Brain and Language 96 (2006) 117128

www.elsevier.com/locate/b&l

Target article

A new empirical angle on the variability debate: Quantitative

neurosyntactic analyses of a large data set from Brocas Aphasia

Dan Drai a, Yosef Grodzinsky b,c,*

b

a
Department of Neurobiology, Weizmann Institute of Science, Rehovot, Israel
McGill University, Department of Linguistics, 1085 Dr. Peneld Avenue, Montreal, Que., Canada PQ H3A 1A7
c
Department of Psychology, Tel Aviv University, Tel Aviv 69978, Israel

Accepted 26 October 2004

Available online 22 August 2005

Abstract
Behavioral variation in Brocas aphasia has been characterized as boundless, calling into question the validity of the syndromebased schema and related diagnostic methods of acquired language disorders. More generally, this putative variability has cast serious doubts on the feasibility of localizing linguistic operations in cortex. We present a new approach to the quantitative analysis of
decient linguistic performance, and apply it to a large data set, constructed from the published literature: Comprehension data of
69 carefully selected Brocas aphasic patients, tested on nearly 6000 stimulus sentences, were partitioned in dierent ways, and subjected to a series of analyses. While a certain amount of variability is indeed evident in the data, our quantitative analyses reveal a
highly robust selective impairment pattern for the group: the patients ability to analyze syntactic movement is severely compromised, in line with the Trace-Deletion Hypothesis. Further analyses suggest that patients performance on no-movement sentence
types exhibits less variation than on sentences that contain movement. We discuss the clinical and theoretical implications of our
results.

2005 Elsevier Inc. All rights reserved.
Keywords: Aphasia; Variability; Syntax; Quantitative analysis; Condence interval; b-Distribution

1. The problem
For over a century, the question of patient grouping has occupied the best minds in behavioral neurology. In many, if not most instances, it took the shape
of a debate over the typology of aphasic syndromes,
and whether or not these are truly distinguishable
from one another. As Mohrs well-known review of
Brocas aphasia (1978) notes, this syndrome has surq
Supported in part by Canada Research Chairs, Canada Foundation for Innovation, and a McGill VP-Research internal grant. D.
Drais research at the Weizmann Institute of Science is supported by
the Golda and Dr. Yehiel Shwartzman and Sara and Haim Medvedi
Families Postdoctoral Fellowship.
*
Corresponding author. Fax: +1 514 398 7088.
E-mail address: [email protected] (Y. Grodzinsky).

0093-934X/$ - see front matter
2005 Elsevier Inc. All rights reserved.
doi:10.1016/j.bandl.2004.10.016

vived a host of classicatory attempts, from Dejerines

to Lurias. Debates regarding diagnosis, localization,
and classication have long been with us, and do
not seem to end. In recent past, many researchers have
focused on putative similarities and dierences between
aspects of Brocas and Wernickes aphasia (i.e.,
whether agrammatism in the speech production of
the former syndrome is a decit dierent from paragrammatism that the latter is said to exhibit, cf. Kolk
& Heeschen, 1992, and whether comprehension patterns in the two groups are truly distinct, cf. Kolk,
2004; Zurif, 1995, for opposing views). Yet the debate
has also touched the question of uniformity within a
syndrome, lately narrowing down to one important issue: is the observed behavior of patients accorded a
diagnostic given label not only distinct from others,

