This article was downloaded by: [University of Victoria]
On: 18 November 2014, At: 12:48
Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Neurocase: The Neural Basis of Cognition
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/nncs20

Complex Graphemes as Functional Spelling Units:
Evidence from Acquired Dysgraphia
a

M. J. Tainturier & B. C. Rapp
a

b

School of Psychology , University of Wales , Bangor, UK

b

Cognitive Science Department , Johns Hopkins University , Baltimore, Maryland, USA
Published online: 02 Feb 2010.

To cite this article: M. J. Tainturier & B. C. Rapp (2004) Complex Graphemes as Functional Spelling Units: Evidence from
Acquired Dysgraphia, Neurocase: The Neural Basis of Cognition, 10:2, 122-131
To link to this article: http://dx.doi.org/10.1080/13554790409609943

PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions

Neurocase
2004, Vol. 10, No. 2, pp. 122–131

Complex Graphemes as Functional Spelling Units:
Evidence from Acquired Dysgraphia
M. J. Tainturier1 and B. C. Rapp2
1

School of Psychology, University of Wales, Bangor, UK and 2Cognitive Science Department,
Johns Hopkins University, Baltimore, Maryland, USA

Downloaded by [University of Victoria] at 12:48 18 November 2014

Abstract
The visual word recognition literature suggests that complex graphemes (or digraphs) such as CK function as units.
This proposal has also been put forward in recent spelling models (Houghton and Zorzi, 2003) and the study we report on
here provides initial empirical support for the claim. We performed detailed analyses of the spelling performance of
two brain-damaged individuals with graphemic buffer deficits. Results revealed that (a) FM and BWN made fewer errors on
consonant digraphs (e.g., CK) than on matched controls clusters (e.g., CR) and (b) BWN produced more transposition
errors on vowel digraphs than on control clusters. These results support the view that digraphs are represented as
units in which the relative order of constituent letters is encoded.

According to most current theories, spelling involves at least
two sets of processes (Fig. 1). First, a spelling can be
generated sublexically by relying on knowledge of the frequent correspondences between phonemes and graphemes.
This sublexical phonology to orthography conversion process
generates plausible spellings for both familiar and unfamiliar
words (or pseudowords). However, the output may not be
entirely accurate in orthographically opaque languages such
as English because many words have an ambiguous or
irregular spelling (e.g., the word ‘‘phone’’ could be spelled
‘‘fone’’). Second, the spelling of familiar words can be
generated lexically, by accessing a stored representation of
the spelling of words in an orthographic lexicon. This would
allow familiar words to be spelled correctly. Although it can
be assumed that the spelling of nonwords is mostly generated
sublexically while that of familiar words is mostly generated
lexically, lexical and sublexical processes will jointly contribute to the activation of a set of abstract graphemic units
(see Fig. 1 and Rapp et al., 2002; Bosse et al., 2003). It has
been proposed (Caramazza et al., 1987) that the abstract
orthographic representations generated by these processes
will be held by a dedicated short-term memory process
(graphemic buffer) during the operation of more peripheral
mechanisms responsible for transforming abstract orthographic representations into form representations that are
specific to the chosen modality and format of output (e.g.,

print, cursive, oral spelling, typing, upper/lower case and so
on). This buffering process is required due to the sequential
nature of letter name/shape selection processes.
But what is the content and organization of the orthographic
representations that are held at the grapheme level? What
information is represented? How is it structured? Minimally,
orthographic representations must encode information about
letter identity and order (e.g., C1A2T3). However, several lines
of evidence now support the idea that orthographic representations are more than linear sequences of abstract letter
identities (for a review, see Tainturier and Rapp, 2001). Most
of this evidence comes from the analysis of the error patterns
of dysgraphic individuals, although there is some converging
evidence from studies of normal spellers. Studies of ‘‘graphemic buffer deficits’’ have proved particularly revealing.
Such deficits typically lead to letter substitutions, omissions,
additions and transpositions that are taken to reflect the
disruption of information about letter identity and order that
occurs when the graphemic buffering capacity has been
compromised. If orthographic output representations were
merely linear sequences of letter identities, one would
not expect the nature and position of letter errors to be affected
by factors that assume a more complex, non-linear organization of the letter identities. Yet, several results indicate
that orthographic representations are, in fact, internally
complex.

Correspondence to: M. J. Tainturier, School of Psychology, Brigantia Building, University of Wales Bangor, Bangor, Gwynedd, LL57 2AS, UK.
Tel: þ44-1-248-382-714; e-mail: m.j.tainturier@bangor.ac.uk

10.1080/13554790490497238$16.00 # Taylor & Francis Ltd.

Downloaded by [University of Victoria] at 12:48 18 November 2014

Spelling complex graphemes

Fig. 1. The spelling process.

First, there is evidence that morphemic structure is encoded
in orthographic representations (for a review, see Allen and
Badecker, 2001). A clear demonstration was provided in a
study of case DH (Badecker et al., 1990). In addition to the
characteristic features of a graphemic buffer deficit, DH’s
spelling showed a serial position effect such that the probability of an error on a given letter increased as a function of
the position of that letter in the word. Strikingly, however, this
pattern only held for monomorphemic words like ‘‘car’’. In
the case of bimorphemic words like ‘‘darkness’’, error rates
increased up to the end of the word stem (‘‘dark’’) but then
dropped on the first letter of the suffix to increase again on
subsequent letters. This indicates that the information
retrieved from the orthographic lexicon and later held in
the graphemic buffer is organized into morpheme-sized units
rather than whole words.
Second, Caramazza and Miceli (1990; see also Ward and
Romani, 2000) have proposed that orthographic representations also encode information about syllabic structure (but see
Jónsdóttir et al., 1996). This conclusion was partly based on
the observation that case LB, an Italian patient with a graphemic buffer deficit, showed a marked tendency to simplify
the syllabic structure of words. For example, letter omissions
were frequent in the context of letter clusters, usually leading
to simpler syllabic structures (e.g., ‘‘creatura’’ ! RETURA).
However, omissions virtually never occurred in simple
CV sequences, where an omission would create a more
complex syllable (that is, LB did not produce errors such as
‘‘creatura’’ ! CREATRA). In a typing study with normal
participants, Zesiger et al. (1994) also demonstrated an
influence of syllabic structure on the temporal course of
spelling: the interval between two strokes was shorter if
the two corresponding letters belonged to the same syllable
than if they did not.
Third, several studies of graphemic buffer deficits have
shown that information about specific letter identities can be
lost while knowledge of the consonant/vowel status of these

