DEVELOPMENTAL NEUROPSYCHOLOGY, 23(1&2), 201–225
Copyright © 2003, Lawrence Erlbaum Associates, Inc.
Implicit Learning in Children and Adults
With Williams Syndrome
Audrey J. Don
Children’s Seashore House
E. Glenn Schellenberg
University of Toronto
Arthur S. Reber
Brooklyn College of City University of New York
Kristen M. DiGirolamo
Children’s Seashore House
Paul P. Wang
Children’s Seashore House and University of Pennsylvania
School of Medicine
In comparison to explicit learning, implicit learning is hypothesized to be a
phylogenetically older form of learning that is important in early developmental
processes (e.g., natural language acquisition, socialization) and relatively impervi-
ous to individual differences in age and IQ. We examined implicit learning in a
group of children and adults (9–49 years of age) with Williams syndrome (WS) and
in a comparison group of typically developing individuals matched for chronologi-
cal age. Participants were tested in an artiﬁcial-grammar learning paradigm and in a
rotor-pursuit task. For both groups, implicit learning was largely independent of age.
Both groups showed evidence of implicit learning but the comparison group outper-
formed the WS group on both tasks. Performance advantages for the comparison
group were no longer signiﬁcant when group differences in working memory or
nonverbal intelligence were held constant.
Requests for reprints should be sent to Paul P. Wang, Pﬁzer Global Research and Development,
50 Pequot Avenue, 6025-B2138, New London, CT 06320. E-mail: email@example.com
DON ET AL.
Williams syndrome (WS) is a genetic disorder caused by a submicroscopic dele-
tion on chromosome 7 (Ewart et al., 1993). WS is typically manifest in mild to
moderate retardation with an unusual cognitive proﬁle, which includes rela-
tively preserved language and music skills, which contrast markedly with ex-
tremely weak visuospatial and visuomotor skills (Bellugi, Wang, & Jernigan,
1994; Dilts, Morris, & Leonard, 1990; Don, Schellenberg, & Rourke, 1999;
Mervis, Morris, Bertrand, & Robinson, 1999; Udwin & Yule, 1990, 1991). Indi-
viduals with WS also tend to be extremely sociable and outgoing (Udwin &
Researchers have suggested that implicit learning plays a crucial role in the ac-
quisition of linguistic, social, and motor skills, and possibly other skills as well
(Gomez & Gerkin, 1999; A. S. Reber, 1992, 1993). In contrast to explicit learning,
implicit learning occurs without conscious awareness and is thought to be a phylo-
genetically older form of learning, which predates consciousness (A. S. Reber,
1992; P. J. Reber & Squire, 1994). Based on this line of reasoning, A. S. Reber
(1992; Abrams & Reber, 1988) proposed that implicit learning should show rela-
tively small variations as a function of individual differences in age and maturity
and be relatively unaffected by neurological or psychological disorder. Moreover,
in contrast to explicit learning, implicit learning should be relatively independent
of measures of higher cognitive functioning (i.e., IQ; A. S. Reber, Walkenfeld, &
Hernstadt, 1991). If implicit-learning processes are indeed largely invariant to in-
dividual differences in age and IQ, then such processes may be relatively preserved
in individuals with WS and underlie their relative strengths in language and social
In this study, we assessed implicit learning in individuals with WS and in a
comparison group of typically developing individuals matched for chronological
age. Our measures were a language-based artiﬁcial grammar learning (AGL) task
and a visuomotor rotor pursuit (RP) task. The contrast between relatively strong
language abilities and weak visuomotor skills in WS invited comparison among
the abstract, nonmotor skills assessed by the AGL task, and the visuomotor skills
measured in the RP task.
A large variety of testing paradigms, including AGL and RP, have been used to
study implicit learning (A. S. Reber, 1993). The ability of individuals with am-
nesia to perform successfully on these tasks is often considered evidence of
their implicit nature (Abrams & Reber, 1988; Corkin, 1968; Sagar, Gabrielli,
Sullivan, & Corkin, 1990). Nonetheless, a few tasks thought to assess implicit
learning in normal individuals, such as the Hebb supraspan digits tasks, cannot
be learned by patients with amnesia (Charness, Milberg, & Alexander, 1988).
