JSLHR

Research Article

Lexical Effects on Children’s Speech Processing:
Individual Differences Reflected in the
Autism-Spectrum Quotient (AQ)
Mitsuhiko Ota,a Mary E. Stewart,b Alexandra M. Petrou,b and Catherine Dickiec

Purpose: This study was undertaken to examine whether children exhibit the same relationship that adults show between lexical influence on phoneme identification and individual variation on the Autism-Spectrum
Quotient (AQ).
Method: Data from 62 4- to 7-year-olds with no diagnosis of autism were analyzed. The main task involved identification of the initial sound in pairs of voice-onset time continua with a real word on one end and a nonword on the other
(e.g., gift–kift, giss–kiss). Participants were also given the children’s version of the AQ and a 2nd instrument related

to autistic-like traits, the Social Responsiveness Scale
(SRS).
Results: The lexical shift was related to the AQ (particularly to its Attention Switching subscale) but not to the SRS.
Conclusions: The size of lexical effects on children’s speech perception can be predicted by AQ scores but not necessarily by other measures of autism-like traits. The results indicate that speech perception in children manifests individual differences along some general dimension of cognitive style reflected in the AQ, possibly in relation to local/global information processing.

U

Ota, 2008; Yu, 2010; Yu, Abrego-Collier, & Sonderegger,
2013). The Autism-Spectrum Quotient (AQ; Baron-Cohen,
Wheelwright, Skinner, Martin, & Clubley, 2001) is a questionnaire designed to test the extent to which behavioral and personality traits associated with autism spectrum conditions
(ASC) can be found across a broad range of populations, including people with no diagnosis of ASC. Adults without
ASC but with high AQ scores (i.e., those who have more traits typically associated with ASC) are less prone to the
Ganong effect, the tendency to shift identification of a sound to fit lexical expectations (e.g., in the lexical context _iss, our phoneme identification is biased toward /k/, which makes the percept a real word [kiss], rather than /g/, which makes the percept a nonword [giss]; see Stewart & Ota,
2008). Women with low AQ scores make fewer adjustments in their speech perception for phonological contexts (e.g., whether a /s/ or /S/ was heard before the vowel /a/ versus /u/) and talker variation (e.g., whether the talker is male or female; Yu, 2010). Individuals with high Attention Switching subscores of the AQ tend to imitate the voice-onset time of the interlocutor more closely (Yu et al., 2013).
In these studies, the linguistic tasks measure how listeners and speakers adjust their processing of acoustic signals (e.g., voice-onset time in consonants) depending on the context in which they occur, such as a potential word, phonological pattern, or speaker identity. Therefore, the

nderstanding and producing language involves the selection and integration of information from multiple sources, including auditory and visual signals, short-term and long-term memory of linguistic forms, and the language-specific organization of categories.
It is therefore unsurprising that speech and language processing are subject to systematic individual differences along dimensions of wider cognitive processing, such as selective attention (Francis & Nusbaum, 2002; Tipper &
Baylis, 1987), suppression (Gernsbacher, 1997), and working memory (Daneman & Merikle, 1996; Frankish, 2008; Just
& Carpenter, 1992). Such individual variation strongly suggests the involvement of nonsensory or central factors in speech and language processing (Holt & Lotto, 2010;
Watson, Qiu, Chamberlain, & Li, 1996).
One measure of individual variation that has been demonstrated to correlate with performance in speech processing tasks is the Autism-Spectrum Quotient (Stewart &

a

University of Edinburgh, United Kingdom
Heriot-Watt University, Edinburgh, United Kingdom c Scottish Government, Edinburgh, United Kingdom
Correspondence to Mitsuhiko Ota: mits@ling.ed.ac.uk b Editor: Rhea Paul
Associate Editor: Linda Watson
Received February 19, 2014
Revision received September 5, 2014
Accepted December 16, 2014
DOI: 10.1044/2015_JSLHR-L-14-0061

422

Disclosure: The authors have declared that no competing interests existed at the time of publication.