123

letters is preserved. In several cases (Miceli et al., 1985;
Caramazza and Miceli, 1990; McCloskey et al., 1994;
Jónsdóttir et al., 1996; Buchwald and Rapp, 2003) substitution errors virtually always occurred within class, with consonants being substituted for consonants and vowels for
vowels. This suggests that letter identity and consonant/vowel
status may be represented separately. In addition, there have
been several reports of patients who made many more errors
on vowels than on consonants (Cubelli, 1991; Cotelli et al.,
2003) and vice versa (Miceli et al., this issue).
Finally, there is strong empirical support for distinct and
dissociable representations of letter identity and of letter
doubling information (Caramazza and Miceli, 1990;
McCloskey et al., 1994; Miceli et al., 1995; Tainturier and
Caramazza, 1996). For example, Italian case SFI (Miceli et al.,
1995) had a selective deficit in the production of double letters
leading to errors like leggo ! LEGO. Case FM (Tainturier
and Caramazza, 1996) showed the reverse pattern, with
doubling information being much better preserved than letter
identity. Consistent with the view that information about letter
doubling and letter identity is dissociable, FM often replaced a
double-letter with another double-letter (e.g., umbrella !
UMMOUCAN, ribbon ! BROLLOW). Furthermore, the
study showed that the representation of double-letters is
distinct from that of other superficially similar groups of
letters such as non-adjacent repeated letters (e.g., CaCtus)
or digraphs (e.g., roCKet).
In summary, there is considerable evidence that orthographic representations are more than simple linear sequences
of abstract letter identities. In particular, various factors such
as morphological and syllabic structure, letter doubling and
CV status can affect the patterns of errors of dysgraphic
subjects or the temporal properties of normal spelling. In this
article, we will examine whether letter sequences that correspond to a single phoneme (such as /k/ ! CK in ROCKET or
/f/ ! PH in PHONE) are represented and processed differently from otherwise similar letter sequences that do not
correspond to a single phoneme (such as ST or PL). We will
refer to sequences such as CK and PH as complex graphemes
or digraphs.
In their connectionist simulation of the spelling process,
Houghton and Zorzi (2003) have proposed that letter
sequences that correspond to a single phoneme are represented as single units, even when they include more than one
letter. Thus, in terms of abstract grapheme representations, a
word such as ‘‘wreath’’ is composed of three grapheme units
(WR þ EA þTH), while ‘‘strict’’ is composed of six
(S þ T þ R þ I þ C þ T). Both do, of course, include six letters,
which becomes relevant at subsequent stages of production
(see Fig. 1). The proposal that complex graphemes differ from
other sequences such as CR or ST is intuitively appealing
because, for sequences such as /f/ ! PH, the correspondence
between orthography and phonology is much more transparent at the level of graphemes than at the level of single letters.
Furthermore, this proposal is supported by the finding of
Houghton and Zorzi (2003) that treating complex graphemes

Downloaded by [University of Victoria] at 12:48 18 November 2014

124

M. J. Tainturier and B. C. Rapp

as units improved both the accuracy and plausibility of
responses of their spelling simulation.
In the written word recognition and reading literature, there
is some empirical support for the view that complex graphemes (other than double letters) are processed as units. Rey
et al. (2000) have shown that it takes longer to detect a single
target letter when it forms part of a complex grapheme than
when it does not (e.g., detecting the letter O in the written
word FLOAT as opposed to SLOPE). Their interpretation of
this finding is that multi-letter graphemes function as perceptual units in reading, making it harder to detect letters that are
embedded within them. Consistent with the view that digraphs
are processed as units, Pring (1981) reported that case-alternated words were recognised more slowly when case alternation disrupted the unity of digraphs (e.g., BREad) than
when it did not (BRead). Similar results were obtained by
Martensen et al. (2003) using slashes to visually separate
letters: words such as ‘bre//ad’ were recognised and named
more slowly than words such as ‘br//ead’. Dickerson (1999),
Joubert and Lecours (2000) and Rastle and Coltheart (1998)
provide further evidence for a distinct processing of complex
graphemes in reading. In addition, Plaut et al. (1996) were
able to significantly improve upon the performance of the
connectionist reading simulation reported by Seidenberg and
McClelland (1989) by representing the input to the reading
system in terms of graphemes.
If complex graphemes are represented at the abstract
grapheme level, what might we observe subsequent to neurological injury affecting the spelling process? Our expectation would be that a process that is sensitive to the number of
elements in a representation (e.g., graphemic buffering)
should show an advantage for complex graphemes over
sequences of simple graphemes. This is because there would
be less information to activate and buffer for a word such as
WREATH (three graphemic units) than for a word such as
STRICT (six units).
On the other hand, complex graphemes may be at a
disadvantage later in the spelling process, given that spelling
ultimately involves the serial production of individual letters.
If complex graphemes have a unitary representation at one
level of processing there must be a later stage of processing at
which their constituent components must be identified or
recovered. In order for this to be accomplished, we must
assume that a digraph’s constituent letters and their ordering
are represented in a manner (e.g., [W1 R2]) which subsequent
processes can interpret or ‘‘unpack’’ (Fig. 1). Consequently,
although complex digraphs may have an advantage over
clusters at one level, they may be more complex than cluster
sequences at the stage that requires the identification and
ordering of their constituent elements. Relevant to this discussion is the proposal that there is a distinction between the
activation and maintenance of orthographic information on
the one hand, and the serially ordered selection of the buffered
letters for production on the other (Houghton et al., 1994;
Rapp and Kong, 2002; Houghton and Zorzi, 2003). Although
digraphs have not been specifically discussed (or modeled)