118

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

but also, suciently similar so as to justify our grouping them together through the traditional syndrome
schema? Discussion, then, has shifted: from inter-syndromic dierences, it has moved to examine intra-syndromic similarities, with Brocas aphasia being the
main case in point. And here, positions dier: some
researchers (e.g., Avrutin, 2001; Goodglass, 1993;
Grodzinsky, 2000; Zurif, 1995) view Brocas aphasia
not only as a distinct clinical entity, but also, as an
important tool for discerning the functional linguistic
role of Brocas area; others (e.g., Berndt, Mitchum,
& Haedinges, 1996; Caramazza, 1986; Caramazza,
Capitani, Rey, & Berndt, 2000) have denied the validity of the traditional syndrome schema, arguing that
no structure can be discerned in the performance of
these patients, when taken as a group. To bolster their
position, the latter researchers have documented performance instability, purportedly showing how a xed
clinical diagnosis leads to seemingly boundless behavioral variation across patients (in a relevant behavioral
domain).
Both groups have claimed that their position is
empirically motivated and based on data that exist in
the experimental literature. It is on this puzzle that the
present paper is focused. Since at most, one of the two
views in the variability debate is correct, it is extremely
important to determine what the empirical record really
is, especially in light of the variation-oriented view: if
true, it speaks against generalizations over these data,
and the validity of traditional diagnostic methods. It
thus undermines both the clinical schema, and the feasibility of localizing linguistic operations in cortex
through lesion data.
Below, we employ novel quantitative methods to a
large database in order to nd structure in the comprehension performance of patients classied as Brocas
aphasics. Our database was compiled from carefully
selected, published test scores of patients with identical
diagnosis of Brocas aphasia. As the variability debate
is concerned with the correct interpretation of a data
set, it is crucial to characterize the data set and its limitations, and also to dene precisely what counts as an
empirical validation of a thesis about the presence (or
absence) of structure in the empirical record. We therefore outline the general logic of our analysis before
embarking upon a detailed description of our
procedure.
Our starting point is that performance variation within a group of patients in itself does not preclude the existence of structure in their decit. Thus in aphasia the
data may present inter-patient variability, but the challenge for us is to try and discover commonalities at
the group level in the face of this variability. That is,
we ask whether group analyses can discover dierent
patterns of performance on dierent tasks. For that,
we constructed a database from raw performance scores

of Brocas aphasicsthe best-studied aphasic syndrome,

hence at the heart of the variability debate. The experiments we discuss measure error rates in forced binary
choice comprehension tasks that involve a variety of
sentence types. The observed performance level of a particular individual on a given type of sentence can thus be
seen to indicate the probability that this patient succeeds. Each patient is identied by a probability parameter for every sentence type. Thus per each sentence
type, our group of Brocas aphasic patients has a distribution of these probabilities. Since we have several types
of sentences, we have several such distributions for the
group, which we can now analyze. We devised a rigorous statistical test to detect dierences between such distributions. In other words, we explicitly study the
distributions of probabilities of successful performance
and seek to infer signicant dierences of group performances on dierent types of sentences. This, we believe,
dispels most of the obscurity as to the structured nature
of the observed performances: if a comparison of performances on two types of sentences yields a signicant difference in the corresponding distributions of
probabilities we conclude that Broca aphasics are dierentially impaired with respect to this contrast.
Before proceeding, let us emphasize that ours is not
a meta-analysis: we devised a method for the analysis
of raw scores, and this is the nature of our database.
Furthermore, our approach does not contradict any
previous nding: while claims regarding variation in
the empirical record do exist, no one has carried out
a large-scale quantitative analysis such as ours. And
while our approach does admit the existence of variability (indeed it models it), it asks whether dierences
between the populations distributions of probabilities
of success can be assessed despite the variation. This
is quite a standard approach in biology, where no
one expects clean dichotomies to appear as a result of
a manipulation. Our analysis thus inheres in a standard
procedure: we ask whether two distributions can be
considered equal in view of the results of a sampling.
We are not contradicting the results of previous studies
through use of a bigger sample; rather, we are enablingand carrying outa principled inquiry into
the existence of a hitherto controversial contrast (see
Drai, in press).
To recapitulate: we have compiled a large data set
of raw data of performance scores, and adopted a statistical methodology which gives a precise quantitative
meaning to the question of the existence of a signicant
dierence in performanceanalyzed at the group levelbetween two types of sentences. Note that the approach is general and can be applied to any subject
group, and to any categorization of the sentences.
We did apply our method to performances of Brocas
aphasics, for whom we have dierent existing categorizations of sentences.

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

We can now forecast our ndings:

(1) Despite performance variability within the group,
we can dene an explicit stochastic model for the
group distribution of probabilities (of success on a
given type of sentence). Moreover, this characterization enables a rigorous statistical test of the difference (or lack thereof) between pairs of sentence
types in the performance of the group.
(2) In our data, this method reveals a highly signicant
performance dierence when the data are categorized by a syntactic principle (whether or not sentences contain a Transformational Movement
relation); more traditional ones (Mood [active/passive], Sentence Complexity [high/low]) yield no
contrast.