Journal of Speech, Language, and Hearing Research • Vol. 58 • 422–433 • April 2015 • Copyright © 2015 American Speech-Language-Hearing Association

performance in these tasks is likely to be related to the general degree to which individuals integrate local information
(e.g., the phonetic signal) and global information (e.g., the broader linguistic context), a dimension of cognitive style difference that has long been implicated in other domains, including visual pattern recognition (Witkin, Dyk, Fattuson,
Goodenough, & Karp, 1962; Witkin & Goodenough, 1981).
If so, the AQ is indexed to individual differences in local/ global processing. Indeed, performance in the speech processing tasks is usually best predicted by the AQ subscale that has a strong construct-based tie to local/global information integration: Attention Switching, a measure of how strongly individuals focus their attention to a single information source at the expense of others (Stewart & Ota,
2008; Yu, 2010; Yu et al., 2013). Furthermore, AQ scores correlate with other cognitive and perceptual differences that can be construed as variation in local/global information processing, for example, recognition of embedded visual patterns (Almeida, Dickinson, Maybery, Badcock, &
Badcock, 2009; Stewart, Watson, Allcock, & Yaqoob,
2009) and taste–color integration (Clark, Hughes, Grube,
& Stewart, 2013).
If the AQ is related to local/global processing, can the
AQ effects found in speech processing tasks still be intrinsically connected to autistic-like traits? Such an interpretation is in keeping with theories of weak central coherence
(WCC; Frith, 1989; Frith & Happé, 1994; Happé, 1996;
Happé & Frith, 2006) and enhanced perceptual functioning in ASC (EPF; Mottron & Burack, 2001; Mottron, Dawson,
Soulières, Hubert, & Burack, 2006). Under these views, bias for local processing is an integral cognitive feature of ASC, which causes either weak top-down processing (WCC) or highly developed low-level processing (EPF). Relatedly, several studies have found individuals with ASC to have enhanced identification and discrimination of isolated acoustic features such as pitch (Bonnel, Mottron, Peretz, Trudel,
& Gallun, 2003; Haesen, Boets, & Wagemans, 2011; Heaton,
2003; Heaton, Hudry, Ludlow, & Hill, 2008). It has been proposed that weak coherence or enhanced local perception is part of a normal distribution of ASC-like cognitive style that encompasses the general population (Happé & Frith,
2006; for the general idea that autistic-like traits are continuously distributed throughout the population, see also
Baron-Cohen et al., 2001; Hoekstra, Verweij, & Boomsma,
2007; Wainer, Ingersoll, & Hopwood, 2011). If this is the case, the relationship between the AQ and speech processing may involve some cognitive mechanisms that crucially underlie ASC.
Alternatively, the AQ effects in speech processing tasks may be related to a general processing style but not necessarily in conjunction with ASC. As mentioned, in studies that find a relationship between the AQ and speech perception patterns, the Attention Switching subscore has emerged as the best predictor. However, recent examination of the AQ has often failed to reliably extract Attention Switching as a subscale in the measure (Austin, 2005;
Hurst, Nelson-Gray, Mitchell, & Kwapil, 2007; Kloosterman,
Keefer, Kelley, Summerfeldt, & Parker, 2011; Stewart &

