Language Processing Modulated by Literacy: A Network Analysis of Verbal Repetition in Literate and Illiterate Subjects

Abstract

Previous behavioral and functional neuroimaging data indicate that certain aspects of phonological processing may not be acquired spontaneously, but are modulated by learning an alphabetic written language, that is, learning to read and write. It appears that learning an alphabetic written language modifies the auditory-verbal (spoken) language processing competence in a nontrivial way. We have previously suggested, based on behavioral and functional neuroimaging data, that auditory-verbal and written language interact not only during certain language tasks, but that learning and developing alphabetic written language capacities significantly modulates the spoken language system. Specifically, the acquisition of alphabetic orthographic knowledge has a modulatory influence on sublexical phonological processing and the awareness of sublexical phonological structure. We have suggested that developing an orthographic representation system for an alphabetic written language, and integrating a phoneme-grapheme correspondence with an existing infrastructure for auditory-verbal language processing, will result in a modified language network. Specifically, we suggest that the parallel interactive processing characteristics of the underlying language-processing brain network differ in literate and illiterate subjects. Therefore, the pattern of interactions between the regions of a suitably defined large-scale functional-anatomical network for language processing will differ between literate and illiterate subjects during certain language tasks. In order to investigate this hypothesis further, we analyzed the observed covariance structure in a PET data set from a simple auditory-verbal repetition paradigm in literate and illiterate subjects, with a network approach based on structural equation modeling (SEM). Based on a simple network model for language processing, the results of the present network analysis indicate that the network interactions during word and pseudoword repetition in the illiterate group differ, while there were no significant differences in the literate group. The differences between the two tasks in the illiterate group may reflect differences in attentional modulation of the language network, executive aspects of verbal working memory and the articulatory organization of verbal output. There were no significant differences between the literate and illiterate group during word repetition. In contrast, the network interactions differed between the literate and illiterate group during pseudoword repetition. In addition to differences similar to those observed in the illiterate group between word and pseudoword repetition, there were differences related to the interactions of the phonological loop between the groups. In particular, these differences related to the interaction between Broca's area and the inferior parietal cortex as well as the posterior-midinsula bridge between Wernicke's and Broca's area. In conclusion, the results of this network analysis are consistent with our previously presented results and support the hypothesis that learning to read and write during childhood influences the functional architecture of the adult human brain. In particular, the basic auditory-verbal language network in the human brain is modified as a consequence of acquiring orthographic language skills.

Full text

Language Processing Modulated by Literacy:
A Network Analysis of Verbal Repetition in
Literate and Illiterate Subjects
Karl Magnus Petersson
Karolinska Institute, Sweden
Alexandra Reis
Karolinska Institute, Sweden and Hospital de Santa Maria, Portugal
Simon Askelo
È
f
Karolinska Institute, Sweden
Alexandre Castro-Caldas
Hospital de Santa Maria, Portugal
Martin Ingvar
Karolinska Institute, Sweden
Abstract
& Previous behavioral and functional neuroimaging data
indicate that certain aspects of phonological processing may
not be acquired spontaneously, but are modulated by learning
an alphabetic written language, that is, learning to read and
write. It appears that learning an alphabetic written language
modifies the auditory-verbal (spoken) language processing
competence in a nontrivial way. We have previously suggested,
based on behavioral and functional neuroimaging data, that
auditory-verbal and written language interact not only during
certain language tasks, but that learning and developing
alphabetic written language capacities significantly modulates
the spoken language system. Specifically, the acquisition of
alphabetic orthographic knowledge has a modulatory influ-
ence on sublexical phonological processing and the awareness
of sublexical phonological structure. We have suggested that
developing an orthographic representation system for an
alphabetic written language, and integrating a phoneme±
grapheme correspondence with an existing infrastructure for
auditory-verbal language processing, will result in a modified
language network. Specifically, we suggest that the parallel
interactive processing characteristics of the underlying lan-
guage-processing brain network differ in literate and illiterate
subjects. Therefore, the pattern of interactions between the
regions of a suitably defined large-scale functional-anatomical
network for language processing will differ between literate
and illiterate subjects during certain language tasks. In order to
investigate this hypothesis further, we analyzed the observed
covariance structure in a PET data set from a simple auditory-
verbal repetition paradigm in literate and illiterate subjects,
with a network approach based on structural equation
modeling (SEM). Based on a simple network model for
language processing, the results of the present network
analysis indicate that the network interactions during word
and pseudoword repetition in the illiterate group differ, while
there were no significant differences in the literate group. The
differences between the two tasks in the illiterate group may
reflect differences in attentional modulation of the language
network, executive aspects of verbal working memory and the
articulatory organization of verbal output. There were no
significant differences between the literate and illiterate group
during word repetition. In contrast, the network interactions
differed between the literate and illiterate group during
pseudoword repetition. In addition to differences similar to
those observed in the illiterate group between word and
pseudoword repetition, there were differences related to the
interactions of the phonological loop between the groups. In
particular, these differences related to the interaction between
Broca's area and the inferior parietal cortex as well as the
posterior-midinsula bridge between Wernicke's and Broca's
area. In conclusion, the results of this network analysis are
consistent with our previously presented results and support
the hypothesis that learning to read and write during
childhood influences the functional architecture of the adult
human brain. In particular, the basic auditory-verbal language
network in the human brain is modified as a consequence of
acquiring orthographic language skills. &
D 2000 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 12:3, pp. 364±382

INTRODUCTION
Several cognitive models for language processing have
been constructed based on clinical and experimental
evidence (e.g., Patterson & Shewell, 1987; Patterson &
Lambon Ralph, 1999; Caplan, 1992). In most cognitive
models of language, written language constitutes a set of
parallel processes interacting with auditory-verbal lan-
guage at different levels. We have previously suggested,
based on behavioral (Reis & Castro-Caldas, 1997) and
functional neuroimaging data (Castro-Caldas, Petersson,
Reis, Stone-Elander, & Ingvar, 1998), the auditory-verbal
and written language interact not only during certain
language tasks, but that learning and developing alpha-
betic written language capacities significantly modulates
the auditory-verbal language system. Additional support
for this hypothesis are data indicating interhemispheric
differences in the posterior parietal cortex related to
literacy (Petersson, Reis, AskeloÈf, Castro-Caldas, & In-
gvar, 1998). These differences are paralleled by morpho-
logical findings indicating that the part of corpus
callosum relaying the interhemispheric connections be-
tween the left and right posterior parietal cortices is
different in literate and illiterate subjects (Castro-Caldas
et al., 1999). Taken together these data suggest that
learning an alphabetic written language influences audi-
tory-verbal language in a nontrivial way. However, it is
still unclear which processes and mechanisms mediate
this modulatory influence and which parts of the lan-
guage system or which processing levels are affected.
To investigate the functional organization of language,
and in particular the modulatory influence of literacy on
language processing, we have chosen to study natural
illiteracy (Castro-Caldas, Reis, & Guerreiro, 1997; Castro-
Caldas et al., 1998; Castro-Caldas et al., 1999; Petersson
et al., 1998; Petersson, Reis, Castro-Caldas, & Ingvar,
1999c; Reis & Castro-Caldas, 1997). Specifically, subjects
are classified as illiterate when they, for social reasons,
have never received any literacy training. The subjects
have never attended school or received any other form
of training in and have no knowledge of reading or
writing. All subjects, literate and illiterate, are recruited
from a similar socio-cultural background in a homoge-
neous fishermen community in southern Portugal. The
subjects are actively working and fully participating in
their community and in other respects are considered
normal. In this context, it is important to distinguish
between natural and functional illiteracy. Functional
illiterates have received literacy training, but lost the
ability to read and write due to lack of practice. Previous
exposure to written language learning and the acquisi-
tion of phonemic±graphemic associations implies the
existence (or at least a residual) of a visuo-graphic
representation system. In contrast, (naturally) illiterate
subjects lack orthographic knowledge and a visuo-gra-
phic representation system for language as well as an
explicit awareness of sublexical segmentation.
There are several reasons for a general network ap-
proach to the understanding of brain function, including
the neuroanatomic organization of the brain itself. The
organization of the brain resembles a hierarchically
structured, recurrently connected network composed
of different brain areas, consisting of several types of
neurons and synaptic connections with different proces-
sing properties (Shepherd, 1997; Felleman & Van Essen,
1991). Generally, information is thought to be repre-
sented as distributed activity in the brain and information
processing, subserving complex cerebral functions, are
hypothesized to emerge from the interactions between
different functionally specialized regions or neuronal
groups (Mesulam, 1998; Arbib, 1995; Macdonald & Mac-
donald, 1995; Koch & Davis, 1994; Mesulam, 1990; Amit,
1989). Independent of whether the net result is serial
symbolic processing (Fodor & Pylyshyn, 1990) or parallel
subsymbolic processing (Macdonald & Macdonald, 1995;
Smolensky, 1988), at one level or another cerebral
functions are implemented in the network architecture
of the brain. A network approach is therefore natural
when trying to describe brain functions. In particular,
analyzing the dynamic patterns of interaction between
different functionally specialized brain regions or neuro-
nal assemblies should contribute to the understanding of
sensorimotor and cognitive functions (Ingvar & Peters-
son, 2000). It has been suggested that three different
language pathways (or processing networks) may be
engaged in verbal repetition: the semantic, the lexical,
and the phonological route. These networks are active
during language processing and may operate as subsys-
tems in a structured network for parallel, distributed,
and interactive processing (Castro-Caldas et al., 1998; see
further Seidenberg & McClelland, 1989; Seidenberg,
1995; Caplan, 1992; Martin & Saffran, 1992; Patterson &
Shewell, 1987; and also Pinker & Prince, 1989; Pinker,
1997; Prince & Smolensky, 1997; Redington & Chater,
1997; Snowling, Hulme, & Nation, 1997; Shastri, 1995;
Shastri & Ajjanagadde, 1993; Rumelhart & McClelland,
1986). We have suggested that developing an ortho-
graphic representation system for an alphabetic written
language, integrating a phoneme±grapheme correspon-
dence with an existing infrastructure for auditory-verbal
(spoken) language processing, will result in a modified
language network. Specifically, we suggest that the par-
allel interactive processing characteristics of the under-
lying language-processing system of the brain differ in
literate and illiterate subjects. Therefore, the pattern of
interactions between the regions of the large-scale func-
tional-anatomical network (for a review see Mesulam,
1990; Mesulam, 1998) for language processing will differ
between literate and illiterate subjects during certain
language tasks. In particular, we hypothesize that the
network interactions during word and pseudoword re-
petition is different in the illiterate group, but more
similar in the literate group.
Petersson et al. 365