IMPLICIT LEARNING IN WILLIAMS SYNDROME
Additional discrepancies among tasks suggest that the processes assessed in the
various implicit-learning paradigms may be related but dissociable (Kosslyn &
Koenig, 1992; Seger, 1994; Squire, Knowlton, & Mussen, 1993).
The AGL task is one of the most widely used measures of implicit learning (for
a descriptive summary, see A. S. Reber, 1993). In the typical paradigm, individu-
als are presented with strings of letters generated from a ﬁnite-state grammar such
as the one illustrated in Figure 1. In an initial learning phase, participants are fa-
miliarized with a set of strings generated from the grammar. In a subsequent test-
ing phase, participants are asked to distinguish between novel strings that follow
the rules of the grammar and those that violate the rules. Ability to distinguish
grammatical from ungrammatical strings is taken as evidence that participants
have learned the rules of the grammar.
Early studies of AGL focused primarily on participants’ ability to distinguish
grammatical from ungrammatical strings, whereas recent research has focused
more on the nature of the learning and the representational form of the informa-
tion learned (e.g., Altmann, Dienes, & Goode, 1995; Knowlton & Squire, 1994;
Manza & Reber, 1997; Servan-Schreiber & Anderson, 1990; Vokey & Brooks,
1992). Some investigators argue that participants’ success on AGL tasks does not
reﬂect an implicit abstraction of the underlying grammatical rules, but, rather,
explicit knowledge of permissible bigrams and trigrams (referred to as chunks) in
Finite-state grammar used by Abrams and Reber (1988). Strings were generated
for this experiment by following the arrows from one node to another. Thus, XXVJ and
XVTVJ are grammatical strings; XVTJ is ungrammatical.
DON ET AL.
the test stimuli (Perruchet & Pacteau, 1990). Such knowledge of permissible
chunks would be stimulus-speciﬁc rather than abstract. Nonetheless, patients with
amnesia who were studied by Knowlton, Ramus, and Squire (1992) performed as
well as controls on AGL tasks. Because the amnesics had impaired declarative
memory for the chunks, explicit knowledge of chunking patterns is unlikely to ac-
count for their performance. Moreover, the same patients showed positive transfer
to a second AGL task that used an identical grammar instantiated in a new letter
set. Again, these ﬁndings indicate that the participants’ knowledge was not
stimulus-speciﬁc (Knowlton & Squire, 1994). Normal controls can also transfer a
learned artiﬁcial grammar across modalities (Altmann et al., 1995) and from one
letter set to additional sets (Matthews et al., 1989). This combination of ﬁndings
provides rather strong support for the proposal that abstract structures (i.e., gram-
mars) are learned implicitly in AGL paradigms.
As with AGL tasks, the RP task has a long history as an implicit-learning
paradigm (e.g., Heindel, Butters, & Salmon, 1988). This task requires ﬁne mo-
tor and visuomotor skills. Participants are asked to maintain contact between a
stylus (a metal pointer) and a target, which is placed near the edge of a horizon-
tally rotating disk. An electric current is established when the stylus is in contact
with the rotating target and accumulated contact time is recorded. Over time,
performance improves, indicating that learning has occurred. Patients with
amnesia demonstrate normal learning on the RP task, which is retained over a
delay (e.g., Brooks & Baddeley, 1976; Milner, Corkin, & Teuber, 1968). By
contrast, patients with damage to the basal ganglia, such as individuals with
Huntington’s disease, show impaired learning even after controlling for their
baseline motor dysfunction (Heindel et al., 1988). Patients with Parkinson’s dis-
ease also have basal ganglia pathology and exhibit similar impairment on the RP
task (Harrington, Haaland, Yeo, & Marder, 1990), but patients with multiple
sclerosis—who have motor impairments not attributable to the basal ganglia—
perform normally (Beatty, Goodkin, Monson, & Beatty, 1990). A single neu-
roimaging study using positron emission tomography also suggests that skilled
performance on the RP task depends on the integrity of the basal ganglia
(Grafton et al., 1992).