Austin, 2009). The implication is that the reported correlations between the AQ and speech processing come from a dimension of cognitive difference that is orthogonal to
ASC.
In the current study, we explored these issues from a developmental perspective, focusing on the AQ-related individual differences in lexical influence on speech perception. Previous research has shown that, like adults, children exhibit a lexical bias in their phoneme identification. For instance, both 5-year-olds and 9-year-olds shift their boundary of vowels (e.g., /I /–/i/) in a continuum that has a real word on one end and a nonword on the other (e.g., bib vs. beeb; Walley & Flege, 1999). However, little is known about the extent to which individual children differ in showing this effect and whether such differences relate to any known measure of cognitive variation. It also remains to be seen whether differences in global versus local processing relate to children’s speech processing, although similar individual variation has been found during infancy and childhood in visual encoding and visual pattern recognition
(Chynn, Garrod, Demick, & DeVos, 1991; Stoecker,
Colombo, Frick, & Allen, 1998).
The main goal of the current study was therefore to examine whether the relationship between AQ scores and the lexical effects on phoneme identification reported for adults by Stewart and Ota (2008) can also be found in children. We administered the AQ to 4- to 7-year-old children without ASC, along with a phoneme identification task in which they classified the initial sound of a syllable in voiceonset time continua spanning the contrast between /k/ and
/g/. The continua were constructed from pairs of words with lexical biases in opposite directions. For example, in the continuum created from kiss and giss, the listener’s perception of the initial sound is likely to shift toward /k/ under the influence of the lexical knowledge that kiss, but not giss, is an existing word. In contrast, in the continuum created from the nonword kift and the real word gift, the perception of the initial sound is likely to shift in the other direction, toward /g/. As the initial portion of the kiss–giss continuum and the kift–gift continuum (i.e., [kI]–[gI]) is acoustically identical, the phoneme identification difference between such a continuum pair can be taken as the amount of lexically induced shift in perception. On the basis of the adult results, we predicted that this perceptual shift (i.e., the
Ganong effect) would be smaller in children with a high
AQ score, who, by hypothesis, are less influenced by lexical information in making judgments about the phonetic dimensions of the stimuli and therefore are more likely to respond without bias.
The relationship between the AQ and the Ganong effect may be mediated by or confounded with several attributes of the children, such as their age, sex, level of auditory perception, and lexical knowledge or access. If we find a link between the AQ and measures of lexical knowledge or access, then it suggests that AQ effects on the phoneme identification task can be a function of better lexical knowledge or faster lexical access rather than local/global information integration. Similarly, if we find a link between the

Ota et al.: Children’s Speech Processing and the AQ

423

AQ and auditory perception abilities, then it suggests that the AQ effects on the phoneme identification task can be attributed to enhanced perception rather than local/global information integration. To eliminate these possibilities, our analysis included measures of lexical knowledge and access (accuracy and latency in a lexical decision task) and auditory perception (performance in a discrimination task) as well as the children’s age and sex as control variables.
A secondary goal of the study was to explore the nature of the correlation that might emerge from the experiment, particularly in order to find out whether the relevant variance in AQ scores is better understood as a manifestation of a cognitive characteristic related to ASC or a general processing style difference in local/global integration that is unrelated to ASC. To this end, we examined the subscales of the AQ to see whether the correlation involved a range of behaviors implicated in ASC-like profiles or only those that relate to local/global processing styles. Furthermore, we measured individual variance in autistic-like traits using another instrument, the Social Responsiveness Scale
(SRS; Constantino & Gruber, 2005), which focuses on social behaviors associated with the ASC, such as reduction in social interaction and communication, and stereotyped patterns of interests and activities. If the individual variance in AQ scores reflected in speech processing differences indeed comes from a tendency toward ASC, similar relationships should be found between the SRS scores and the speech task performance.

Method
Participants
The experiment was completed by a total of 81 children between the ages of 4 and 7, who were recruited from nurseries and primary schools in Edinburgh, United Kingdom.
None were diagnosed with ASC. All families of the children were informed of the purpose of the study and provided written consent to participate. Of these children, 62 passed the fidelity check (described later) and contributed data that were further analyzed. The resulting participant pool ranged in age from 4;2 (years;months) to 7;7, with a mean of 5;11 (fifteen 4-year-olds, fourteen 5-year-olds, seventeen 6-yearolds, and sixteen 7-year-olds). Twenty-eight of them were female and 34 were male.