In everyday conversations, spoken language appears
similar in literate and illiterate subjects. However, closer
examination reveals that the groups differ when tested
on several language tasks, in particular if awareness of
the phonological structure of words is required to solve
the task (Castro-Caldas et al., 1998; Reis & Castro-Caldas,
1997; Adrian, 1993; Morais, Cary, Alegria, & Bertelson,
1979; Morais, 1993). Some of these differences appear to
be related to aspects of sublexical phonological proces-
sing and the explicit awareness of sublexical structure. In
order to further investigate the hypothesis that the
pattern of interactions in the language processing net-
work differ between literate and illiterate subjects, we
analyzed the covariance structure or functional connec-
tivity (Friston, Frith, Liddle, & Frackowiak, 1993; Friston,
1994; Aertsen, Gerstein, Habib, & Palm, 1989; Aertsen &
Preissl, 1991) of our previously reported PET data
(Castro-Caldas et al., 1998). We used a network analysis
approach based on structural equation modeling (SEM)
(McIntosh & Gonzalez-Lima, 1991; McIntosh & Gonza-
lez-Lima, 1994; McIntosh et al., 1994; Horwitz & McIn-
tosh, 1993; Bollen, 1989; Hayduk, 1987) in conjunction
with a simple network model for language processing
including simple models for attentional modulation and
verbal working memory. In the following section, we
briefly describe the basis for the functional-anatomical
network model that we used in this study.
Verbal Working Memory
Verbal repetition engages verbal working-memory pro-
cesses (Gathercole & Baddeley, 1993). Repetition of
both words and pseudowords have been used when
testing the language capabilities of brain-damaged sub-
jects (e.g., Glosser, Kohn, Friedman, Sands, & Grugan,
1997; Costlett, Roeltgen, Rothi, & Heilman, 1987) and to
investigate different mechanisms involved in language
processing of normal subjects (e.g., Patterson & Shewell,
1987; Patterson & Lambon Ralph, 1999; Caplan, 1992).
Pseudoword repetition, that is, repeating phonologically
plausible sequences of phonemes lacking lexico-seman-
tic representation may be used to test phonological
processing relatively independent (or at least less de-
pendent) of lexico-semantic processing. When repeating
pseudowords, the language system is confronted with
new phonological sequences that are stored in the
phonological loop until the articulatory output se-
quence is organized and executed. In contrast to well-
known words, which can access the lexico-semantic
system directly and can be analyzed as a whole, pseudo-
words have to be correctly parsed into phonemes or
other sublexical units (input phonology) in order to
correctly organize the articulatory output sequence
(output phonology) and to be correctly repeated. It
should be noted that the ability to correctly repeat
pseudowords has also been related to the capacity of
the verbal short-term working memory (Gathercole &
Baddeley, 1993; Gathercole & McCarthy, 1994; Gather-
cole, Willis, Baddeley, & Emslie, 1994).
The Baddeley±Hitch model of verbal working memory
(Baddeley, 1992) include a phonological store for hold-
ing phonologically coded information on-line and an
articulatory control process based on inner speech that
can update the information in the phonological store.
The phonological loop (the phonological store + the
articulatory control process) is behaviorally the most
well characterized component of the verbal working
memory (Baddeley, 1997). Recently, it was suggested
that the verbal working memory has an important
functional role as a language learning device during
language acquisition (Baddeley et al., 1998). The differ-
ence in performance when repeating pseudowords be-
tween literates and illiterates (Castro-Caldas et al., 1998;
Reis & Castro-Caldas, 1997) may indicate that the acqui-
sition of written language modulates the phonological
loop or other components of verbal working memory.
As an additional prelude to the construction of the
functional-anatomical language network, we review
some additional functional neuroimaging and anatomi-
cal data. Functional neuroimaging data (Paulesu, Frith,
& Frackowiak, 1993; replicated in Paulesu et al., 1996
and Salmon et al., 1996) indicate that the phonological
loop is subserved by the posterior third of the superior
temporal (Wernicke's area, BA 22), the inferior frontal
(Broca's area, BA 44/6), and the inferior parietal (BA 39/
40) cortices as well as the insula. In particular, it has
been hypothesized (Paulesu et al., 1993) that the pho-
nological store is related to the inferior parietal region
(De
Â
monet, Price, Wise, & Frackowiak, 1994) while the
articulatory control process is related to Broca's area. In
addition, it has been suggested that Wernicke's and
Broca's areas subserve different aspects of phonological
processing (Paulesu et al., 1996). Wernicke's area is
thought to subserve auditory word form or input pho-
nology (Paulesu et al., 1996; De
Â
monet et al., 1992), while
Broca's area subserves segmented output phonology
(Price, Moore, Humphreys, & Wise, 1997; De
Â
monet,
Fiez, Paulesu, Petersen, & Zatorre, 1996; Paulesu et al.,
1996). Furthermore, results from PET activation studies
indicate that both irregular (compared to regular words)
and nonword reading (compared to regular and irregu-
lar words) activate inferior frontal regions (Herbster,
Mintun, Nebes, & Becker 1997). These activations may
be related to processes that generate articulatory repre-
sentations and finally verbal output. Consistently, a task
minimizing semantic and maximizing phonological pro-
cessing activated inferior parietal and left precentral/
inferior frontal region (Price et al., 1997). Functional
neuroimaging studies have also indicated that some
central executive aspects of verbal working memory
are localized to the left middle-inferior prefrontal cortex
(BA 45/46) (D'Esposito et al., 1995; Petrides, Alivisatos,
Meyer, & Evans, 1993; Petrides, 1995; Petrides, Alivisatos,
& Evans, 1995).
366 Journal of Cognitive Neuroscience Volume 12, Number 3

It has been hypothesized that the posterior insula may
be a neural relay for language (Mesulam & Mufson,
1985), in particular that automatic language processing
is relayed via the insula (Raichle, 1994; Raichle et al.,
1994; see also the neurological models of language in
Posner & Raichle, 1994, in particular the dual-pathway
model for word generation). This suggestion is consis-
tent with PET data indicating that activation of the
posterior insula is associated with repetition of words
under conditions minimizing semantic processing and
with the development of automaticity in learning a
language task (Raichle, 1994; Raichle et al., 1994; Peter-
sen, Fox, Posner, Mintun, & Raichle, 1989; Petersen &
Fiez, 1993). Furthermore, additional PET data indicate
that the insula may mediate interaction between the
(left) inferior parietal region and Broca's area (Paulesu
et al., 1996). Some clinical evidence also indicate that
lesions of the left insula can cause conduction aphasia in
which the translation from heard, read or self-generated
words into appropriate phonemic sequences is defect
(Damasio & Damasio, 1980).
Most of the cortical and subcortical connections of the
insula are reciprocal. The anterior-midinsular connec-
tions include inferotemporal, temporopolar, and medial
temporal cortices. The anterior and midcingulate cor-
tices connect most prominently to the midportion of the
insular cortex, while the posterior-midinsular connec-
tions include inferotemporal, cingulate, and prefrontal
cortices (Mesulam & Mufson, 1985). In addition, the
posterior-midinsula receives projections from the pos-
terior auditory association area of the superior temporal
gyrus and sends projections to the opercular paramotor
cortex (Mesulam & Mufson, 1985). There are also wide-
spread intrainsular connections, which strongly inter-
connect the various sectors of the insular regions
(Mesulam & Mufson, 1985).
Word and Pseudoword Repetition in Literate and
Illiterate Subjects
The results of our previously reported PET study (Cas-
tro-Caldas et al., 1998) comparing literate and illiterate
subjects repeating words and pseudowords are briefly
summarized here (Figure 1). In the word versus pseudo-
word comparison, there were similar bilateral activations
in the posterior parietal regions (BA 7, 19, 39) in both
groups, with a greater left-sided posterior parietal dom-
inance in the literate compared to the illiterate group. In
the literate group, posterior midline activations were
limited to the right posterior cingulate cortex (BA 23),
and in the illiterate group, the precuneate region (BA 7,
31) extending into posterior cingulate region (BA 23)
was activated. In the literate group there were also
activations in the left superior and middle frontal
region (BA 8, 9), in the right posterior parieto-occipi-
to-temporal region (BA 39, 37, 19), and left occipito-
temporal region (BA 19, 37). When thresholding at a
lower level in word versus pseudoword comparison,
the pattern of activation in literates and illiterates
tended to converge toward similar patterns. The only
activation in words versus pseudowords that was sig-
nificantly greater in the literate than in the illiterate
group was the more prominent left inferior parietal
activation (BA 40).
In the reverse comparison (pseudoword vs. word),
the literate group displayed a significant activation in the
anterior insular (BA 14, 15) bilaterally and right inferior
frontal/frontal opercular (BA 44, 45, 47, 49), left anterior
cingulate cortices (BA 24, 32), as well as subcortical
structures, including left basal ganglia (putamen, globus
pallidus extending into head of caudate nucleus) and
midline cerebellum. In the illiterate group, significant
activation was only seen in the middle frontal/frontopo-
lar region (BA 10). In general, the interaction analysis
confirmed these findings.
The behavioral results showed that the illiterate sub-
jects performed pseudoword repetition less well than
the literate subjects (Castro-Caldas et al., 1998; Reis &
Castro-Caldas, 1997). In this context, it should be em-
phasized that all subjects only produced words or
pseudowords and no other type of speech was pro-
duced during the PET scanning. When subjects failed to
repeat a pseudoword correctly, they still repeated a
pseudoword (except for some lexico-semantic analogies,
Figure 1. Maximum intensity projections of the SPM[Z] thresholded
at Z = 2.58 (omnibus significance p  .005) in the words±pseudowords
contrast in the literate (A) and in the illiterate (B) groups. The reverse
contrast (pseudowords±words) in the literate (C) and in the illiterate
(D) groups. For further details and results see Castro-Caldas et al.
(1998).
Petersson et al. 367

2% of the errors in the literate and 11% of the errors in
the illiterate group). The fact that the subjects repeated
pseudowords (albeit sometimes incorrectly) makes it
less likely that the differences in activation pattern is
related to performance differences as such. In support
of this there were no or only weak correlations between
performance and regional cerebral blood flow in either
group or condition using the local Z maximum test
statistic (overall p  .2). In addition, including the
performance as a confounding covariate in the general
linear model, the patterns of activations were similar. In
particular, the differences between the literate and
illiterate group were generally independent of whether
the performance covariate was included in the analysis
or not (Table 1).
Functional Organization of Verbal Repetition
and SEM
SEM provides an opportunity to explicitly test hypoth-
eses relating to functional-anatomical models subserving
different cognitive functions in terms of which regions
are involved and how they interact in a given network
model. SEM has been used to investigate the covariance
structure of functional neuroimaging data in a given
cognitive state or during a specific task (Buchel &
Friston, 1997; Buchel, Coull, & Friston, 1999; Nyberg
et al., 1996; McIntosh et al., 1994; Horwitz & McIntosh,
1993). For example, SEM has been used to investigate
the effective connectivity (Friston, 1994) in visual pro-
cessing (McIntosh et al., 1994), in assessing the effects
of pallidotomy in Parkinson patients (Grafton, Sutton,
Couldwell, Lew, & Waters, 1994), in working memory
(McIntosh et al., 1996), episodic memory retrieval
(Nyberg et al., 1996), and visual attention (Buchel &
Friston, 1997). The network approach characterizes the
functional organization in terms of effective connec-
tions between regions in a specific functional-anatomi-
cal model. This model is used in conjunction with SEM
to model the observed covariance structure between
the different regions of interest included in the net-
work model.
Functional-Anatomical Network Model
The regions included in the functional-anatomical net-
work model were represented as spherical regions of
interests (ROIs, see Table 2) in the Karolinska Compu-
terized Brain Atlas of Greitz (Greitz, Bohm, Holte, &
Eriksson, 1991). The ROIs were chosen based on the
neuropsychology of language literature (Kolb &
Whishaw, 1996; Caplan, 1992; Mesulam, 1990; Patterson
& Shewell, 1987), functional neuroimaging data (Price
et al., 1996; Price et al., 1997; De
Â
monet et al., 1992;
De
Â
monet, Wise, & Frackowiak, 1993; De
Â
monet et al.,
1996; Frackowiak, 1994; Liotti, Gay, & Fox, 1994; Peter-
sen et al., 1989; Petersen & Fiez, 1993; Wise et al., 1991;
Table 1. Group Specific Differences in Activation Patterns (A)
in Words versus Pseudowords and (B) Pseudowords versus
Words. Anatomical structures and Brodmann's areas (BA)
refer to the Talairach and Tournoux (1988) atlas except when
indexed * referring to the Karolinska Computerized Brain
Atlas of Greitz (Greitz et al. 1991). Local Z score maxima with
Z >2.58(p < .005, uncorrected) are reported. For further
details and more results see Castro-Caldas et al. (1998).
Region Z score [xyz]
(A) Greater activation in literates compared to illiterates in
words versus pseudowords
Left frontal operculum/anterior insula region, BA 47/49*/14*
With performance covariate 2.64 30 30 0
Left inferior/superior parietal region, BA 40
With performance covariate 3.22 44 28 40
2.90 36 50 48
Without performance covariate 3.00 36 52 48
(B) Greater activation in literates compared to illiterates in
pseudowords versus words
Right frontal operculum/anterior insula region, BA 49*/45*/
14*
With performance covariate 2.70 20 22 32
Without performance covariate 2.78 22 22 28
Left anterior cingulate region, BA 24
With performance covariate 3.05 14 28 16
Without performance covariate 3.59 16 26 12
Left putamen/pallidum
With performance covariate 2.70 24 0 4
Without performance covariate 3.20 20 0 4
Anterior thalamus/hypothalamus
With performance covariate 3.03 2 80
Without performance covariate 3.04 0 60
Pons
Without performance covariate 2.69 2 34 16
Medial cerebellum
With performance covariate 2.67 10 70 16
Without performance covariate 2.35 4 40 20
368 Journal of Cognitive Neuroscience Volume 12, Number 3

