihood ratios of this function and a saturated (unconstrained) model over 1500 simulations (bootstrap analysis), a goodness-of-fit measure (p-value for deviance: pDEV: 0-1) was generated for each psychometric curve. The critical threshold for pDEV is 0.05, and lower values are considered to reflect unacceptably poor fit (Kingdom & Prins, 2009;Prins & Kingdom, 2018).
The means and standard deviations were subsequently extracted from each curve ( Fig. 2A). The mean represents the point of subjective equality (PSE), which is the level of visual-tactile asynchrony for which the participant is equally likely to choose the lRH or the rRH. A positive PSE indicates that the participant perceives the rubber hands as equally theirs, even if the lRH is not stroked synchronously with their real hand. This means that there is an overall preference for ownership towards the lRH. In contrast, a negative PSE indicates that there is an ownership overall preference towards the rRH. Thus, when comparing different conditions, a change in PSE reflects a change in the ownership illusion's relative strength perceived for the two model hands. As explained in the introduction, we expected our occlusion conditions (approach-occluded and touch-occluded) to result in a weaker RHI experienced for the occluded rRH than the baseline condition with full stimulus visibility for this model hand. Thus, we expected to observe an increased PSE in the Fig. 2. A. Psychometric curves for one representative participant with the projection of the mean (point of subjective equality) for each condition. The PSE represents the level of visual-tactile asynchrony for which the lRH and rRH are equally likely to be experienced as one's own hand. For instance, in the approach-occluded condition, this participant perceived the rubber hands to be equally theirs when rRH (approaching movement occluded) was stroked synchronously with their own hand and the fully visible lRH was touched with a delay of 78 ms. This finding indicates that there is a stronger ownership towards the lRH in this condition. B. Mean PSE (+SD) for each condition showing a significant increase in PSE in the approach-occluded condition compared to the two other conditions. C. Individual-subject PSE for each condition. The gray lines join the PSEs in each of the three conditions individually obtained for each participant. *** = p < .001.
approach-occluded condition compared to the no-occlusion condition, i.e., a preference for the lRH as the latter baseline condition allowed the participants to use the visual information from seeing the robot moving towards the model hand to form visuotactile predictions. We also expected to see an increased PSE in the touch-occluded condition compared to the no-occlusion condition, as the participant lost the visual confirmation of touch and had to rely solely on visually induced tactile predictions to experience the RHI.
The standard deviation (σ; sigma) reflects the magnitude of visualtactile asynchrony that the participant needs to make a clear distinction between the two rubber hands. A change in this parameter between conditions would mean that the occlusion of the robot movement, or the contact on the rubber hand, or none of it, depending on the condition, changed the participants' ability to discriminate between the rubber hands. This would mean an increased or decreased sensitivity to visuotactile asynchrony for body ownership perception depending on the occlusion. We only expected our experimental manipulation to affect the strength of the RHI (i.e., PSE), and not the temporal discrimination threshold (sigma), i.e., ability to discriminate between two rubber hands.

Statistical analysis
One-way repeated measures ANOVA with post hoc comparisons (Bonferroni corrected) was used to analyze the differences between the PSEs in each condition. Complementary Bayesian analyses were executed to quantify the evidence in favor of the alternative hypothesis, regardless of the post hoc comparisons' significance. Bayes factors were obtained using Bayesian paired t-tests with an unspecified Jeffreys prior (r = 1) (Rouder, Speckman, Sun, Morey, & Iverson, 2009). A nonparametric Friedman's test was used to analyze differences in the sigma, as the normality assumption was violated in two conditions. All analyses were executed in SPSS 25.0, except for the Bayesian analyses conducted in RStudio (BayesFactor package). Unless stated otherwise, α = 0.05 (two-tailed).

Results
The cumulative Gaussian function fit the data well, with an average pDEV of 0.40 for the no-occlusion condition (SD = 0.33), 0.36 for the approach-occluded condition (SD = 0.31) and 0.40 for the touch-occluded condition (SD = 0.31). This is consistent with previous studies (Chancel and Ehrsson, 2020).