in terms of the proposed activation/selection processes, it
would seem that the process responsible for selecting letters
for production must not only preserve the order of the elements that have been activated and buffered but that it would
also need to recover the order of letters within digraphs. Given
the proposed internal structure of digraphs, if the selection/
unpacking process is disrupted, we might expect digraphs to
be particularly vulnerable to transpositions of their constituent
elements (e.g., ‘‘wreath’’ ! WRAETH). Blanken et al. (1999)
present evidence which is generally consistent with this view.
Also consistent with the view that digraphs might be more
vulnerable at some level of the spelling process, Caramazza
and Miceli (1990) reported that Italian patient LB tended to
write the sequence ‘‘sc’’ less accurately when it corresponded
to a single phoneme as in ‘‘scena’’ /sena/ (29.2% errors) than
when it corresponded to two phonemes as in ‘‘scusa’’ /skuza/
(17.9% errors). However, these particular findings should be
taken with caution, as digraphs and control clusters were not
matched on a number of variables which strongly affect the
performance of individuals with graphemic buffer deficits,
such as word length and position of the sequence within the
word.
In this study, we will provide empirical support for the
view that complex graphemes (such as those in WR þ EA þ
TH) are represented in a unitary manner at some level of the
spelling system, as proposed in Houghton and Zorzi (2003).
This will be done through a detailed analysis of the spelling
errors of two individuals with acquired graphemic buffer
deficits: FM and BWN. We will also present evidence of
the internal organization of digraphs based on an analysis of
the transposition errors produced by case BWN.

Methods and results
This section will be organized as follows: we first present case
histories of the two individuals who are the subjects of this
report, FM and BWN; we then present evidence that they
suffer from deficits affecting the buffering of abstract graphemes; finally, we present experimental findings regarding
their spelling of digraphs and control segments.

Case histories
FM
At the time the data for this study were collected, FM was a 46
year-old male who had suffered a stroke ten years earlier. He
was right-handed and a high school graduate who had been
working as a truck driver prior to the stroke. A CT-scan
performed two years post-stroke revealed a large area of
damage in the territory of the left middle cerebral artery.
More specifically, damage involved the infero-posterior frontal lobe, the inferior parietal lobe and the anterior temporal
lobe, as well as the underlying white matter and lateral basal
ganglia; there was also cortical atrophy of the remainder of the
left frontal convexity.

Downloaded by [University of Victoria] at 12:48 18 November 2014

Spelling complex graphemes

125

At the time FM’s spelling was investigated, FM showed right
hemianopia, hemiparesis mostly affecting the upper limb, and
he was still severely aphasic. His language disorder would be
clinically characterized as agrammatic Broca’s aphasia. That
is, his speech was laborious and ‘telegraphic’, with word
finding difficulties, phonetic/phonological errors and morphological errors. Spoken comprehension was very well preserved, except for syntactically complex sentences (see
Badecker et al., 1991). FM’s reading aloud was also severely
impaired and characteristic of deep dyslexia. Specifically, he
was most impaired for long, abstract and low-frequency words,
and his performance was affected by grammatical category,
such than nouns were read better than adjectives, which were
read better than verbs, which were read better than function
words. FM could not read aloud pseudowords and, when
reading words, he produced visual, semantic, visual/semantic
and morphological errors. More details on FM’s reading
impairment can be found in Badecker and Caramazza (1987).

BWN
BWN was a 79 year-old, right-handed man who had suffered
two strokes, four years and one year prior to the onset of this
investigation. He held a Ph.D., had worked as a teacher, a
high-level school system administrator and was president of
the city chapter of a prestigious civic organization. Premorbidly, BWN was an excellent public speaker and speller.
While in school he had been a finalist in regional spelling
competitions and his editorials and letters to local newspapers
reveal a man who was particularly gifted in written expression. CT scans indicate an older left parietal lesion and a more
recent right parietal lesion (see Fig. 2). The left hemisphere
lesion is the larger of the two, extending in the superiorinferior dimension from the superior parietal gyrus to the
superior temporal gyrus and, in the anterior-posterior dimension, from the post-central gyrus to the angular gyrus. The
right hemisphere lesion affects the posterior parietal lobe,
including the angular and superior parietal gyri.
As a result of his strokes, BWN suffered mild right-sided
weakness and some somatosensory loss in the right hand but
he showed no evidence of visuo-spatial neglect. He had
moderate difficulties in spoken language production, primarily characterized by word finding difficulties and phonological errors. Although BWN suffered mild hearing loss and
had some auditory discrimination difficulties (word discrimination 90%, nonword discrimination 91%, mean normal
performance 98%), his auditory word and sentence comprehension was very good. Written comprehension was also good
(e.g., 100% correct in synonym matching for abstract and
concrete words) and he read the newspaper for pleasure,
although not as easily as he had prior to the strokes. His
spelling difficulties are described below.

Graphemic buffer deficits
Characteristics of a selective graphemic buffer deficit
(Caramazza et al., 1987) include: (a) a significant effect of

Fig. 2. BWN’s CT scan.

word length (measured in letters rather than phonemes); (b)
errors consisting of letter substitutions, deletions, transpositions and additions (and not phonologically plausible errors or
semantic errors); (c) comparable performance across modality of input (e.g., in spelling to dictation, written picture
naming, delayed copy transcoding); (d) comparable performance across modality of output (oral or written spelling); (e)
comparable performance across stimulus types (words/nonwords, regular/irregular, abstract/concrete); (f) absence of an
effect of lexical frequency; (g) a bow-shaped accuracy function across letter positions such that performance is most
accurate with initial and final letters and least accurate with
letters in the center of words.

