Action Enhances Predicted Touch

Method

We created a contact condition in Experiment 1 to determine whether we could replicate typical action-attenuation effects within our setup. Participants therefore moved their right index finger to make contact with a button, generating a simultaneous mechanical force to their left index finger below (see Procedure and Fig. 1a). We thereby examined whether left-hand stimulation is reported as less forceful during right-hand action (active trials) than when the right hand remains still (passive trials). We were also interested in the nature of effects in a no-contact condition, in which a similar downward motion of the right index finger triggered the same left-hand stimulation even though the right-hand finger did not make contact with a button. Stimulation to the left hand was instead triggered when a motion tracker detected movement of the right index finger. Given that intransitive actions frequently produce sensory effects, active finger contact should not be required to form sensorimotor predictions per se, and this generation of an active-finger sensory event simultaneous with target stimulation especially complicates interpretation. Specifically, studies examining predictive attenuation have focused on perception of events on passive effectors due to potential confounds of “generalized gating” when measuring perception on active effectors. Generalized gating is thought to reduce perception of any events delivered to moving effectors (Williams & Chapman, 2000) without exhibiting specificity to predicted consequences because it is hypothesized to occur at the earliest relay in the spinal cord (Seki & Fetz, 2012). One possible reason that active finger contact complicates interpretation of attenuation in contact setups is that concurrent gated sensory events on active effectors could bias responses about stimulation on passive effectors, for instance, because of response biases (Firestone & Scholl, 2016). If this is the case, we would expect to observe effects in the contact condition that are different from those observed in the no-contact condition, considering that this nonpredictive influence was removed in the latter.

OPEN IN VIEWER

Fig. 1.

Paradigm and results from Experiment 1. On each trial (a), participants made downward movements with their right index finger either over a motion tracker (no-contact condition) or to make a button press (contact condition). Each movement elicited a tactile punctate event to the left index finger positioned directly below. Points of subjective equality (PSEs; b) were calculated for each participant and indexed the point at which participants responded that the target stimulus intensity was the same as the subsequent reference stimulus intensity. Data from an example participant are shown for active trials (in which participants were cued to move their right index finger; dark blue slope) and passive trials (in which participants were cued to not move their finger; light blue slope) in the no-contact condition. Dashed lines indicate the PSEs for this participant. Mean PSEs in the contact and no-contact conditions (c) are shown for active trials (darker bars) and passive trials (lighter bars). Lower PSEs indicate more intense target percepts. Asterisks indicate significant differences between trial types (*p < .05, **p = .001). Error bars show standard errors of the mean. The PSE effect of movement (passive trials – active trials) for the contact and no-contact conditions (d) is plotted with rain-cloud plots. The shaded areas display probability density estimates. Box plots denote lower and upper quartiles (left and right sides, respectively), along with the middle quartile (median). Whiskers denote 1.5 times the interquartile range, and dots denote difference scores for each participant (N = 30). Positive effects of movement indicate more intensely perceived active target events relative to passive events, but negative values indicate the reverse—less intensely perceived active events.

Results

PSE values were analyzed in a 2 (contact: contact vs. no contact) × 2 (movement: active vs. passive) within-participants analysis of variance, which revealed no main effect of contact, F(1, 29) = 3.11, p = .089, η _p ² = .10, or movement, F(1, 29) = 1.24, p = .274, η _p ² = .04. However, there was a significant interaction between contact and movement, F(1, 29) = 15.39, p < .001, η _p ² = .35. This effect was driven by lower force judgments (higher PSEs) in active (M = 4.73, SD = 1.22) compared with passive (M = 4.35, SD = 0.80) trials in the contact condition, t(29) = 2.07, p = .047, d = 0.38, but higher force judgments (lower PSEs) in active (M = 3.82, SD = 1.11) compared with passive (M = 4.45, SD = 0.97) trials in the no-contact condition, t(29) = −3.80, p = .001, d = 0.69 (see Fig. 1c).

Method

Experiment 1 replicated previous findings (Bays et al., 2005; Kilteni et al., 2019) that tactile events on a stationary left finger are perceived less intensely during active right-hand movement, but only when the active finger makes contact with a button, as in the typical version of the paradigm. When there is no contact, events are perceived more intensely during movement. One possible explanation for the difference between conditions is that perception of gated stimulation on the active effector contributed to the active–passive difference in the contact condition.