Materials
Phoneme identification. The materials for the phoneme identification task consisted of three pairs of word-tononword and nonword-to-word voice-onset time (VOT) continua along the /k/–/g/ contrast. One pair was based on the words kiss and gift, with one continuum ranging from the real word kiss to the nonword giss and the other from the nonword kift to the real word gift. The other two continuum pairs were based on keep (to geep) and geese (to keese) and on kept (to gept) and guess (to kess). These continua were produced by digitally cross-splicing naturally spoken tokens
(e.g., gift, kift, kiss, and giss) read by a female speaker of

424

Standard Scottish English and recorded at a sampling rate of 48 kHz. The initial proportions of the /k/-initial nonword and /g/-initial real word (e.g., kift and gift) were replaced by those with the same onset (e.g., kiss and giss, respectively) such that the endpoint pairs were acoustically identical up to 100 ms after the onset. These tokens were then crossspliced to produce pairs of seven-point continua with equal steps, with minor adjustments made to enable splicing at zero crossings. The VOTs of the stimuli are given in Table 1.
The stimuli were down-sampled to 44 kHz and mounted on a stimulus presentation program (E-Prime).
In order to make the phoneme identification task more accessible to the young participants, visual stimuli were created to anchor the target sounds /k/ and /g/. The sound /k/ was associated with a kangaroo called Keeka (/kikə/) and the sound /g/ with a gorilla called Geega (/gigə/). These animated characters appeared next to each other with equal distance from the center of the screen and moved in reaction to the children’s response as described in the Procedure section. A screenshot of these characters is available in the online supplemental materials.
Auditory lexical decision. This task was used to measure the general speed of word retrieval involving voice contrasts in the initial position. The stimuli were 48 recorded tokens of real words and nonwords, read by the same female speaker who provided the base tokens for the phoneme identification task. Half the tokens were experimental items that were designed to be similar to the word–nonword pairs in the identification task. They were all monosyllabic words with an initial stop onset, with the real word members matched for frequency with the real words used in the phoneme identification task, based on lemma counts in the British National Corpus (BNC Consortium, 2007). The stimulus set was balanced for place of articulation ([p/b],
[t/d], [k/g]) and voice/voiceless direction. The mean age of acquisition of these words was 4.1 years, according to estimates from Kuperman, Stadthagen-Gonzalez, and Brysbaert
(2012). All experimental items are given in the online supplemental materials. The remaining 24 tokens were fillers, half of which were real words and the other half nonwords.
They were all monosyllabic words beginning in a singleton consonant (e.g., cheese, muft) or a consonant cluster (e.g., bring, fleague).
Nonword XAB discrimination. This task was used to measure children’s perceptual sensitivity to VOT differences.
Table 1. Voice-onset time (ms) of auditory stimuli used in the phoneme identification task.

Continuum pair gift–kift giss–kiss geese–keese geep–keep guess–kess gept–kept Step
1

2

3

4

5

6

7

25.5

32.6

40.8

49.3

55.5

64.0

76.6

23.6

30.0

36.9

47.7

57.0

66.1

77.1

19.6

28.1

36.2

44.8

52.6

60.8

69.4

Journal of Speech, Language, and Hearing Research • Vol. 58 • 422–433 • April 2015

