Developmental Science 2017; 20: e12435 DOI: 10.1111/desc.

12435

PAPER
Children on the autism spectrum update their behaviour in
response to a volatile environment
Catherine Manning,1* James Kilner,2* Louise Neil,3 Themelis Karaminis3
and Elizabeth Pellicano3,4
1. Department of Experimental Psychology, University of Oxford, UK
2. Wellcome Trust Centre for Neuroimaging, UCL Institute of Neurology, University College London, UK
3. Centre for Research in Autism and Education (CRAE), UCL Institute of Education, University College London, UK
4. School of Psychology, University of Western Australia, Australia

Abstract
Typical adults can track reward probabilities across trials to estimate the volatility of the environment and use this information
to modify their learning rate (Behrens et al., 2007). In a stable environment, it is advantageous to take account of outcomes
over many trials, whereas in a volatile environment, recent experience should be more strongly weighted than distant experience.
Recent predictive coding accounts of autism propose that autistic individuals will demonstrate atypical updating of their
behaviour in response to the statistics of the reward environment. To rigorously test this hypothesis, we administered a
developmentally appropriate version of Behrens et al.’s (2007) task to 34 cognitively able children on the autism spectrum aged
between 6 and 14 years, 32 age- and ability-matched typically developing children and 19 typical adults. Participants were
required to choose between a green and a blue pirate chest, each associated with a randomly determined reward value between 0
and 100 points, with a combined total of 100 points. On each trial, the reward was given for one stimulus only. In the stable
condition, the ratio of the blue or green response being rewarded was fixed at 75:25. In the volatile condition, the ratio
alternated between 80:20 and 20:80 every 20 trials. We estimated the learning rate for each participant by fitting a delta rule
model and compared this rate across conditions and groups. All groups increased their learning rate in the volatile condition
compared to the stable condition. Unexpectedly, there was no effect of group and no interaction between group and condition.
Thus, autistic children used information about the statistics of the reward environment to guide their decisions to a similar extent
as typically developing children and adults. These results help constrain predictive coding accounts of autism by demonstrating
that autism is not characterized by uniform differences in the weighting of prediction error.

Research highlights Introduction

• Predictive coding models have recently been pro- The decisions we make in a given moment are informed
posed to account for the complex autism phenotype. by expectations derived from the outcomes of similar
• Here, we test a key prediction from predictive coding decisions we have made in the past (Behrens, Woolrich,
accounts using a probabilistic learning task. Walton & Rushworth, 2007; Louie & Glimcher, 2012;
• Autistic children did not have generally elevated Summerfield & Tsetsos, 2015). Rather than assigning
learning rates compared to typically developing equal weight to all previous outcomes, neurotypical
children, and updated their learning rate in a volatile adults can track the statistics of the environment in order
reward environment. to determine how much weight should be given to new
• These results suggest that autism is not characterized information. When the reward environment is stable,
by uniformly high and inflexible weighting of predic- people take account of previous outcomes over many
tion errors. trials to guide their decisions (Behrens et al., 2007;

Address for correspondence: Catherine Manning, Department of Experimental Psychology, University of Oxford, 9 South Parks Road, Oxford OX1
3UD, UK; e-mail: [email protected]
*Catherine Manning and James Kilner contributed equally to this study.

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
This is an open access article under the terms of the Creative Commons Attribution License, which permits use,
distribution and reproduction in any medium, provided the original work is properly cited.
2 of 13 Catherine Manning et al.

O’Reilly, 2013). Yet, when the reward environment is 2008). This balance may be atypical in autistic individ-
volatile (i.e. fluctuating over blocks of trials), people uals, who may have a low precision of prior information
weight their recent experience more strongly than their relative to that of the sensory information (Friston,
distant experience (Behrens et al., 2007). The relative Lawson & Frith, 2013; Lawson, Rees & Friston, 2014).
weighting given to recent and distant trials is reflected in Van de Cruys, Evers, Van der Hallen, Van Eylen, Boets
a person’s learning rate (Dayan, Kakade & Montague, et al. (2014) made a more specific proposal that the
2000). In a stable environment, neurotypical adults precision of prediction errors is uniformly high and
demonstrate a low learning rate (Behrens et al., 2007), inflexible in individuals with autism. Finally, Sinha,
as the history of outcomes is more predictive of the Kjelgaard, Gandhi, Tsourides, Cardinaux et al. (2014)
current state of the environment than the outcomes in proposed that autistic individuals have impairments in
the most recent trials. In a volatile environment, adults estimating the conditional probability of future events or
demonstrate an increased learning rate, whereby they states given a previous state occurring, particularly when
‘take notice’ of the outcomes of more recent trials and the relationship between events is probabilistically weak
use these to modify their behaviour (Behrens et al., and/or when events are separated by long temporal
2007). intervals. Yet, autistic individuals may excel at learning
The ability to take account of previous information rules, due to heightened learning of novel stimuli (Sinha
has been the focus of recent advances in theories of et al., 2014).
autistic perception and cognition (see Brock, 2014, for Atypical deployment of previous experience has been
review). Applying a Bayesian framework, Pellicano and suggested to relate to a range of autism symptoms, such
Burr (2012a) suggested that autistic1 individuals make as insistence on sameness, repetitive behaviours, and
less use of prior information than typical individuals. impaired social functioning (Lawson et al., 2014; Van de
Under the Bayesian framework, the percept (or poste- Cruys et al., 2014). Yet, experimental studies are
rior) results from a combination of incoming sensory required to rigorously test the mechanisms proposed by
information (the likelihood) and previous information such approaches. If autistic individuals do not use
(priors), the weighting of which depends on their previous experience in the same way as typical individ-
respective precision. Pellicano and Burr suggested that uals – either through weakened priors or atypical
autistic individuals have attenuated (broader) priors, predictive mechanisms – they may behave differently in
meaning that their perception is more dominated by the a task that involves tracking the statistics of the reward
incoming sensory information. While the theory initially environment, such as that used by Behrens et al. (2007).
focused on perception, the weaker influence of prior No previous studies have assessed whether children on
information could potentially account for a range of the autism spectrum increase their learning rate in
aspects of the autism phenotype, such as social func- response to environmental volatility. Yet, reversal learn-
tioning (Pellicano & Burr, 2012b). ing studies give some insight into how individuals with
Elaborations of this account have used the predictive autism deal with probabilistic information. Most of these
coding framework, which provides a biologically plausi- studies suggest that individuals with autism can learn
ble implementation of Bayesian inference (see Clark, initial reinforcement probabilities, but demonstrate dif-
2013, for review), replacing priors and sensory evidence ficulties in switching when these probabilities reverse (e.g.
with predictions and prediction errors, respectively. Solomon, Smith, Frank, Ly & Carter, 2011; South,
According to the predictive coding approach, the brain Newton & Chamberlain, 2012) or in maintaining new
aims to predict what will happen next and attempts to response probabilities (D’Cruz, Ragozzino, Mosconi,
minimize prediction error – the discrepancy between the Shrestha, Cook et al., 2013). Such difficulties in reversal
prediction and reality. Predictions emerging from higher learning may reflect executive functioning difficulties in
brain areas are used to attempt to ‘explain away’ the autism (Pennington & Ozonoff, 1996; see Pellicano,
input from lower brain areas and prediction errors are 2012, for review). Yet, these studies do not address how
passed up the hierarchy to inform higher-level expecta- higher-order statistics about the environment – such as
tions. The influence of prior beliefs relative to sensory volatility – affect autistic individuals’ responses.
evidence is controlled by the precision assigned to Robic, Soni
e, Fonlupt, Henaff, Touil et al. (2015)
prediction errors at each level of the hierarchy (Friston, presented a cued decision-making task to 14 autistic
adults and 15 neurotypical adults whilst manipulating the
1
The term ‘autistic person’ is the preferred language of many people
volatility of the environment (stable, unstable). A social
on the spectrum (see Sinclair, 1999; Kenny, Hattersley, Molins, Buckley, cue (video of a human actor) or non-social cue (arrow)
Povey et al., 2016). In this article, we use this term as well as person- was presented before participants chose between one of
first language to respect the wishes of all individuals on the spectrum. two options, with both the reliability of the cue and the

