Weakened Bayesian Calibration for Tactile Temporal Order Judgment in Individuals with Higher Autistic Traits

Individuals with TD and ASD participated in our experiments (details are described below). Informed consent was obtained from all participants prior to the experiments, and the ethics committees of the National Rehabilitation Center for Persons with Disabilities approved the study. All experiments were performed in accordance with the relevant regulations and guidelines of Japan’s Ministry of Health, Labour, and Welfare.

Thirty-two paid volunteers participated in the experiments. Two were excluded because their responses could not be fitted to the cumulative Gaussian function (Eq. 1) (R² of the psychometric function < 0.6). We excluded one additional participant who exhibited inaccurate responses [σ = 191 ms > 180 ms (3rd quartile + 2 times the interquartile range)]. Consequently, we conducted the subsequent analyses on data from the remaining 29 participants [15 females, 14 males; 19–43 (26.4 ± 5.77, mean ± standard deviation) years]. There was no such exclusion in previous studies (Miyazaki et al., 2006; Nagai et al., 2012). Some low-reliability fittings would occur because of the smaller number of trials in the present experiments (248 trials/condition) compared to preceding experiments (e.g., 1000 trials/condition). In this study, we implemented fewer trials as an ethical consideration since individuals with ASD are easily fatigued by sensory stimuli.

The autistic traits of each participant were assessed using the Japanese version of the autism spectrum quotient (AQ) score (Wakabayashi et al., 2004) which is based on the English version (Baron-Cohen et al., 2001). The mean AQ score of the TD participants was 16.8 ± 9.22 (mean ± standard deviation; range: 2–35). The general status of the participants was as follows: the handedness of all participants was assessed using the Edinburgh Handedness Inventory [laterality quotient (LQ)] (Oldfield, 1971). There was one left-handed TD participant (LQ = − 73) and one ambidextrous TD participant (LQ = 50); the remaining TD participants were right-handed (LQ = 93.0 ± 9.44; range: 60–100). All participants had normal or corrected-to-normal vision (e.g., glasses, contacts).

Twelve paid volunteers with ASD participated in the experiments. Four were excluded because their responses could not be fitted to the Gaussian cumulative function. Consequently, we conducted subsequent analyses on the remaining eight participants with ASD. These participants were diagnosed by a medical doctor; diagnoses included ASD and pervasive developmental disorders (PDD) (Table 1). Several participants had additional diagnoses of attention-deficit hyperactivity disorder (ADHD) or learning disorders (LD).

All participants with ASD received the Japanese version (Kuroda & Inada, 2015) of the Autism Diagnostic Observation Schedule Component, Second Edition (ADOS-2) (Lord et al., 2012) and a Japanese version (Fujita et al., 2006) of the Wechsler Adult Intelligence Scale-III (Wechsler, 1997). The ADOS-2 scores were evaluated by an occupational therapist (M. S.) who had a research license for the ADOS-2. Table 1 presents the profiles of each participant with ASD, including any prescribed psychiatric medication.

Participants judged the orders of two successive tactile stimuli that were delivered to each hand separately. The participants placed their hands palm down on the desk and received tactile stimuli on the dorsal surface of both index fingers, which were delivered by a computer-controlled tactile stimulator using solenoids (Uchida Denshi, Tokyo, Japan) (see inset in Fig. 2b). To produce each tactile stimulus, a rectangular voltage pulse (11 V, 10 ms) was applied to each solenoid stimulator. The distance between the index fingers was 20 cm. A response button was placed under each index finger pad. During the experiments, participants closed their eyes while white noise (≈ 80 dB) was played through headphones placed over the participants’ ears with earplugs to prevent them from hearing sounds created by the tactile stimuli.

The experiment consisted of three conditions: left-first biased, right-first biased, and unbiased. The order of the conditions was counterbalanced among participants, and the three conditions were conducted on the same day.

In the left-first and right-first biased conditions, stimulus onset asynchronies (SOAs) were sampled from discrete approximations of Gaussian distributions with a mean of ± 80 ms and a standard deviation of 80 ms (Fig. 2a). Positive SOAs indicated that the right hand was stimulated first. The SOA for each trial was randomly assigned from ten possible SOAs ranging from − 260 to 100 ms in the left-first biased condition (− 260, − 220, − 180, − 140, − 100, − 60, − 20, 20, 60, and 100 ms) or from − 100 to 260 ms in the right-first biased condition (− 100, − 60, − 20, 20, 60, 100, 140, 180, 220, and 260 ms) with appearance ratios of 1, 3, 6, 9, 12, 12, 9, 6, 3, and 1 (for a total of 62 trials/epoch), respectively. Each participant experienced four epochs of SOAs for each condition. Therefore, each participant performed 62 × 4 = 248 trials for each condition.