Importantly, there was a significant main effect of experimental condition on the PSE (F(2, 48) = 28.42, p < .001; Fig. 2B and C). The average PSEs in each condition are displayed in Fig. 2B. Post hoc comparisons revealed that the PSE in the approach-occluded condition (M = 89.64, SD = 49.49) was significantly higher than those in the no-occlusion condition (M = 26.36, SD = 50.79; p < .001) and the touch-occluded condition (M = 27.32, SD = 63.98; p < .001). In other words, there was a significantly stronger preference for the non-occluded lRH in the approach-occluded condition than in the other two conditions. This result indicated that eliminating the visual feedback from the robot's approaching movement towards the rRH significantly reduced the illusion, which was in line with our hypothesis. In contrast, there was no significant difference between the PSE in the touch-occluded and no-occlusion conditions, p = 1. This unexpected finding suggested that eliminating the visual feedback from the touch event on the rRH did not reduce the ownership-illusion for this fake hand. This negative result further underscores the critical roles of visuotactile predictions and object permanence in the RHI. Subsequent Bayesian analyses further supported the results from the frequentist statistics. The Bayes factors for both the comparison between the no-occlusion and approach-occluded (BF 10 = 36,445) and the approach-occluded and touch-occluded conditions (BF 10 = 9907) were strongly in favor of the alternative hypothesis (Jarosz & Wiley, 2014). The Bayes factor for comparing no-occlusion and touch-occluded conditions (BF 10 = 0.15) was moderately in favor of the null hypothesis (Jarosz & Wiley, 2014).
Shapiro-Wilk tests showed that data were not normally distributed for the sigma in the no-occlusion condition (M = 184, SD = 88; p = .036) and the approach-occluded condition (M = 183, SD = 164; p < .001). A subsequent Friedman's analysis demonstrated no significant difference between the sigmas in each condition, χ 2 (2) = 4.160, p = .125. Thus, as expected, the basic ability to discriminate the temporal asynchronies did not change across conditions.

Discussion
In the present study, we examined the extent to which visually induced tactile predictions are involved in body ownership using the RHI and a recently developed 2-AFC psychophysical hand-ownership discrimination task (Chancel & Ehrsson, 2020). By occluding, or not occluding, different phases of the visual stimulation from view, we assessed the relative contribution of the visual information regarding the object moving towards the rubber hand and the visual feedback from the object stroking the rubber hand for the emergence of the RHI. As hypothesized, the participants' ownership perception was significantly impaired when the visual information about the object's motion trajectory towards the rubber hand was eliminated, and the formation of visuotactile predictions about upcoming physical contact was therefore disrupted. Surprisingly, no significant difference in terms of ownership perception was observed regardless of whether the participants could see the rubber hand being touched or not, as long as they could see the movement of the robot arm holding the object approaching the rubber hand. This observation suggests that the visuotactile predictions formed by seeing the object approaching were very precise in terms of the exact timing of the anticipated touch event behind the occluding screen. These predictions were then combined with the actual somatosensory signals from the real hand to elicit the RHI. These results collectively suggest an important role of anticipatory perceptual processes in body ownership and multisensory body representation.

Multisensory predictions, body ownership, and peripersonal space
The current consensus is that predictive processes influence perception (de Lange et al., 2018;Den Ouden et al., 2012;Engel, Fries, & Singer, 2001;Friston, 2005;Kandula et al., 2015;O'Callaghan et al., 2017;Pinto et al., 2019;Rao & Ballard, 1999;Trapp & Bar, 2015). Therefore, it is reasonable to assume that sensory predictions, such as tactile predictions, are also involved in the perception of one's own body. Consistent with this idea, the present findings suggested that the effect of visual-tactile stimulation in the RHI depends partially on tactile predictions that are being generated by the sight of the approaching object (typically a brush) towards the model hand and not only from the brief movement of physical contact.
This result echoes the observation that body ownership is deeply linked to multisensory processing (Ehrsson, 2012) and that sensory predictions modulate the processing of bottom-up visuotactile feedback in many other perceptual tasks (Cléry et al., 2015;Kandula et al., 2015;van Ede, de Lange, Jensen, & Maris, 2011;van Ede, Jensen, & Maris, 2010). Critically, the tactile stimulation on the hidden real hand was perceptually combined with the expected impact of the unseen object as it moved behind the occluding screen, and this integration process depends on the temporal congruence between the expected contact between the object and the model hand and the actual tactile feedback, as revealed by our psychophysics model fitting over the range of asynchronies tested. Thus, the RHI corresponds to a naturally evolving temporally extended multisensory event, where the participant experiences a unified perception of an object moving towards one's hand and stroking it in a single episode. Blocking or limiting the generation of tactile predictions as done in the approach-occluded condition when the participant could not see the robotic arm approach the rubber hand thus weakened the entire resulting perception of the episode and led to a reduced RHI. These results highlight that the embodiment of the rubber hand in the RHI emerges from an interaction between top-down sensory predictions and bottom-up afferent sensory signals. Note that this interplay can occur within the perceptual systems, and do not necessarily require higher cognitive process influencing perception (e.g., semantic, conceptual knowledge). In multisensory models of body ownership, such top-down and bottom-up information are combined in a single causal inference process to decide whether the rubber hand is one's own or not (Ehrsson & Chancel, 2019;Fang et al., 2019;Samad, Chung, & Shams, 2015). Finally, we should add that while our study investigated sensory predictions elicited from visual information, it is theoretically possible that signals from other sensory modalities can also generate predictions relevant to body ownership, such as proprioception (Ehrsson, 2005) or audition (Radziun & Ehrsson, 2018).