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
 Response to volatility in autism 3 of 13

reward probabilities of each option being manipulated Alongside these primary hypotheses, we were also
throughout the experiment. The authors recorded the interested in examining possible relationships between
proportion of correct choices made by participants, and task performance and anxiety. Anxiety commonly co-
reported that fewer individuals with autism met a criterion occurs with autism (White, Oswald, Ollendick & Scahill,
of 60% correct than typical individuals when reward 2009), and high levels of trait anxiety have been linked to
probabilities fluctuated (unstable condition), but not reduced updating of behaviour in response to volatility in
when reward probabilities were fixed (stable condition), an aversive version of Behrens et al.’s task (Browning,
and that individuals with autism performed particularly Behrens, Jocham, O’Reilly & Bishop, 2015). This link is
poorly in response to the social cue. The authors particularly interesting given a suggestion that atypical
concluded that individuals with autism have particular predictive mechanisms may give rise to increased anxiety
difficulties in learning reward probabilities when the in autistic individuals (Sinha et al., 2014). To this end,
reward environment is unstable and involves a social we investigated relationships between behaviour updat-
aspect. Yet, importantly, half of the participants with ing and parent-reported anxiety in our children with and
autism did meet a criterion of 60% in the unstable without autism.
condition, suggesting considerable individual differences
in this performance metric. In this study, we used a more
fine-grained measure to characterize behavioural Materials and methods
responses to environmental volatility in autism.
We administered a child-friendly version of Behrens
Participants
et al.’s (2007) task to 34 cognitively able autistic
children, 32 typically developing children and 19 typical Three groups of participants were tested: 34 autistic
adults. Participants were required to choose between two children, 32 typically developing children and 19 typical
stimuli under two conditions: in the stable condition, the adults (see Table 1 for participant characteristics). The
probability of each stimulus being rewarded was fixed. In
the volatile condition, the probability of each stimulus
being rewarded alternated every 20 trials. A crucial Table 1 Participant characteristics
difference between this paradigm and those used in Typically
previous reversal learning studies in autism (e.g. D’Cruz Children developing Typical
et al., 2013; Solomon et al., 2011; South et al., 2012; with autism children adults
Robic et al., 2015) was that each stimulus was associated
N 34 32 19
with a reward value that varied trial by trial. This Gender (n males: 29: 5 22: 10 7: 12
paradigm allowed us to model the learning rates of n females)
participants (cf. Behrens et al., 2007), and in turn to Age (years; months)
Mean (SD) 9;11 (2;0) 9;2 (1;10) 24;2 (3;9)
address specific predictions derived from Bayesian and Range 7;0–14;3 6;6–13;2 18;5–33;1
predictive coding accounts of autism. Performance IQ
Pellicano and Burr (2012a) suggested that individuals Mean (SD) 105.44 (14;94) 104.84 (14.11)
Range 79–141 78–131
with autism are less influenced by information presented Verbal IQ
in the past, due to reduced priors. Van de Cruys et al. Mean (SD) 100.15 (17.47) 108.03 (11.04)
(2014) suggested that individuals with autism have very Range 71–130 86–132
Full-scale IQ
precise prediction errors, which would mean that they Mean (SD) 102.94 (15.46) 107.28 (10.66)
should heavily weight violations to their expectations Range 76–129 89–131
and update their behaviour accordingly. Sinha et al. SCQ score
Mean (SD) 23.78 (7.35) 5.77 (4.04)
(2014) also predicted heightened learning to novel Range 5–35 1–14
stimuli. Thus, according to these three accounts, children Spence Children’s Anxiety
on the autism spectrum may demonstrate a generally Mean (SD) 33.52 (20.35) 19.78 (10.67)
Range 6–76 6–43
elevated learning rate, emphasizing the outcomes of ADOS total score
more recent trials (and thus, violations to their predic- Mean (SD) 10.45 (4.64)
tions) more than typical children. Van de Cruys et al. Range 2–21
also made a further specific proposal that individuals
with autism do not flexibly weight their prediction Notes: SCQ = Social Communication Questionnaire (Rutter et al.,
errors, which may mean that they do not modify their 2003). ADOS = Autism Diagnostic Observation Schedule (Lord et al.,
1999, 2012). Verbal, Performance and Full-Scale IQ scores were derived
learning rate in the volatile condition compared to the from the Wechsler Abbreviated Scales of Intelligence (WASI-II;
stable condition to the same extent as typical individuals. Wechsler, 2011).