2. Data selection
Aphasiological studies are typically conducted on
samples of small size (rarely exceeding n = 10). To give
our analysis a serious quantitative character we constructed a large data set of sentence level comprehension
scores of Brocas aphasics. Furthermore, the analytic
methods we developed for the evaluation of raw behavioral test results from multiple experiments presuppose a
high degree of homogeneity among studies. To ensure
homogeneity in our database, we limited the scope of
our investigation to a single type of result from one syndrome: scores obtained from Brocas aphasics in sentence-to-scenario matching experiments that measure
error rates in sentence comprehension. While task and
clinical criteria were kept constant, sentence stimuli varied in systematic ways that are at the heart of our
analysis.
We used clinical diagnosis for patient selection. Did
this selection bias our results? While logically possible,
we do not see a reason to suppose that this was the case.
Our main demonstration below is that one partition
method discovers structure in the data, whereas other
methods do not. No clinical diagnostic test we use is
geared toward such distinctions. These tests are designed
to evaluate general communicative skills, and thus focus
mostly on speech production, whereas comprehension
abilities are evaluated in a very general way. It thus
seems that the way clinical tests select patients is orthogonal to our demonstration.
Diagnosis was done on (mostly) behavioral grounds
(typically the BDAE, WAB, AAT, and the like), and
we thus restrict ourselves to the analysis of behavior.
We now move to describe how our database was constructed: scores were selected from the published literature (cf. Appendix A for a list of studies) by a sieve of
ve highly restrictive criteria. A patients score was

119

admitted into the study only if all of the following

requirements were met:1
(a) Positive diagnosis of Brocas aphasia by a well-established test,2 and sound neurological considerations
(i.e., imaging when available, clinical workup
otherwise);
(b) Tests investigate comprehension at the sentence
level, through a forced binary-choice task (who did
what to whom) with multiple semantically reversible sentences, and with scenario pairs that depict
thematicreversals<adoesXtob,bdoesXtoa>;
(c) Detailed descriptions of experimental conditions
and procedures are available;
(d) Raw individual patient scores are available;
(e) Identifying information is available (to avoid multiple counting of patients who participated in more
than one study).
This rigorous sieve produced a uniform data matrix
in terms of patient selection and data acquisition procedures.3 It contained raw scores from 21 studies, with 233
data points (each being a single patients raw score on a
particular sentence type), which were constructed from
5934 trial sentences of 14 types. Tests were conducted
in dierent locations and times, with 69 aphasic speakers
of seven languages (Dutch, English, German, Hebrew,
Japanese, Korean, and Spanish). Experiments diered
only in number of trials (not always equal per contrast
within, let alone between, experiments), and number of
successes per experimental condition.4
3. A three-way data partition
The data set contains comprehension test scores from
two main classes of sentence stimuli: bi-clausal relatives,
and mono-clausal actives/passives. The systematic richness thus obtained enabled quantitative comparisons
(always among sets of minimal pairs) between three data
categorization methods:
(A) The psycholinguistic measure of perceptual Complexity (HIGH/LOW): a behavioral property of a
sentence that seems to correlate with one or more
syntactic properties. Widely viewed as a central
factor in processing, complexity is thought to rely
1
The complete data matrix may be downloaded from http://
freud.tau.ac.il/~yosef1.
2
E.g. Boston Diagnostic Aphasia Examination, Western Aphasia
Battery, Aachen Aphasia Test, etc.
3
We focused on behavior because it is the element under re in
recent years. Neuroanatomically, we remained at the level of clinical
diagnostic tests and structural MR images, that have a reasonably
good localizing value.
4
This dispenses with the need for the use of a weighted eect size
based meta-analysis.

