Proprioceptive Precision and Degree of Visuo-Proprioceptive Discrepancy Do Not Influence the Strength of the Rubber Hand Illusion

Introduction

There is widespread consensus that passively induced rubber hand illusion (RHI; Botvinick & Cohen, 1998) arises as a result of the integration of visual, tactile, and proprioceptive signals (Samad, Chung, & Shams, 2015). Vision provides information both about the hand’s position in external space as well as temporal and spatial properties of tactile stimulation. On the skin, spatiotemporal parameters are also estimated by touch. Finally, proprioception gives information about the hand’s orientation and position in space. Due to the superior precision of visual signals in providing spatial information, proprioceptive representation of one’s hand shifts onto the visual reference frame, carrying tactile sensations with itself, provided that the seen and felt brushstrokes are spatiotemporally congruent.

Proprioceptive signals seem to constrain RHI in many ways: The illusion was reported to be stronger at shorter distances between the real and fake hand (Kalckert & Ehrsson, 2014; Lloyd, 2007), sensitive to angular discrepancies in hand orientation (Costantini & Haggard, 2007), and anatomically implausible arrangements of the rubber hand prevent the illusion from occurring (Tsakiris & Haggard, 2005). However, the literature offers mixed evidence. The distance between hands affects proprioceptive estimations of one’s hand position rather than body ownership (Abdulkarim & Ehrsson, 2016; Hinz, Lanillos, Mueller, & Cheng, 2018). Another study reported that comparably strong RHI may be elicited for various angular positions of a fake hand (Ide, 2013).

Here, we aimed to test the predictions of a recent computational Bayesian bottom-up model of RHI (Samad et al., 2015). In this model, a cognitive system carries out Bayesian inference to determine whether visual, tactile, and proprioceptive signals share a causal origin (illusion present) or not (illusion absent). The outcome experience reflects the hypothesis with higher posterior probability derived from the prior probability of a common cause and the likelihood of sensory signals. Samad et al. (2015) propose that only two factors influence likelihood distributions: the degree of (a) discrepancy between visual and proprioceptive spatial estimates of the hand’s position and (b) temporal (in)coherence of seen and felt brushstrokes. As, in their model, the prior does not favour either alternative, the inference relies on the properties of sensory signals: unimodal estimates and their relative precisions (noisiness). Thus, the model predicts that the illusion is stronger the nearer the fake and real hand are to each other, the noisier the proprioception modality is, (and) the more congruent the temporal pattern of stroking is across visual and tactile modalities (p. 19).

This model (Samad et al., 2015) reproduced Lloyd’s (2007) findings of attenuated RHI for greater visuo-proprioceptive discrepancies. However, such effect was not observed in the study by Abdulkarim and Ehrsson (2016) in which RHI was elicited during shifts in position of one’s real hand—either towards or away from the rubber hand. Notably, as end positions were in the same place for both drifting away and drifting towards conditions, the time-averaged distance between hands was shorter in the former than in the latter. Therefore, the lack of differences might have resulted from contradictory effects of the hand’s average position and direction of the shift cancelling each other out. To account for this, we employed a slightly modified procedure with the displacement of one’s hand either to the position closer or farther from the rubber hand (starting from the same position and taking place prior to elicitation). Furthermore, we assessed interindividual differences in proprioceptive accuracy and precision with a task requiring active reproduction of one’s arm position (Lubiatowski et al., 2013). In this way, we aimed to test the predictions of the Bayesian bottom-up model (Samad et al., 2015), that the illusion should be stronger for (a) shorter distance between the hands and (b) lower precision of proprioceptive signals. Also, we expected to observe the interaction of these factors, as precise proprioception should allow detection of a smaller spatial discrepancy leading to greater attenuation of the illusion at larger distances.

Method

Procedure

Proprioceptive accuracy/precision assessment

Proprioception was assessed using active joint position reproduction task (Lubiatowski et al., 2013) in which blindfolded participants tried to reproduce their hand’s position as accurately as possible. Two movements (abduction/flexion of the right arm) and three target positions (60°, 90°, and 120° deviation from a vertical axis) were utilised. After setting up the target position, participants were asked to stabilise their arm and press a button on a device held in their other hand. Next, the experimenter slowly pulled down their arm to the default position (0°, lowered arm). Then, participants were asked to reproduce the target position and report it with a button press. Participants performed the task 5 times in each position, which resulted in 30 reproductions per participant. Proprioceptive accuracy was operationalised as inversely related to mean difference between target and reproduced positions (error) and precision as inversely related to the variance of errors distribution. The discrepancies between target and reproduced positions were measured using the Propriometer R system, including an arm-mounted electric goniometer with 0.1° precision measurement (Lubiatowski et al., 2013).