FM
FM’s spelling performance has been presented in detail elsewhere (Tainturier and Caramazza, 1996) and will only be
summarised here. FM’s spelling was tested extensively 10 years
after his stroke. Over a period of approximately 15 months, a
large corpus of errors was collected in spelling to dictation (823
stimuli), written picture naming (379 stimuli), and delayed copy
transcoding (842 stimuli). In the latter task, FM was asked to
reproduce in uppercase a stimulus that had been presented in

Downloaded by [University of Victoria] at 12:48 18 November 2014

126

M. J. Tainturier and B. C. Rapp

lowercase and then hidden from view. This task is assumed to
require graphemic buffering to maintain the activation of letter
identities while the forms of the letters are produced.
Performance in these three tasks showed the characteristics
of a graphemic buffer deficit. First, the deficit affected all
stimulus types (regular words, irregular words and nonwords)
and was very similar across the three spelling tasks which,
despite differing in input, all make use of the graphemic
buffer. Second, errors increased markedly with stimulus
length (see Fig. 3). Third, there was a serial position effect
that has been reported for a number of graphemic buffer
deficits, such that letters from medial positions were less
likely to be produced correctly than letters in initial and final
positions. Finally, the type of errors reflected a disruption of
information about letter identity and order. Thus, the majority
of errors were nonword responses that differed in their degree
of similarity to the targets (pea ! PAE; grapes ! GPAGE).
More specifically, most errors could be described in terms of
single letter substitutions (panel ! PADEL), deletions (apricot ! APICOT), misorderings (ground ! GRONUD), additions (beauty ! BEAUNTY), or various combinations of the
above. As can be seen in Fig. 4, the distribution of single-letter

Fig. 3. Case FM: The effect of word length in spelling to dictation (STD),
written picture naming (WPN) and delayed copy transcoding (DCT).

errors was very similar across all three tasks. This set of
characteristics is clearly indicative of a deficit at the level of
the graphemic buffer.
FM’s performance in spelling to dictation and written
picture naming also suggested an additional deficit to the
lexical semantic system and/or the orthographic output lexicon, given that FM produced semantic errors in those tasks
(11% and 16%, respectively). In delayed copy transcoding,
which involves the graphemic buffer but does not require
lexical mediation, FM produced fewer semantic errors (2%),
his errors tended to be closer to the targets, and overall
accuracy was higher (54% errors in delayed copy transcoding,
versus 76% in spelling to dictation and 66% in written picture
naming). This is consistent with the view that performance in
delayed copy transcoding reflected almost exclusively the
graphemic buffer deficit, while performance in the other
two tasks was influenced by the additional lexical deficit.

BWN
BWN’s spelling performance clearly indicated that his spelling errors arose at the level of the graphemic buffer (see also
Buchwald and Rapp, 2003). BWN was administered 3947
words (between 4 and 10 letters in length) for written spelling.
Because of his mild hearing loss, BWN was asked to either
repeat each stimulus or indicate its meaning prior to spelling
the word. If a stimulus was misperceived, it was repeated until
BWN demonstrated that he had correctly understood it and
only then did he spell it. BWN spelled 24% of words
incorrectly and (see Table 1) his error rate for long words
(8–10 letters) was significantly greater than his error rate for
short words (4–6 letters), whether accuracy is calculated in
terms of word errors (w2 ¼ 539.2, p < .0001) or letter errors
(w2 ¼ 594.8, p < .0001). Furthermore, as indicated in Fig. 5,
BWN exhibited the characteristic bow-shaped accuracy function across letter positions (n ¼ 30 at each length), regardless
of word length.
BWN’s spelling errors were almost entirely letter errors:
substitutions (22.5%; 213/948); deletions (39.3%; 373/948)
additions (3.9%; 37/948); and exchanges (10.2%; 97/948).
The remaining errors (24.1%; 228/948) were some combination of these single-letter errors (e.g., ‘claw’ ! CRACW).
Only 79 (8.3%) of the 948 word errors were phonologically

Table 1. BWN’s spelling performance according to word length (length
measured in letters)

Fig. 4. Case FM: The distribution of single-letter errors in three spelling tasks.

Length

Word errors

%

Letter errors

%

4
5
6
7
8
9
10

16/382
125/926
121/879
217/844
239/486
158/293
72/137

4.2
13.5
13.8
25.7
49.2
53.9
52.6

16/1528
146.5/4630
166/5274
298/5908
376/3888
356/2637
189/1370

1.0
3.2
3.1
5.0
9.7
13.5
13.8

Totals

948/3947

24.0

1547.5/25235

6.1

Downloaded by [University of Victoria] at 12:48 18 November 2014

Spelling complex graphemes

Fig. 5. Case BWN: Percentage of letters correctly produced as a function of
letter position and word length.

plausible spellings of the target word, with 76 of the 79
(96.2%) consisting of single letter substitutions (e.g.,
‘facet’ ! FACIT), deletions (e.g., ‘chrome’ ! CROME),
insertions (e.g., ‘mistress’ ! MISSTRESS), or transpositions
(e.g., ‘deuce’ ! DUECE). The three remaining phonologically plausible errors involved some combination of single
letter errors. Thus, although the phoneme/grapheme conversion system cannot be ruled out as the source of some of these
responses, they are all consistent with a graphemic buffer
deficit. In addition, BWN did not produce semantic errors in
spelling to dictation.
BWN’s performance across modalities of input was evaluated with spelling to dictation and delayed copy transcoding
on the same stimulus set of sixty 7–10 letter words. As
indicated in Table 2, his level of performance was very similar
across the two tasks (word accuracy: w2(1) ¼ 0.106, ns; letter
accuracy: w2(1) ¼ 0.006, ns) and there were comparable
length effects. Furthermore, BWN’s performance was virtually identical across output modalities when oral and written
spelling were evaluated with the same set of seventy 4–8 letter
words, yielding word error rates of 23% (oral spelling) and
21% (written spelling; w2 ¼ 0.00, ns) and letter error rates of
5% in both modalities.