for a review see Habib & Demonet, 1996), lesion data
(Basso, Lecours, Moraschini, & Vanier, 1985; Damasio &
Damasio, 1980), and from the activations reported in
Castro-Caldas et al. (1998) as well as other functional
neuroimaging studies (Carter et al., 1998; Paulesu et al.,
1993; Paulesu et al., 1996; Salmon et al., 1996; Vanden-
berghe, Price, Wise, Josephs, & Frackowiak, 1996; Buck-
ner et al., 1995; D'Esposito et al., 1995; Petrides et al.,
1993; Petrides, 1995; Petrides et al., 1995).
The objective of constructing the functional-anatomi-
cal model for language processing during verbal repeti-
tion was to generate a simple network that could
explain a sufficient part of the observed covariance,
both in the literate and illiterate group during word as
well as pseudoword repetition. At the same time, we
required that the network model should be both
theoretically and empirically plausible. The functional-
anatomical network model used included several ROIs
(Table 2, Figure 2) and interconnections (Figure 3). For
more details on the construction of the functional-
anatomical network model, see the Methods section.
The right hemisphere language representation is less
well known compared to the left hemisphere organiza-
tion of language, in particular in illiterate subjects. There
are some indications that language processing in illiter-
ate subjects may, under certain circumstances, recruit
bilateral brain regions to a greater extent than literate
subjects (Petersson et al., 1999c; Castro-Caldas et al.,
1998). In addition, some brain lesion data have been
interpreted to indicate that the functional architecture
of language is more bilaterally organized in illiterate
compared to literate subject (Lecours et al., 1987a;
Lecours et al., 1987b; Wechsler, 1976; Cameron, Currier,
& Haerer, 1971). However, this issue is complex and
there are other data indicating that this may not always
be the case (Dama
Â
sio, Castro-Caldas, Grosso, & Ferro,
1976a; Dama
Â
sio, Castro-Caldas, Grosso, & Ferro, 1976b).
Furthermore, it is unclear how the right and left hemi-
spheric language networks interact. Taken together
and for simplicity we decided to restrict our analysis
to the left hemisphere. However, given the possibility
that the bilateral language processing in the illiterate
Table 2. The Spherical Regions of Interest Used in the
Functional-Anatomical Network Model were Located with the
Help of the Karolinska Computerized Brain Atlas of Greitz
(Greitz et al., 1991)
Region of interest
Brodmann's
area
Diameter
(mm)
Primary auditory cortex, left BA 41/42 12
Wernicke's area, left BA 22 12
Angular/supramarginal gyrus, left BA 39/40 16
Posterior-mid insula, left BA 13/14 8
Broca's area, left BA 44 12
Lentiform nucleus, left 12
Lateral cerebellum, right 12
Primary motor region
(mouth area), left
BA 4 8
Anterior cingulate cortex, left BA 24/32 12
Prefrontal cortex, left BA 45/46 12
See also Figure 2.
Figure 2. The anatomical locali-
zation of the regions of interest
displayed in the anatomical space
of Greitz (Greitz et al., 1991). The
numbers displayed on the atlas
correspond to the Brodmann's
areas. S = primary/secondary
auditory cortex, W = Wernicke's
area, iPC = inferior parietal
(angular/supramarginal gyrus)
cortex, pI = posterior-mid insula,
B = Broca's area, M = primary
motor region of the mouth and
larynx, Cdx = right lateral cere-
bellum, ACC = anterior cingulate
cortex, and PFC = middle-inferior
prefrontal cortex. The location of
the lentiform nucleus is not
shown.
Petersson et al. 369

and literate subjects differ, it is not inconceivable that
the inclusion of a right hemispheric network would
accentuate the differences between the two literacy
groups.
In the following, ! indicates unidirectional connec-
tions while () indicates bidirectional or recurrent
connections, that is, a feedforward and feedback con-
nection,whichareallowedtodifferinconnection
weights. The network model (cf. Figure 3) includes a
simplification of the Wernicke±Geschwind model (e.g.,
Kolb & Whishaw, 1996) represented by the Wernicke's
area (W, posterior third of left superior temporal gyrus,
BA 22) connected to Broca's area (B, posterior part of
left inferior frontal gyrus BA 44, W!B) with input from
the left primary/secondary auditory cortex (S, BA 41/42,
S!W) and a simple motor output circuit (left lenticular
nucleus, NcL, and left primary motor cortex for articu-
lation (the mouth and larynx area) M, BA 4; B!M,
B!NcL!M). This core was extended to include the
anterior cingulate cortex (ACC) hypothesized to be
related to focused attention, error detection and re-
sponse competition/selection (Carter et al., 1998; Vogt,
Finch, & Olson, 1992; Pardo et al., 1990; Posner &
Petersen, 1990). The network model was also extended
to include the phonological loop as described by
Paulesu et al. (1993, 1996). This introduced the inferior
parietal cortex (iPC, on the border between angular
and supramarginal gyrus, BA 39/40) and the posterior-
midinsula (pI, BA 13 on the border of BA 14). The
connections of the phonological loop were generally
recurrent (W () iPC () B). Since the insula has been
hypothesized to be a neural relay for automatic lan-
guage processing the connections to and from the
posterior-midinsula were feedforward (W!pI!B).
The interactions between the ACC and the phonologi-
cal loop were represented by recurrent connections
(W () ACC () B and pI () ACC () iPC). In addi-
tion, the left middle-inferior dorsolateral prefrontal re-
gion (PFC, on the border between BA 45 and 46)
suggested to subserve the central executive aspects of
verbal working memory (D'Esposito et al., 1995; Petrides
et al., 1993; Petrides, 1995; Petrides et al., 1995), was
included with input from the ACC, the Wernicke's area,
and the inferior parietal cortex (ACC!PFC, W!PFC,
iPC!PFC) and with outputs modulating the organiza-
tion of the articulatory motor output (PFC!B, PFC!M,
PFC!NcL, PFC!Cdx). Finally, the right lateral cerebel-
lar region (Cdx) was included since this region has been
related to certain aspects of language processing (Van-
denberghe et al., 1996; Buckner et al., 1995; Raichle,
1994; Raichle et al., 1994) with inputs from cortical
motor regions (PFC! Cdx, B!Cdx, M!Cdx!NcL; cf.,
Figure 3). For a general empirical review of PET studies
of cognition see Cabeza and Nyberg (1997) and for a
review of the cerebellar contribution to cognition see
Schmahmann (1996).
Figure 3. The connectivity of
the functional-anatomical net-
work model for language pro-
cessing with the sub-networks
outlined; the auditory input, the
phonological loop, the articula-
tory motor subnetwork, the
attention subnetwork, the cen-
tral executive subnetwork. See
text for details.
370 Journal of Cognitive Neuroscience Volume 12, Number 3

RESULTS
The goodness-of-fit of the language network indicated
that the network models fitted the observed covariance
structure reasonably well in both groups and conditions
(overall the goodness-of-fit corresponded to p  .14 in
all groups and states, see Appendix). We used a hier-
archical approach (Table 3 and Table 4) to test for
differences between states and groups. First we tested
if the network interactions in both conditions and
groups could be explained by the same model using a
stacked analysis (cf. the Methods section); this was not
the case ( p = .003). We then tested for differences
either between groups in a given state or between states
in a given group. There was no significant difference in
the pattern of network interactions between word and
pseudoword repetition in the literate group ( p = .10).
Neither was there any significant difference between the
literate and illiterate group repeating words ( p = .10).
In contrast, there were significant differences in the
network interactions between word and pseudoword
repetition in the illiterate group ( p = .002), and
consistently, there were significant differences between
the literate and illiterate group repeating pseudowords
( p = .038).
In a second step, we characterized the differences
between word and pseudoword repetition in the illit-
erate group (Table 4, Figure 4) as well as the differ-
ences between the literate and the illiterate group
repeating pseudowords (Table 4, Figure 5). Specifically,
we localized the observed differences within the differ-
ent subnetworks as far as the sensitivity in this study
allowed to (Figure 3), and only the connections show-
ing relative differences greater than or equal to .15
(standardized units) in connection strength were
further investigated.
The differences between word and pseudoword repeti-
tion in the illiterate group were broadly related to atten-
tional modulation (i.e., ACC connections, p = .024),
executive aspects of processing (i.e., PFC connections,
p = .047), and the motor output circuit ( p = .052),
while the differences relating to the phonological loop
were nonsignificant ( p = .103). More specifically, the
differences related to attentional modulation could be
localized to the interaction between the ACC and the
inferior parietal cortex (ACC () iPC, p = .043) as well
as the posterior-midinsula (ACC () pI, p = .044). In
addition, there were differences related to the prefron-
tal influence on the articulatory output of the language
network and the motor subnetwork (PFC!B+PFC!
Cdx+PFC!NcL, p = .027), finding its clearest expres-
sion in the connection PFC!NcL ( p = .050). The
differences in connections strengths of the articulatory
output circuit mainly related to the influence of Broca's
area on primary motor, lenticular nucleus and right
lateral cerebellar regions ( p = .045)
In repeating pseudowords, the literate and the illit-
erate group differed in attentional modulation of the
auditory input and phonological loop of the language
network ( p = .034). These differences related to the
interactions between the ACC and the auditory input
( p = .069) as well as Broca's area and the posterior-
midinsula ( p = .044). More specifically, it appeared
that these differences related to the interactions be-
tween the ACC and the posterior-midinsula (AC-
C () pI, p = .049) and the interaction between the
ACC and Broca's area (ACC () B, p = .074). The
interactions of the phonological loop differed between
groups ( p = .036). These differences could be further
localized to the interaction between Broca's area and
the inferior parietal cortex ( p = .042). In addition,
there were differences related to the posterior-midin-
sula (W!pI!B, p = .027), with contributions from
Table 3. The Results from the Omnibus Stacked Model
Comparisons Between the Different Literacy Groups and Tasks
Omnibus comparison p value
Literate group, word/pseudoword,
illiterate, word/pseudoword repetition
.003
Literate group,
word versus pseudoword repetition
ns (.10)
Illiterate group,
word versus pseudoword repetition
.002
Word repetition,
literate versus illiterate group
ns (.10)
Pseudoword repetition,
literate versus illiterate group
.038
ns = nonsignificant.
Table 4. The Results from the Sub-Network Stacked Com-
parisons
Omnibus subnetwork comparison p value
Illiterate group, word versus pseudoword repetition
Attention subnetwork .024
Central executive subnetwork .047
Phonological loop ns (.10)
Articulatory motor output .052
Pseudoword repetition, literate versus illiterate group
Attention subnetwork .034
Central executive subnetwork .045
Phonological loop .036
Articulatory motor output .085
For more detailed results see the Results section. Only connections
showing differences  .15 in standardized units were investigated,
ns = nonsignificant.
Petersson et al. 371