RHI elicitation and subliminal displacement of participants’ hands

To minimise the possibility of cognitive bias resulting from knowledge of one’s hand position (Erro, Marotta, Tinazzi, Frera, & Fiorio, 2018), we used subliminal hand displacement procedure similar to the one designed by Abdulkarim and Ehrsson (2016; Figure 1(b), see Supplemental Material for a video and additional details). The rubber hand was placed on the top surface directly in front of the participant’s right shoulder, while the real hand was hidden underneath on the hand displacement sheet. The hands were separated by 12.5 cm on the vertical axis and by 16 cm on the horizontal axis (starting position). The real hand was displaced with the use of an external driver either towards or away from the rubber hand—to a position 8 cm (close condition) or 24 cm from the rubber hand (far condition) on the horizontal axis. The velocity of displacement was set to 0.9 mm per second and was, therefore, undetectable (Pickett & Konczak, 2009). The Bayesian bottom-up model predicts 100% probability of the illusion for the former distance and ca. 50% for the latter (Samad et al., 2015; Figure 3), which makes the distances appropriate for testing the hypotheses.

OPEN IN VIEWER

Figure 1.

Experimental procedure. (a) Active joint reproduction task. The picture shows flexion at 90°. (b) Experimental manipulation. (c) Illusion elicitation.Note: Please refer to the online version of the article to view the figures in colour.

For illusion elicitation, we used a realistic, life-sized right rubber hand. The space between the participant’s neck and the rubber hand was occluded. In contrast to the original study (Abdulkarim & Ehrsson, 2016), the illusion was elicited after, rather than during, the displacement, to isolate exact hand positions in both close and far conditions. In a short post hoc questionnaire, none of the participants reported having been aware of the displacement. After the real hand reached its final position, visuotactile stimulation was used to elicit RHI (Figure 1(c)). In the synchronous condition, both hands were brushed simultaneously, while in the control asynchronous condition, strokes were approximately 500 ms out of phase. As asynchronous condition is supposed to disrupt visuo-proprioceptive recalibration (Abdulkarim & Ehrsson, 2016), it always followed synchronous condition to preclude the possibility that the participants would notice the displacement. Thus, the difference in RHI strength between these conditions served only as a manipulation check and synchrony factor was not included in further analyses. The stimulation in both conditions lasted 2 minutes and was standardised using auditory beeps signalling timings of the strokes, audible only to the experimenter. We utilised audio tracks developed by Fuchs, Riemer, Diers, Flor, and Trojan (2016) with slightly extended beep durations. The pattern of brushstrokes was pseudo-random and irregular, as unpredictable stimulation results in stronger illusion (Armel & Ramachandran, 2003).

To measure the subjective strength of the illusion, we adapted questionnaire items (on a 7-point Likert-type scale) and the computation of an illusion strength index from Abdulkarim and Ehrsson’s (2016) study. Items S1 to S3 concerned the illusion, whereas S4 to S6 served to control for task demand or suggestibility effects (Items S7 to S9 were not used in our study; see Table 1 in Supplemental Material for the full questionnaire). The illusion strength index was defined as the difference between the means of the illusion-related questions (S1–S3) and the control items (S4–S6) and was calculated separately for synchronous and asynchronous conditions.

EDA measurement and analysis

For a psychophysiological measure of the strength of the illusion, we utilised event-related electrodermal activity (skin conductance response; SCR) in a situation endangering the rubber hand (Armel & Ramachandran, 2003). Electrodermal activity was recorded using a Biopac EDA100C amplifier and AcqKnowledge 5.0 software. The data sampling rate was 2,000 data points per second with the gain set to 10 µS/V. The electrodes were placed on nonadjacent distal phalanges of the participant’s left hand. Subsequently, subjects waited 5 minutes until their electrodermal activity reached baseline state and then proceeded to the experiment. In both conditions, the dummy was threatened 5 times: The first knife attack occurred 40 seconds after the stimulation started and the following attacks were separated by 20-second intervals.

Before analysis, following Braithwaite, Watson, and Dewe (2017), we normalised the data using log (SCR + 1) correction. Then, the amplitude of SCR was standardised into z-scores based on each participant’s mean and standard deviation derived from the recordings from both synchronous and asynchronous conditions. During each attack, SCR amplitude was analysed in a time window encompassing 1 to 6 seconds after stimulus onset (Figner & Murphy, 2010). To account for the possibility that scores could also reflect a continually declining signal, cases with maximum value within the first second were labelled as no evoked response, with amplitude value set to 0 (this concerned 14 attacks from 9 participants in the synchronous condition and 72 attacks from 40 participants in the asynchronous one). The strength of the illusion was operationalised as mean SCR amplitude of five knife attacks. All preprocessed data and the data analysis code are available on GitHub at https://github.com/Pawel-Motyka/RHI_proprioception.