Table 2. Case BWN: Spelling performance as a function of word length and
input modality (spoken versus visual)
Length

Spelling to dictation

Delayed copy transcoding

Word errors

Letter errors

Word errors

Letter errors

7
8
9
10

40% (6/15)
60% (9/15)
73% (11/15)
93% (14/15)

9% (10/105)
16% (20/120)
21% (29/135)
33% (50/150)

53% (8/15)
60% (9/15)
80% (12/15)
93% (14/15)

11% (12/105)
17% (20/120)
27% (36/135)
26% (39/150)

Totals

67% (40/60)

21% (109/510)

72% (43/60)

21% (107/510)

127

Further evidence that BWN’s spelling errors did not arise
from damage to the lexical process comes from the apparent
lack of an effect of frequency on his spelling performance.
BWN was administered a list of 60 words of high (mean ¼ 55)
and low-frequency (mean ¼ 5.3) (Francis and Kucera, 1982),
matched for length. There was no significant difference in his
spelling performance on these lists either in terms of word
accuracy (w2(1) ¼ 1.0, ns) or letter accuracy (w2(1) ¼ 2.9,
p < .09). BWN was also administered three lists of words
contrasting the extremes of the frequency continuum (high
frequency > 100; low frequency < 10) (Francis and Kucera,
1982). On two of these three lists, there was no significant
difference in his performance with high- and low-frequency
words. These results, in combination with the relatively small
number of phonologically plausible errors, the absence of
semantic errors, no significant effects of concreteness/
abstractness (w2 ¼ 0.0) or of regularity (w2 ¼ 1.53, p < .22),
all indicate that BWN is unlikely to suffer from an additional
lexical impairment.
The only area that we were unable to adequately assess was
BWN’s spelling of nonwords. This assessment was complicated by the combination of his mild hearing loss and
phonological errors in spoken production. Consequently, if
he repeated a nonword stimulus incorrectly we couldn’t be
certain if he had misperceived it and then correctly repeated
his misperception or if, instead, he correctly perceived the
stimulus but incorrectly repeated it, or both. Note that the
situation was different for words as with these BWN was able
to communicate the meaning of any word he repeated incorrectly, allowing us to be confident of his perception. As a
result, BWN was not tested extensively with nonwords. Out of
56 nonwords, he spelled 21 correctly (37.5%), with a letter
accuracy rate of 82.1%. Accuracy was poorer with nonwords
as compared to words, although error types were the same,
consisting of letter substitutions (‘‘scart’’ ! SCARK), deletions (‘‘cantree’’ ! CATRE), insertions (‘‘foit’’ ! FOINT)
and misorderings (‘‘remun’’ ! RUMEN). Given the difficulties in evaluating nonword spelling, we cannot be certain if the
accuracy difference between words and nonwords is due to
errors in perception or an additional deficit affecting phonology-orthography conversion. However, even if the latter were
to be the case, it would not have any bearing on the interpretation of the experimental work we present below.
In sum, BWN’s performance is precisely that expected
from a graphemic buffer deficit which compromises an individual’s ability to maintain the activation of orthographic
representations.

Experiment 1: Analysing the relative
preservation of digraphs and control clusters
In earlier analyses of FM’s spelling deficit (Tainturier and
Caramazza, 1996), there was some indication that digraphs
such as SH behaved differently from control clusters such as
ST. More specifically, FM was less likely to produce only one
of the two target letters in digraphs than in control sequences

Downloaded by [University of Victoria] at 12:48 18 November 2014

128

M. J. Tainturier and B. C. Rapp

(22/68 vs. 32/68 in spelling to dictation, 7/20 vs. 10/20 in
written picture naming, and 15/78 vs. 23/78 in delayed copy
transcoding; non-significant differences). That is, FM produced fewer errors such as ‘‘rocket’’ ! ROCET or ROCLET
than errors such as ‘‘basket’’ ! BASET or BASPET. However, this aspect of FM’s performance was not pursued further
at the time since the main goal of the earlier study was to
demonstrate that double-letter sequences (as in RABBIT)
were processed differently from other types of letter
sequences, including digraphs, something which was clearly
shown to be the case.
In this experiment, we further investigated FM’s processing
of digraphs using a larger and better controlled set of stimuli
presented in a delayed copy transcoding task. For BWN we
examined his performance in spelling to dictation a list of
consonant digraphs and matched controls that was developed
specifically for assessing digraph spelling.

ple, in this task FM correctly produced 31/80 six-letter words
with frequencies ranging from 0–20 (Mean ¼ 7) and 30/80
six-letter words with frequencies ranging from 21–442
(Mean ¼ 105).
For BWN we developed a list of 30 pairs of 7–10 letter
words matched pairwise for length and frequency with the
consonant digraphs and control clusters in the central (more
error-prone) positions (e.g., PURCHASE/CONTRACT). For
this stimulus set, the digraphs and clusters were not matched
for bigram frequency, digraphs having a higher bigram frequency than the clusters (1330 vs. 427; t ¼ 2.98, p < .006;
Mayzner and Tresselt, 1965). We address this point in the
General Discussion. BWN was administered this 60 stimulus
list twice for spelling to dictation; the results of each administration were extremely similar and, therefore, the data from
the two administrations were combined. All chi-square analyses use a Yates correction for continuity.

Methods
For FM we undertook a post hoc analysis of a large corpus of
spellings obtained in a delayed copy transcoding task. In this
task, the experimenter showed words typed in lowercase
characters, the stimulus was then covered and FM was asked
to write it using uppercase letters. As we have seen, performance in delayed copy transcoding was very similar to that in
other spelling tasks. We chose to analyse data from this task,
as opposed to spelling to dictation, because it provided a
sufficient number of items suitable for the digraph analysis
and because performance in this task was not confounded by
FM’s lexical deficits (as explained above), and therefore, is a
more direct reflection of the graphemic buffer deficit.
From the existing corpus of 842 stimuli, we selected 5 to 9
letter words that included a consonant digraph (e.g., sh, ck)
and matched them to words from the corpus that included
consonant clusters (e.g., st, cr) but no digraphs. Only consonant digraphs were included as there were very few possible
controls for vowel digraphs in the corpus (i.e., there were very
few words such as ‘‘lion’’ in which the vowel sequence does
not correspond to a single phoneme). Words with digraphs and
clusters were matched pairwise for (a) word length in letters,
(b) the position of the target sequence in the word (word
initial, word final, and medial positions), and (c) the consonant/vowel context of the target sequence (whether the
target sequence is preceded or followed by a vowel or a
consonant). Digraph and control sequences were also matched
listwise for bigram frequency (Mayzner and Tresselt, 1965).
This was to ensure that any potential advantage of digraphs
over control sequences could not be attributed to the frequency with which the two component letters co-occur in
English. The application of these criteria yielded an experimental list of 232 words (116 matched pairs). Examples include: SHOWER-PLANET, OSTRICH-HARVEST,
MORPHINE-COMPLETE (target sequences underlined).
Note that word pairs were not matched for lexical frequency
because there was no evidence that this factor affected FM’s
performance in the delayed copy transcoding task. For exam-