Wernicke's area (W!pI, p = .057) and output to Broca's
area (pI!B, p = .085). There were also differences
related to the connections with the PFC ( p = .045). In
particular, the interactions between the inferior parietal
and Broca's area via the PFC (iPC!PFC!B, p = .042;
iPC!PFC, p = .064; PFC!B, p = .064) differed between
the literate and illiterate group. The differences related
to the articulatory output circuit were statistically weaker
(Cdx!NcL!M, p = .085).
DISCUSSION
Amongst others (Buchel & Friston, 1997; McIntosh &
Gonzalez-Lima, 1994; McIntosh et al., 1996; Nyberg et al.,
1996; Friston, 1994; Grafton et al., 1994), we have
demonstrated the benefit of a network approach to
the analysis of functional neuroimaging data, providing
complementary information to the more common acti-
vation or general linear approach (Friston et al., 1995).
The results of the network approach naturally have to be
interpreted in the context of its limitations (cf. the
Methods section).
The results of this study indicate that the network
interactions were relatively similar during word and
pseudoword repetition in literate subjects and also
when literate and illiterate subjects repeated words. In
contrast, there were significant differences in the net-
work interactions when the illiterate subjects repeated
words compared to pseudowords. Given the hypothe-
sized functional role of the different regions included in
the network model of language processing (cf. above),
these differences were mainly related to those parts of
the language network that may be viewed as general
support and control systems, that is, central executive
aspects of processing as well as attentional modulation.
In addition, there were differences related to the motor
circuit hypothesized to support the organization of
articulatory output. There were no significant differ-
ences related to the phonological loop. The results also
indicated significant differences between the literate and
illiterate subjects when repeating pseudowords. These
differences were related to the phonological loop, in
particular the interactions between Broca's area and the
inferior parietal region. There were also differences
related to the posterior-midinsular bridge between Wer-
nicke's and Broca's area. Additional differences related
to attentional modulation of the auditory input, the
Broca's and the posterior-midinsular regions as well as
Figure 4. The relative differ-
ences between word and pseu-
doword repetition in the
illiterate group, that is, the
difference between the corre-
sponding connection strengths
(word±pseudoword). Only dif-
ferences  .15 in standardized
units are shown. The omnibus
comparison indicated that the
difference were significant
( p = .002).
372 Journal of Cognitive Neuroscience Volume 12, Number 3

executive aspects of language processing and articula-
tory output.
The absence of significant difference between word
and pseudoword repetition in the literate group indi-
cates that the network interactions are relatively similar
in word and pseudoword repetition. This can be inter-
preted to indicate that the literates automatically recruit
a phonological processing network with sufficient com-
petence for sublexical processing and segmentation
duringsimpleimmediateverbalrepetition,whether
words or pseudowords, while this is not the case for
the illiterate group. Especially since the network inter-
actions differed significantly between word and pseudo-
word repetition in the illiterate group, this indicates that
the illiterate subjects process words and pseudowords
differently during verbal repetition. Consistently, the
network interactions were significantly different when
the literate and illiterate group repeated pseudowords,
lending support to the interpretation that the illiterate
subjects process pseudowords qualitatively differently
compared to literate subjects.
Behavioral data indicate that illiterates are not fully
competent in explicit manipulation of sublexical phono-
logical structures (Castro-Caldas et al., 1998; Reis &
Castro-Caldas, 1997; Morais et al., 1979; Morais, 1993).
Pseudowords cannot be repeated exclusively using a
lexico-semantic processing system or the type of (im-
plicit) phonological system recruited by illiterate sub-
jects. Instead, this seems to require an organization of
the phonological system found in literate subjects. We
have previously suggested that when literate subjects
repeat pseudowords, they engage attentional/awareness
components of phonological processing, which illiterate
subjects fail to engage (Castro-Caldas et al., 1998). The
production of motor sequences based on new phono-
logical sequences not previously learned, as in pseudo-
word repetition, may depend on these aspects of
phonological processing. This indicates that phonologi-
cal attention/awareness are necessary for the de novo
sequential arrangement of verbal output. In short, the
present results in conjunction with previous data sug-
gest that learning an alphabetic visual representation of
language entails the development of new auditory-
verbal processing capacities. Such acquired capacities
may explain the differences in interaction pattern ob-
served in the illiterate subjects that was not observed in
literate subjects. In particular, the differences in the
network interactions between groups when repeating
Figure 5. The relative differ-
ences during pseudoword re-
petition between the literate
and the illiterate group, that is,
the difference between the
corresponding connection
strengths (literate±illiterate).
Only differences  .15 in stan-
dardized units are shown. The
omnibus comparison indicated
that the difference were signifi-
cant ( p = .038).
Petersson et al. 373

pseudowords were broadly related to general support
and control (i.e., attentional and central executive
aspects of language processing) as well as the phono-
logical loop and the organization of articulatory motor
output.
Given our previous hypothesis and the suggested
importance of inferior parietal, Wernicke's, Broca's,
and insular regions subserving different aspects of pho-
nological processing (Price et al., 1997; De
Â
monet et al.,
1996; Paulesu et al., 1996), one might expect differences
in the connection strengths between the two literacy
groups repeating pseudowords, reflecting the interac-
tions within the phonological loop and between the
phonological loop and the other subnetworks, that is,
attentional mechanisms and executive or control aspects
of verbal working memory. The phonological loop sub-
network was based on anatomical (Mesulam & Mufson,
1985; Mesulam, 1990) and functional neuroimaging data
(Salmon et al., 1996; Paulesu et al., 1993; Paulesu et al.,
1995; Raichle, 1994; Raichle et al., 1994). This model of
the phonological loop includes Wernicke's area (sup-
porting input phonology), Broca's area (supporting out-
put phonology), inferior parietal lobe (supporting the
phonological store) and the posterior-midinsula, hy-
pothesized to provide a ``bridge'' between Wernicke's
and Broca's area (Raichle et al., 1994; Mesulam &
Mufson, 1985). In addition, based on functional neuro-
imaging data indicating that dyslectic subjects manifest
phonological processing disturbances that parallels the
inability to coordinate the interactions between the
different regions of the phonological loop, it has been
suggested that the (anterior) insula serve as a ``bridge''
between the inferior parietal region and Broca's area
(Paulesu et al., 1996). In our previous report (Castro-
Caldas et al., 1998), the anterior insular and frontal
opercular activations were observed in the literate group
during pseudoword repetition (compared to word re-
petition), but not in the illiterate group. The lack of
insular activation in the illiterate group may indicate the
absence of an efficient link between the different com-
ponents of phonological processing subserved by Wer-
nicke's, Broca's, and inferior parietal regions. This is
consistent with the interpretation of the insular role
suggested by Paulesu et al. (1996). Furthermore, Raichle
et al. (1994) have suggested that the (posterior) insula
subserve more automatic aspects of language proces-
sing. In this context, it should be pointed out that there
are widespread intrainsular connections that strongly
interconnect the various sectors of the insular regions
(Mesulam & Mufson, 1985). Along the lines of Raichle
(1994), Paulesu et al. (1993), and Paulesu et al. (1996),
the results of the present network analysis indicate that
illiterate subjects are not able to coordinate the interac-
tions between the different regions of the phonological
loop as literate subjects do during verbal repetition (see
below). As previously noted, these differences were
particularly related to the interactions between Broca's
area and the inferior parietal region as well as the
posterior-midinsular bridge between Wernicke's and
Broca's area.
The ACC have been related to focused attention, error
detection and response competition/selection (Carter
et al., 1998; Vogt et al., 1992; Pardo et al., 1990; Posner &
Petersen, 1990), while the central executive aspects of
verbal working memory have been related to the mid-
dle-inferior prefrontal region (BA 45/46) (D'Esposito
et al., 1995; Petrides et al., 1993; Petrides, 1995; Petrides
et al., 1995). The level of attentional processing, as
indicated by the level of ACC activation, was greater in
pseudoword compared to word repetition in the literate
group (Castro-Caldas et al., 1998). This may reflect a
greater need for attentional control (Carter et al., 1998;
Cohen, Dunbar, & McClelland, 1990; Cohen, Servan-
Schreiber, & McClelland, 1992) to achieve a more effec-
tive phonological analysis during pseudoword repetition
compared to word repetition. However, the absence of
significant differences in the literate group between
word and pseudoword repetition, both at the network
level and at the behavioral level, indicate that the pattern
of network interactions related to attentional modula-
tion or executive control was qualitatively similar in both
tasks. In contrast, the lack of significant increase of ACC
activity in pseudoword compared to word repetition in
the illiterate subjects may be related to the lack of
explicit sublexical phonological awareness (Castro-Cal-
das et al., 1998). In addition, the pattern of network
interactions related to attentional processing was differ-
ent both in word compared to pseudoword repetition in
the illiterate subjects as well as between the literate and
illiterate group during pseudoword repetition. The dif-
ferences between word and pseudoword repetition in
illiterate subjects were mainly related to the interactions
ACC () iPC and ACC () pI. In addition to these
findings, group differences during pseudoword repeti-
tion were also related to the interaction between the
ACC and Broca's area as well as attentional modulation
of the auditory input regions. This is consistent with the
suggestions that the ACC subserve aspects of error
detection and on-line monitoring of behavior (Carter
et al., 1998). Interestingly, the difference in pseudoword
repetition related to the interaction ACC () pI repre-
sented decreases in connection strengths (both in lit-
erates vs. illiterates and in word vs. pseudoword
repetition in the illiterate group), possibly reflecting
the need to inhibit automatic language processing re-
layed via the insula (Raichle, 1994; Raichle et al., 1994;
Mesulam & Mufson, 1985). At the same time, there was a
corresponding increase in connection strengths be-
tween the ACC and the inferior parietal and Broca's
region, possibly reflecting a compensatory mechanism.
In the illiterate group, the differences in the interac-
tion pattern related to the prefrontal region (word vs.
pseudoword repetition) were mainly focused to the
connections modulating the articulatory motor circuit.
374 Journal of Cognitive Neuroscience Volume 12, Number 3