Results

Concerning the results of proprioception assessment, mean error in the active joint position reproduction task—the (inverse) index of accuracy—equalled 4.93° (SD = 1.59°), whereas variance of the errors—the (inverse) measure of precision—was 5.34° (SD = 1.61°). Proprioceptive accuracy and precision were found to be strongly positively correlated, r_s = .85, p < .001 (see Figure S1 in Supplemental Material). As both variables appeared collinear, we considered only proprioceptive precision in further analyses (as directly related to our hypotheses).

As for manipulation check, we successfully induced the illusion: Subjective body ownership was greater for synchronous (M = 2.05, SD = 1.94) as compared to asynchronous stimulation (M = 0.27, SD = 1.60), t(48) = 5.12, p < .001, Cohen’s d = .73. Similarly, electrodermal response was stronger in the synchronous (M = 1.15, SD = 0.55) than in the asynchronous (M = 0.34, SD = 0.39) condition, Z = 5.82, p < .001, r = .82.

To test our hypotheses, mixed linear regression models were run separately for subjective and physiological measures of illusion strength, each with proprioceptive precision (continuous variable), distance (between-subjects factor: close vs. far), and the interaction between these variables as predictors. The regression model for subjective ratings of body ownership was not significant, R² = .009, F(3, 45) = .14, p = .94. Neither the main effects of distance, β = .16, t = .30, p = .51, and proprioceptive precision, β = .15, t = .59, p = .56, nor interaction, β = −.19, t = −.32, p = .75, were significant predictors of illusion strength (Figure 2(a); for the full questionnaire results, see Figure S2 in Supplemental Material).

OPEN IN VIEWER

Figure 2.

Subjective body ownership (a) and electrodermal activity (b) as a function of proprioceptive precision (operationalised as variance of proprioceptive errors: x-axis) and distance between the real and fake hand (separate lines for each condition: 8 cm vs. 24 cm). For both measures, neither these factors nor their interaction were significant predictors of RHI strength.Note: Please refer to the online version of the article to view the figures in colour.

To assert that non-significant results reflected null effects rather than experimental insensitivity, we computed Bayes factors (BF; Dienes, 2014) using the BayesFactor R package (Morey & Rouder, 2018 with the default Jeffreys--Zellner--Siow (JZS) prior; Rouder, Morey, Speckman, & Province, 2012) for two alternative linear models—with and without interaction—against the intercept only model and found strong evidence for null hypothesis for both of them (BF₁₀ = 0.040 and BF₁₀ = 0.091, respectively; Figure 3). Next, Bayes factors were calculated for the effects of distance and precision separately. We found substantial evidence for null hypothesis for both of these factors (distance: BF₁₀ = 0.285 and precision: BF₁₀ = 0.325; Figure 3).

OPEN IN VIEWER

Figure 3.

Null effects of proprioceptive precision, distance between hands, and the interaction of these factors for both subjective and physiological measures of illusion strength. The y-axis presents alternative linear models against the intercept only model. The x-axis depicts Bayes factors obtained for these models. The dotted lines mark boundaries between proposed categories of evidence strength (Jeffreys, 1961).Note: Please refer to the online version of the article to view the figures in colour.

As for electrodermal activity, the regression model also did not significantly predict illusion strength, R² = .078, F(3, 45) = 1.27, p = .30. Neither main effects of distance, β = .20, t = .38, p = .35, and proprioceptive precision, β = .23, t = .94, p = .56, nor interaction: β = −.04, t = −.07, p = .94, were statistically significant (Figure 2(b)). Here, we also report substantial evidence for null hypothesis for the model with interaction (BF₁₀ = 0.144) and anecdotal support for the model without interaction (BF₁₀ = 0.385) and unifactorial models (distance: BF₁₀ = 0.524 and precision: BF₁₀ = 0.794; Figure 3).

Discussion

In this article, we tested two predictions based on the computational model proposed by Samad et al. (2015); namely, that the strength of the illusion is inversely proportional to (a) the distance separating rubber hand from the real one and (b) the precision of proprioceptive signals. Contrary to these predictions, we report evidence that neither the distance between the hands, nor proprioceptive precision, nor interaction of these two factors influence illusion strength measured with subjective ratings of body ownership. Tentative evidence for the lack of influence of these factors on RHI was also found for an indirect physiological measure (electrodermal response to a situation threatening the rubber hand).