Results
Regardless of the accuracy of non-target letters in the stimuli,
responses to each of the target sequences were scored as
belonging to one of three categories. The first category
(sequence present) includes responses in which the target
sequence (digraph or control cluster) was preserved, i.e., both
target letters were produced, adjacent to one another and in
the correct relative order. In the second category (sequence
absent), neither of the two target letters appeared anywhere in
the response. The third category (sequence ‘‘broken’’)
includes cases in which (a) one and only one of the two
letters was produced or (b) the two letters were produced but
in non-adjacent positions. Examples of each type of errors are
presented in Table 3. If complex graphemes such as ‘‘CK’’
(digraphs) function as units, one would expect that there
should be fewer cases of ‘broken’ sequences for these
sequences as compared to control letter clusters such as ‘‘ST’’.
This is precisely what the analysis revealed. The distribution
of each category of response for each type of target sequence is
presented in Fig. 6. For both FM and BWN, there were fewer
cases of ‘‘broken’’ sequences in the case of digraphs than in the
case of control clusters (FM: w2(1) ¼ 7.09, p < .01; BWN:
w2(1) ¼ 5.7, p < .02). In addition, and also consistent with the
hypothesis that they are processed as units, digraph sequences
were more often produced correctly than control sequences
(FM: w2(1) ¼ 6.23, p < .01; BWN: w2 ¼ 4.58, p < .04).
Table 3. Case FM: Examples of the three error types analysed in Experiment
1 (target letters in boldface)
Digraph targets

Control targets

Sequence present

knife ! KNIFE
radish ! RISDISH

crime ! CRIME
blister ! BISTEN

Sequence absent

ostrich ! OSTRIP
machine ! MAEMISE

anglican ! ACNINT
hermit ! HINDOE

Sequence broken

pocket ! POCTER
wrench ! WENCH

drama ! DMANA
parent ! PANAET

Spelling complex graphemes

129

Downloaded by [University of Victoria] at 12:48 18 November 2014

frequency effect on BWN’s spelling) and therefore the results
for the two lists are combined for the analyses reported below
(total n ¼ 266 words, 133 of each type).
As indicated earlier, there were not enough instances of
vowel clusters in FM’s corpus to compare the performance of
digraph vowels with matched control clusters. Nevertheless,
we will present his performance on a set of 78 words with
vowel digraphs (e.g., ‘beaver’) and 28 words with vowel
clusters (e.g., ‘onion’), ranging in length from 5 to 9 letters.

Fig. 6. Experiment 1: Distribution of responses on target sequences
(expressed in percentages) for Case FM and Case BWN.

Thus, the results of Experiment 1 indicate that digraphs and
clusters are represented in a manner that is sufficiently distinct
that they can be affected differently by a graphemic buffering
deficit. More specifically, the results are consistent with the
view that digraphs are represented as single units while
clusters are not.

Experiment 2: Analysing transposition
errors in vowel digraphs
As indicated in the Introduction, even if digraphs are units at
some level of the spelling process, such units will need to be
‘unpacked’ and their constituents serially produced. This
involves the specific challenge of selecting and producing
the constituents in their correct relative order. Although earlier analyses did not reveal any specific ordering difficulties
for consonant digraphs (see General Discussion), preliminary
evidence suggested that BWN produced a particularly high
rate of transposition errors on vowel digraphs. Therefore,
BWN and FM’s spelling of vowel digraphs was investigated
further.

Methods
For BWN, we prepared two sets of word pairs to examine
the spelling of vowel digraphs and clusters. Within each set,
the word pairs were matched for length and the position of the
critical segments (e.g., REACH/REACT; DAIRY/DIARY;
LOIN/LION; FAINT/GIANT). For one of the lists, words
were also matched for frequency, while for the other list they
were not. The results for the two lists were virtually identical
(which is unsurprising given the general absence of a word

Results
The results indicate that, in contrast to what was observed
with consonant digraphs and clusters, BWN spelled vowel
digraphs (74%; 99/133) significantly less accurately than
vowel clusters (89%; 118/133) (w2(1) ¼ 6.1, p < .02). This
occurred despite the fact that the digraphs had a higher
average bigram frequency than control clusters (371 vs.
122; t ¼ 2.96, p < .006). Furthermore, BWN’s rate of transposition errors was quite striking for vowel digraph elements.
In fact, 85% (29/34) of his errors to vowel digraphs were
adjacent letter transpositions (e.g., ‘‘flaunt’’ ! FLUANT;
‘‘again’’ ! AGIAN; ‘‘tabloid’’ ! TABLIOD). In contrast,
transpositions made up only 53% (8/15) of BWN’s errors
to vowel clusters (w2(1) ¼ 4.2, p < .05). Further evidence of
the predominance of transpositions in vowel digraphs comes
from an evaluation of errors on an additional set of 3200
words that BWN was administered for spelling to dictation
during the same time period. Out of the 88 transposition errors
he produced on this word set, 68% occurred on digraphs.
Although this is not a controlled comparison of digraphs with
other letter types, it is quite remarkable, nonetheless, especially if one considers that the vast majority of letters pairs
in this 3200 word set are certainly not digraphs.
FM’s performance with vowel digraphs does not follow the
pattern observed in BWN. Although we could not match the
two types of letter sequences, it is nonetheless noteworthy
that there is no indication that vowel digraphs are more
difficult than control clusters or than consonant digraphs.
Indeed, FM spelled 75% of vowel digraphs correctly, as
opposed to 32% of vowel clusters, an advantage consistent
with what was observed with consonant digraphs. In addition,
unlike BWN, FM produced only 5% transposition errors on
digraphs and 4% on clusters. Possible reasons for these
differences between FM’s and BWN’s performance are discussed below.