Differences were also related to the articulatory motor
output circuit interactions themselves. This was also the
case during pseudoword repetition in the literate versus
illiterate comparison, indicating that the need to recruit
executive aspects of verbal working memory to modu-
late and control the organization of articulatory output
may differ between the literacy groups during pseudo-
word and also between word and pseudoword repeti-
tion in illiterate subjects. In addition, the organization of
articulatory output itself may differ, requiring different
compensatory processing when literate and illiterate
subjects repeat pseudowords and when words and
pseudowords are repeated by illiterate subjects.
In summary, a system for phonological processing and
verbal repetition modulated by orthographic knowledge
is available to literate subjects, including capacities
needed for more demanding phonological tasks, e.g.,
successful pseudoword repetition. In contrast, illiterate
subjects seem not to have such a formatted system for
phonological processing. The results of the present
network analysis are consistent with our hypothesis that
the language networks in literate and illiterate subjects
subserve qualitatively different parallel distributed and
interactive processing reflecting the modulatory influ-
ence of having once received formal education and
acquired the skills of reading and writing. The results
are also consistent with our previous suggestion that
these differences are related to attentional modulation,
phonological processing, and the organization of articu-
latory output (Castro-Caldas et al., 1998; Petersson et al.,
1998; Reis & Castro-Caldas, 1997). However, it should be
pointed out that the results of this study do not exclude
the possibility that there are differences in network
interactions between the two tasks in the literate group
as well as between the literate and illiterate group
repeating words. Since, given the network model, it is
possible that the present results indicate that the mag-
nitude of possible differences was not large enough, in
relation to the observed variability, to be significant. The
statistical power of the network approach to functional
neuroimaging data is generally unknown (for a further
discussion of the limitations of the network approach
see the Methods section). In addition, the absence of
significant differences between word and pseudoword
repetition in literates may be related to the suggestion
that the network model used captures more of the
phonological rather than the semantic aspects of lan-
guage processing.
Conclusions
Previous behavioral and functional neuroimaging data
indicate that certain aspects of explicit phonological
processing may not be acquired spontaneously, but
are modulated by learning an alphabetic written lan-
guage, that is, learning to read and write. It appears that
learning an alphabetic written language modifies the
auditory-verbal (spoken) language processing compe-
tence in a nontrivial way. Based on behavioral and
functional neuroimaging data, we have suggested that
spoken and written language interact not only during
certain language tasks, but that learning and developing
alphabetic written language capacities significantly mod-
ulates the spoken language system. Specifically, the
acquisition of orthographic knowledge has a modulatory
influence on sublexical phonological processing. This
may result in a language network with different parallel
interactive processing characteristics in literate and illit-
erate subjects. Specifically, the pattern of interactions in
the language-processing network may differ between
literate and illiterate subjects. In order to investigate this
hypothesis, we analyzed the observed covariance struc-
ture in a PET data set from a simple auditory-verbal
repetition paradigm (immediate repetition of words and
pseudowords) in literate and illiterate subjects, from the
same socio-cultural background in a fishermen commu-
nity in southern Portugal, with a network approach
based on SEM.
Based on a simple network model for language pro-
cessing, the results of the present network analysis
indicate that the network interactions during word and
pseudoword repetition in the illiterate group differ,
while there were no significant differences in the literate
group. The differences between the two tasks in the
illiterate group may reflect differences in attentional
modulation of the language network, executive aspects
of verbal working memory and the articulatory organiza-
tion of verbal output. There were no significant differ-
ences between the literate and illiterate group during
word repetition. In contrast, the network interactions
differed between the literate and illiterate group during
pseudoword repetition. In addition to differences similar
to those observed in the illiterate group between word
and pseudoword repetition, there were differences re-
lated to the interactions of the phonological loop. In
particular, these differences between the two literacy
groups related to the interaction between Broca's area
and the inferior parietal cortex as well as the posterior-
midinsula bridge between Wernicke's and Broca's area.
The results of this network analysis are consistent
with our previously presented results and support the
hypothesis that learning to read and write during child-
hood influences the functional architecture of the adult
human brain. In particular, the basic auditory-verbal
language network in the human brain is modified as a
consequence of acquiring orthographic language skills.
METHODS
A detailed account of the PET study, the selection and
experimental procedures has been published in Castro-
Caldas et al. (1998). In brief, subjects were classified as
illiterate when they, for social reasons, had never en-
teredschoolandhadnoknowledgeofreadingor
Petersson et al. 375

writing. Twelve right-handed women (six literate and six
illiterate subjects) from similar socio-cultural back-
ground in a homogeneous fishermen community in
southern Portugal were included in the study. Literate
and illiterate subjects were selected if they performed in
the normal range (1 SD, norms according to age and
literacy group) on all subtasks of a test-battery adapted
for this population (Garcia & Guerreiro, 1983). Previous
diseases potentially involving the brain were ruled out
by clinical assessment, previous clinical information
provided by the local doctor and diagnostic MRI scans.
The literate women had received 4 years of schooling
and performed normally on reading comprehension and
writing tests. Six lists of 20 high frequency three-syllable
words were constructed based on frequency of use in
common Portuguese spoken language (Nascimento,
Rivenc, & Cruz, 1987). Six lists of pseudowords were
constructed based on the real words by changing the
consonants and maintaining the vowels as well as the
word length. The subjects were instructed to repeat
words or pseudowords and to avoid any other type of
speech production.
PET Scanning and Data Preprocessing
Repeated measurements of regional cerebral blood
flow (rCBF, 6 word + 6 pseudoword repetition tasks/
subject) were made with an ECAT Exact HR PET
scanner in the 3-D acquisition mode (Wienhard et al.,
1994) and bolus injections of [
15
O]butanol (Berridge,
Cassidy, & Terris, 1990). The PET images were rea-
ligned, spatially normalized and transformed into a
common stereotactic anatomical space (Talairach &
Tournoux, 1988), 3-D isotropic Gaussian filtered and
proportionally scaled to account for global confoun-
ders. In order to minimize the confounding contribu-
tion of filter-dependent spatial autocorrelation to the
covariance of closely located regions of interest, a filter
of 10 mm full width at half maximum (FWHM) was
used (McIntosh et al., 1994). Adjusted data were gen-
erated using the SPM96 software (Wellcome Depart-
ment of Cognitive Neurology, London, [Friston et al.,
1995], see also http://www.fil.ion.bpmf.ac.uk/) remov-
ing subject effects modeled as block-confounders in the
general linear model. The adjusted data were then
transformed into the anatomical space of the Karolinska
Computerized Brain Atlas of Greitz (Greitz et al., 1991).
SEM and Data Analysis
For adequate modeling it is required that the variations
in the activity of the regions of interest (ROIs) are
representative of respective functional region. For each
scan and for each ROI, the activity was averaged across
the voxels of the ROI, generating the ROI data. The
covariance matrix of the ROI data was computed across
subjects and repetitions (6 subjects  6 repetitions, i.e.,
36 observations) for each group and condition (literate/
illiterate, word/pseudoword), giving in total four covar-
iance matrices. Correlation matrices were also computed
and checked for high correlations, but none were de-
tected, indicating that the amount of filtering was not
too large in relation to the distance between the ROIs.
In general, SEM uses a linear system of equations to
describe the interrelations of activities between the
regions of the functional-anatomical model (McIntosh
& Gonzalez-Lima, 1994). SEM models are parameterized
by the connection strengths (path coefficients or con-
nection weights), the residuals, and the variances/covar-
iances among and between these parameters (McIntosh
& Gonzalez-Lima, 1994; Bollen, 1989; Hayduk, 1987).
Restrictions on these parameters in terms of constraints
(e.g., fixed values, symmetry or more complex con-
straints) may be specified. In the models considered in
this study, the residuals were fixed and the covariances
between residuals were set to zero (McIntosh & Gonza-
lez-Lima, 1994). The residuals of network components
with more than one input were fixed to 66% while the
residuals of components with only one input were fixed
to 100% of the corresponding observed variance. Fixing
residual parameters is to some extent arbitrary and the
residuals may generally be thought of as the combined
influences of regions not explicitly modeled and the
influence of a brain region upon itself as well as mea-
surement errors (McIntosh & Gonzalez-Lima, 1994). In
our case, the choice of values represents the hypothesis
that the network model is a significant simplification of
language processing, primarily taking into account pho-
nological processing, and that the model is only ex-
pected to explain parts of the observed covariance
structure. The standardized solutions, that is, the matrix
of connection strength for each group and state, are
reported in the Appendix. The connection strengths
were estimated with an iterative maximum likelihood
optimization process, using the LISREL software (version
7, Scientific Software, Mooresville, IN) for SEM modeling
(JoÈreskog & SoÈrbom, 1989). The optimization process
may informally be viewed as a procedure that recreates
the observed covariance between regions as close as
possible by finding the optimal values of the connection
strengths. In LISREL, starting values are estimated using
instrumental variables and a two-stage least square
approach. The resulting estimates are statistically con-
sistent, and with reasonably good fitting models, the
initial estimates are close to the (iterative) maximum
likelihood estimate (JoÈreskog & SoÈrbom, 1989).
Different network models were specified and con-
nections strengths were estimated minimizing the
maximum-likelihood norm between the empirically
observed covariance structure (covariance matrix) and
the covariance structure of the model network using
least square estimates as initial values. For each model,
the optimization process yields estimated connection
strengths, goodness-of-fit estimates and modification
376 Journal of Cognitive Neuroscience Volume 12, Number 3

indices indicating which connections may be introduced
or freed to enhance the goodness-of-fit. The goodness-
of-fit index used is the maximum-likelihood distance
between implied and observed covariance structure. In
the context of multivariate normally distributed vari-
ables, the goodness-of-fit index is x
2
distributed with
(q/2)(q+1)P degrees of freedom, where P is the
number of free parameters and q is the number of
observed variables (Bollen, 1989). The goodness-of-fit
index can be translated into a p value that indicates
how well the model can explain or represent the
observed covariance structure. Given that the model
is correct, the p value represents the probability of
making an error when rejecting the null-hypothesis
that the observed data could have been generated by
the underlying model. This implies that a nonsignifi-
cant p value (>.05) is desirable indicating that a
sufficient part of the observed covariance structure
can be explained by the network model. Modification
indices indicate ways that a given network model may
be modified to increase the goodness-of-fit or the
amount of observed covariances that can be explained
(McIntosh & Gonzalez-Lima, 1994; Bollen, 1989; Hay-
duk, 1987). The modification indices are of particular
interest when the modifications suggested are consis-
tent with theoretical or empirical considerations relat-
ing to the cognitive processes being modeled or known
anatomical pathways. In this way, modification indices
can be used to guide the model selection process
(McIntosh & Gonzalez-Lima, 1994). In the present
study, there was a reasonable correspondence between
good-fitting and theoretically plausible models in the
sense that models with random connections often
resulted in a marked drop in goodness-of-fit.
The described network approach, using SEM, gener-
ates quantitative estimates of the connection strengths
between regions. The pattern of connection strengths can
be compared between groups or conditions and using a
stacked models analysis (McIntosh & Gonzalez-Lima,
1994; Bollen, 1989; Hayduk, 1987). In a stacked analysis
the connections strengths are estimated under the con-
straint that the corresponding connections are equal
across groups or states. In a stacked analysis, if the
underlying assumptions are valid, then the p value reflects
the goodness-of-fit of the hypothesis that the same set of
network connections can explain the observed covar-
iance structures across groups or states. Expressed more
informally, that the cognitive processing in different
groups or different conditions is qualitatively and quanti-
tatively similar, as reflected in the network interactions of
the functional neuroimaging data. Stacked analyses may
also be performed on subnetworks by letting the com-
plementary part of the network be free while the con-
nection strengths of the subnetwork of interest are
constrained to be the same. In this way, using a stacked
analysis, it can be tested if the pattern of network
interactions is different between groups or states.
Duringtheinitialmodelselectionstagewealso
tested, to some extent, the stability of the results in
relation to parameters such as the amount of filtering,
small variations in the locations of the ROIs, different
initial values, and the effects of different goodness-of-fit.
To test the dependence on the amount of filtering, the
data were also analyzed when Gaussian filtered (isotro-
pic) at 16 mm FWHM. This yielded similar results in
terms of the magnitude of path-coefficients and signifi-
cance levels in the preliminary network models during
the initial model selection stage. The results also seemed
reasonably stable when small changes in the locations of
the ROIs were introduced. Different initial values gave
the same results. To control for effects of different
goodness-of-fit of the individual (group or condition)
models when running stacked models, networks with
comparable fit for all conditions and groups was also
analyzed, yielding similar results.
Some Assumptions and Limitations of the
Network Approach
The SEM network approach described here is an exam-
ple of a covariance-based approach (Horwitz & McIn-
tosh, 1993; Horwitz, Soncrant, & Haxby, 1992). The basic
hypothesis of different covariance based approaches is
that the intrinsic variability in the neural response,
during a cognitive state or particular task, will emulate
the relevant functional interactions or effective connec-
tivity (Friston, 1994; Friston, 1995). It is assumed that
these interactions and the underlying (dynamic) func-
tional architecture are reflected in the observed covar-
iance structure. Two major ways of estimating the
covariances within a cognitive state have been de-
scribed: across subjects (Horwitz, McIntosh, Haxby, &
Grady, 1995), over time/repetitions within subject (Bu-
chel & Friston, 1997) or both (Buchel et al., 1999). In
this context, it should be noted that the sources of
within-state interregional covariances are beyond experi-
mental control.
Covariance Sources
Several sources of interregional covariances have been
proposed (Horwitz et al., 1992) and the actual sources
of the observed covariances are largely unknown.
However, some of the proposed sources may give rise
to spurious correlations that are necessarily con-
founded with correlations arising because of the effec-
tive connectivity or the functional architecture (e.g.,
adaptation, fatigue, attentional drift; cf. Horwitz et al.,
1992 and Buchel & Friston, 1997). Obvious potential
confounders in the study of interregional covariances
are global effects. Variability in any global signal will
introduce correlations that most often are of no inter-
est. It has been suggested that the problem of global
effects may be discounted by using partial correlation
coefficients or proportional scaling, both of which yield
Petersson et al. 377