In the unbiased condition, the SOA for each trial was randomly assigned from 14 possible SOAs ranging from − 260 to 260 ms (− 260, − 220, − 180, − 140, − 100, − 60, − 20, 20, 60, 100, 140, 180, 220, and 260 ms) with a uniform distribution (i.e., 14 trials/epoch). In this condition, 18 epochs were repeated, and thus, the total number of trials in this condition was 14 × 18 = 252.

The TOJ task was conducted in a two-alternative forced-choice manner. The participants were instructed to report their judgments by pressing the button on the side that was stimulated first as quickly as possible, as in a previous study using tactile TOJ (Yamamoto & Kitazawa, 2001). The participants were also requested to perform the TOJ task naively without any strategy. They were instructed not to intentionally anticipate the orders to be in line with those from prior trials, and not to intentionally anticipate the orders to be opposed to those in prior trials. For example, in our preliminary experiment, one participant selectively aimed to detect opposite orders to those of prior trials. The participant had attended many psychological experiments and had a preconception that her responses should not be biased towards a specific side. This instruction aimed to prevent potentially biased preconceptions and strategies.

When the reaction time was longer than 5000 ms or when the reaction was earlier than the onset of the second stimulus, we replaced the trial with the same SOA at the end of each condition. No feedback was provided to the participants.

TD participants were separated into low-AQ and high-AQ groups. The low-AQ group comprised TD participants with AQ scores of less than 26 (n = 22). The high-AQ group comprised TD participants with AQ scores of 26 or more (n = 7). This grouping was based on previous evidence that showed an AQ score of ≥ 26 is indicative of the possibility that an individual has ASD (Woodbury-Smith et al., 2005). Participants with ASD (n = 8) had diverse backgrounds (i.e., additional diagnoses and medications), as shown in Table 1.

We sorted the response data from each condition based on the SOA to calculate the proportions of the trials in which participants judged the right hand as being stimulated first. The judgment proportion (p) can be fitted to a psychometric function using a cumulative Gaussian function (Yamamoto & Kitazawa, 2001):where t, d, and σ denote the SOA, mean, and standard deviation, respectively, in the fitted Gaussian function. p_max and p_min denote the upper and lower asymptotes of the judgment proportion, respectively. In this study, d, σ, p_max, and p_min were constrained to − 400 to + 400 ms, 1–1000 ms, 0.99–1.0, and 0.0–0.01, respectively. MATLAB, with the optimization toolbox (MathWorks, Natick, MA, USA), was used for fitting using the maximum likelihood estimation method.

The bias and temporal resolution in the TOJ can be evaluated by the mean and standard deviation of the psychometric function (d and σ in Eq. 1), respectively. The mean implies the point of subjective simultaneity (PSS). The standard deviation implies the temporal uncertainty (i.e., the inverse of the temporal resolution).

Bayesian calibration (Miyazaki et al., 2006) should appear as shifts of the PSS in directions opposite to the mean SOAs (peaks in frequency) of prior distributions (Fig. 1a, b). The PSS shifts are expressed by Eq. (2) [see Eqs. (1–16) in the Supplementary Methods (https://​www.​nature.​com/​articles/​nn1712#Sec2) of Miyazaki et al. (2006) for derivation]:where σ_sensed denotes the sensory uncertainty (i.e., the inverse of the sensory temporal resolution) in the TOJ. According to the Bayesian estimation model of the TOJ, σ_sensed is not affected by prior experience [see Eq. (1), (14), and (15) in the Supplementary Methods of Miyazaki et al. (2006)]. In the present experiments, there was no difference in the σ according to differences in prior experience (i.e., left-first biased, right-first biased, or unbiased conditions; p = 0.40, see the last paragraph in Results for details), as predicted by the Bayesian estimation model. Therefore, σ_sensed can be measured as the standard deviation of the psychometric function (σ in Eq. 1). σ_prior and μ_prior are the standard deviation and mean of the prior distribution, respectively. In our experiments, we used 80 ms for σ_prior and ± 80 ms for the μ_prior.

We evaluated whether Bayesian calibration occurred using ΔPSS (see Fig. 2b). The ΔPSS was calculated by subtracting the PSS under the right-first biased condition from that under the left-first biased condition. If Bayesian calibration occurs, the ΔPSS should be positive.

Moreover, we evaluated the consistency of the observation values of ΔPSS with the theoretical predictions using the Bayesian estimation model. The greater positive values of ΔPSS imply a higher dependency on the prior but do not always imply an execution of the ‘optimal’ Bayesian estimation. Excessive dependency on prior experience leads to larger errors. To generate an optimal estimation, it is necessary to moderately depend on prior experience according to the degree of sensory temporal resolution of each participant. To evaluate this, we calculated the deviation rate of the observed ΔPSS relative to the theoretical ΔPSS (%ΔPSS_obs−theor, Eq. 3) for each participant.