Our findings are consistent with the demonstrated interaction between peripersonal space and body ownership. The interplay between body ownership and peripersonal space has been widely investigated in both behavioral and neuroimaging studies (Brozzoli et al., 2014(Brozzoli et al., , 2012Grivaz, Blanke, & Serino, 2017;Guterstam et al., 2016;Lloyd, 2007;Makin et al., 2008;Noel, Blanke, & Serino, 2018). For example, multisensory neurons in regions encoding peripersonal space are particularly sensitive to dynamic stimuli (Colby, Duhamel, & Goldberg, 1993). These neurons respond specifically to looming visual stimuli and encode the time and location of the predicted impact on the observer's body (Avillac et al., 2005;Graziano et al., 1997aGraziano et al., , 1997bGraziano & Cooke, 2006). These results are especially relevant in the context of RHI paradigms used to investigate body ownership, including the present study, since in most cases, predictable movements are used to approach the fake limb or body with an object to deliver tactile stimulation to elicit the illusion. The generation of visuotactile predictions based on these movements occurring in the peripersonal space could thus be a mechanism through which multisensory integration for body ownership perception is favored within the peripersonal space.
Our results inform us about how the integration of actual somatosensory signals and precisely predicted visuotactile events contributes to the RHI. This integration of predictions and sensory feedback is different from a possible general tactile anticipation effect. In previous experiments, when the participant's real hand was never touched, and the participant only observed an object very slowly (2 cm per second) moving towards the rubber hand but stopping 15 cm away from it and never touching it, visuotactile predictions were very weak we theorize; and indeed the RHI was significantly weaker compared to when synchronous visuotactile stimulation was delivered to the real hand and the artificial hand (Guterstam et al., 2019(Guterstam et al., , 2016but see Ferri, Ambrosini, Pinti, Merla, & Costantini, 2017;Ferri, Chiarelli, Merla, Gallese, & Costantini, 2013;Smit et al., 2018). The possibility that the mere anticipation of touch could boost the RHI irrespectively of somatosensory feedback (Ferri et al., 2013;Guterstam et al., 2019;Samad et al., 2015) is an issue we did not directly investigate, and our result does not resolve this debate. Such a general tactile anticipation effect has been proposed to operate under a different mechanism from the one studied here (Ferri et al., 2013), based on a top-down selection of competing sensory representations by attentional mechanisms (Engel et al., 2001). This theoretical mechanism is independent of errors or matches between predictions and actual sensory feedback (Ferri et al., 2013). In contrast, the present experiments were designed to probe the degree of temporal match or mismatch between precise multisensory predictions and somatosensory feedback; this temporal congruence or mismatch detection is thought to be critical for to RHI in both classic multisensory integration models (Ehrsson, 2012;Ehrsson, 2020) and predictive coding models of body ownership (Apps & Tsakiris, 2014;Limanowski & Blankenburg, 2013).
Importantly, suppose a general tactile expectation independent of sensory feedback is also active in the current experiments. In that case, it could not explain our key findings as we found that illusory hand-ownership depended on a significant systematic relationship between the fine-grained degree of asynchrony between the anticipated unseen multisensory event and the actual tactile feedback. At a more general cognitive level, the subject could anticipate touch on their hand based on the starting cue (beep) at the beginning of the trial in all our conditions. Therefore, in terms of associative learning and attention towards the hand (Carlsson, Petrovic, Skare, Petersson, & Ingvar, 2000;Roland, 1981), all our conditions were perfectly matched in terms of general statistical predictability of touch (same delay between starting cue and tactile stimulation). However, what differed was the specific information that could be extracted from observing the object's movement trajectory in peripersonal space and used to predict the contact event and compare this anticipated event to the actual somatosensory feedback from the real hand. Thus, the present results cannot be explained by a general anticipation of touch or the related heightened attention to the rubber hand or the real hand, but by the integration of precisely time multisensory predictions and somatosensory feedback.