similar results (Horwitz & Rapoport, 1988). It has been
indicated that accounting for global effects with simple
approaches such as ANCOVA, proportional scaling or
the use of partial correlation coefficients can in some
circumstances yield biased results by introducing spur-
ious correlations (Ford, 1986). However, it has been
argued that such correlations often are small and that
their presence can, to some extent, be tested for
(Horwitz & Rapoport, 1988). In any case, the presence
of spurious correlations will bias the results to greater
or lesser extent, unless properly removed or accounted
for. Another possible source of confounding, when the
covariances are calculated across subjects, is related to
partial volume effects. This may be particularly relevant
for networks including bilateral homologous structures.
Homologous right±left structures are often relatively
symmetric in size, shape, and position. Partial volume
effects may thus introduce correlations between corre-
sponding right±left structures. In addition, when the
covariances are estimated over subjects, it is necessary
to assume that the subjects implement a sufficiently
similar functional organization. In this case, speaking
informally, the covariance structure may reflect an
average common functional architecture. In PET stu-
dies the number of intrasubject observations is limited
(restricted by radiation exposure) and the data is often
pooled over subjects in order to increase sensitivity.
With FMRI data, it is possible to study functional and
effective connectivity in single subjects (Buchel &
Friston, 1997). A distinct advantage with several sin-
gle-subject studies is that these provide information on
the generalizability of the results and the results may
also be used in a meta-analytic approach.
Specific Limitations of Structural Equations Modeling
The results of SEM analysis are potentially difficult to
interpret for several reasons. For example, there is no
guarantee that the connections modeled reflect direct
effective connections; it is possible that they are
mediated through areas or connections not included in
the model. Similarly, observed changes in the weights
between states or groups may reflect common input
from regions not modeled. This touches on the problem
of model selection, that is, the general problem of
matching model and data complexity. In the case of
SEM, model selection may be performed in a data-driven
manner, guided by goodness-of-fit values, modification
indices or using a hierarchical model building approach
when subsets of weights are estimated recursively (McIn-
tosh & Gonzalez-Lima, 1994). The data-driven approach
is vulnerable to over-fitting since sample specific char-
acteristics may be modeled (e.g., noise and outliers).
Investigating over-fitted models can limit the general-
izability of results. Alternatively, model selection may be
theory driven, running the risk of investigating incom-
plete models. Unless reasonable goodness-of-fit can be
achieved with a given model in all states or groups
investigated, the results of a stacked models comparison
can be difficult to interpret. For example, using an under-
parameterized model to test differences between states
or groups in a stacked analysis may yield results due to
an ill-fitting model in one of the states or groups. The
effect of using under-parameterized models has been
investigated to some extent in a simple network model
(McIntosh & Gonzalez-Lima, 1994). This simulation
study indicate that the results from analyzing moderately
reduced models are fairly stable and that the modifica-
tion indices can to some extent provide indications of
omissions of connections or regions in the chosen
model. It should also be noted that the maximum like-
lihood estimation of the SEM approach assumes that a
global optimum has been reached. The standard imple-
mentation in LISREL uses instrumental variables and a
two-stage least square approach in combination with the
Davidon±Fletcher±Power algorithm and line search to
find a (local) optimum of the objective function, which in
practice often is close to the global optimum (cf. Bollen,
1989; JoÈreskog & SoÈrbom, 1986). In general, the cost (or
objective) function to be optimized is a complicated
nonlinear function of the model parameters and often
an explicit global optimum solution is not known (Bol-
len, 1989). In addition, it is possible that the objective
function has several local optima. In such a case, there is
no guarantee that the global optimum will be reached
with deterministic gradient descent algorithms or non-
exhaustive search procedures. Alternatively, a simulated
annealing approach to optimization can be used (Kirk-
patrick, Gelatt, & Vecchi, 1983; Kirkpatrick & Sorkin,
1995). These and other issues are discussed further in
Petersson et al., 1999a and Petersson et al., 1999b.
Appendix
The connection strength matrices of the literate and illiterate
group in the word and pseudoword repetition tasks.
Literate, word repetition:
Goodness-of-fit corresponding to p =.16
ACC B Cdx iPC PFC M NcL pI S W
ACC .00 .24 .00 .13 .00 .00 .00 .11 .00 .09
B .18 .00 .00 .05 .09 .00 .00 .23 .00 .26
Cdx .00 .26 .00 .00 .06 .29 .00 .00 .00 .00
iPC .19 .19 .00 .00 .00 .00 .00 .00 .00 .19
PFC .01 .00 .00 .14 .00 .00 .00 .00 .00 .37
M .00 .52 .00 .00 .07 .00 .01 .00 .00 .00
NcL .00 .02 .00 .00 .39 .00 .00 .00 .00 .00
pI .12 .00 .00 .00 .00 .00 .00 .00 .00 .12
S .15 .00 .00 .00 .00 .00 .00 .00 .00 .00
W .06 .00 .00 .09 .00 .00 .00 .00 .53 .00
378 Journal of Cognitive Neuroscience Volume 12, Number 3

Acknowledgments
This study is part of the EU Biomed 1 programme (BMHI CT94-
1261) and was financed in part by grants from the Swedish
Medical Research Council (8276, 12716), the Karolinska
Institute, the Knut and Alice Wallenberg Foundation, the
Swedish Bank Tercentenary Foundation and project STRIDE
(no 352/92-JNICT). The authors also want to thank the
volunteers for their participation, Dr Fransisco Reis who
assisted in the recruitment of subjects, and the help of the
Portuguese Embassy in Stockholm.
Reprint requests should be sent to Dr. Karl Magnus Petersson,
Cognitive Neurophysiology R2-01, Department of Clinical
Neuroscience, Karolinska Institute, Karolinska Hospital, 171
76 Stockholm, Sweden, email: karlmp@neuro.ks.se.
REFERENCES
Adrian, J. A. (1993). Habilidade metafonoligica en sujetos
analfabetos y malos lectores. Boletin de Psicologia, 39,
7±19.
Aertsen, A. M. H., Gerstein, G. L., Habib, M. K., & Palm, G.
(1989). Dynamics of neuronal firing correlation: Modulation
of ``effective connectivity''. Journal of Neurophysiology, 61,
900±917.
Aertsen, A. M. H., & Preissl, H. (1991). Dynamics of activity and
connectivity in physiological neuronal networks. In: H. G.
Schuster (Ed.), Nonlinear Dynamics and Neuronal Net-
works (pp. 281±302), New York: VHC Publishers.
Amit, D. J. (1989). Modeling brain function: The world of
attractor neural networks, Cambridge, UK: Cambridge
Universiy Press.
Arbib, M. A. (Ed.) (1995). The handbook of brain theory and
neural networks, Cambridge, MIT Press.
Baddeley, A. (1992). Working memory. Science, 255, 556±559.
Baddeley, A. (1997). Human memory: Theory and practice,
Revised ed. Erlbaum.
Baddeley, A., Gathercole, S., & Papagno, C. (1998). The pho-
nological loop as a language learning device. Psychology
Review, 105, 158±173.
Basso, A., Lecours, A. R., Moraschini, S., & Vanier, M. (1985).
Anatomo-clinical correlations of the aphasias as defined
through computerized tomography: On exceptions. Brain
and Language, 26, 201±229.
Berridge, M. S., Cassidy, E. H., & Terris, A. H. (1990). A
routine, automated synthesis of oxygen-15-labeled butanol
for positron tomography. Journal of Nuclear Medicine, 31,
1727±1731.
Bollen, K. A. (1989). Structural equations with latent vari-
ables. New York: Wiley.
Buchel, C., Coull, J. T., & Friston, K. J. (1999). The predictive
value of changes in effective connectivity for human learn-
ing. Science, 283, 1538±1541.
Buchel, C., & Friston, K. J. (1997). Modulation of connectivity
in visual pathways by attention: Cortical interactions evalu-
ated with structural equation modelling and fMRI. Cerebral
Cortex, 7, 768±778.
Buckner, R. L., Petersen, S. E., Ojemann, J. G., Miezin, F. M.,
Squire, L. R., & Raichle, M. E. (1995). Functional anatomical
studies of explicit and implicit memory retrieval tasks.
Journal of Neuroscience, 15, 12±29.
Cabeza, R., & Nyberg, L. (1997). Imaging cognition: An em-
pirical review of PET studies with normal subjects. Journal
of Cognitive Neuroscience, 9, 1±26.
Cameron, R. F., Currier, R. D., & Haerer, A. F. (1971). Aphasia
and literacy. British Journal of Disorder Communication, 6,
161±163.
Illiterate, word repetition:
Goodness-of-fit corresponding to p =.22
ACC B Cdx iPC PFC M NcL pI S W
ACC .00 .08 .00 .25 .00 .00 .00 .23 .00 .05
B .26 .00 .00 .06 .22 .00 .00 .18 .00 .22
Cdx 0.00.15 .00 .00 .34 .15 .00 .00 .00 .00
iPC .15 .16 .00 .00 .00 .00 .00 .00 .00 .18
PFC.05 .00 .00 .36 .00 .00 .00 .00 .00 .33
M .00 .60 .00 .00 .08 .00 .17 .00 .00 .00
NcL .00 .09 .14 .00 .12 .00 .00 .00 .00 .00
pI .20 .00 .00 .00 .00 .00 .00 .00 .00 .06
S .32 .00 .00 .00 .00 .00 .00 .00 .00 .00
W .06 .00 .00 .19 .00 .00 .00 .00 .33 .00
Literate, pseudoword repetition:
Goodness-of-fit corresponding to p =.14
ACC B Cdx iPC PFC M NcL pI S W
ACC .00 .03 .00 .02 .00 .00 .00 .14 .00 .01
B .01 .00 .00 .07 .26 .00 .00 .19 .00 .04
Cdx .00 .28 .00 .00 .07 .04 .00 .00 .00 .00
iPC .10 .09 .00 .00 .00 .00 .00 .00 .00 .15
PFC .08 .00 .00 .34 .00 .00 .00 .00 .00 .13
M .00 .32 .00 .00 .26 .00 .08 .00 .00 .00
NcL .00 .03 .12 .00 .30 .00 .00 .00 .00 .00
pI .11 .00 .00 .00 .00 .00 .00 .00 .00 .31
S .17 .00 .00 .00 .00 .00 .00 .00 .00 .00
W .10 .00 .00 .03 .00 .00 .00 .00 .32 .00
Illiterate, pseudoword repetition:
Goodness-of-fit corresponding to p =.18
ACC B Cdx iPC PFC M NcL pI S W
ACC .00 .30 .00 .16 .00 .00 .00 .14 .00 .10
B .16 .00 .00 .11 .12 .00 .00 .35 .00 .09
Cdx .00 .26 .00 .00 .02 .32 .00 .00 .00 .00
iPC .19 .34 .00 .00 .00 .00 .00 .00 .00 .04
PFC .01 .00 .00 .10 .00 .00 .00 .00 .00 .28
M .00 .40 .00 .00 .07 .00 .26 .00 .00 .00
NcL .00 .11 .06 .00 .44 .00 .00 .00 .00 .00
pI .19 .00 .00 .00 .00 .00 .00 .00 .00 .06
S .20 .00 .00 .00 .00 .00 .00 .00 .00 .00
W .04 .00 .00 .11 .00 .00 .00 .00 .33 .00
Petersson et al. 379