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
4 of 13 Catherine Manning et al.

autistic children had previously received an independent room. The WASI and ADOS were administered in
clinical diagnosis of an autism spectrum condition further sessions.
according to ICD-10 (World Health Organization,
1993) or DSM-IV (American Psychiatric Association,
Familiarization phase
2000) criteria. Typically developing children had no
parent-reported diagnoses of developmental conditions Participants were initially introduced to images on cards
and typical adults reported no previous diagnoses of showing two pirate chests, one green and one blue, each
developmental conditions. Children’s parents completed with a flag displaying a reward value between 0 and 100
the Social Communication Questionnaire (SCQ; Rutter, points (or ‘coins’), with a combined total of 100 points
Bailey & Lord, 2003) and children with autism were (Figure 1). The experimenter explained to participants
administered the Autism Diagnostic Observation Sched- that only one chest would contain treasure on each trial
ule (ADOS-G or ADOS-2; Lord, Rutter, DiLavore & (the rewarded stimulus); the other chest would be empty
Risi, 1999; Lord, Rutter, DiLavore, Risi, Gotham et al., (the non-rewarded stimulus). Critically, participants were
2012) using the revised algorithm (Gotham, Risi, told that the chest containing treasure could change
Pickles & Lord, 2007; Gotham, Risi, Dawson, Tager- throughout the game. Next, participants viewed 20 trials
Flusberg, Joseph et al., 2008). All autistic children in which the chests were opened to reveal which chest
scored above threshold for an autism spectrum condi- was empty and which contained treasure (Figure 1A).
tion on one or both of these measures and all typically The ratio of the blue or green chest being rewarded was
developing children scored below the cut-off for autism
on the SCQ (< 15; Rutter et al., 2003) (see Table 1 for
scores).
The groups of autistic and typical children were
matched in terms of age, t(64) = 1.65, p = .10, performance
IQ, t(64) = .17, p = .87 and full-scale IQ, t(58.78) = 1.33,
p = .19, as assessed by the Wechsler Abbreviated Scales of
Intelligence, Second Edition (WASI-II, Wechsler, 2011).
The children with autism had lower verbal IQ scores than
typical children, consistent with their clinical profile,
t(56.17) = 2.21, p = .03. Anxiety was measured using the
parent-report version of the Spence Children’s Anxiety
Scale (SCAS-P; Nauta, Scholing, Rapee, Abbott, Spence
et al., 2004), which was returned by 58 parents. The
children with autism (n = 31) had significantly higher
scores than typical children (n = 27), t(46.58) = 3.28,
p = .002.

Procedure
This study was approved by the UCL Institute of
Education’s Research Ethics Committee and was con-
ducted in accordance with the principles of the Decla-
ration of Helsinki. Parents and adult participants gave
their written informed consent and children provided
their verbal assent prior to participation.
We adapted Behrens et al.’s task into a child-friendly,
pirate-themed game. The task began with an initial
familiarization phase, designed to introduce the partic-
ipants to the task, followed by the experimental phase.
Both phases were completed in a single session lasting Figure 1 Familiarization phase. (A) Example of a trial in the
approximately 15 minutes. Adult participants also com- familiarization phase, where participants passively viewed
pleted a longer version of the experiment in a separate stimuli. (B) After viewing 20 trials, participants were asked to
session lasting approximately 25 minutes (see below). estimate the ratio in which the green or blue chest contained
Children and adults were seen individually in a quiet treasure.

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
 Response to volatility in autism 5 of 13

fixed at 80:20. Whether the blue or green chest was points were added to an accumulated total and displayed
rewarded most often (i.e. 80% of the time) was random- on a bar chart (Figure 2). If participants chose the
ized across participants. The values on the flags were incorrect (non-rewarded) stimulus, an empty chest was
randomly selected on each trial, with the constraint that revealed and participants did not receive any coins.
the total number of points was always 100. After viewing Visual feedback was provided (‘Better luck next time!’).
all 20 trials, participants were asked to estimate the ratio When participants accumulated enough points to exceed
at which the blue and green chests contained treasure, by the limits of the bar chart, they were shown a screen that
moving a yellow indicator along a scale ranging from told them they had reached the next ‘level’ (e.g.
‘All green’ to ‘All blue’ (Figure 1B). The familiarization ‘Congratulations, you are now a level 1 pirate!’). The
phase was repeated for participants who misunderstood bar chart was then emptied and participants accumu-
the task (two typical children; e.g. those who responded lated points to reach the next level.
that the treasure was always in the green chest when it During the task, the first 80 trials belonged to the
was mostly blue, and vice versa). stable condition, where the ratio of the blue or green
response being rewarded was fixed at 75:25. The next 80
trials belonged to the volatile condition, whereby the
Experimental phase
ratio alternated between 80:20 and 20:80 every 20 trials.
Immediately following the familiarization phase, partic- The conditions followed on from each other without a
ipants completed the experimental task. As in the break and participants were na€ıve to the reward struc-
familiarization phase, participants were presented with ture. At the end of the task, the participant’s total
a green and a blue pirate chest on each trial. Here, number of points (or ‘coins’) was displayed, along with
participants were required to actively choose either the the final ‘level’ s/he had reached in the game (e.g. ‘Wow!
green or blue pirate chest on each trial using response You reached level 10 and collected 1500 coins!’).
pads (Figure 2). If participants chose the correct (re-
warded) stimulus, the chest containing treasure was
Additional experiment for adult participants
revealed and they were awarded the number of points (or
‘coins’) indicated on the flag. Participants were given The experimental task described above contained fewer
visual feedback (‘Well done, you chose correctly!’) and trials than that used by Behrens et al. (2007) to make it
auditory feedback (the sound of a coin dropping). The suitable for child participants. To allow direct

Figure 2 Experimental task. Example of a trial in the test phase in which the rewarded stimulus is the blue pirate chest. If the
participant chose the green chest, an empty chest was revealed and no points were awarded. If the participant chose the blue chest,
the treasure was revealed and the participant received the number of points (or ‘coins’) displayed on the flag (i.e. 27).

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
6 of 13 Catherine Manning et al.