Having established that we can observe typical attenuation during action—but that attenuation becomes enhancement in a no-contact condition—in Experiment 2, we isolated the particular functional influence of prediction mechanisms by manipulating conditional probabilities between actions and outcomes. This is particularly important for establishing the role of action predictions in determining perception because—as outlined in the introduction—an active–passive comparison does not isolate predictive influences of action even in a no-contact condition, for example, because it confounds the number of tasks. Therefore, we compared perception of tactile events when they were expected or unexpected on the basis of learned action–outcome probabilities established in a preceding training session but always in the presence of action. Although downweighting accounts predict that expected tactile events will be rated less intensely than unexpected events, upweighting theories predict that expected events will be rated more intensely.

There were several procedure changes relative to Experiment 1. Each participant now performed one of two movements that predicted one of two tactile effects (see Fig. 2a). Participants were positioned with their left index and middle fingers making contact with independent solenoids (see Fig. 2a). At the beginning of each trial, an arbitrary cue (either a square or circle) instructed participants to move their right index finger either upward or downward from the metacarpophalangeal joint, tracked by an infrared motion sensor. This action triggered delivery of the target stimulus to either solenoid in the same way as in the no-contact condition of Experiment 1 (for more details, see the Supplemental Material). During training, participants’ right index-finger action (e.g., downward movement) was 100% predictive of the location of left-hand tactile events (e.g., index finger). In a test session 24 hr later, the action–outcome relationship was degraded to measure perception of expected and unexpected events—the expected finger was stimulated on 66.7% of trials, and the unexpected finger was stimulated on the remaining 33.3% of trials.

OPEN IN VIEWER

Fig. 2.

Paradigm and results from Experiments 2 and 3. On each trial in Experiment 2 (a), participants made a downward or an upward movement with their right index finger over a motion tracker, which elicited tactile punctate events to the left index or middle finger. In Experiment 3 (b), participants made only downward movements with either their right index or middle finger, which again elicited tactile punctate events to the left index or middle finger. An example training procedure (c) is shown for Experiment 2. Movements were 100% predictive of tactile events during the training session, and 66.7% predictive in the test session. This procedure was also adopted in Experiment 3 but with different action types. Flash symbols illustrate locations of stimulation. Mean points of subjective equality (PSEs) in Experiments 2 and 3 (d) are shown for expected trials (darker bars) and unexpected trials (lighter bars). Lower PSEs indicate a more intensely perceived target stimulus. Asterisks indicate significant differences between trial types (*p < .05). Error bars show standard errors of the mean. The PSE expectation effect (unexpected trials – expected trials) for each experiment (e) is plotted with rain-cloud plots. The shaded areas display probability density estimates. Box plots denote lower and upper quartiles (left and right sides, respectively), along with the middle quartile (median). Whiskers denote 1.5 times the interquartile range, and dots denote difference scores for each participant (N = 30 for each experiment). Positive expectation-effect values indicate more intensely perceived expected events relative to unexpected events.

Presenting two action types and two stimulation types also allowed us to compare perception of expected and unexpected events while controlling for repetition effects. It should be noted that any action predictions should determine where stimulation will be received, rather than its intensity. However, in Bayesian models, it is assumed that enhanced detection and intensity of expected events relates to the precision of the estimate (Brown et al., 2013; e.g., a force precisely estimated to have occurred on a certain region of tactile space should feel more intense because of the precise estimate of spatial information rather than an estimate of the force per se). Because these models assume that predictions enhance the precision of resultant estimates, they would also predict enhancements in perceived force (and indeed other sensory attributes, such as brightness or loudness).

There were 420 trials in each session. Trial order was randomized, and the action–stimulus mapping was counterbalanced across participants. The cue–action mapping was reversed halfway through each session to account for effects resulting from possible learning of cue–outcome associations instead of action–outcome associations. The training and test sessions were carried out at the same time on consecutive days. Participants completed 12 practice trials before the main session trials.

Results

Psychometric functions were modeled to participants’ responses similarly as in Experiment 1 but now separately for expected and unexpected events. PSE values were lower on expected trials (M = 3.72, SD = 0.96) than unexpected trials (M = 3.93, SD = 0.80), t(29) = −2.13, p = .041, d = 0.39 (see Fig. 2d), demonstrating more forceful perception of expected than unexpected action outcomes.