Caplan, D. (1992). Structure, processing and disorders, Cam-
bridge: MIT Press.
Carter, C. S., Braver, T. S., Barch, D. M., Botvinick, M. M., Noll,
D., & Cohen, J. D. (1998). Anterior cingulate cortex, error
detection, and the online monitoring of performance.
Science, 280, 747±749.
Castro-Caldas, A., Miranda Cavaleiro, P., Carmo, I., Reis, A.,
Leote, F., Ribeiro, C., & Ducla-Soares, E. (1999). Influ-
ence of learning to read and write on the morphology
of the corpus callosum. European Journal of Neurology,
6, 23±28.
Castro-Caldas, A., Petersson, K. M., Reis, A., Stone-Elander, S.,
& Ingvar, M. (1998). The illiterate brain: Learning to read and
write during childhood influences the functional organiza-
tion of the adult brain. Brain, 121, 1053±1063.
Castro-Caldas, A., Reis, A., & Guerreiro, M. (1997). Neuropsy-
chological aspects of illiteracy. Neuropsychological Rehabi-
litation, 7, 327±338.
Cohen, J. D., Dunbar, K., & McClelland, J. L. (1990). On the
control of automatic processes: A parallel distributed pro-
cessing model of the Stroop effect. Psychological Review,
99, 45±77.
Cohen, J. D., Servan-Schreiber, D., & McClelland, J. L. (1992). A
parallel distributed processing approach to automaticity.
American Journal of Psychology, 105, 239±269.
Costlett, H. B., Roeltgen, D. P., Rothi, L. G., & Heilman, K.
(1987). Transcortical sensory aphasia: Evidence for subtypes.
Brain and Language, 32, 362±378.
Dama
Â
sio, A. R., Castro-Caldas, A., Grosso, J. T., & Ferro, J. M.
(1976). Brain specialization for language does not depend
on literacy. Archives of Neurology, 33, 300±301.
Dama
Â
sio, A. R., Castro-Caldas, A., Grosso, J. T., & Ferro, J. M.
(1976). Letter to the Editor. Archives of Neurology, 33, 662.
Damasio, H., & Damasio, A. R. (1980). The anatomical basis of
conduction aphasia. Brain, 103, 337±350.
De
Â
monet, J.-F., Chollet, F., Ramsay, S., Cardebat, D., Nespou-
lous, J.-L., Wise, R., Rascol, A., & Frackowiak, R. S. J. (1992).
The anatomy of phonological and semantic processing in
normal subjects. Brain, 117, 1753±1768.
De
Â
monet, J.-F., Fiez, J. A., Paulesu, E., Petersen, S. E., & Za-
torre, R. J. (1996). PET Studies of phonological processing: A
critical reply to Poeppel. Brain and Language, 55, 352±379.
De
Â
monet, J.-F., Price, C., Wise, R., & Frackowiak, R. S. J. (1994).
A PET study of cognitive strategies in normal subjects during
language tasks: Influence of phonetic ambiguity and se-
quence processing on phoneme monitoring. Brain, 117,
671±682.
De
Â
monet, J.-F., Wise, R., & Frackowiak, R. S. J. (1993). Lan-
guage functions explored in normal subjects by positron
emission tomography: A critical review. Human Brain
Mapping, 1, 39±47.
D'Esposito, M., Detre, J. A., Alsop, D. C., Shin, R. K., Atlas, S., &
Grossman, M. (1995). The neural basis of the central ex-
ecutive system of working memory. Nature, 378, 279±281.
Felleman, D. J., & Van Essen, D. C. (1991). Distributed hier-
archical processing in the primate cerebral cortex. Cerebral
Cortex 1, 1±47.
Fodor, J. A., & Pylyshyn, Z. W. (1990). Connectionism and
cognitive architecture: A critical analysis. Cognition, 28,
3±71.
Ford, I. (1986). Confounded correlations: Statistical limitations
in the analysis of interregional relationships of cerebral me-
tabolic activity. Journal of Cerebral Blood Flow and Meta-
bolism, 6, 385±388.
Frackowiak, R. S. J. (1994). Functional mapping of of verbal
memory and language. TINS, 17, 109±115.
Friston, K. (1994). Functional and effective connectivity. Hu-
man Brain Mapping, 2, 56±78.
Friston, K., Holmes, A., Worsley, K., Poline, J.-B., Frith, C., &
Frackowiak, R. (1995). Statistical parametric maps in func-
tional imaging: A general linear approach. Human Brain
Mapping, 2, 189±210.
Friston, K. J. (1995). Commentary and opinion: II. Statistical
parametric mapping Ð Ontology and current issues. Jour-
nal of Cerebral Blood Flow and Metabolism, 15, 361±370.
Friston, K. J., Frith, C. D., Liddle, P. F., & Frackowiak, R. S.
(1993). Functional connectivity: The principal-component
analysis of large (PET) data sets. Journal of Cerebral Blood
Flow and Metabolism, 13, 5±14.
Garcia, G., & Guerreiro, M. (1983). Pseudo-dementia in illiter-
acy. 6th European meeting of the International Neuropsy-
chological Society.
Gathercole, S. E., & Baddeley, A. D. (1993). Working memory
and language, Hove, UK: Erlbaum.
Gathercole, S. E., & McCarthy, R. A. (1994). Memory tools for
the researcher and clinician. Memory, 2, 97±101.
Gathercole, S. E., Willis, C. S., Baddeley, A. D., & Emslie, H.
(1994). The children's test of nonword repetition: A test of
phonological working memory. Memory, 2, 103±127.
Glosser, G., Kohn, S. E., Friedman, R. B., Sands, L., & Grugan,
P. (1997). Repetition of single words and nonwords in Alz-
heimer's disease. Cortex, 3, 653±666.
Grafton, S. T., Sutton, J., Couldwell, W., Lew, M., & Waters, C.
(1994). Network analysis of motor system connectivity in
Parkinson's disease: Modulation of thalamocortical interac-
tions after pallidotomy. Human Brain Mapping, 2, 45±55.
Greitz, T., Bohm, C., Holte, S., & Eriksson, L. (1991). A com-
puterized brain atlas: Construction, anatomical content and
some applications. Journal of Computer Assisted Tomogra-
phy, 15, 26±38.
Habib, M., & De
Â
monet, J.-F. (1996). Cognitive neuroanatomy of
language: The contribution of functional neuroimaging.
Aphasiology, 10, 217±234.
Hayduk, L. A. (1987). Structural equation modeling with LIS-
REL, Baltimore, MD: Johns Hopkins.
Herbster, A. N., Mintun, M. A., Nebes, R. D., & Becker, J. T.
(1997). Regional cerebral blood flow during word and non-
word reading. Human Brain Mapping, 5, 84±92.
Horwitz, B., & McIntosh, A. R. (1993). Network modeling and
functional neuroimaging for mapping brain cognitive func-
tion. In: K. Uemura, N. A. Lassen, T. Jones & I. Kanno (Eds.),
Quantification of Brain Function, (pp. 589±596), Amster-
dam, Netherlands: Elsevier.
Horwitz, B., McIntosh, A. R., Haxby, J. V., & Grady, C. L.
(1995). Network analysis of brain cognitive function using
metabolic and blood flow data. Behavioural Brain Re-
search, 66, 187±93.
Horwitz, B., & Rapoport, S. I. (1988). Partial correlation
coefficients approximate the real intrasubject correlation
pattern in the analysis of interregional relations of cere-
bral metabolic activity. Journal of Nuclear Medicine, 29,
392±399.
Horwitz, B., Soncrant, J. V., & Haxby, J. V. (1992). Covariance
analysis of functional interactions in the brain using meta-
bolic and blood flow data. In: F. Gonzalez-Lima, T. Finken-
staedt & H. Scheich (Eds.), Advances in Metabolic Mapping
Techniques for Brain Imaging of Behavioral and Learning
Functions, (pp. 189±217). Dordrecht, Netherlands: Kluwer
Academic Publishing.
Ingvar, M., & Petersson, K. M. (2000), forth coming. Functional
maps and brain networks. In: A. W. Toga & J. C. Mazziotta
(Eds.), Brain Mapping: The Systems, San Diego, CA: Aca-
demic Press.
JoÈreskog, K., & SoÈrbom, D. (1986). LISREL VI: Analysis of lin-
ear structural relationships by maximum likelihood and
least square methods, Mooresville, IN: Scientific Software.
380 Journal of Cognitive Neuroscience Volume 12, Number 3