comparisons with Behrens et al.’s original paradigm, the previous trial if that option had been rewarded, and
adults also completed a longer version of the experiment, choosing the opposite option if their response on the
which included the same number of trials as presented by previous trial had not been rewarded. The reward value
Behrens et al. The adults completed the longer version in regressor was the reward value for the blue option.
a separate session after the short version, to ensure that The main aim of this study was to test whether the
the adults’ performance on the short version of the task different participant groups modulated their learning
was directly comparable to that of the children’s rate based on the volatility of the environment. To this
performance. In this longer version, the stable condition end we estimated the learning rate for each participant in
consisted of 120 trials with a fixed ratio of 75:25, and the the stable and volatile conditions of the task. For the
volatile condition consisted of 170 trials with a ratio that shorter version of the task the learning rate was
switched between 80:20 and 20:80 every 30 or 40 trials estimated using the 61 trials from trial 20 to trial 80 in
(i.e. 30 – 40 – 30 – 40 – 30). The order of stable and the stable condition, and the 61 trials from trial 90 to
volatile conditions was counterbalanced in this task to trial 150 in the volatile condition. For the longer version
confirm that the change in learning rate was not of the task, the corresponding windows were the last 81
dependent on order of presentation (as originally trials of each of the stable and volatile phases, following
demonstrated by Behrens et al.). Behrens et al. (2007). A reinforcement-learning model
was fitted to each participant’s decisions in each window.
The model has two parts: a ‘predictor’, which estimates
Analysis
the current reward rate given past observations, and a
Prior to testing whether participants modulated their ‘selector’, which generates actions on the basis of these
learning rates as a function of volatility, an initial estimates. The predictor is in the form of a simple delta-
analysis was performed to determine whether the learning rule (Rescorla & Wagner, 1972), which has a
participants’ behaviour could be reliably explained by single free parameter: the learning rate.
the ideal Bayesian observer model, and whether this
model was the best model to explain the participants’
behaviour in each group. We fitted each individual Results
participant’s choices (green or blue) across all trials with
a logistic generalized linear model. The design matrix Figure 3 shows the mean choices made by each group
contained a constant term and four different models: (i) (upper panel) and the ideal observer model (lower
ideal observer, (ii) alternating choices, (iii) win-stay lose- panel). The ideal observer model provided the best fit
shift, and (iv) reward value. to the data in all groups, compared to the alternating
The ideal observer regressor was estimated from the choices, win-stay lose-shift and reward value models
sequence of rewarded stimuli across all trials using the (Figure 4A). The ideal observer model provided a
model described in Behrens et al. (2007). In short, significant fit (p < .05) to the choices from the majority
optimal behaviour requires participants to estimate the of participants in all groups (Figure 4B). We used chi-
probability of reward for each stimulus and to compute squared tests to determine whether the proportion of
the expected value as reward probability multiplied by participants fit by each model varied (i) between child
reward size. Given that the amplitude of the reward was participants and adults, and (ii) between autistic and
random across trials, the ideal observer of the underlying typically developing children. The only significant dif-
probability of reward success is modelled only by the ference was that more child participants were fit by the
reward probabilities. In the task, the reward probability win-stay lose-shift model than adults, v2(1) = 5.77,
varies across trials and is dependent upon the volatility, p = .02. All other comparisons between children and
which changes across the experiment. Therefore, the adult participants were non-significant, ps ≥ .68, and
optimal observer relies on the parameter estimates of the there were no significant differences in the proportions of
reward probability, the volatility and the confidence in autistic versus typically developing children fit by each
the volatility estimate from the preceding trial, and the model, ps ≥ .23.
latest trial outcome in order to determine decision and Having demonstrated that the ideal observer model
learning on the next trial. provided the best fit to the data, we next assessed the
The alternating choices regressor tested whether par- resulting learning rate estimate. First, we analysed the
ticipants simply switched between blue and green options data from the adult participants on the long version of
on every consecutive trial. The win-stay lose-shift the experiment (Figure 5) to ensure that we could
regressor modelled whether the participant’s behaviour replicate Behrens et al.’s (2007) pattern of results. All
could be explained by them choosing the same option as learning rates were log-transformed prior to analysis to

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
 Response to volatility in autism 7 of 13

Figure 3 Mean performance of ideal and real observers. The upper panel shows the mean choices 1 standard error of the mean
made by autistic children (green), typically developing (TD) children (red) and adults (blue). The lower panel shows mean ideal
observer performance 1 SEM based on the reward probabilities and values presented to all participants. The dotted lines represent
the underlying reward probabilities. Trials 1–80 belong to the stable condition, in which the reward probability is fixed at 75:25.
Trials 81–160 belong to the volatile condition, in which the reward probability fluctuates between 80:20 and 20:80 every 20 trials.
The probability of choice refers to the probability of choosing the option that was most frequently rewarded in the initial stable
condition (which was counterbalanced among participants to be either the blue or green pirate chest). Boxes represent the window
of trials over which learning rates were estimated for each individual.

meet the assumption of normality required for general F(1, 18) = 54.81, p < .001, ɳp2 = .75, whereby participants
linear models.2 A mixed-design ANOVA with condition increased their learning rate in the volatile condition. In
(stable, volatile) as a within-participants factor and order addition, there was an effect of length, with higher
of presentation (stable first, volatile first) as a between- learning rates obtained in the long version of the
participants factor confirmed that participants increased experiment than the short version, F(1, 18) = 24.87,
their learning rate in the volatile condition, F(1, 17) = p < .001, ɳp2 = .58. These effects were qualified with an
51.86, p < .001, ɳp2 = .75. In line with Behrens et al., the interaction between condition and length, F(1, 18) =
order of presentation did not have a significant effect 19.21, p < .001, ɳp2 = .52. While the increase in learning
on learning rate estimates, F(1, 17) = .91, p = .35, and rate was more pronounced in the long version, post hoc
did not interact with condition, F(1, 17) = 2.54, p = .13. t-tests confirmed that the learning rate increased in both
Thus, participants had a higher learning rate in the the short, t(18) = 3.40, p = .003, and long version of the
volatile condition than the stable condition even when experiment, t(18) = 8.04, p < .001. While it is not
the volatile condition was presented first. possible to determine whether the elevated learning rates
Next, we compared the adults’ learning rates in the in the long condition are a result of the increased number
short and long versions of the experiment (Figure 5) of trials, or increased familiarity with the task (as the
using a within-participants ANOVA with condition long condition was presented after the short version), we
(stable, volatile) and length of session (short, long) as have established that a clear effect of volatility is present
factors. Again, we found an overall effect of condition, with a reduced number of trials in adult participants.
Next, we aimed to assess whether typically developing
children and autistic children increased their learning
2
Note that the same pattern of results was obtained without rate as a function of volatility, in the same way as adults.
transformation. Learning rates for the three groups are shown in

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
8 of 13 Catherine Manning et al.

Figure 4 Results of model fitting to participant data. (A) Mean t-statistic of fit for four models to the data of autistic children (green),
typically developing (TD) children (red), and typical adults (blue): (1) an ideal observer model, (2) a model which alternates between
responses trial-by-trial, (3) a win-stay lose-shift model (maintaining responses after successes and switching responses after failures),
and (4) a model based on the reward value. Error bars represent 1 standard error of the mean. (B) Proportion of participants in each
group whose data were significantly fit with each of the four models (p < .05).

Figure 5 Learning rates of adult observers. Estimated learning rates of adult observers as a function of the order in which stable and
volatile conditions were presented (left panel) and the length of the testing session (right panel). Error bars represent 1 standard
error of the mean.