120

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

on Brocas region as its anatomical substrate (Caplan & Waters, 1999; Just, Carpenter, Keller,
Eddy, & Thulborn, 1996; Roder, Stock, Neville,
Bien, & Rosler, 2001). Bi-clausal sentences in our
data matrix were (high complexity) Center Embedded, and (low complexity) Right-Branching, relative clauses.
(B) The traditional grammar distinction Mood (ACTIVE/PASSIVE): arguably, brain damaged individuals experience diculty in processing bi-clausal
sentences in general. If so, then contrast (A) may lead
to perturbations. Luckily, the best studied pair in
aphasiology (active/passive) contains mono-clausal
declarative sentences, and may provide a more distilled contrastMood. In (agentive) active sentences,
the Subject is agent-of-action, and passive morphology is absent; in passive, the Subjects being recipientof-action correlates with morphological marking of
passive voiceEnglish-en or its analogs. Time and
again has the passive construction featured in investigations into the nature of Brocas aphasia (Goodglass, 1968; Grodzinsky, 1984a, 1984b, 1986;
Schwartz, Saran, & Marin, 1980; and many, many
others). Early results come almost exclusively from
English, a language in which passive tends to correlate with deviation from canonical ordering of major constituentsa notion often thought to be
equivalent to syntactic movement. More recent studies have obtained cross-linguistic data, from language
that enable to tease the active/passive distinction
apart from word order changes. We thus chose Mood
as a dimension along which to analyze the data we
compiled.
(C) The linguistic notion Movement (): simply put,
Movement involves intra-sentential action at a distance, which cuts the data orthogonally to each of
the two contrasts above. Consider the active/passive distinction: in the active sentence the horse
kicked the rider, the verb kick determines the
semantic roles of the argument immediately preceding it (the horse), and of the one immediately following it (the rider).5 In the corresponding
passive the rider was kicked c by the horse, however, the elements <kick, the rider> are non-adjacent, and their sequential order is reversed. Still,
semantic roles are preserved under this major
changethe rider remains recipient-of-action.
The verb kick assigns a patient role rightwards just

For simplicity of exposition, we gloss over string-vacuous movement, that is, over movement from VP-internal position into subject
position, as well as movement from subject position in relative clauses.
These distinctions are suppressed as they are not presently relevant, see
Grodzinsky (2004) for further discussion.

like in the active, to the position marked by c (cf.

e.g., Marantz, 1984). This means that phonetically,
the rider in passive is phonetically present sentenceinitially, but its semantic role is downstream, in c.
The two positions <the rider, c> must be related,
to ensure correct interpretation during comprehension. This relation is captured by a Transformational
Movement rule. Similar considerations hold when
two types of relative clauses are compared: the nurse
helped the horse that kicked the rider vs. the nurse
helped the rider that the horse kicked c. Some aphasia and fMRI studies have linked Movement to Brocas region (e.g., Ben-Shachar, Hendler, Kahn, BenBashat, & Grodzinsky, 2003; Grodzinsky, 2000; Zurif, 1995).
Having established these three categorization methods, we proceeded to study them quantitatively. First,
we categorized the bi-clausal relative clause subset of
the data by Complexity, and then by Movement; second,
we categorized the mono-clausal sentences by Mood,
and then by Movement. Doing so enabled us to test
the relative merit of the dierent data partitioning
methods.
4. Data analysis: Synopsis of methods
The raw data take the following shape: for a given
patient, performance on a particular sentence type is
given as the couple (number of trials, number of successes). As each patient has several data points, scores can
be categorized to obtain dierent contrasts (A or B or
C above). Our analysis rst categorizes the raw data in
one way, analyzes them quantitatively at the individual
and the group levels, and then reanalyzes precisely the
same data set after re-categorization. This enables us
to carry out numerical comparisons between dierent
partitioning of the data, to determine the categorization
of choicethe one in which structure is best discerned.
Our analysis proceeds in two rounds (see Appendix B
for formal details):
(i) As a rst step we look at individual patient scores,
all obtained through binary-choice tasks. We represent each individual data point not as a number
(proportion correct), but rather, as a condence
interval on a binomial distribution as follows: we
take the couple (nXA, mXA), where nXA is the total
number of trials for a patient A on type X sentences (e.g., +Movement, or
Mood) and mXA is the
corresponding number of successes. We assume
independence between trials, therefore interpreting
the whole experiment, for a given patient, as
repeated Bernoulli trials with probability parameter PXA. We compute a 99% condence interval
for PX,Aspanning from Lower to Upper value