General discussion
Our principal findings are that: (a) both FM and BWN spelled
consonant digraphs significantly more accurately than clusters and (b) BWN produced significantly more transposition
errors on vowel digraphs than on control vowel clusters.
With regard to the greater accuracy of consonant digraphs
over clusters, is important to emphasize that this difference
cannot be attributed to the bigram frequency of the target
segments. In FM’s case this is a straightforward conclusion as

Downloaded by [University of Victoria] at 12:48 18 November 2014

130

M. J. Tainturier and B. C. Rapp

consonant clusters and digraphs were matched for frequency.
In BWN’s case a bigram frequency account is unlikely
because, although for both consonants and vowels the
digraphs were more frequent than the clusters, consonants
yielded greater accuracy for digraphs than clusters, while for
vowels the reverse pattern was observed.
We can account for findings a and b if we assume that
digraphs are not represented as simple letter sequences but
instead that they are represented as units with an internal
structure that specifies their constituent letters and their
relative order (see Fig. 1). Specific support for this proposal
is as follows. First, evidence of the unitary nature of digraphs
comes from: (a) the finding of superior performance with
digraphs than clusters and (b) the tendency for digraphs to
retain their ‘‘unitary’’ character as evidenced by the fact that
they are less likely than clusters to be broken up through the
deletion or non-adjacent movement of constituents. Second,
evidence of the internal complexity of digraphs comes from
BWN’s extremely high rate of transposition errors for vowel
digraphs. These errors indicate that the relative ordering of the
digraph’s constituents is more vulnerable to disruption than
the relative ordering of any other letter pairs.
Although this proposal can account for the primary findings
of this investigation, a number of outstanding questions
remain. First, we found that BWN was more accurate with
consonant digraphs than clusters while the reverse was
observed for vowels. Furthermore, vowel digraphs gave rise
to transposition errors while consonants digraphs did not. This
contrast between consonant and vowel digraphs may result
from differences in the orthotactic legality of the letter pairs
that would result from letter transposition in each case. In fact,
the transposition of the elements of consonant digraphs generally results in orthotactically illegal or highly dispreferred
sequences, while this is not the case for consonant clusters. For
example, transpositions of digraphs such as CH ! HC or
CK ! KC are unattested in English and even PH ! HP and
SH ! HS are extremely unusual. However, transpositions
within consonant clusters such as BR ! RB, GR ! RG or
TR ! RT are all attested and fairly common (mean bigram
frequency ¼ 117; Mayzner and Tresselt, 1965). In contrast,
transpositions within vowel digraphs and clusters are legal and
quite common (e.g., OI ! IO, AI ! IA or OU ! UO), with the
mean bigram frequency for the transpositions of digraphs and
clusters equal to 83 and 165, respectively. Thus, with regard to
the orthotactic acceptability of transposition errors for
digraphs vs. clusters, there is a very striking asymmetry for
consonants, but not for vowels. This may result in consonant
digraphs being ‘‘protected’’ against transposition errors.
A second question concerns the fact that transpositions
were observed for vowel digraphs in BWN’s case but not in
FM’s. This may reflect differences in the extent or severity of
their deficits. For example, if one assumes that graphemic
buffering involves both the activation and the serial selection
of letters (Houghton et al., 1994; Rapp and Kong, 2002), then
it may be the case that BWN’s deficit affects both the
activation and selection components, while FM’s affects only

the activation of graphemes. As discussed in the Introduction,
the serial selection of letters that form part of a digraph may be
particularly error-prone if one assumes that the grapheme unit
would need to be ‘‘unpacked’’. This view of digraphs as units
which, subsequent to activation, must be unpacked, resembles
proposals regarding the representation of compound words
(e.g., Badecker, 2001). Consistent with this view and with
the observation of letter transpositions in BWN’s spelling
of vowel digraphs, transpositions of the elements of compounds have been reported both in normal slips of the tongue
(Fromkin, 1973) as well as subsequent to neural injury
(Badecker, 2001).
A final question concerns the representational relationship
between digraphs and double letters, such as BB in RABBIT.
As indicated, an earlier study of FM (Tainturier and
Caramazza, 1996) showed that doubles were produced more
accurately than digraphs and also yielded different error
patterns. For example, FM often substituted double letters
with other doubled letters (e.g., rabbit ! RAPPIT), but this
was not observed for digraphs. These various observations can
be accommodated if one considers that while digraphs and
doubles are both represented in a unitary manner at some level
of representation and processing, they differ in terms of
subsequent processing. Doubles require a process that recognizes the doubling feature and produces a repetition of the
letter to be doubled, whereas digraphs require a process that
identifies their constituents and their ordering. These processes may be differentially affected by neural injury, leading
to the various patterns of impairment that have been attested
and which have provided such insight into the nature and
internal organization of orthographic representations.

Summary and conclusions
In sum, we have reported two cases of acquired dysgraphia
that reveal that digraphs are represented in a different manner
than other grapheme sequences. In order to account for the
pattern of results, we have proposed that digraphs are represented as units with an internal structure that specifies the
identity and ordering of their constituents. This study provides
the first empirical evidence in support of this proposal. More
generally, these findings further strengthen the view that the
orthographic representations used in spelling are complex
mental objects rather than simple linear representations of
letter strings.

Acknowledgements
We are very grateful to FM and BWN for their participation in
this study. We also thank Joanne Egan, who helped with the
scoring of some of the data. This research was partly supported by the FRSQ (Fonds de la recherche en santé du
Québec) while the first author was a postdoctoral fellow at
Johns Hopkins University, by an ARASH grant to MJ
Tainturier and Brenda Rapp, by NIH grant NS22201 to
Alfonso Caramazza, and by NIMH grant R29MH55758 to

Spelling complex graphemes

Brenda Rapp. A related preliminary report has been published
elsewhere (Tainturier et al., 2000).