JoÈreskog, K., SoÈrbom, D., (1989). LISREL 7 user's reference
guide, Mooresville, IN: Scientific Software.
Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983). Optimi-
zation by simulated annealing. Science, 220, 671±680.
Kirkpatrick, S., & Sorkin, G. B. (1995). Simulated annealing. In:
M. A. Arbib (Ed.), The handbook of brain theory and neural
networks, (pp. 876±879). Cambridge: MIT Press.
Koch, C., & Davis, J. L. (1994). Large-scale neuronal theories of
the brain, Cambridge: MIT Press.
Kolb, B., & Whishaw, I. Q. (1996). Fundamentals of human
neuropsychology, New York: W. H. Freeman and Company.
Lecours, A. R., Mehler, J., Parente, M. A., Aguiar, L. R., Silva, A.,
B., Caetano, M., Camarotti, H., Castro, M. J., Dehaut, F.,
Dumais, C., Gauthier, L., Gurd, J., Leitao, O., Maciel, J., Ma-
chado, S., Melaragno, R., Oliveira, L. M., Paciornik, J., Sanvito,
W., Silva, E. S., Silifrandi, M., & Torne
Â
, C. H. (1987a). Illiteracy
and Brain Damage: 2. Manifestations of unilateral neglect in
testing ``auditory comprehension'' with iconographic mate-
rials. Brain and Cognition, 6, 243±265.
Lecours, A. R., Mehler, J., Parente, M. A., Beltrami, M. C., Silva, A.,
B., Tolipan, L. C., Cary, L., Castro, M. J., Carrono, V., Chagas-
telles, L., Dehaut, F., Delgado, R., Evangelista, A., Fajgenbaum,
S., Fontoura, C., Karmann, D. F., Gurd, J., Torne
Â
, C. H., Jaku-
bovicz, R., Kac, R., Lefevre, B., Lima, B., Maciel, J., Mansur, L.,
Martinez, R., Nobrega, M. C., Osorio, Z., Paciornik, J., Papa-
terra, F., Penedo, M. A. J., Saboya, B., Scheuer, C., Silva, A. B.,
Spinardi, M., & Teixeira, M. (1987b). Illiteracy and Brain Da-
mage: 3. A contribution to the study of speech and language
disorders in illiterates with unilateral brain damage (initial
testing). Neuropsychologia, 26, 575±589.
Liotti, M., Gay, C. T., & Fox, P. T. (1994). Functional imaging
and language: Evidence from positron emission tomogra-
phy. Journal of Clinical Neurophysiology, 11, 175±190.
Macdonald, C., & Macdonald, G. (1995). Connectionism: De-
bates on Psychological Explanation, Oxford, UK: Blackwell
Publishers.
Martin, N., & Saffran, E. M. (1992). A computational account of
deep dysphasia: Evidence from a single case study. Brain
and Language, 43, 240±274.
McIntosh, A. R., & Gonzalez-Lima, F. (1991). Structural mod-
eling of functional neural pathways mapped with 2-deoxy-
glucose: Effects of acoustic startle habituation on the
auditory system. Brain Research, 547, 295±302.
McIntosh, A. R., & Gonzalez-Lima, F. (1994). Structural equa-
tion modeling and its application to network analysis in
functional brain imaging. Human Brain Mapping, 2, 2±22.
McIntosh, A. R., Grady, C. L., Haxby, J. V., Ungerleider, L. G., &
Horwitz, B. (1996). Changes in limbic and prefrontal func-
tional interactions in a working memory task for faces. Cer-
ebral Cortex, 6, 571±584.
McIntosh, A. R., Grady, C. L., Ungerleider, L. G., Haxby, J. V.,
Rapoport, S. I., & Horwitz, B. (1994). Network analysis of
cortical visual pathways mapped with PET. Journal of Neu-
roscience, 14, 655±66.
Mesulam, M.-M., & Mufson, E. J. (1985). The insula of Reil in
man and monkey: Architectonics, connectivity, and function.
In: A. Peters & E. G. Jones (Eds.), Cerebral cortex, (pp. 179±
226). New York: Plenum.
Mesulam, M. M. (1990). Large-scale neurocognitive networks
and distributed processing for attention, language, and
memory. Annals of Neurology, 28, 597±613.
Mesulam, M. M. (1998). From sensation to cognition. Brain,
121, 1013±1052.
Morais, J. (1993). Reading disabilities: Diagnosis and com-
ponent processes. In: R. M. Joshi & C. K. Leong (Eds.),
(pp. 175±184). Dordrecht, Netherlands: Kluwer Academic
Publishers.
Morais, J., Cary, L., Alegria, J., & Bertelson, P. (1979). Does
awareness of speech as a sequence of phones arise sponta-
neously? Cognition, 7, 323±331.
Nascimento, M. F., Rivenc, P., & Cruz, M. L. (1987). Portugues
fundamental, metodos e documentos. Instituto Nacional de
Investigatio Cientifica, Centro de Linguistica da Universidade
de Lisboa.
Nyberg, L., McIntosh, A. R., Cabeza, R., Nilsson, L.-G., Houle, S.,
Habib, R., & Tulving, E. (1996). Network analysis of positron
emission tomography regional cerebral blood flow data:
Ensemble inhibition during episodic memory retrieval.
Journal of Neuroscience, 16, 3753±3759.
Pardo, J. V., Pardo, P. J., Janer, W., & Raichle, M. E. (1990).
The anterior cingulate cortex mediates processing selec-
tion in the Stroop attentional conflict paradigm. Pro-
ceedings of the National Academy of Sciences, U.S.A, 87,
256±259.
Patterson, K., & Lambon Ralph, M. A. (1999). Selective dis-
orders of reading? Current Opinion in Neurobiology 9,
235±239.
Patterson, K., & Shewell, C. (1987). Speak and spell: Dissoca-
tions and word-class effects. In: S. G. Coltheart, G. Sartori &
R. Job (Eds.), The cognitive neuropsychology of language,
(pp. 273±294). Hillsdale, NJ: Erlbaum.
Paulesu, E., Connelley, A., Frith, C. D., Friston, K. J., Heather, J.,
Meyers, R., Gadian, D. G., & Frackowiak, R. S. J. (1995).
Functional MR imaging, correlations with positron emission
tomography: Initial experience using a cognitive activation
paradigm on verbal working memory. Neuroimaging Clinics
of North America, 5, 207±225.
Paulesu, E., Frith, C. D., & Frackowiak, R. S. (1993). The neural
correlates of the verbal component of working memory.
Nature, 362, 342±344.
Paulesu, E., Frith, U., Snowling, M., Gallagher, A., Morton, J.,
Frackowiak, R. S. J., & Frith, C. D. (1996). Is developmental
dyslexia a disconnection syndrome? Evidence from PET
scanning. Brain, 119, 143±157.
Petersen, S. E., & Fiez, J. A. (1993). The processing of single
words studied with positron emission tomography. Annual
Review of Neurology, 16, 509±530.
Petersen, S. E., Fox, P. T., Posner, M. I., Mintun, M., & Raichle,
M. E. (1989). Functional imaging and language: Evidence
from positron emission tomography. Journal of Cognitive
Neuroscience, 1, 153±170.
Petersson, K. M., Nichols, T. E., Poline, J.-B., & Holmes, A. P.
(1999a). Statistical limitations in functional neuroimaging I:
Noninferential methods and statistical models. Philosophi-
cal Transactions of the Royal Society of London, B 354,
1239±1260.
Petersson, K. M., Nichols, T. E., Poline, J.-B., & Holmes, A. P.
(1999b). Statistical limitations in functional neuroimaging
II: Signal detection and statistical inference. Philosophical
Transactions of the Royal Society of London, B 354,
1261±1282.
Petersson, K. M., Reis, A., AskeloÈf, S., Castro-Caldas, A., & In-
gvar, M. (1998). Differences in inter-hemispheric interactions
between literate and illiterate subjects during verbal repeti-
tion. NeuroImage, 7, S217.
Petersson, K. M., Reis, A., Castro-Caldas, A., & Ingvar, M.
(1999c). Effective auditory-verbal encoding activates the left
prefrontal and the medial temporal lobes: A generalization
to illiterate subjects. NeuroImage, 10, 45±54.
Petrides, M. (1995). Functional organization of the human
frontal cortex for mnemonic processing. Evidence from
neuroimaging studies. Annals of the New York Academy of
Science, 769, 85±96.
Petrides, M., Alivisatos, B., & Evans, A. C. (1995). Func-
tional activation of the human ventrolateral frontal cortex
during mnemonic retrieval of verbal information. Pro-
Petersson et al. 381

ceedings of the National Academy of Sciences, U.S.A., 92,
5803±5807.
Petrides, M., Alivisatos, B., Meyer, E., & Evans, A. C. (1993).
Functional activation of the human frontal cortex during
the performance of verbal working memory tasks. Pro-
ceedings of the National Academy of Sciences, U.S.A., 90,
878±852.
Pinker, S. (1997). Words and rules in the human brain. Nature,
387, 547±548.
Pinker, S., & Prince, A. (1989). Rules and connections in human
language. In: R. G. M. Morris (Ed.), Parallel distributed
processing: Implications for psychology and neurobiology,
(pp. 182±199). Oxford: Clarendon Press.
Posner, M. I., & Petersen, S. E. (1990). The attention system
of the human brain. Annual Review of Neuroscience, 13,
25±42.
Posner, M. I., & Raichle, M. E. (1994). Images of mind, New
York: Scientific American Library.
Price, C. J., Moore, C. J., Humphreys, G. W., & Wise, R. J. S.
(1997). Segregating semantic from phonological processes
during reading. Journal of Cognitive Neuroscience, 9,
919±931.
Price, C. J., Wise, R. J. S., Warburton, E. A., Moore, C. J., Ho-
ward, D., Patterson, K., Frackowiak, R. S. J., & Friston , K. J.
(1996). Hearing and saying: The functional neuro-anatomy
of auditory word-processing. Brain, 119, 919±931.
Prince, A., & Smolensky, P. (1997). Optimality: From neural
networks to universal grammer. Science, 275, 1604±1610.
Raichle, M. E. (1994). Images of the mind: Studies with mod-
ern imaging techniques. Annual Review of Psychology, 45,
333±356.
Raichle, M. E., Fiez, J. A., Videen, T. O., MacLeod, A.-M. K.,
Pardo, J. V., Fox, P. T., & Petersen, S. E. (1994). Practice-
related changes in human brain funtional anatomy during
nonmotor learning. Cerebral Cortex, 4, 8±26.
Redington, M., & Chater, N. (1997). Probabilistic and distribu-
tional approaches to language acquisition. TICS, 1, 273±281.
Reis, A., & Castro-Caldas, A. (1997). Illiteracy. A bias for cog-
nitive development. Journal of the International Neurop-
sychological Society, 3, 444±450.
Rumelhart, D., & McClelland, J. L. (1986). Parallel distributed
processing: Explorations in the microstructures of cogni-
tion. Cambridge: MIT Press.
Salmon, E., Van der Linden, M., Collette, F., Delfiore, G., Ma-
quet, P., Degueldre, C., Luxen, A., & Franck, G. (1996). Re-
gional brain activity during working memory tasks. Brain,
119, 1617±1625.
Schmahmann, J. D. (1996). From movement to thought: Ana-
tomical substrates of the cerebellar contribution to cognitive
processing. Human Brain Mapping, 4, 174±198.
Seidenberg, M. (1995). Linguistic morphology. In: M. A. Arbib
(Ed.), (pp. 546±549). Cambridge: MIT Press.
Seidenberg, M., & McClelland, J. L. (1989). A distributed, de-
velopmental model of word recognition and naming. Psy-
chology Review, 96, 523±568.
Shastri, L. (1995). Structured connectionist models. In: M. A.
Arbib (Ed.), The handbook of brain theory and neural
networks, (pp. 949±952). Cambridge: MIT Press.
Shastri, L., & Ajjanagadde, V. (1993). From simple associations
to systematic reasoning: A connectionist encoding of rules,
variables, and dynamic bindings using temporal synchrony.
Behavioural Brain Science, 16, 417±494.
Shepherd, G. M. (1997). The synaptic organization of the
brain, 4th ed. New York: Oxford University Press.
Smolensky, P. (1988). On the proper treatment of connec-
tionism. Behavioural Brain Science, 11, 1±74.
Snowling, M., Hulme, C., & Nation, K. (1997). A connectionist
perspective on the development of reading skills in children.
TICS, 1, 88±91.
Talairach, J., & Tournoux, P. (1988). Co-planar stereotaxic
atlas of the human brain. Stuttgart: George Thieme Verlag.
Vandenberghe, R., Price, C., Wise, R., Josephs, O., & Fracko-
wiak, R. S. J. (1996). Functional anatomy of a common
semantic system for words and pictures. Nature, 383,
254±256.
Vogt, B. A., Finch, D. M., & Olson, C. R. (1992). Functional
heterogeneity in cingulate cortex: The anterior executive
and posterior evaluative regions. Cerebral Cortex, 2,
435±443.
Wechsler, A. F. (1976). Crossed aphasia in an illiterate dextral.
Brain and Language, 3, 164±172.
Wienhard, K., Dahlbom, M., Eriksson, L., Michel, C., Bruck-
bauer, T., Pietrzyk, U., & Heiss, W.-D. (1994). The ECAT
EXACT HR: Performance of a new high resolution positron
scanner. Journal of Computer Assisted Tomography, 18,
110±118.
Wise, R. J. S., Chollet, F., Hadar, U., Friston, K., Hoffner, E., &
Frackowiak, R. S. J. (1991). Distribution of cortical neural
networks involved in word comprehension and word re-
trieval. Brain, 114, 1803±1817.
382 Journal of Cognitive Neuroscience Volume 12, Number 3