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

[PLXA, PUXA].6 The relevant set of condence

intervals is now computed for the dierent methods of categorization. Informal, visual inspection
of the results provides a preliminary indication as
to the categorization of choice.
(ii) To get a numerical evaluation of our contrasts, we
apply the appropriate hypothesis testing method
at the group level: we assume that for each subject,
the binary response (success or failure) forms a set
of Bernoulli trials whose probability parameter is
a random variable varying from subject to subject
following a certain probability law. We then assess
the hypothesis that performances on sentence types
X, Y, come from the same underlying distribution
of individual probability parameters. Since a t test
approach cannot be used here,7 we assume that
the probability of success follows a probability law
given by the parametric family B (a,b) (the b-distribution). This form of distribution is capable of a
variety of shapes as its parameters vary, so that no
severe limitation is imposed on the way the probability uctuates (Skellam, 1948). Moreover, it has
been widely used on biological data (Williams,
1975). Implementation was carried out in Mathematica, using the Find Minimum function on the
negative log-likelihood to nd the parameters of
the b-distribution. We validated our implementation by reproducing the numerical results of a
published biological test (Williams, 1975).8
4.1. Complexity vs. Movement in relative clauses
Applying this two-pronged method, we rst compared the categorizations Complexity and Movement
on the relative clause data set (Fig. 1).

6
Notice that the condence interval summarizes information for each
patient about both the percentage of successful trials and the overall
number of trials. This is necessary in view of the fact that the number of
trials for the two branches of a contrast are not always equal in a given
experiment (if it was not specically designed to test the contrast under
scrutiny), let alone in dierent experiments. To display all the results for a
given contrast we have chosen to show each patient as a vertical line
spanning the corresponding [PLXA, PUXA] interval.
7
We have three reasons to disqualify a t-test approach: (A)
summarizing the performances of each patient as a percentage of
success gives an empirical distribution of numbers in the [0, 1] ranges,
which does not square with the range of the normal distribution
assumption underlying the t-test. (B) This move is unwarranted also
because the number of trials for dierent patients (and even for
dierent contrasts for a given patient) is not xed. This means that the
percentage values cannot be pooled safely. (C) More conceptually, we
are trying to determine the distribution over the population of Px for a
given X, and there is no reason to assume a priori that this distribution
is bell shaped (Fig. 4 shows a case in which the best tting distribution
is J-shaped).
8
See supplementary materials for a formal presentation, implementation, and validation.

121

Fig. 1 presents two levels of contrast in English relative

clauses. Each sentence contains a main clause (bolded),
and a relative clause (bracketed, courier subscript). Rows
constitute the Complexity level: the relative can either
modify the subject of the main clause (resulting in the
more complex Center Embedding a, b), or the object (in
the less complex Right Branching c, d). Columns constitute the Movement level: in the [bracketed] relative clause,
the grammatical function of the head (the woman, the
man) can either be Subject b (a, c) or Object c (b, d). In
Subject relative constructions (a, c), the blue relative head
connects to an empty proximal position b, located on the
left edge of the relative clause. The link <the woman-b>
crosses no part of the relative, and is analyzed as
Movement.9 In Object-relative constructions (b, d), by contrast,
the red relative head connects to c, a distal, non-left-edge,
empty position. The +Movement link <the man-c>,
depicted by the arrow, crosses both the verb and the subject of the relative clause.10
The 2 2 data matrix has 64 data points from 23
English speaking Brocas aphasics, with results arranged
according to the contrasts given in Fig. 2:
In panel (A), the Center Embedding (perforated red
line) vs. Right Branch (full black line) contrast is represented by a vertical display of the condence intervals
for all patients. Each vertical line represents the condence interval for a given patient. Panel (B) is same as
panel (A), this time
Movement = perforated red line,
+Movement = full black line. Panel (C) represents the
b-model for CE (Red) and RB (Black). The x-axis represents possible values of the probability parameter,
the y-axis units are such that the area under the curve
is 1. Panel (D) is same as panel (C), this time
Movement = red, +Movement = black.
The complexity contrast (A and C) discerns no structure; Movement was vastly superiorthe contrast was
highly signicant (Fig. 2; see Table 1 for all numerical
results), indicating that Movement, but not Complexity,
is a linguistic factor supported by Brocas region.