The auditory stimuli for the XAB discrimination task were generated from the initial 100-ms portions of the endpoint stimuli of the kiss–gift series used in the phoneme identification task. These base stimuli were cross-spliced to yield a
[kI]–[gI] continuum that ranged in VOT from 10 to 70 ms with a step size of 10 ms. As a pilot study indicated that many young children find it difficult to discriminate a 20-ms difference in these stimuli, the XAB materials were created by concatenating two stimuli 30 ms apart (e.g., 10 and 40 ms) with an interstimulus interval of 1 s. All four permutations for each combination (i.e., AAB, ABA, BAB, and BBA) were included in the stimulus set.
To make the XAB task accessible to the young participants, a visual paradigm was used to anchor the X, A, and
B stimuli to three animated frogs. A large light-green frog
(mother frog) appeared in the center of the screen, flanked by two small frogs (baby frogs), one orange and the other dark green. The X auditory stimulus was synchronized with a croaking movement of the mother frog, and the A and
B stimuli with the croaking of each of the two baby frogs.
A screenshot of the animation is available in the online supplemental materials.
AQ-Child. The first instrument used to measure traits associated with ASC was a version of the AQ adapted for children 4–11 years of age (AQ-Child; Auyeung, BaronCohen, Wheelwright & Allison, 2008). The AQ-Child is a parent-report questionnaire with 50 items designed to assess five areas: Social Skills (“Good at social chit-chat”), Attention Switching (“Can switch back after an interruption”),
Attention to Detail (“Notices numbers of strings of information”), Communication (“Does not let others get a word in edgeways”), and Imagination (“Finds making up stories easy”). Research has shown that three of the original 50 items have questionable validity (Auyeung et al., 2008). These items, all from the Attention to Detail subscale, were therefore excluded from the present analysis. The questionnaire items were rated by the parent on a 4-point Likert scale— definitely agree (3), slightly agree (2), slightly disagree (1), and definitely disagree (0)—with reverse scoring applied to polarity-reversed items. The total range of possible scores on the AQ-Child is 0 to 147 (after the exclusion of the three items mentioned), where a higher score indicates more autistic-like traits.
SRS. The second measure of autistic-like traits was the Social Responsiveness Scale (SRS). The SRS is a
65-item parent/teacher-report questionnaire designed to measure ASC-related behaviors in individuals from 4 to
18 years of age, with a particular focus on the ability to engage in an emotionally appropriate social interaction with other individuals. Like the AQ-Child, items in the SRS are broken down into five components: Social Awareness
(“Knows when he/she is too close to someone or invading someone’s space”), Social Information Processing (“Concentrates too much on parts of things rather than ‘seeing the whole picture,’ for example, if asked to describe what happened in a story, child may talk only about the kind of clothes the characters were wearing”), Capacity for Reciprocal Social Responses (“When under stress, child seems to

go on ‘auto-pilot’”), Social Anxiety/Avoidance (“Does not join group activities unless told to do so”), and Characteristic Autistic Preoccupations/Traits (“Has repetitive, odd behaviors, such as hand flapping or rocking”). Responses are scored on a Likert scale of never true (0), sometimes true
(1), often true (2), and almost always true (3), with appropriate reverse coding. The overall score can range from 0 to
195, with a score of 75 to best discriminate children with and without ASC (Constantino & Gruber, 2005).

Procedure
Each child was tested separately in a quiet room by an experimenter who administered the three experimental tasks on a computer. The child wore a headset to listen to the recorded material. The phoneme identification task was given first, followed by the lexical decision task and then the XAB task. As an incentive, the child received a sticker after completing each task. The AQ-Child and the SRS were completed by the parent(s) of each child. Procedural details of the three experimental tasks follow.
Phoneme identification. The child first saw a video in which the two animated characters introduced themselves:
“Hi! I’m Geega the gorilla. I like the sound /g/, like in goat and giggle.” “Hi! I’m Keeka the kangaroo. I like the sound
/k/, like in kitten and kick.” The voice-over was provided by the same person who read the base test stimuli for the identification task. The experimenter then asked the child what kind of sound Geega the gorilla and Keeka the kangaroo liked, respectively. If the child responded correctly, the experimenter proceeded to the practice session. If not, she played the introduction sequence again. The replay was done only once. The practice session consisted of four items: cat, gas, gat, and cas, with the last two items created by cross-splicing the first two. The children were told that they were to choose the gorilla if they heard something that started with the sound /g/ and the kangaroo if they heard something that started with the sound /k/. Children responded by pressing one of the two buttons on a serial response box that had a printout of the two characters attached. When a button was pressed, the corresponding character was highlighted by a changing background and a sound effect was played. This sequence took 1 s, after which the next stimulus was played. Upon completion of the practice session, the experimenter once again checked with the child to see whether she or he had understood the task before starting the main task. The main task contained 42 unique stimuli (three pairs of seven-step continua) presented twice each, for a total of 84 trials. The trials were broken into four blocks, with each unique stimulus played once within the first two blocks and once in the second two blocks. Trials were randomized within each block. Children were allowed to take a brief break between blocks.
Lexical decision task. Prior to the practice session, the experimenter explained to the child that her or his task was to press the designated yes button on the response box if she or he heard a real word and the no button if she or he heard a made-up word. Each trial began with a standby