Downloaded by [University of Victoria] at 12:48 18 November 2014

References
Allen M, Badecker W. Morphology: The internal structure of words. In:
Rapp B, editor. The handbook of cognitive neuropsychology: What
deficits reveal about the human mind. Philadelphia: Psychology Press,
2001.
Badecker W. Lexical composition and the production of compounds: Evidence
from errors in naming. Language and Cognitive Processes 2001; 16:
337–66.
Badecker W, Caramazza A. The analysis of morphological errors in a case of
acquired dyslexia. Brain and Language 1987; 32: 278–305.
Badecker W, Hillis AE, Caramazza A. Lexical morphology and its role in the
writing process: Evidence from a case of acquired dysgraphia. Cognition
1990; 35: 205–43.
Badecker W, Nathan P, Caramazza A. Varieties of sentence comprehension
deficits: A case study. Cortex 1991; 27: 311–21.
Blanken G, Schafer C, Tucha O, Lange KW. Serial processing in graphemic
encoding: Evidence from letter exchange errors in a multilingual patient.
Journal of Neurolinguistics 1999; 12: 13–39.
Bosse ML, Valdois S, Tainturier MJ. Analogy without priming in early
spelling development. Reading and Writing 2003; 16: 693–716.
Buchwald A, Rapp B. The orthographic representation of consonant-vowel
status: Evidence from two cases of acquired dysgraphia. Brain and
Language 2003; 87: 120–1.
Caramazza A, Miceli G. The structure of graphemic representations.
Cognition 1990; 37: 243–97.
Caramazza A, Miceli G, Villa G, Romani C. The role of the graphemic buffer
in spelling: Evidence from a case of acquired dysgraphia. Cognition 1987;
26: 59–85.
Cotelli M, Abutalebi J, Zorzi M, Cappa SF. Vowels in the buffer: A case study
of acquired dysgraphia with selective vowel substitutions. Cognitive
Neuropsychology 2003; 20: 99–114.
Cubelli R. A selective deficit for writing vowels in acquired dysgraphia.
Nature 1991; 353: 258–60.
Dickerson J. Format distortion and word reading: The role of multiletter units.
Neurocase 1999; 5: 31–6.
Francis WN, Kucera H. Frequency analysis of english usage. Boston:
Houghton Mifflin, 1982.
Fromkin VA. Slips of the tongue. Scientific American 1973; 229:
110–7.
Houghton G, Glasspool DW, Shallice T. Spelling and serial recall: Insights
from a competitive queuing model. In Brown GDA, Ellis NC, editors.
Handbook of spelling: Theory, process and intervention. New York: John
Wiley and Sons Ltd., 1994.
Houghton G, Zorzi, M. Normal and impaired spelling in a connectionist dualroute architecture. Cognitive Neuropsychology 2003; 20: 115–62.

131

Jónsdóttir MK, Shallice T, Wise R. Phonological mediation and the graphemic
buffer disorder in spelling; cross-language differences? Cognition 1996; 59:
169–97.
Joubert SA, Lecours AR. The role of sublexical graphemic processing in
reading. Brain and Language 2000; 72: 1–13.
Martensen H, Maris E, Dijkstra T. Phonological ambiguity and context
sensitivity: On sublexical clustering in visual word recognition. Journal of
Memory and Language 2003; 49: 375–95.
Mayzner MS, Tresselt ME. Tables of single-letter and digram frequency
counts for various word-length anf letter-position combinations. Psychonomic Monograph Supplements 1965; 1: 13–32.
McCloskey M, Badecker W, Goodman-Shulman RA, Aliminosa D. The
structure of graphemic representations in spelling: Evidence from a case of
acquired dysgraphia. Cognitive Neuropsychology 1994; 2: 341–92.
Miceli G, Benvegnú B, Capasso R, Caramazza A. Selective deficit in
processing double letters. Cortex 1995; 31: 161–71.
Miceli G, Capasso R, Benvegnú B, Caramazza A. Working memory interacts
with independent consonant vowel representations in writing. Neurocase
[this issue].
Miceli G, Silver C, Caramazza A. Cognitive analysis of a case of pure
dysgraphia. Brain and Language 1985; 25: 187–221.
Plaut DC, McClelland JL, Seidenberg MS, Patterson KE. Understanding
normal and impaired word reading: Computational principles in quasiregular domains. Psychological Review 1996; 103: 56–115.
Pring L. Phonological codes and functional spelling units: Reality and
implications. Percept Psychophysics 1981; 30: 573–8.
Rapp B, Epstein C, Tainturier, MJ. The integration of information across
lexical and sublexical processes in spelling. Cognitive Neuropsychology
2002; 19: 1–29.
Rapp B, Kong D. Revealing the component functions of the graphemic buffer.
Brain and Language 2002; 83: 112–3.
Rastle K, Coltheart M. Whammy and double whammy: Length effects in
nonword naming. Psychonomic Bulletin Review 1998; 5: 277–82.
Rey A, Ziegler JC, Jacob AM. Graphemes are perceptual reading units.
Cognition 2000; 75: B1–B12.
Seidenberg MS, McClelland JL. A distributed developmental model of word
recognition and naming. Psychological Review 1989; 96: 523–68.
Tainturier MJ, Caramazza A. The status of double letters in graphemic
representations. Journal of Memory and Language 1996; 35: 53–73.
Tainturier MJ, Egan J, Rapp B. Structure and texture in graphemic
representations: Further evidence from dysgraphia. Brain and Language
2000; 74: 376–8.
Tainturier MJ, Rapp B. The spelling process. In: Rapp B, editor. The handbook
of cognitive neuropsychology: What deficits reveal about the human mind.
Philadelphia: Psychology Press, 2001.
Ward J, Romani C. Consonant-vowel encoding and orthosyllables in a case of
acquired dysgraphia. Cognitive Neuropsychology 2000; 17: 641–63.
Zesiger P, Orliaguet JP, Boë LJ, Mounoud P. The role of syllabic structure in
handwriting and typing production. In: Faure C, Keuss P, Lurette G, Vinter
A, editors. Advances in handwriting: A multidisciplinary approach. Paris:
Europia, 1994.