5. Mood vs. Movement in actives and passives

A second quantitative comparison that we carried out
was between Mood and Movement, for which we used
the active and passive data. Consider what is needed
to test this contrast: we must nd an array of monoclausal sentences in which word-order changes (i.e.

The relative pronoun who is viewed here as being in between the

clauses, belonging to neither; we assume irrelevance of so-called
vacuous movement for the perspective developed here (or a
vacuous movement contrast).
10
Some data from three additional langugages (Chinese, Japanese
and Korean) exist in the data set, but was not included in the analysis
for logistical reasons.

122

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

Fig. 1. Movement vs. complexity in relative clauses.

Fig. 2. Individual condence intervals and b-distribution for relative clauses. (See text for interpretation.)

Table 1
Summary of parameters of the b-models tted to each contrast, together with ratios and p value of the test for the null hypothesis (i.e., no
performance dierence between the branches of the contrast)

M
P
reasy/rhard
p

Mood vs. movement (English

included)

Mood vs. movement (English

excluded)

Complexity vs. movement

All movement

Passive/Active

+Mov/
Mov

Passive/Active

+Mov/
Mov

HIGH (CE)/
LOW(RB)

+Mov/
Mov

+Mov/
Mov

0.713
0.804
0.194
0.164
1.183
0.015

0.610
0.849
0.172
0.123
1.398
4.7 10
13

0.768
0.791
0.168
0.179
0.938
0.57

0.594
0.866
0.118
0.110
1.072
2.9 10
15

0.716
0.687
0.160
0.170
0.941
0.84

0.573
0.835
0.119
0.09
1.322
2.4 10
8

0.594
0.842
0.154
0.117
1.316
<10
16

Movement) and morphological (i.e., Mood) changes are

independent of one another. Ideally, this yields a 2 2
design: one axis in which word order is kept constant
and morphology is changed; and another axis in which
morphology is kept constant and word order is changed.

How can morphological change and word order

change be teased apart in mono-clausal sentences?
Looking at English (or Hebrew, or Spanish), this task
appears nearly impossible: as we pointed out above,
English passive constructions contain passive morpholo-

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

123

Fig. 3. Movement vs. mood in simple sentences. (For interpretation of the references to colors in this gure legend, the reader is referred to the web
version of this paper.)

gy, an inserted auxiliary verb, which in most cases are

accompanied by a change in word order relative to the
active counterpart. English (or Hebrew, or Spanish)
passive is thus derived by syntactic movement. These
are SVO (SubjectVerbObject) languages, in which
simple active declarative sentences feature the subject
on the left of the verb, and the direct object on its right.
In passive, morphological changes and auxiliary insertion co-occur with the movement of the object from
the right of the verb leftward. The object thereby crosses
the verb [__Saw Bill Bill was seen__]. This complex
cluster of active/passive dierences makes it dicult to
tease Mood apart from Movement through these constructions in this language (although see Wasow, 1977,
for a classical paper on passive constructions that are
not derived by movement; see Grodzinsky, Pierce, &
Marakovitz, 1991, for an investigation of these in
aphasia).
Yet this picture changes if we extend our eld of view
to other language types. SOV (SubjectObjectVerb)
languages like German or Dutch avail us of the desired
contrast. There, passivization does not necessarily cooccur with verb crossing by the (internal) argument that
becomes the subject of the passive sentence. To illustrate, the Dutch analogs of the English fragments above
are [__heeft het meisje gekust het meisje wordt__ gekust].11 Het meisje (the girl) maintains its position relative to the participle gekust, despite auxiliary change
(cf. the bottom left quadrant of Fig. 3). Mood change
in Dutch/German does not necessarily entail word order
change (auxiliary notwithstanding). This provides one
half of the desired contrast (constant order, varying
morphology), which we can now explore quantitatively.
The other half of the contrast (varying order, constant morphology) would be a Movement contrast that
does not entail Mood change (as reected in morphological change). This we nd in English, Hebrew, Spanish,