Findings of relatively intact AGL abilities in patients with Parkinson’s disease
raise the possibility that AGL and RP performance may be dissociable. For exam-
ple, Meulemans and Van der Linden (1998) reported that immediately after the
training phase of their AGL procedure, patients with Parkinson’s disease per-
formed as well as controls, although their performance deteriorated to chance
levels during the second half of the testing phase. P. J. Reber and Squire (1997) re-
ported more compelling evidence for preserved AGL skills in Parkinson’s disease.
In their study, participants with Parkinson’s disease performed similarly to con-
trols and demonstrated positive transfer to a novel letter set (a new instantiation of
the same grammar).
IMPLICIT LEARNING IN WILLIAMS SYNDROME
IMPLICIT LEARNING ACROSS DEVELOPMENT AND IN INDIVIDUALS
WITH MENTAL RETARDATION
Most studies of implicit learning have been conducted with adults. Although a
slight decline in performance is noted in very old adults, implicit learning appears
to be remarkably stable across most of adulthood (Curran, 1997; Howard &
Howard, 1997; for a review, see A. S. Reber & Allen, 2000). Studies of implicit
learning in younger participants indicate that children can learn artiﬁcial gram-
mars, but it is uncertain whether their performance matches that of adults. In one
study, good AGL performance was evident in children 9 to 11 years of age
(Fischer, 1997). In another study, Gomez and Gerkin (1999) used a head-turn pref-
erence procedure to assess whether 1-year-old infants could distinguish strings that
conformed to an artiﬁcial grammar from strings that violated the grammar. Testing
was conducted after less than 2 min of exposure to examples of grammatical
auditory strings (sequences of nonsense syllables). Infants successfully distin-
guished between grammatical and ungrammatical strings, and they also
transferred their knowledge to a second task in which the same grammar was
instantiated in a new vocabulary. Performance could have been influenced,
however, by explicit as well as implicit strategies, and the authors did not make
any claim about the cognitive strategy used by their infant participants.
Nonetheless, other studies with younger infants provide converging evidence
of implicit learning in infancy. For example, when presented with structured
sequences of nonsense syllables for brief periods of time, 7- and 8-month-old
infants subsequently exhibit knowledge of the transitional probabilities
between consecutive syllables (Saffran, Aslin, & Newport, 1996) and the gram-
matical rules that were used to construct the stimulus sequences (Marcus,
Vijayan, Rao, & Vishton, 1999). Presumably, such learning in young infants is
implicit rather than explicit.
Other studies provide additional support for the idea that implicit-learning
abilities are well developed in early childhood and do not vary signiﬁcantly with
increased age. For example, Mecklenbraeuker, Wippich, and Schulz (1998)
assessed memory for picture puzzles and found no difference in performance
between younger (6–7 years of age) and older (9–10 years of age) children.
Meulemans and Van der Linden (1998) found that sequence learning on a serial
reaction-time task was equivalent for children (age 6–10) and young adults (age
18–27) and remained equivalent at follow-up 1 week later.
By contrast, Maybery, Taylor, and O’Brien-Malone (1995) observed that per-
formance on an implicit contextual-learning task varied with age but not with IQ.
Participants were children from two age groups (5–7 years and 10–12 years) sub-
divided into three IQ subgroups ranging from the borderline to the superior
range. The older children performed at above-chance levels regardless of IQ,
whereas the younger groups performed at chance. Another study from the same
DON ET AL.
laboratory examined implicit-learning abilities across an even wider range of IQ
(Fletcher, Maybery, & Bennett, 2000). Participants who were diagnosed with
mental retardation performed less well on the contextual learning task than did
participants with IQ scores in the normal range.