Method

Experiment 3 was designed to provide a conceptual replication of Experiment 2 using a setup more closely aligned with typical action paradigms (e.g., Experiment 1)—whereby one always makes a movement toward another effector. We additionally controlled for possibilities that the expectation effect in Experiment 2 resulted from cue–outcome learning by removing the cue stimulus and requiring free selection of action. The explicit reference stimulus was also removed, and comparisons were made against an implicit reference, eliminating the possibility that effects would be determined by forming predictions about the reference stimulus.

There were several procedure changes relative to Experiment 2. Independent solenoids were now attached to the left index and middle fingers via adhesive tape (diameter of metal rod = 4 mm, diameter of solenoid = 18 mm; TactAmp 4.2 solenoid controller box; Dancer Design, St. Helens, UK). The foot pedals were positioned at either a 45° angle or 90° angle (for stronger and weaker responses, respectively) relative to participants’ right foot to record responses to account for any spatial biases resulting from positioning foot pedals as “left” and “right.” At the start of each trial, participants made a downward movement with either their right index or middle finger. These action types were selected to ensure that effects in Experiment 2 were not specific to those action types and to determine whether similar effects could be observed with actions that are always made toward another effector. Participants’ hands, and therefore index and middle fingers, were spatially aligned with each other (see Fig. 2b). Actions were freely selected, and the frequency of index- and middle-finger movements was monitored to ensure approximately equal numbers of both action types. Participants’ actions (e.g., downward movement of right index finger) were still perfectly predictive of the location of tactile events (e.g., left index finger) during training, and this contingency was again degraded to 66.7% in the following test session. Participants were asked whether they perceived the test force to be more or less forceful than the average force intensity. An example of the average force was presented to each finger once at the end of short breaks every 21 trials (note that the average force was identical to the intensity of the reference force).

The experiment consisted of two training blocks followed by a test block, all occurring in the same session of testing. In the first training block, participants responded “yes” or “no” to the question “Tap on index or middle finger?” and in the second training block, they were asked about the force, as in Experiment 2 and in subsequent test blocks. For half of the participants, moving the right index finger resulted in stimulation of the left index finger, and moving the right middle finger resulted in stimulation of the left middle finger. For the other half of the participants, moving the right index finger resulted in stimulation of the left middle finger, and moving the right middle finger resulted in stimulation of the left index finger. There were 210 trials in each session.

Results

As in Experiment 2, PSE values were lower in expected (M = 4.08, SD = 1.00) than unexpected (M = 4.28, SD = 0.99) trials, t(29) = −2.56, p = .016, d = 0.47 (see Fig. 2d), again demonstrating more forceful perception of expected than unexpected action outcomes. Additional post hoc analyses revealed that the specific kinematics of action were similar in expected and unexpected trials and that the PSE expectation effect was comparable at the start and end of 21-trial miniblocks (see the Supplemental Material).

Method

The present findings are consistent with predictive-upweighting theories of perception, which propose that observers combine sensory evidence with prior knowledge—biasing perception toward what we expect (Kersten et al., 2004). This may be achieved mechanistically by altering the weights on sensory channels, increasing the gain of expected relative to unexpected signals (de Lange et al., 2018). However, expectation effects may instead reflect biasing in response-generation circuits if action biases people to respond that expected events are more intense rather than altering perception itself (Firestone & Scholl, 2016).

Perceptual and response biases can be dissociated in computational models that conceptualize perceptual decisions as a process of evidence accumulation. Perceptual biases are modeled as growing across time—every time response units sample from perceptual units, they will be sampling from a biased representation, therefore increasing the magnitude of biasing effects across a larger number of samples (Yon et al., 2021). In contrast, response biases are modeled as operating regardless of current incoming evidence and to be present from the outset of a trial. According to this logic, we can model the decision process with drift-diffusion modeling (Ratcliff & McKoon, 2008) to identify the nature of the biasing process. We can thus establish whether action expectations shift the starting point of evidence accumulation toward a response boundary (“start biasing”; z parameter; see Fig. 3a) or instead bias the rate of evidence accumulation (“drift biasing”; db parameter; see Fig. 3b).

OPEN IN VIEWER

Fig. 3.