Another issue that is worth discussing is how our results relate to visual perceptual mechanisms related to estimating time to collision with looming visual stimuli. Such mechanism may not involve a transformation of visual information into tactile predictions but simply computing the time remaining before an approaching visual object makes contact with the observer using visual information alone, for example, by using the optical variable tau, defined as the inverse of the relative rate of expansion of the incoming object's image on the retina (Lee, 1976;Yan, Lorv, Li, & Sun, 2011). Indeed, an extensive literature has investigated human capacity to generate such visual motion prediction, inside as well as outside the peripersonal space (Iachini, Ruotolo, Vinciguerra, & Ruggiero, 2017;Vagnoni, Andreanidou, Lourenco, & Longo, 2017;Vagnoni, Lingard, Munro, & Longo, 2020;Yan et al., 2011). These time-to-collision studies showed that sensory-driven predictive mechanisms, and not only top-down interactions, could reflect how efficient individuals are at predicting the time and place of contact for a looming object (Iachini et al., 2017). However, according to the same study, time-to-collision judgments in the peripersonal space are more influenced by the temporal parameters than in the extra-personal space, which especially relevant in our study. Moreover, as mentioned earlier, specific multisensory neural responses are elicited by visual movement in the peripersonal space (Avillac et al., 2005;Avillac et al., 2007). Critically, the optical variable tau was always constant within our conditions as the probe touching the rubber hand does not get closer to the participant's face; only the precise temporal match or mismatch of the tactile feedback was manipulated. Altogether, this led us to conclude that our results are due visuotactile predictions within the peripersonal space related to body representation, not purely visual predictions based on optical information alone (e.g. tau).

If the touch is predicted, does seeing it add anything?
A somewhat unexpected finding was the absence of a significant difference between the RHI strength in the touch-occluded and no-occlusion conditions. We initially expected to observe a stronger preference for the rubber hand that the participants could see being touched compared to the rubber hand for which participants could only predict the touch, since decreasing the availability of sensory information generally leads to a less vivid perceptual experience (Stein & Meredith, 1993). As shown in many studies, the RHI can emerge from the integration of visual and tactile signals, so it seemed reasonable to assume that visual feedback from the actual touch event should significantly contribute to the illusion. Notably, the present results suggested that the participants could represent the moving object behind the occluding screenbased on its observed trajectory before disappearingand use this information to predict the exact impact on the rubber finger. This visuotactile prediction was then integrated with the afferent tactile signals into a coherent multisensory episode. Of course, this result does not mean that the visual confirmation of touch is not involved in the RHI or body ownership. In the approach-occluded condition, the participants still felt ownership for the rRH even if they could not see the robot arm approach the hand and saw only the touch events. This condition just induced a weaker RHI. For example, in the approach-occluded condition, in the trials when both rubber hands were touched synchronously, most of our participants still chose the rRH as their hand in at least one of 12 trials and half of them at least twice (see Supplementary material, Fig. S3). Of course, some of these "rRH" answers could be a mistake from the participant; an error lapse is always to be considered in psychophysics. However, this lapse is generally assumed to explain less than 6% of participant responses (Wichmann & Hill, 2001). If the confirmation of touch (without visually induced sensory predictions) was insufficient to elicit the RHI on the rRH, we should have observed significantly fewer rRH choices than we do. In this specific condition, i. e., the synchronous stimulation in the approach-occluded condition, the participants chose the rRH over the lRH in 17 +/− 18% of the trials on average; in the approach-occluded condition in general, i.e., across all asynchronies, they chose the rRH in 30 +/− 11% of the trials. Thus, these ratios are far above the 6% error lapse (a descriptive posthoc t-test included purely for this discussion showed a greater number of observed rRH answer than explained by the lapse: in the synchronous condition: t = 3.2106, df = 27, p = .002; for all asynchrony confounded: t = 11.814, df = 27, p < .002).
However, our results suggested that the visual information about the object moving towards the rubber hand was relatively more important for the elicitation of the RHI compared to visual information about the actual touch. One possible explanation for this finding could be a ceiling effect: the visually induced tactile predictions carry enough information to elicit a maximal RHI, i.e., the visual and tactile information allowed the participant to perceive and predict the entire multisensory episode, including the unseen touch on the occluded object with minimal uncertainty. The visual information about the actual touch did not increase the multisensory event precision, as this perception was already optimal by integrating the visuotactile predictions and afferent tactile signals. Such a relative weighting of sensory information occurs in other perceptual tasks (Blanchard, Roll, Roll, & Kavounoudias, 2011, 2013Fetsch, Turner, DeAngelis, & Angelaki, 2009). In line with this interpretation, predictions about impact location were unaffected by the inability to see the complete trajectory of an approaching object (Neppi-Mòdona et al., 2004). Additionally, tactile sensitivity was enhanced at the predicted time and impact location, regardless of whether the participant could see the moment of contact (Kandula et al., 2015). Thus, according to this view, the central nervous system needs only a sufficient amount of visual information to effectively represent the dynamic episode in a predictive manner. Visual information that is "filled in" (inferred and extrapolated) can be integrated with somatosensory signals and lead to flexible updating of the multisensory body representation.