Other studies of individuals with mental retardation have used visual-priming
paradigms to assess implicit memory rather than implicit learning. In priming
tasks, participants are exposed to stimuli without explicit instructions to memo-
rize the stimuli. Priming is inferred if subsequent identiﬁcation of the same
stimuli (often degraded) is facilitated. Priming can result from a single prior pres-
entation, whereas implicit-learning paradigms involve multiple presentations of
the information that is to be learned.
Results from priming studies of individuals with mental retardation are equiv-
ocal. In a large-sample study, Wyatt and Connors (1998) found that 6- to 17-year-
olds with mental retardation performed similarly to a control group (matched for
age) on a visual priming task (picture-fragment completion). In other words, per-
formance on this task appeared to be independent of IQ. Nonetheless, age-related
increases in performance were observed for both groups. Similarly, Vicari,
Bellucci, and Carlesimo (2000) reported that repetition priming was preserved in
14 individuals with Down syndrome (M age = 21 years), who were compared to
a control group matched for mental age (M chronological age = 5 years). These
investigators also administered a simpliﬁed version of Nissen’s serial reaction-
time test (Nissen, Willingham, & Hartman, 1989) and found preserved perform-
ance on that task as well. By contrast, Mattson and Reilly (1999) found that
mentally retarded children with Down syndrome showed signiﬁcantly less prim-
ing than two control groups (both with higher mean IQ scores), who performed
In sum, the existing literature leads to competing hypotheses about the
implicit-learning abilities of individuals with WS. On the one hand, A. S. Reber
(1992) contended that implicit learning should be robust in the face of neuro-
logical disorder. The ﬁndings of some investigators that implicit learning is
preserved in populations with mental retardation are consistent with this pro-
posal. On the other hand, the basal ganglia appear to subserve some forms of
implicit learning, and MRI ﬁndings indicate that basal ganglia volumes are di-
minished in WS (Jernigan, Bellugi, Sowell, Doherty, & Hesselink, 1993).
Moreover, other investigations have reported implicit-learning and priming
deﬁcits in samples of mentally retarded individuals. In other words, predictions
about implicit learning in WS were unclear. Nonetheless, because explicit
learning is impaired in WS, we predicted that the comparison group would at-
tend more than the WS group to the chunk strength of the stimuli in the AGL
task. Moreover, marked visuospatial and motor impairments in WS led us to
expect that implicit learning would be better preserved in the AGL task than in
the RP task.
IMPLICIT LEARNING IN WILLIAMS SYNDROME
The WS group consisted of 27 individuals (14 male, 13 female) between 9 and
49 years of age (M = 23 years, 7 months; SD = 13 years, 6 months). The age-
matched comparison group consisted of 27 normally developing adolescents and
adults (11 male, 16 female), who ranged in age from 9 to 50 years (M = 23 years,
7 months; SD = 13 years, 4 months). Matching was within 6 months for partici-
pants 9 to 13 years of age, within 1 year for 14- to 29-year-olds, and within
2 years for those 30 years and over. Participants with WS were recruited through
the local chapter of the Williams Syndrome Association and through national and
regional meetings of the same organization. The comparison group consisted of
siblings of the WS participants, employees of the institution where the research
was conducted, and others recruited by word of mouth. All participants spoke
English as their primary language and were without signiﬁcant sensory or phys-
Implicit learning was assessed with an AGL task and an RP task. Short-term
memory, working memory, receptive vocabulary, and nonverbal reasoning were
also assessed to investigate the relationship between these variables and the
For the AGL test, we generated 36 letter strings
from the ﬁnite-state grammar shown in Figure 1 (Abrams & Reber, 1988). This
grammar was used to create 16 training strings and 20 test strings, each of which
was two to ﬁve letters in length. In addition, 20 ungrammatical letter strings of two
to ﬁve letters were created. Each ungrammatical string violated the rules of the
grammar at only one position in the string, and such violations occurred at all po-
sitions. The 20 grammatical and ungrammatical strings were paired by length.