Korean, and Japanese, where simple active declarative

sentences manifest two varieties: either the subject precedes the object (S..O..), or the object precedes the subject (O..S..). This can happen through either scrambling,
or topicalization, depending on the language and the
operation chosen. Such an order change (in which
meaning remains unmodied) is mostly viewed as an instance of Movement in both SVO and SOV languages
(for recent discussion, cf. Fanselow, 2001, for German;
Saito & Fukui, 1998, for Japanese). Crucially, while
the subject and the object switch positions, the verb remains in active voice, keeping Mood constant (e.g., Fig.
3 below, top right quadrant). This Movement operation
thus occurs within a single active clause, and without
visible morphological or Mood changes. Movement in
simple sentences, then, can be independent of change
in Mood. We used this contrast in our test.
We have thus illustrated how all four possibilities are
realized, and it important to note that the syntactic analysis we assume is not controversial. The resulting array
is summarized in Fig. 3, with new example sentences.
And while there are complications that go beyond the
scope of this discussion, our discussion is empirically
justied (though somewhat simplied for ease of exposition, see Grodzinsky, in press, for detailed elaboration
and justication of some of the necessary linguistic
assumptions).12 Fig. 3 is set up to reect this distinction:
sentences may contain Movement without passive morphology (cf. top right corner of Fig. 3), as well as Passive
morphology without Movement (bottom left corner).
The same gure presents the logical structure of our
second test, which sought to see which distinction is
apparent in Brocas aphasia: the traditional active/passive (Mood) distinction or Movement. Fig. 3 thus presents two levels of contrast in monoclausal sentences.
Rows constitute the traditional grammar Mood level:
12

11

For simplicity, we ignore here an extra complication: SOV

conguration in Dutch and German is evident in simple sentence only
when there is an auxiliary in addition to main verb (cf. Bill sah Peter
vs. Bill hat Peter gesehen). For our purposes, this issue can be ignored,
as the SVOSOV contrast is maintained throughout our sample. See
Grodzinsky, 2004a, for further elaboration.

The position of the subject relative to the by-phrase in German can

be reversed (i.e., the order door het meisje wordt de jongen b gekust is
also possible). One experiment tested both orders, with the same results
(Friederici & Graetz, 1987). The near-perfect performance on both
types of sentences, as well as linguistic considerations pertaining to the
(adjunct) status of the by-phrase, allows us to abstract away from this
contrast.

124

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

Fig. 4. Individual condence intervals and b-distribution for non-English active/passive. (For interpretation of the references to colors in this gure
legend, the reader is referred to the web version of this paper.)

actives (a, b) lack passive morphology, and their logical

object is in Object position. Passives (c, d) contain an
auxiliary verb (e.g., English is, Dutch wordt), and a passive morpheme (e.g., English-en, Dutch ge-V-t); the
logical object is located in Subject position. Columns
constitute the Movement level:
Movement trivial links:
(a) Object (blue) is exactly in the position of recipient-ofaction in Hebrew active; (c) Logical object (blue) is
adjacent to the position of recipient-of-action in Dutch
passive <de jongen-b>. Non-trivial +Movement links:
(b) The arrow-marked link between the topicalized Object (red) and the position of recipient-of-action <harofe, c> crosses the Subject and the verb (d) The arrow-marked link <the boy
c> in the passive crosses
the verb.
5.1. The quantitative analysis
The complete active/passive data subset contained
155 points, from 69 Brocas aphasic speakers of ve languages (Fig. 3).13 Recall that in English, Movement and
Mood mostly overlap.14 For this reason, we analyzed
our data twicewith and without the English data.
We rst analyzed it as a whole, namely with 57 English
data points included. A signicance dierence was found
for both contrastsnamely Mood and Movement (Table
13
Some data from three additional languages (Chinese, Japanese and
Korean) exists in the data set, but was not included in the analysis.
14
Except special cases, such as adjectival passive, cf. Grodzinsky,
Pierce and Marakovitz (1991).