Figure 6. As above, log-transformed learning rates were however, there was neither a significant effect of group,
used in the analysis. A mixed-design ANOVA was F(2, 82) = .02, p = .98, ɳp2 < .001, nor a significant
conducted on learning rate estimates with condition interaction between group and condition, F(2, 82) = .22,
(stable, volatile) as a within-participants variable and p = .81, ɳp2 = .005. Thus, all groups increased their
group (children with autism, typical children, and typical learning rate in the volatile condition to a similar extent.
adults) as a between-participants variable. As expected, In order to judge the extent of a potential undetected
there was a main effect of condition, with higher learning effect, we compared the difference in log learning rates in
rates in the volatile condition than the stable condition, the stable and volatile conditions between autistic and
F(1, 82) = 32.52, p < .001, ɳp2 = .28. Unexpectedly, typical children, and found that the 95% confidence

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
 Response to volatility in autism 9 of 13

Figure 6 Learning rates in the stable and volatile conditions. Figure 7 Gradients of regression lines fitted to participant
Estimated learning rates in the stable and volatile conditions choices after a switch in reward values. Mean regression slopes
for autistic children (green), typically developing (TD) children (1 standard error of the mean) for the running average of four
(red) and typical adults (blue). Error bars represent 1 standard consecutive choices made by autistic children (green),
error of the mean. typically developing (TD) children (red) and adults (blue) in the
first 10 trials after the reward probabilities switched. For
example, if the reward probability for the blue to the green
intervals of the group difference were tightly distributed chest switched from 80:20 to 20:80 at trial 100, participants
became less likely to choose the blue chest (i.e. a negative
around zero [ .03, .02]. Learning rates were not related
slope). When the reward probability reversed again at trial
to age, performance IQ or verbal IQ (all ps ≥ .39).
120, the participants became more likely to choose the blue
Similarly, the difference in learning rates between the chest (i.e. a positive slope).
conditions was unrelated to age and ability (ps ≥ .58).
To investigate whether increased anxiety (as rated by 150) as the within-participants factor, and group (autism,
parents) is related to reduced updating in the volatile TD, adult) as the between-participants factor. Mauchly’s
condition (cf. Browning et al., 2015), we investigated the test demonstrated that sphericity could be assumed,
relationship between scores on the SCAS-P and the v2(2) = 3.91, p = .14. There was a significant within-
difference in learning rates between the volatile and participants effect of trial interval on slope coefficients,
stable conditions for each participant. Parent ratings F(2, 164) = 8.57, p < .001, ɳp2 = .10.3 Repeated contrasts
were not related to difference scores, r = .12, p = .35, nor revealed that the slopes in the first interval (trials 101–
to learning rates in either the stable or the volatile 110) were approximately as steep as slopes in the second
condition (ps ≥ .15). interval (trials 121–130), F(1, 82) = 3.54, p = .06, ɳp2 =
In order to align our paradigm with previous studies .04, while slopes in the third interval (trials 141–150)
of reversal learning in autism (cf. D’Cruz et al., 2013; were significantly steeper than those in the second
Solomon et al., 2011; Robic et al., 2015), we investigated interval, F(1, 82) = 16.16, p < .001, ɳp2 = .17. Impor-
whether the autistic children could learn new reward tantly, however, there was no overall group difference
probabilities as quickly as typically developing individ- in slope coefficients, F(2, 82) = 2.82, p = .07, ɳp2 = .06,
uals. We calculated a running average of choices made by nor an interaction between trial interval and group,
each participant over four consecutive trials, and fitted F(4, 164) = .52, p = .72.
regression lines to the first ten running averages after the
reward probabilities switched (i.e. trials 101–110, 121–
130, and 141–150). The mean regression slopes for each Discussion
group are shown in Figure 7.
We applied a square-root transformation to the In this study, children with autism, typical children and
absolute slope coefficients in each condition in order to typical adults completed a decision-making task under
meet the normality assumption required for an ANOVA
(Shapiro-Wilk test, ps ≥ .06).2 Next, we conducted a
mixed-design ANOVA on the transformed absolute slope 3
Note that this effect remained significant (p < .001) after applying a
coefficients with trial interval (101–110, 121–130, 141– Greenhouse-Geisser correction.

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
10 of 13 Catherine Manning et al.