When computing the theoretical ΔPSS, we used the root mean square (RMS) of the σ values calculated from the psychometric functions of the left-first biased, right-first biased, and unbiased conditions. Then, we assigned RMS σ to σ_sensed in Eq. (2). This process is reasonable because prior experience has no effect on σ_sensed, as described above. In addition, this averaging process also contributed to stabilizing the σ_sensed values and the resultant theoretical ΔPSS values. For the other parameters in Eq. (2), we assigned 80 ms to σ_prior and ± 80 ms to the μ_prior. Thus, we obtained a ΔPSS value that was theoretically optimal for the sensory temporal resolution of each participant. We then calculated the %ΔPSS_obs−theor for each participant as:where ΔPSS_obs denotes the observed ΔPSS, and ΔPSS_theor denotes the theoretical value. The %ΔPSS_obs−theor is zero when a positive aftereffect occurs, as predicted by the Bayesian estimation model. The %ΔPSS_obs−theor should be larger than zero when the participant exhibited a larger positive aftereffect relative to the theoretical prediction, whereas it should be smaller than zero when the participant exhibited a smaller aftereffect relative to the theoretical prediction.

In our experiment, Bayesian calibration was weakened in participants with higher autistic traits. According to the Bayesian estimation model (Körding & Wolpert, 2004; Miyazaki et al., 2006; Pellicano & Burr, 2012), there are two possible mechanisms for the result. The first is ‘hypo-priors,’ which is an impairment of learning and/or the utilization of prior distributions leading to weaker Bayesian calibration in participants with higher autistic traits. The second is ‘hyper-sensors,’ in which a higher sensory temporal resolution leads to weaker Bayesian calibration in participants with higher autistic traits. Our results do not support the latter hypothesis. This is due to the lack of a significant difference for the σ values among groups, and that there was no significant correlation between the σ values and AQ scores (Table 2, Supplementary Tables S1 and S2). Moreover, participants with higher AQ scores exhibited smaller %ΔPSS_obs−thoer values (i.e., weaker Bayesian calibration), each of which was adjusted by the σ value of each participant (Figure 5). Accordingly, the series of our results suggest that the weakened Bayesian calibration observed herein is attributable to the impairment of learning and/or utilization of the prior distributions.

Thus, our results are consistent with the ‘hypo-priors’ hypothesis (Pellicano & Burr, 2012). Hypo-priors have also been observed in the interval timing of children with autism (Karaminis et al., 2016). Interval-timing responses are biased to the mean of the target time intervals (central tendency effect), which can be accounted for by the Bayesian estimation model (Jazayeri & Shadlen, 2010; Miyazaki et al., 2005; Roach et al., 2017). Karaminis et al. reported that in children with autism, the central tendency was much less than that predicted by the theoretical model that took into account the sensory temporal resolution of each participant. Previous findings and our results suggest that hypo-priors generally occur during human sensorimotor processing. Furthermore, we infer that hypo-priors may also be observed in motor controls. Psychophysical studies using motor control tasks first reported that the human brain generates a Bayesian estimation by learning the prior distribution (Körding & Wolpert, 2004; Körding et al., 2004). Moreover, an fMRI study suggested that tactile TOJs share neural bases with motor controls (Miyazaki et al., 2016). Individuals with ASD frequently exhibit developmental coordination disorder (Cacola et al., 2017), some of which may be explained by impairments in the learning of Bayesian priors.

Several psychophysical studies also reported hypo-priors in audiovisual lag adaptation (or called audiovisual temporal recalibration) (Noel et al., 2017; Stevenson et al., 2017; Turi et al., 2016), although they did not show a mathematical model to specifically explain lag adaptation. Lag adaptation has been observed as a negative aftereffect in audiovisual simultaneity judgment (SJ) (Fujisaki et al., 2004) and TOJs (Hanson et al., 2008; Miyazaki et al., 2006; Vroomen et al., 2004; Yamamoto et al., 2012). The negative aftereffects in SJ and TOJ suggest that after repeated exposure to audiovisual stimuli with a lag, participants eventually perceived the repeated lag to be shorter than that in reality. Lag adaptation is considered to reflect brain adaptation, which compensates for differences in physical and neuronal conduction times between audio and visual signals (Fujisaki et al., 2004). Furthermore, Van der Burg demonstrated that a similar negative aftereffect occurred in response to only one prior experience (rapid lag adaptation) (Van der Burg et al., 2013). Recent studies have reported impairments in the rapid type of lag adaptation in participants with ASD (Noel et al., 2017; Turi et al., 2016).