The most common theoretical approach used to explain the weighting of different kinds of sensory information in multisensory perceptual tasks is maximum likelihood estimation (MLE). The MLE principle states that each sensory input is proportionally weighted to its relative reliability in the combination process that leads to a multisensory percept (Ernst & Banks, 2002). In the present experimental context, we could conclude that the visuotactile predictions generated by the view of the approaching robot's probe were so reliable that the added value of the visual confirmation of touch was minimal. This interpretation relies on the concept of object permanence, i.e., humans' (and many animals) capacity to represent the existence and trajectory of a disappearing object (Piaget, 1952). Not only do visuotactile cortical neurons continue to track an object's spatial position in relation to the body after it disappears from view (Graziano et al., 1997b) but also there is evidence that the predictive tracking of an occluded object is accurate if the trajectory is short and linear (Bertenthal, Longo, & Kenny, 2007;Gredeback & von Hofsten, 2004), which was the case in our experiment (touch-occluded condition).

Remaining questions and limitations
Our study constitutes only a first step in the understanding of the role of multisensory predictions in body ownership, and several interesting follow-up studies can be envisioned. In the present study, the movement trajectories were always natural, stable, and predictable, in line with the ecological validity we were striving for. However, future studies could manipulate the accuracy, uncertainty, and predictability of movementrelated information to investigate how such factors influence the RHI. Another question worth further investigation is the minimum viewing time of the approaching movement that is required for effective predictions to be generated. One could expect that the shorter the viewing period, the less reliable the predictions, leading to reduced integration of afferent sensory information and predictions, which in turn would result in weaker RHI.
Further, experimental manipulation of prediction reliability and prediction errors is needed to examine the learning processes that mediate the updating of multisensory hand representation in the RHI. It has been suggested that as a general strategy, the brain infers the cause of the sensory information it receives and "avoids surprises" by minimizing the prediction errors (Friston, 2005(Friston, , 2010 at many levels, including the perceptual level (Den Ouden et al., 2012). Applied to multisensory body representation and the feeling of ownership (Apps & Tsakiris, 2014;Limanowski & Blankenburg, 2013), the same goal of minimizing prediction error should be observable in the RHI as well. In the current study, a sequence of six touch episodes was used to elicit the RHI over 12 s, during which time the RHI gradually developed (Ehrsson et al., 2004) based on accumulation of evidence based on the multisensory correlations (Ehrsson, 2020). The visually induced tactile predictions from observing the movement of the object towards the rubber hand probably develop gradually during this time course, and as the rubber hand feels more like one's own, the stronger the tactile predictions of self-touch. Thus, it would be interesting in future studies to examine the time course of the development of visuotactile predictions in the RHI, their relationship to the subjective experience of body ownership over time, and the effect of manipulating prediction errors during this phase, for example, by omitting expected tactile feedback.

Conclusion
To conclude, the current study demonstrates that visually induced visuotactile predictions contribute to the RHI. The influence of these sensory predictions on body ownership is so significant that just seeing the object moving towards the rubber handbut without witnessing the actual contactis sufficient to elicit the RHI in combination with synchronous tactile feedback from the real hand. The illusion triggered in this way is equally strong as the classic version with full visibility but stronger than when seeing the actual touch on the rubber hand without the incoming object's preceding visual impressions. These findings suggest that the multisensory integration processes that support body ownership and the RHI are anticipatory. Both predicted sensory events and actual sensory feedback are combined to generate coherent multisensory precepts of dynamic episodes involving one's own body. Thus, the current study advances our understanding of the basic multisensory mechanisms that underpin body ownership and the balance between top-down predictions and afferent sensory signals in the construction of the sense of bodily self.

Funding
The project was funded by The Swedish Research Council, Torsten Söderbergs Stiftelse, and Göran Gustafssons Stiftelse, and M.C. was supported by a postdoctoral grant from the Wenner-Gren foundation.