two probabilistic reward schedules: a stable condition Dawson, Mottron & Gernsbacher, 2008, for review).
where the reward probabilities were fixed, and a volatile Indeed, previous research has suggested that individuals
condition where the reward probabilities fluctuated. with autism learn reward probabilities as well as typical
Based on recent Bayesian and predictive coding accounts individuals in simple learning tasks (D’Cruz et al., 2013;
of autism, we predicted that autistic children would Faja, Murias, Beauchaine & Dawson, 2013; Solomon
assign more weighting to the outcome of recent trials et al., 2011), perform successfully in implicit learning
than typical children and adults, and that they would not tasks (Brown, Aczel, Jim
enez, Kaufman & Plaisted-
flexibly update their behaviour in response to volatility. Grant, 2010; Nemeth, Janacsek, Balogh, Londe, Min-
In our task, these predictions would be manifest, gesz et al., 2010; Foti, De Crescenzo, Vivanti, Menghini
respectively, as an increased learning rate compared to & Vicari, 2015), and even demonstrate enhanced visual
typical children, and a reduced tendency to increase their statistical learning (Roser, Aslin, McKenzie, Zahra &
learning rate in the volatile condition. Contrary to these Fiser, 2015).
predictions, we found that children with autism had a Yet, differences between children with autism and
similar learning rate to typical children and adults typical children may become apparent using more
overall, and that they modified their learning rate to a complex learning tasks. For example, Pellicano, Smith,
similar extent as typical children and adults. Cristino, Hood, Briscoe et al. (2011) showed that chil-
These results appear to be at odds with Bayesian and dren with autism were slower than typical children to
predictive coding accounts of autism. Children with learn reward probabilities in a large-scale foraging task,
autism employed the recent history of trial outcomes in a which required children to continuously update their
similar way to typical children, contrasting both Pelli- spatial representations while remembering what loca-
cano and Burr’s (2012a) hypothesis of reduced use of tions they had already searched. Social situations may
priors in autism, and Van de Cruys et al.’s (2014) also pose particularly pronounced predictive processing
suggestion of highly precise prediction errors in autism, challenges for autistic people (Gomot & Wicker, 2012;
meaning that violations to expectations should be heavily Lawson et al., 2014), as there is no simple one-to-one
weighted (see also Sinha et al., 2014). Furthermore, the mapping between a cause and the sensory input. Thus,
children with autism were able to flexibly weight their differences may well arise in a social version of our
prediction errors in order to update their learning rate in learning task (Behrens, Hunt, Woolrich & Rushworth,
response to environmental volatility (cf. Van de Cruys 2008). In such a task, autistic children may demonstrate
et al., 2014). Why, then, do we find no differences in difficulties in tracking the probability that a confederate
performance? would give the correct advice. Preliminary support for
To date, Bayesian predictive models have arguably had this suggestion comes from a recent study demonstrating
most success in explaining how individuals with autism reduced use of social information during a similar
process sensory information. For example, these decision-making task in typical people with high levels
accounts have been linked to reports of reduced adap- of autistic traits (Sevgi, Diaconescu, Tittgemeyer &
tation to high-level sensory attributes, such as faces Schilbach, 2016), although this will require replication in
(Ewing, Pellicano & Rhodes, 2013; Fiorentini, Gray, participants with a clinical diagnosis of autism. Predic-
Rhodes, Jeffery & Pellicano, 2012; Pellicano, Jeffery, tive coding accounts have been proposed to explain a
Burr & Rhodes, 2007; Pellicano, Rhodes & Calder, 2013; range of high-level social abilities that may be affected in
Rutherford, Troubridg & Walsh, 2012; but see also autism, such as theory of mind (Koster-Hale & Saxe,
Cook, Brewer, Shah & Bird, 2014) and numerosity (Turi, 2013), interpersonal inference (Moutoussis, Fearon, El-
Burr, Igliozzi, Aagten-Murphy, Muratori et al., 2015), as Deredy, Dolan & Friston, 2014) and interoception
well as reduced use of contextual information in the (Quattrocki & Friston, 2014; Seth, Suzuki & Critchley,
rubber-hand illusion (Palmer, Paton, Kirkovski, Enticott 2011). Thus, further studies are required to probe the
& Hohwy, 2015) and reduced filtering of signal-from- limits of atypical predictive processing in autism, and
noise in motion displays (Manning, Tibber, Charman, theoretical accounts will need to be updated to account
Dakin & Pellicano, 2015; Zaidel, Goin-Kochel & Ange- for these. In particular, differences between children with
laki, 2015). Thus, it is conceivable that atypical predictive autism and typical children may be manifest in more
mechanisms account for perception in autism, but may complex situations where the one-to-one mapping
not extend to learning tasks, as in the current study. This between events is even less clear, when higher levels of
proposal may not be too surprising, given that sensory the predictive coding hierarchy are required (see also
symptoms are a core feature of the autistic phenotype Qian & Lipkin, 2011). While our task was the ideal
(American Psychiatric Association, 2013), while general candidate for assessing decision-making in an uncertain
difficulties in learning have not been established (see environment (Van de Cruys et al., 2014), it might not

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
 Response to volatility in autism 11 of 13

have been sufficiently challenging to reveal differences The discrepancy between the current study and
between children with and without autism. previous studies of reversal learning raise the intriguing
A previous study with adults showed that anxious traits possibility that a certain amount of environmental
were related to reduced updating of behaviour in an volatility may actually help children to adapt to new
aversive version of the task, in which incorrect responses situations. Previous studies of reversal learning in autism
were followed by electric shocks (Browning et al., 2015). have been unable to dissociate whether difficulties
In this study, we did not find a relationship between levels changing responses after a switch result from slower
of parent-reported trait anxiety and task performance. switching per se, or from individuals with autism
Yet, the children with autism were reported by their building a more stable state prior to the switch. When
parents to have elevated levels of anxiety overall. Thus, there is no reward value information, the optimal
even those children with high levels of anxiety were able to strategy is to consistently choose the option with the
update their behaviour to environmental volatility in a highest reward probability. Under these conditions,
typical fashion. The lack of the predicted relationship may individuals with autism may reach a very stable state,
have been because we did not use an aversive reinforcer in which means that they will need more evidence to reject
this task, unlike Browning et al. Future research could their current model and shift to a new model. Yet, the
investigate relationships between task performance and addition of randomly fluctuating reward value informa-
child-reported anxiety, or measures of state anxiety (e.g. tion, as in our paradigm, may prevent children with
heart rate, saliva cortisol levels). autism getting into such a stable initial state. Thus, it is
As well as addressing specific proposals from Bayesian possible that reward value fluctuations in the stable
predictive coding models, our results add more generally condition help children with autism to deal with the
to the reward learning literature in autism. Previous bigger change that occurs when they move into the
studies have revealed subtle differences in reversal volatile phase of the experiment. While speculative, this
learning between individuals with and without autism. may suggest that some divergence from routines may be
For example, D’Cruz et al. (2013) reported that individ- beneficial to autistic children in adapting to everyday
uals with autism needed more trials to achieve a criterion situations. To investigate this possibility further, it will be
of 8 correct out of 10 consecutive responses after a important to compare different learning paradigms in
switch, and were more likely to return to the previously the same samples of participants.
reinforced response after having selected the new correct In sum, our study shows that children with and
choice. Robic et al. (2015) reported that fewer autistic without autism can learn about the statistics of the
individuals reached a criterion of 60% correct than reward environment in a similar way to typical adults, by
typical individuals when reward probabilities fluctuated, updating their learning rate when the environment
and that autistic individuals were more likely to maintain becomes volatile. The typical performance of children
their response after a failure than typical individuals. with autism contrasts hypotheses emerging from recent
Furthermore, Solomon et al. (2011) reported that indi- Bayesian and predictive coding accounts of autism.
viduals with autism were less likely to maintain their While we believe that there is much mileage in these
response after a success, and Solomon, Frank, Ragland, approaches, they need to be explicitly laid out in order to
Smith, Niendam et al. (2015) reported selective difficul- make clear predictions from behavioural studies. Van de
ties with learning high-probability pairings but not Cruys et al. (2014) made one such prediction: ‘We
lower-probability pairings. The participants involved in predict that things will go awry in ASD [autism spectrum
the current study differed substantially from those tested disorder] when the probabilistic structure changes during
by Robic et al. (2015) and Solomon et al. (2011, 2015), the experiment, for instance when the predictability of a
who tested adult participants, and D’Cruz et al. (2013), cue changes across blocks’ (p. 655). Our results do not
who tested participants of a wide age range (between 8 support this prediction and argue against the suggestion
and 44 years). However, it is worth noting that the that precision is set uniformly high in individuals with
current paradigm is not directly comparable to these autism. Bayesian and predictive coding accounts may be
previous studies which did not manipulate the reward best suited to explaining atypical sensation and percep-
value. It does not make sense to assess the proportion tion in autism, and may not generalize to learning tasks.
correct in this study in the same way as in these previous Yet, future research is needed in order to determine
studies, as on a given trial, the optimal choice may not whether atypical predictive processing may be apparent
have been the same as the most frequently rewarded for more complex learning tasks, such as those requiring
choice. However, the children with autism appeared to individuals with autism to learn about social aspects of
change their behaviour just as quickly as typical children the environment. Further insights into learning mecha-
and adults after a switch in the current experiment. nisms in autism may be gleaned by using implicit