Ota et al.: Children’s Speech Processing and the AQ

425

visual stimulus (a line drawing of a girl listening to a portable recorder), which lasted for 500 ms. The stimulus was then played. When the button on the serial response box was pressed, a picture of a big smiley face with a thumbs-up was shown for 750 ms regardless of the accuracy of the response. The practice session consisted of four items: kiss, keep, gat, and geep. The experimenter checked the child’s comprehension of the task requirements and, if necessary, explained the task again before moving on to the main items. The main task contained 48 unblocked randomized trials. XAB task. Prior to the practice session, the experimenter explained to the child that there are three frogs: mother frog and two baby frogs. Mother frog croaks first, and the two baby frogs try to croak just like mother frog.
The task was to choose the baby frog that sounded exactly like mother frog for that round. The child made the response by pressing one of the two buttons on the response box that were marked with the pictures of the baby frogs.
No feedback was given, and the next item set was played
1 s after the button was pressed. The experiment started with a practice session consisting of four XAB items with VOTs of 10/10/70 ms, 10/70/10 ms, 70/10/70 ms, and 70/70/10 ms, respectively. The main session consisted of two blocks, each containing a randomized set of 16 unique trials. Children were allowed to take a short break between the blocks.

Results
Fidelity Check
In order to exclude from the main analysis children who failed to understand the procedural requirements of the phoneme identification task, signal detection theory was used to measure the participants’ fidelity to the voice distinction in the onset of the stimuli (i.e., /k/ vs. /g/), the distinction to which the children were meant to be responding. For each child, a d 0 value adjusted for two-alternative forced choice was calculated on the response data at the two ends of each continuum (e.g., kiss vs. giss), where the
VOT values matched those of the naturally spoken tokens
(e.g., of kiss and gift). The d 0 values ranged from 3.83 (consistently selecting the correct sound) to −0.121 (slightly biased to pick the wrong sound). Children with a d 0 value below
0.2 showed little sensitivity to the onset difference of these end-of-continuum stimuli and were likely to have misunderstood the nature of the task. These children (N = 19) were excluded from further analysis. As a group, the excluded children (mean age: 5;2) were younger than the included children (mean age: 5;11), F(1, 79) = 9.38, p < .01. Their mean
AQ and SRS scores were lower (35.1 and 19.9, respectively) than those of the included children (37.6 and 23.6, respectively), but these differences were not statistically significant,
AQ: F(1, 79) = 0.80; SRS: F(1, 79) = 0.84.

Main Analysis
For the auditory discrimination task, trials with reaction times longer than 2.5 standard deviations above each