These results suggest that passive proprioceptive signals may be of minimal importance for RHI induced by visuotactile stimulation. However, this does not imply that proprioception is irrelevant for sense of ownership in general. Various other combinations of coherent multisensory signals have been used for passive elicitation of bodily illusions—tactile-proprioceptive (as in the somatic rubber hand illusion; Ehrsson, Holmes, & Passingham, 2005) or visuo-proprioceptive (illusion arises without/prior to delivery of tactile signals; Samad et al., 2015; Walsh, Moseley, Taylor, & Gandevia, 2011). Surprisingly, even passive movement signals from muscle spindles, without any visual or tactile feedback, may be sufficient to elicit an illusory sense of ownership over an artificial finger (Héroux, Walsh, Butler, & Gandevia, 2013). It seems that a cognitive system may use various sources of information, depending on their availability, to determine where its boundaries are.

Thus, our findings highlight the need to nuance contemporary Bayesian models of RHI to better identify the conditions in which proprioceptive signals contribute to feelings of body ownership. Based on our results, we argue that, when all sources are available, visuotactile spatiotemporal coherence may override visuo-proprioceptive incongruency regardless of the magnitude of the latter—at least for distances within peripersonal space. Speculatively, proprioceptive signals are relatively less precisely pinpointed (and, therefore, less weighted) than trajectories of visual and tactile stimulation, both mapped on high resolution spatial fields. Therefore, the detected spatiotemporal (in)congruence of visuotactile stimulation decisively affects the likelihood that both signals were caused by the same external event. Also, the visuotactile coupling prior is assigned a high value, in line with the emerging understanding of peripersonal space as a space of increased tendency to couple body-related exteroceptive signals (i.e., visual) as these approach the body (i.e., the source of tactile and proprioceptive signals) (Noel et al., 2018, p. 2). Therefore, the posterior probability distribution favours the a common cause hypothesis, regardless of the exact magnitude of the discrepancy between visual and proprioceptive estimates and individual proprioceptive abilities.

However, when tactile information is not available, proprioception may gain in relevance. In that case, likelihood will be derived solely from the degree of convergence of visual and proprioceptive spatial estimates and their relative precisions. Then, reduced distance between hands and the weighting of noisy proprioception should yield a multisensory estimate of the hand’s location closer to the rubber (observed) hand and strengthen the illusion. This was predicted and experimentally confirmed by Samad et al. (2015). Thus, our findings do not necessarily contradict Bayesian models of RHI; they do not rule out the possibility that proprioception shapes body ownership feelings when direct (tactile) information regarding body boundaries is absent. We argue that the gamut of factors influencing likelihood distributions should be extended to account for multisensory interactions giving rise to RHI. For example, spatial properties of visuotactile stimulation (i.e., trajectories of brushstrokes in the hand-centred reference frame; Costantini & Haggard, 2007) should be introduced to the model, as its presence seemingly limits the relative influence of proprioceptive signals on RHI.

Previously reported attenuation of RHI for larger distances between hands, both on horizontal (Lloyd, 2007) and vertical (Kalckert & Ehrsson, 2014) axes, may result from methodological choices. Both of these studies harnessed much simpler (single-finger, unvarying, and repetitive) stimulation patterns—brushstrokes extending from proximal to distant phalanx of the finger (Lloyd, 2007) or touches only at the proximal phalanx (Kalckert & Ehrsson, 2014). Such stimulation lessens spatiotemporal complexity of information conveyed by tactile signals, which could reduce their relative weighting in favour of proprioception. Moreover, RHI strength was calculated using different measures: median scores and number of responders in each condition (Kalckert & Ehrsson, 2014), and a single statement that was picked because it was understood most coherently and consistently and more likely to evoke affirmative responses (Lloyd, 2007, p. 106). In contrast to these measures, the index used in our study should not be prone to factors threatening internal validity of the study (e.g., task compliance or suggestibility). Finally, Kalckert and Ehrsson (2014) actually did not find differences in RHI strength between shorter distances (12 vs. 27.5 cm)—the illusion was attenuated only at the farthest distance (43 cm). The study by Lloyd (2007) remains the only one showing a linearly decreasing illusion strength with increasing horizontal distance between hands (within PPS); in addition, (a) contribution of other factors cannot be conclusively ruled out and (b) contrary evidence can also be found in the literature (e.g., Hinz et al., 2018; Kalckert & Ehrsson, 2014).

To sum up, our findings are contrary to the predictions of the Bayesian RHI model (Samad et al., 2015): We showed that the distance separating hands and proprioceptive precision do not influence the strength of visuo-haptically induced illusion. These results challenge the notion that visuo-haptically driven RHI relies on the coherence of visuo-proprioceptive estimates and highlight the need for extension of Bayesian models, so they could account for the exchangeability of multisensory signals giving rise to feelings of body ownership.

Supplemental Material