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
12 of 13 Catherine Manning et al.

measures (e.g. eyetracking or neuroimaging) in conjunc- D’Cruz, A.M., Ragozzino, M.E., Mosconi, M.W., Shrestha, S.,
tion with our behavioural task. Cook, E.H., et al. (2013). Reduced behavioral flexibility in
autism spectrum disorders. Neuropsychology, 27, 152–160.
Ewing, L., Pellicano, E., & Rhodes, G. (2013). Atypical
updating of face representations with experience in children
Acknowledgements
with autism. Developmental Science, 16, 116–123.
Faja, S., Murias, M., Beauchaine, T.P., & Dawson, G. (2013).
We are extremely grateful to all the participants and Reward-based decision making and electrodermal respond-
families who participated in this research, to Timothy ing by young children with autism spectrum disorders during
Behrens for supplying the scripts for analysing the a gambling task. Autism Research, 6, 494–505.
learning rates and volatility parameters and for helpful Fiorentini, C., Gray, L., Rhodes, G., Jeffery, L., & Pellicano, E.
discussions about the analysis and the results, and to (2012). Reduced face identity aftereffects in relatives of
Abigail Croydon for help with testing. This research was children with autism. Neuropsychologia, 50, 2926–2932.
funded by a Medical Research Council grant awarded to Foti, F., De Crescenzo, F., Vivanti, G., Menghini, D., & Vicari,
EP (MR/J013145/1) and a Scott Family Junior Research S. (2015). Implicit learning in individuals with autism
Fellowship awarded to CM at University College, spectrum disorders: a meta-analysis. Psychological Medicine,
45, 897–910.
Oxford. Research at CRAE (EP, LN, TK) is supported
Friston, K. (2008). Hierarchical models in the brain. PLoS
by the Clothworkers’ Foundation and Pears Foundation.
Computational Biology, 4, e1000211.
Friston, K.J., Lawson, R., & Frith, C.D. (2013). On hyperpriors
and hypopriors: comment on Pellicano and Burr. Trends in
References Cognitive Sciences, 17, 1.
Gomot, M., & Wicker, B. (2012). A challenging, unpredictable
American Psychiatric Association (2000). Diagnostic and sta- world for people with autism spectrum disorder. Interna-
tistical manual of mental disorders: DSM-IV-TR (4th edn., tional Journal of Psychophysiology, 83, 240–247.
text revision). Washington, DC: American Psychiatric Asso- Gotham, K., Risi, S., Dawson, G., Tager-Flusberg, H., Joseph,
ciation. R., et al. (2008). A replication of the Autism Diagnostic
American Psychiatric Association (2013). Diagnostic and sta- Observation Schedule (ADOS) revised algorithms. Journal of
tistical manual of mental disorders (5th edn.). Arlington, VA: the American Academy of Child and Adolescent Psychiatry,
American Psychiatric Publishing. 47, 642–651.
Behrens, T.E., Hunt, L.T., Woolrich, M.W., & Rushworth, M.F. Gotham, K., Risi, S., Pickles, A., & Lord, C. (2007). The
(2008). Associative learning of social value. Nature, 456, 245– Autism Diagnostic Observation Schedule: revised algorithms
249. for improved diagnostic validity. Journal of Autism and
Behrens, T.E., Woolrich, M.W., Walton, M.E., & Rushworth, Developmental Disorders, 37, 613–627.
M.F. (2007). Learning the value of information in an Kenny, L., Hattersley, C., Molins, B., Buckley, C., Povey, C.,
uncertain world. Nature Neuroscience, 10, 1214–1221. et al. (2016). Which terms should be used to describe autism?
Brock, J. (2014). Combining the old and the new. The Perspectives from the UK autism community. Autism, 20,
Psychologist, 27, 750–753. 442–462.
Brown, J., Aczel, B., Jim
enez, L., Kaufman, S.B., & Plaisted- Koster-Hale, J., & Saxe, R. (2013). Theory of mind: a neural
Grant, K. (2010). Intact implicit learning in autism spectrum prediction problem. Neuron, 79, 836–848.
conditions. Quarterly Journal of Experimental Psychology, 63, Lawson, R.P., Rees, G., & Friston, K.J. (2014). An aberrant
1789–1812. precision account of autism. Frontiers in Human Neuro-
Browning, M., Behrens, T.E., Jocham, G., O’Reilly, J.X., & science, 8, 302.
Bishop, S.J. (2015). Anxious individuals have difficulty Lord, C., Rutter, M., DiLavore, P.C., & Risi, S. (1999). Autism
learning the causal statistics of aversive environments. Nature Diagnostic Observation Schedule. Los Angeles, CA: Western
Neuroscience, 18, 590–596. Psychological Services.
Clark, A. (2013). Whatever next? Predictive brains, situated Lord, C., Rutter, M., DiLavore, P., Risi, S., & Gotham, K. et al.
agents, and the future of cognitive science. Behavioral and (2012). Autism Diagnostic Observation Schedule, Second
Brain Sciences, 36, 181–204. Edition (ADOS-2). Torrance, CA: Western Psychological
Cook, R., Brewer, R., Shah, P., & Bird, G. (2014). Intact facial Services.
adaptation in autistic adults. Autism Research, 7, 481–490. Louie, K., & Glimcher, P.W. (2012). Efficient coding and the
Dawson, M., Mottron, L., & Gernsbacher, M.A. (2008). neural representation of value. Annals of the New York
Learning in autism. In J.H. Byrne & H. Roediger (Eds.), Academy of Sciences, 1251, 13–32.
Learning and memory: A comprehensive reference (pp. 759– Manning, C., Tibber, M.S., Charman, T., Dakin, S.C., &
772). New York: Elsevier. Pellicano, E. (2015). Enhanced integration of motion infor-
Dayan, P., Kakade, S., & Montague, P.R. (2000). Learning and mation in children with autism. Journal of Neuroscience, 35,
selective attention. Nature Neuroscience, 3, 1218–1223. 6979–6986.