426

participant’s mean latency were excluded from the analysis
(3.4% of the data). Discrimination level was indexed to mean accuracy for the 10/40-ms and 20/50-ms XAB stimuli, where the correct response rates were the highest. The reaction times in the lexical decision task were based on correct responses only and did not include trials with reaction times longer than 2.5 standard deviations above each participant’s mean latency (3.2% of the data).
Descriptive statistics for the independent factors are summarized in Table 2. The score distributions for the AQ
(M = 37.6, SD = 16.4) and the SRS (M = 23.6, SD = 16.6) were typical of children with no diagnosis of ASC. By way of comparison, the control participants in the Auyeung et al. (2008) study of the AQ-Child had a mean of 41.7 and a standard deviation of 18.6. The control participants in Constantino and Todd’s (2003) examination of the SRS had a mean of 29.9 and a standard deviation of 15.0.
Correlations among the main independent factors and among the AQ and SRS subscales are given in Tables 3 and 4, respectively. There was a strong positive correlation between the overall AQ and SRS scores, confirming the high convergent validity of the two measures (see Armstrong
& Iarocci, 2013). As can be seen in Table 4, correlations were particularly high between the AQ’s Social Skills and
Communication subscales on the one hand and the SRS’s
Social Information Processing and Capacity for Reciprocal
Social Responses subscales on the other. There was also a negative correlation between the children’s age and speed of lexical decision, presumably in reflection of age-related familiarity with the lexical items. Subscales of the two autisticlike trait measures were highly interrelated except for the
Attention to Detail AQ subscale, which did not correlate with any other AQ subscale scores.
We first examined the extent to which lexical information generally affected the response pattern in the phoneme identification task. Figure 1 shows the mean /g/ response rate for the two types of continua at each VOT step. A two-way repeated measures analysis of variance was performed on the mean proportion of /g/ responses as the dependent factor and Condition (word-to-nonword vs. nonword-to-word) and VOT step (1 through 7) as independent factors. A significant main effect of condition was found, F(1, 61) = 34.70, p < .001, as well as a main effect of
VOT step, F(6, 366) = 73.78, p < .001. There was also a marginal interaction between the two sources, F(6, 366) =
1.90, p = .08. Items in the word-to-nonword continua (e.g., gift–kift) were more likely to be judged as being /g/-initial than the VOT-corresponding items in the nonword-to-word
Table 2. Descriptive statistics of measured variables.
Factor
Autism-Spectrum Quotient
Social Responsiveness Scale
Auditory discrimination (%)
Lexical decision
(response time in ms)

Journal of Speech, Language, and Hearing Research • Vol. 58 • 422–433 • April 2015

M (SD)

Range

37.6 (16.4)
23.6 (16.6)
71.2 (15.5)
2,330 (1,544)

5–69
0–77
32–94
1,086–12,180

Table 3. Correlations between main independent factors.
SRS
total
AQ total
SRS total
Age (months)
Discrimination

Age
(months)

.696**

.113
.249

slopes, for trial block, condition, and continuum words
(gift–kiss vs. geese–keep vs. guess–kept). These models were obtained through backward elimination from initial models that additionally contained Sex (male versus female), Discrimination (mean accuracy), and Lexical decision (reaction time) as predictors, none of which was significant in the larger models and all of which were eliminated from the final models.
The results of the model including the AQ are summarized in Table 5. Positive b values indicate that a higher value in a predictor makes a /g/ response more likely.
Confirming the results of the analysis of variance reported earlier, we found significant main effects of step and condition. The odds of a child giving a /g/ response decreased as a function of an increase in the VOT step and also when the stimuli were taken from nonword-to-word continua (e.g., giss–kiss) compared to word-to-nonword continua (e.g., gift–kift). Both step and condition showed a significant interaction with age, such that older children exhibited larger differences in their response pattern depending on the VOT and lexical status of the stimuli. The interactions between age and step and between age and condition are illustrated in Figure 2 and Figure 3, respectively. Most important, there was also a significant interaction between AQ and condition.
The regression weight for the interaction was slightly positive, b = 0.0207 (or odds ratio = 1.02), counteracting the much larger main effect of condition in the opposite direction, b = −1.341 (or odds ratio = 0.26). In other words, as can be seen in Figure 4, the difference in the proportion of /g/ responses due to the lexical status of the stimuli was attenuated in children with high AQ scores compared to those with low AQ scores.
The results of the model with the SRS are summarized in Table 6. They confirm the main effects of step and condition as well as the interactions between age and step and between age and condition. However, the SRS showed a trend toward significance only as a main effect and had no significant interaction with condition. This model was not significantly different from a model without the
SRS as a factor, c2 = 8.32, df = 8, ns. Thus, there was no

Lexical decision Discrimination
−.037
−.024
.145

.122
−.003
−.401**
−.226

Note. AQ = Autism-Spectrum Quotient; SRS = Social
Responsiveness Scale.
**p < .01.