Grammatical strings and foils were also paired on the basis of chunk strength, a
term that refers to the presence of potentially familiar fragments within the
Chunk strength was calculated in the manner of Knowlton and Squire (1996).
The 16 training stimuli were examined to determine the frequency with which
each possible bigram and trigram appeared across the training set. This frequency
parameter has been termed the associative strength of each chunk. Each test stim-
ulus was then examined to determine the number of bigrams and trigrams that
appeared in each stimulus. For example, in the stimulus string XXVJ, the bigrams
DON ET AL.
XX, XV, and VJ appear, as do the trigrams XXV and XVJ. The chunk strength of
each stimulus was calculated by averaging the associative strength of all the bi-
grams and trigrams that it contained. Half of the targets and half of the foil strings
were created as high-chunk-strength strings. The other half had low chunk
strength. The grammatical high-chunk-strength stimuli had a mean chunk strength
of 6.81 (SD = 1.13); the low-chunk-strength stimuli averaged 3.60 (SD = 0.74).
The ungrammatical, high-chunk-strength strength stimuli averaged 6.14
(SD = 1.26); the ungrammatical, low-chunk-strength strength stimuli averaged
3.42 (SD = .58).
Five pairs of each possible combination of high-chunk-strength and low-chunk-
strength targets and foils were included in the set of 20 grammatical/ungrammatical
test pairs (i.e., ﬁve sets each of high/high, low/low, high/low, or low/high target-foil
combinations). The number of letters per string was matched in each pair.
The 16 training strings were printed in large bold letters on 3 × 5-in. cards and
decorated with dinosaur stickers to create interest. Participants were shown the
complete deck of training stimuli three times (48 training trials in total) in stan-
dard order and asked to spell the stimuli (called “dinosaur words”) out loud each
time. Testing immediately followed training. Test items consisted of 20 pairs of
target–foil strings printed the same size as the training stimuli and placed one pair
per page. One string was placed at the top of the page; the other was placed at the
bottom. The placement of the target and the foil on the page was pseudorandom,
with the target never appearing more than four times consecutively in one posi-
tion. Participants were shown the new words, in pairs, and asked to spell both
words out loud. They were then asked to identify the dinosaur word in each pair.
This “forced-choice” design is a departure from the more typical classiﬁcation
task that has been used in other AGL studies. The method was modiﬁed to mini-
mize yes or no response biases that might occur for individually presented stim-
uli. The entire block of 20 pairs was repeated immediately after the ﬁrst block with
string placement reversed and page sequence randomized. Thus, participants
identiﬁed dinosaur words in a total of 40 string pairs. The outcome measure was
the number of items answered correctly.
The motor-learning task was a standard RP task. Participants were asked to
maintain contact between a stylus and a target on a rotating disk. On each trial, the
disk rotated at 30 RPM for 20 sec. Trials were presented in blocks of four, with a rest
period of 20 sec between trials. Duration of contact was recorded for each block. Six
blocks were completed in a single testing session with a rest of approximately 1 min
after blocks 1, 3, and 5 and a rest of about 3 min after blocks 2 and 4. The primary
outcome measure was duration-of-contact on the sixth (ﬁnal) block. We also exam-
ined participants’ improvement in performance across the six blocks.
The Peabody Picture Vocabulary Test–Revised
(PPVT–R; Dunn & Dunn, 1981) was used to assess receptive vocabulary. The
IMPLICIT LEARNING IN WILLIAMS SYNDROME
Matrices subtest of the Kaufman Brief Intelligence Test (K–BIT; Kaufman &
Kaufman, 1990) was used to estimate nonverbal intelligence.