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.
 Response to volatility in autism 13 of 13

Moutoussis, M., Fearon, P., El-Deredy, W., Dolan, R.J., & Rutherford, M.D., Troubridge, E.K., & Walsh, J. (2012). Visual
Friston, K.J. (2014). Bayesian inferences about the self (and afterimages of emotional faces in high functioning autism.
others): a review. Consciousness and Cognition, 25, 67–76. Journal of Autism and Developmental Disorders, 42 (2), 221–
Nauta, M.H., Scholing, A., Rapee, R.M., Abbott, M., Spence, 229.
S.H., et al. (2004). A parent-report measure of children’s Rutter, M., Bailey, A., & Lord, C. (2003). Social Communica-
anxiety: psychometric properties and comparison with child- tion Questionnaire. Los Angeles, CA: Western Psychological
report in a clinic and normal sample. Behavior Research and Services.
Therapy, 42, 813–839. Seth, A.K., Suzuki, K., & Critchley, H.D. (2011). An intero-
Nemeth, D., Janacsek, K., Balogh, V., Londe, Z., Mingesz, R., ceptive predictive coding model of conscious presence.
et al. (2010). Learning in autism: implicitly superb. PLoS Frontiers in Psychology, 2, 395.
ONE, 5, e11731. Sevgi, M., Diaconescu, A.O., Tittgemeyer, M., & Schilbach, L.
O’Reilly, J.X. (2013). Making predictions in a changing world: (2016). Social Bayes: using Bayesian modelling to study
inference, uncertainty, and learning. Frontiers in Neuro- autistic trait-related differences in social cognition. Biological
science, 7, 105. Psychiatry, 80(2), 112–119.
Palmer, C.J., Paton, B., Kirkovski, M., Enticott, P.G., & Sinclair, J. (1999). Why I dislike ‘person-first’ language.
Hohwy, J. (2015). Context sensitivity in action decreases Available at: http://autismmythbusters.com/general-public/
along the autism spectrum: a predictive processing perspec- autistic-vs-people-with-autism/jim-sinclair-why-i-dislike-per-
tive. Proceedings of the Royal Society B, 282, 20141557. son-first-language/ (accessed 5 January 2016).
Pellicano, E. (2012). The development of executive function in Sinha, P., Kjelgaard, M.M., Gandhi, T.K., Tsourides, K.,
autism. Autism Research and Treatment, 2012, 146132. Cardinaux, A.L., et al. (2014). Autism as a disorder of
Pellicano, E., & Burr, D. (2012a). When the world becomes ‘too prediction. Proceedings of the National Academy of Sciences,
real’: a Bayesian explanation of autistic perception. Trends in USA, 111, 15220–15225.
Cognitive Sciences, 16, 504–510. Solomon, M., Frank, M.J., Ragland, J.D., Smith, A.C.,
Pellicano, E., & Burr, D. (2012b). Response to Brock: noise and Niendam, T.A., et al. (2015). Feedback-driven trial-by-trial
autism. Trends in Cognitive Sciences, 16, 574–575. learning in autism spectrum disorders. American Journal of
Pellicano, E., Jeffery, L., Burr, D., & Rhodes, G. (2007). Psychiatry, 172, 173–181.
Abnormal adaptive face-coding mechanisms in children with Solomon, M., Smith, A.C., Frank, M.J., Ly, S., & Carter, C.S.
autism spectrum disorder. Current Biology, 17, 1508–1512. (2011). Probabilistic reinforcement learning in adults with
Pellicano, E., Rhodes, G., & Calder, A.J. (2013). Reduced gaze autism spectrum disorders. Autism Research, 4, 109–120.
aftereffects are related to difficulties categorising gaze South, M., Newton, T., & Chamberlain, P.D. (2012). Delayed
direction in children with autism. Neuropsychologia, 51, reversal learning and association with repetitive behaviour in
1504–1509. autism spectrum disorders. Autism Research, 5 (6), 398–406.
Pellicano, E., Smith, A.D., Cristino, F., Hood, B.M., Briscoe, J., Summerfield, C., & Tsetsos, K. (2015). Do humans make good
et al. (2011). Children with autism are neither systematic nor decisions? Trends in Cognitive Sciences, 19, 27–34.
optimal foragers. Proceedings of the National Academy of Turi, M., Burr, D.C., Igliozzi, R., Aagten-Murphy, D., Mura-
Sciences, USA, 108, 421–426. tori, F., et al. (2015). Children with autism spectrum disorder
Pennington, B.F., & Ozonoff, S. (1996). Executive functions show reduced adaptation to number. Proceedings of the
and developmental psychopathology. Journal of Child Psy- National Academy of Sciences, USA, 112, 7868–7872.
chology and Psychiatry, 37, 51–87. Van de Cruys, S., Evers, K., Van der Hallen, R., Van Eylen, L.,
Qian, N., & Lipkin, R.M. (2011). A learning-style theory for Boets, B., et al. (2014). Precise minds in uncertain worlds:
understanding autistic behaviors. Frontiers in Human Neuro- predictive coding in autism. Psychological Review, 121, 649–
science, 5, 77. 675.
Quattrocki, E., & Friston, K. (2014). Autism, oxytocin and Wechsler, D. (2011). Wechsler Abbreviated Scale of Intelligence
interoception. Neuroscience and Biobehavioral Reviews, 47, (2nd edn.). San Antonio, TX: Psychological Corporation.
410–430. White, S.W., Oswald, D., Ollendick, T., & Scahill, L. (2009).
Rescorla, R.A., & Wagner, A.R. (1972). A theory of Pavlovian Anxiety in children and adolescents with autism spectrum
conditioning: variations in the effectiveness of reinforcement disorders. Clinical Psychology Review, 29, 216–229.
and nonreinforcement. In A.H. Black & W.F. Prokasy (Eds.), World Health Organization (1993). The ICD-10 classification of
Classical conditioning II: Current research and theory (pp. 64– mental and behavioural disorders: Diagnostic criteria for
99). New York: Appleton-Century Crofts. research. Geneva: World Health Organization.
Robic, S., Soni
e, S., Fonlupt, P., Henaff, M.A., Touil, N., et al. Zaidel, A., Goin-Kochel, R.P., & Angelaki, D.E. (2015). Self-
(2015). Decision-making in a changing world: a study in motion perception in autism is compromised by visual noise
autism spectrum disorders. Journal of Autism and Develop- but integrated optimally across multiple senses. Proceedings
mental Disorders, 45, 1603–1613. of the National Academy of Sciences, USA, 112, 6461–6466.
Roser, M.E., Aslin, R.N., McKenzie, R., Zahra, D., & Fiser, J.
(2015). Enhanced visual statistical learning in adults with Received: 2 October 2015
autism. Neuropsychology, 29, 163–172. Accepted: 15 March 2016

© 2016 The Authors. Developmental Science Published by John Wiley & Sons Ltd.