continua (e.g., giss–kiss), although this tendency was less pronounced in the short-VOT end of the continua. There was a bias toward /g/ responses (mean /g/ response rate = 64.6%), which may have been an effect of the cross-splicing of the stimuli or the attractiveness of the gorilla character in the animation. This response bias did not differ by age, F(3, 58) =
2.36, ns, or sex, F(1, 60) = 0.07, ns. Despite this bias, the result clearly demonstrates a Ganong effect in this group of children.
In order to examine how the responses in the identification task may be related to the AQ and SRS scores, we used mixed-effects models with a logistic link function. The models were fitted with the lmer4 package of R. As the AQ and the SRS were highly correlated with each other, separate models were built for these measures. For both models, the dependent variable was the participant’s response
(coded “1” for /g/ and “0” for /k/), indicating the likelihood of /g/ response. Each model contained four main effects:
Condition (word-to-word vs. nonword-to-word), Step (1 through 7, where 1 corresponded to the shortest VOT), Age of participant (in days, to avoid false convergence), and
AQ/SRS. AQ/SRS, Age, and Step were centered. Condition was coded numerically as −0.5 (word-to-nonword) and 0.5 (nonword-to-word). The models also contained the full set of interactions between Step, Condition, Age, and AQ/SRS. In addition, the models contained a random by-participant intercept and three by-participant random

Table 4. Correlations between Autism-Spectrum Quotient (AQ) and Social Responsiveness Scale (SRS) subscales.
Subscale
AQ
1. Social Skills
2. Attention Switching
3. Attention to Detail
4. Communication
5. Imagination
SRS
6. Social Awareness
7. Social Information Processing
8. Capacity for Reciprocal Social Responses
9. Social Anxiety/Avoidance
10. Characteristic Autistic Preoccupations / Traits

1

2

3

4

5

6

7

8

9

10

—

.61**
—

−.02
.12
—

.68**
.61**
.17
—

.64**
.56**
.14
.53**
—

.36*
.31*
.12
.40**
.22

.66**
.50**
.27*
.68**
.61**

.65**
.55**
.30*
.70**
.66**

.60**
.48**
.22
.44**
.55**

.57**
.61**
.44**
.57**
.56**

—

.64**
—

.62**
.82**
—

.53**
.66**
.65**
—

.53**
.70**
.75**
.57**
—

*p < .05. **p < .01.

Ota et al.: Children’s Speech Processing and the AQ

427

Figure 1. Proportion of /g/ responses in the phoneme identification task. The x-axis represents steps in voice-onset time (VOT) from the shortest (1) to the longest (7). Word-to-nonword continua had a real word on the lower VOT end
(e.g., gift–kift); nonword-to-word continua had a nonword on the lower VOT end (e.g., giss–kiss).

evidence that differences in the SRS had an impact on the effects of continuum type.
Turning back to the AQ effects, we further explored the relationship between children’s responses in the identification task and their scores in the AQ subscales (i.e.,
Social Skills, Attention Switching, Attention to Detail,
Table 5. Predictor estimates (with the Autism-Spectrum Quotient
[AQ]) for responses in the identification task.
Predictor
Intercept
Condition
Step
AQ
Age (days)
Condition × Step
Condition × AQ
Step × AQ
Condition × Age
Step × Age
AQ × Age
Condition × Step × AQ
Condition × Step × Age
Condition × AQ × Age
Step × AQ × Age
Condition × Step ×
AQ × Age
**p < .01. ***p < .001.

428

β

SE(β)

z

p

0.661
−1.341
−0.480
−0.002
−0.145
−0.047
0.024
0.002
−1.058
−0.270
0.003
0.000
0.012
0.016
0.004
−0.005

0.147
0.142
0.022
0.009
0.338
0.043
0.009
0.001
0.326
0.051
0.019
0.003
0.099
0.018
0.003
0.005

4.509
−9.462
−21.882
−0.191
−0.429
−1.094
2.693
1.618
−3.242
−5.336
−0.153
0.075
0.121
0.886
1.633
−0.957