Illustration of how the drift-diffusion model could explain expectation biases (top row) and results of computational modeling (bottom row). For an unbiased decision process (black lines in a and b), sensory evidence integrates toward the upper response boundary when stimuli are stronger than average (solid lines) and toward the lower response boundary when weaker than average (dotted lines). In a start-bias model (blue lines in a), baseline shifts in decision circuits could shift the start point of the accumulation process nearer to the upper boundary for expected events (influencing the parameter z). Alternatively, in a drift-bias model (red lines in b), selectively altering the weights on sensory channels could bias evidence accumulation in line with expectations (influencing parameter db). The simulated start-plus-drift-bias (winning deviance information criterion [DIC] model) expectation effect is plotted against the empirical expectation effect in (c). The simulated drift-bias expectation effect is plotted against the empirical expectation effect after accounting for simulated start-bias effects (plotted as the residuals from a model in which the simulated start-bias effect predicts the empirical effect) in (d). All expectation effects were calculated by subtracting expected points of subjective equality (PSEs) from unexpected PSEs.

We fitted drift-diffusion models to participant choice and reaction time data from Experiment 3 using the hDDM package (Version 0.6.0; Wiecki et al., 2013) implemented in Python (note that reaction times were not collected in Experiments 1 and 2). In the hDDM, model parameters for each participant are treated as random effects drawn from group-level distributions, and Bayesian Markov chain Monte Carlo sampling is used to estimate group- and participant-level parameters simultaneously. We specified four different models: (a) a null model in which no parameters were permitted to vary between expected and unexpected trials, (b) a start-bias model in which the start point of evidence accumulation (z) could vary between expectation conditions, (c) a drift-bias model in which a constant added to evidence accumulation (db) could vary according to expectation, and (d) a start-plus-drift-bias model in which both parameters could vary according to expectation.

All models were estimated with Markov chain Monte Carlo sampling, and parameters were estimated with 30,000 samples (burn-in = 7,500). Model convergence was assessed by inspecting chain posteriors and simulating reaction time distributions for each participant. Models were compared using the deviance information criterion (DIC) as an approximation of Bayesian model evidence, a common method used to determine model fit. Lower DIC values relative to a baseline, or null, model are indicative of a better model fit.

A posterior predictive check was conducted using the hDDM package to establish how well each model was able to reproduce the patterns in our data. The posterior-model parameters for the start-bias, drift-bias, and start-plus-drift-bias models were used to simulate a distribution of 500 reaction times and choices for each trial for each participant. From these simulated data, we calculated the probability that a stronger than average response was given at each intensity level, separately for expected and unexpected trials. This allowed us to model simulated psychometric functions for expected and unexpected trials, exactly as we had done for empirical decisions. Performing this procedure for each model yielded separate simulated expectation effects (unexpected PSE – expected PSE) for each participant under the start-bias, drift-bias, and start-plus-drift-bias models.

Results

Fitting the drift-diffusion model to the behavioral data revealed that the model allowing both start and drift biases to vary according to expectation provided the best fit (DIC relative to null = −234.8) relative to both the start-bias (DIC relative to null = −191.06) and drift-bias (DIC relative to null = −8.62) models. This finding may suggest that observed biases are a product of both start and drift-rate biasing. However, although the DIC measure does include a penalty for model complexity, it is thought to be biased toward models with higher complexity (Wiecki et al., 2013), and it indeed favored the most complex model here.

We conducted a posterior predictive check to evaluate how well simulated data from each of the models could reproduce key patterns in our data. Correlations were calculated to quantify how well simulated expectation effects reproduced empirical expectation effects, which revealed significant relationships for all three models—start-bias model: r(30) = .39, p = .034; drift-bias model: r(30) = .43, p = .017; start-plus-drift-bias model: r(30) = .53, p = .003 (see Fig. 3c).

Given that we were interested in whether any of the PSE expectation effects were generated by sensory biasing—rather than possible additional contributions of response biasing—we examined whether drift biasing accounted for any further variance in expectation effects than start biasing alone by conducting a stepwise linear regression to predict the empirical expectation effect (unexpected PSE – expected PSE). In the first step, we included the simulated expectation effect from the start-bias model to predict the empirical expectation effect. The simulated start-bias data were able to predict the empirical expectation effect, R² = .15, F(1, 28) = 4.96, p = .034. In the second step, we included the simulated expectation effect from the drift-bias model as an additional predictor of the empirical expectation effect, importantly providing a significant improvement to the model fit, ΔF(1, 27) = 6.72, p = .015; R² = .32, F(2, 27) = 6.34, p = .006. This analysis revealed that a model implementing a drift-biasing mechanism better predicted empirical effects of expectation on perceptual decisions by explaining unique variance in participant decisions that cannot be explained by start biasing.