A more recent study (Stevenson et al., 2017) reported that long-term lag adaptation (Fujisaki et al., 2004) was weakened in TD participants with higher scores of attention to detail (subscale of AQ score); however, there was no correlation between the total AQ score and other subscales. In the present experiments, the ΔPSS (amplitude of positive aftereffect) exhibited a significant negative correlation with the total AQ score in the TD participants (Fig. 4). Moreover, ΔPSS had significant (p < 0.05) or marginally significant (p < 0.1) negative correlations with the subscales, including attention to detail, with the exception of social skills and imagination (Supplementary Table S3). In addition, after including the ASD participants (Supplementary Tables S4–S6), similar results were observed when excluding ASD participant #3 (Supplementary Table S5) or participants #3 and #6 (Supplementary Table S6). The difference between our results and those of Stevenson et al. suggests that audiovisual lag adaptation and tactile Bayesian calibration reflect distinctive features in autistic spectrum conditions.

We should also note two participants with ASD (#3 and #6) who displayed particular results. ASD participant #3 exhibited an irregularly large positive aftereffect. Notably, participant #3 had an additional diagnosis of ADHD and was prescribed methylphenidate (Table 1). Methylphenidate is a norepinephrine and dopamine reuptake inhibitor (Madras et al., 2005). A dose of methylphenidate increased the learning rate in a probabilistic learning task in healthy human participants (Howlett et al., 2017). The increase in learning rate implies that the participants quickly adjusted their responses. In theory (Howlett et al., 2017; Yu and Huang, 2014), this strategy is optimal for tasks with high volatility, but is not always optimal for static or coherent tasks, such as our TOJ, which is systematically biased to a specific order. Notably, a positive aftereffect itself should appear even when TOJ was biased to the orders in recent prior trials, although it did not lead to optimal judgment in our task. Methylphenidate may have increased the learning rate of participant #3, resulting in an excessive bias to the priors.

ASD participant #6 revealed an irregularly large negative deviation in his aftereffect relative to the theoretical prediction (Fig. 5). Participant #6 did not display a prominent irregular value in the aftereffect (#6 in Fig. 4). This apparent discrepancy was due to his irregularly high sensory temporal resolution (i.e., extremely small σ: 8.52 ms, see Table 1). Such extremely high sensory temporal resolution implies that the participant has an extremely high information density per unit of time. In such participants, if a perceptual change occurred in the TOJ, it appeared small in the observed ΔPSS value. Since we used the %ΔPSS_obs−thoer, each of which value was standardized by the σ value of each participant, we detected the irregularly large negative aftereffect of participant #6 relative to their extremely high sensory temporal resolution. Notably, participant #6 was diagnosed with PDD and prescribed atomoxetine (Table 1). A psychopharmacological study in mice reported that an acute administration of atomoxetine improved temporal precision (Balci et al., 2008). Thus, the small σ of participant #6 may have been due to atomoxetine. Moreover, atomoxetine is a norepinephrine reuptake inhibitor (Wong et al., 1982). Therefore, atomoxetine may also increase learning rates. However, at this stage, it is unclear whether atomoxetine caused a negative aftereffect instead of a positive aftereffect, even if atomoxetine was truly involved. Our results may suggest a contrasting pharmacological effect on perceptual/cognitive adaptations between methylphenidate and atomoxetine; however, more pharmacological evidence is needed to ascertain our suggestions.

As described above, our results were discussed based on Bayesian explanations for autistic perceptions (Pellicano & Burr, 2012). However, there are limitations to the present study. First, the number of participants with ASD was small (n = 8). The results of the ASD participants suggested not only generality along with those of TD participants but also the particularity of certain ASD participants (#3, #6). Although we discussed the possible mechanisms and implications for the particular results of the two ASD participants based on their medications, each explanation was only an ex-post speculative hypothesis based on only one case. The relationship between their behavior and medications may just be due to chance. Large-sample and interventional studies are needed to test these hypotheses. Our hypotheses may be rejected in future studies. However, repetitions of such hypothesis testing will clarify the mechanisms to systematically account for both general and diverse aspects of autistic perception and behaviors.

Compared to the ASD participants, the present study showed more coherent results for TD participants with a larger sample size (n = 29). Here, we should note that if a participant without an ASD diagnosis shows a small or no Bayesian calibration, it does not always imply that the participant has ASD. Although the AQ score is a reliable index to evaluate autistic traits, TD individuals with high AQ scores are not always diagnosed with ASD (Wakabayashi et al., 2004). Our results provide evidence to support the Bayesian explanation of autistic perception and behavior (Pellicano & Burr, 2012), but we should not automatically apply our results to the diagnosis of ASD.