The Number Recall subtest of the Kaufman Assessment Battery for Children
(K–ABC; Kaufman & Kaufman, 1983) provided a measure of short-term verbal
memory. This subtest requires participants to repeat increasingly longer se-
quences of digits presented at a rate of one digit per sec. The Spatial Memory
subtest from the K–ABC was also included in the testing protocol but the com-
parison group performed at ceiling levels. Thus, it was excluded from further
Working memory was assessed with the Counting Span Test, an experimental
measure adapted from Case, Kurland, and Goldberg (1982). Participants were
shown an array of large blue and yellow dots arranged randomly on a page and
asked to touch and count all of the blue dots on a page (e.g., 5). They were then
shown another page of blue and yellow dots and again asked to count (beginning
at 1) all of the blue dots on the page (e.g., 3). Finally, the participant was shown a
page with a question mark and asked to recall the number of blue dots on each
page in the order they were presented (e.g., 5, 3). Testing began with a block of
ﬁve two-page trials. Additional pages were added for subsequent blocks until tri-
als included a maximum of ﬁve pages. Testing was terminated when participants
failed two or more trials within a block.
Participants were tested at their convenience, either in the research laboratory of a
children’s hospital, at meetings of the Williams Syndrome Association, or at home.
Testing took place in a quiet room that was free from distractions. Tasks were
administered in a ﬁxed order. Matrices and the RP task were administered ﬁrst.
Number Recall and Spatial Memory were tested during rest periods of the RP task.
The ﬁnal three tests were Counting Span, the AGL task, and PPVT–R. Testing took
approximately 75 min to complete. For 1 participant with WS, results from the sec-
ond half of the AGL task were unavailable because he became fatigued and refused
to complete the second half of the task. For 1 participant in the comparison group,
RP scores were excluded because of technical difﬁculties.
A set of preliminary analyses examined differences between the WS and compar-
ison groups on the raw scores obtained on the supplementary measures.
DON ET AL.
Descriptive statistics are provided in Table 1. The comparison group performed bet-
ter than the WS group on each measure, PPVT–R: t(52) = 3.26, p = .002; Matrices:
t(51) = 10.31, p < .001; Number Recall: t(52) = 5.41, p < .001; Counting Span:
t(51) = 8.16, p < .001. Whereas group membership accounted for only 17% of the
variance (i.e., eta-squared) in PPVT–R scores, it explained at least 36% of the vari-
ance in each of the other three measures (Number Recall 36%; Counting Span 57%;
Matrices 68%). This pattern is consistent with other results showing that the vocab-
ulary knowledge of individuals with WS is relatively spared compared to their
marked impairments in other domains. Indeed, the WS group had higher standard
scores on our test of receptive vocabulary (PPVT–R) than on our test of nonverbal
reasoning (Matrices), t(22) = 2.10, p = .048. By contrast, the comparison group
performed better on Matrices than on PPVT–R, t(26) = 2.19, p = .038.
Implicit Learning: Artiﬁcial Grammar
Scores on the AGL task represent the total number of correct responses
(maximum = 40). Descriptive statistics are provided in Table 2. One-sample t tests
(one-tailed) were used to compare performance with chance levels (20 correct).
The comparison group performed signiﬁcantly better than chance, t(26) = 8.39,
p < .001, with performance levels (66% correct) similar to that reported in other
AGL studies with normal individuals (Knowlton et al., 1992; McAndrews &
Moscovitch, 1985; A. S. Reber & Allen, 2000). The WS group was only marginally
better than chance, t(25) = 1.36, p = .093. Whereas 56% of participants in the
comparison group (15 of 27) performed signiﬁcantly better than chance as individ-
uals (binomial test), only 15% did so in the WS group (4 of 26). An independent-
samples t test conﬁrmed that the difference among groups was reliable,
Descriptive Statistics for Scores on the Supplementary Measures
PPVT–R: raw score
PPVT–R: standard score
Matrices: raw score
Matrices: standard score
Number Recall: raw score
Counting Span: raw score
PPVT–R = Peabody Picture Vocabulary Test–Revised.
*Williams syndrome and comparison groups signiﬁcantly different, ps < .005.