1). Next, we excluded the English cases, and only kept

data from the four languages that featuring a Movement contrast within actives, and a Movement contrast
in passive within and between languages (98 scores, 42
Brocas speakers of Dutch, German, Hebrew, and
Spanish).
Our results are presented in Fig. 4 as follows: Panel
(A) the Active (perforated red line) vs. Passive (full
black line) contrast is represented by a vertical display
of the condence intervals for all patients. Each vertical
line represents the condence interval for a given patient. Panel (B) is same as panel (A), this time
Movement = perforated red, +Movement = full black. Panel
(C) the b-model for Active (Red) and Passive (Black).
The x-axis represents possible values of the probability
parameter, the y-axis units are such that the area under
the curve is 1. Panel (D) is same as panel (C), this time

Movement = red, +Movement = black. The Mood
contrast (A and C) discerns no structure, while the
Movement contrast discerns structure, and yields highly
signicantly dierent probability distributions, providing further evidence of the strong relation between a
Movement decit and Brocas aphasia, and indicating
that Mood is not a natural class vis-a`-vis the syntactic
comprehension decit in this syndrome.
To underscore this last point, we present a comparison
between English and Dutch/German passive, that is, a
subset of the data in Fig. 4 that include +Movement passives in English (n = 34), and the
Movement passives
in Dutch/German (n = 18). A robust dierence is apparent (Fig. 5).

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

Fig. 5. Individual condence intervals of Movement passive: Dutch/

German (perforated red line), vs. English (full black line).

Finally, a combined analysis, in which all our data

were pooled together and partitioned by the Movement contrast, was highly signicant15 ("all Movement"
in Table 1), strengthening the view that the two separately investigated contrasts fall under a single syntactic
generalization (e.g., Chomsky, 1995).

6. Remaining variation
While the movement eect is robust, and helps discern structure in the data, variation no doubt exists in
our data set, as can be revealed by mere inspection of
the individual condence intervals, or, more formally,
by the r values of each of the b-distributions (Table 1;
see Appendix B for the details of the calculation of for
this distribution). We do not know whether this variation exceeds the variation documented through other
experimental methods, in other populations or cognitive
domains. Still, it may be worthwhile to try and investigate the variation in our data in more detail. In a preliminary attempt to do so, we examined the r values of
the dierent curves generated for the comparisons
above. Unstructured variation would lead to similar r
values across curves. If we were to expect structured variation, then harder experimental conditions would likely
lead to r values that are greater than those of the
easier conditions, because the latter would be more
likely to reach ceiling performance. This means that
rhard/reasy > 1. If our reasoning is valid, then the r-ratio
might be used to point to the correct hard/easy contrast
in the data.
The evidence we have in this respect is weak, but
suggestive: it appears that +Movement experimental
conditions produce data whose variance is greater
15

An additional 18 data points, from related syntactic contrasts

(Subject- vs. Object-questions, Subject- vs. Object-Topicalization) were
included in the last analysis.

125

than the
Movement conditions. That is, in all

instances, r+Mov/r
Mov > 1 (1.398 for the Movement
contrast in active/passive with English; 1.072 for the
same contrast in active/passive without English;
1.322 for the Movement contrast in relative clauses;
1.316 for the same contrast for all sentence types).
The other contrasts, by comparison, are non-uniform
in this respect: the active/passive (with English) case,
the hard/easy ratio is 1.183; the same comparison
without English is <1 (0.938); nally, as is the rCE/
rRB ratio in the branching contrast of the relative
clauses (0.941). These are, admittedly, weak eects,
but they do point to a direction consistent to our
claims.

7. A concluding remark
The new method for quantitative analysis we propose shows that for the 69 patients we studied, Brocas aphasia leads to a robust Movement decit,
cutting across other elements that have been thought
to generate comprehension diculties (Mood, Complexity). Although the group we have studied is large,
it is, of course, dicult to assert that the selection criteria we used are universally valid; yet, they do dene
a clinical population that exhibits stable behavioral
decits, and that these decits characterize what clinicians have in mind when they think of Brocas aphasia. Neurologically, then, existent diagnostic methods
appear ecacious for this syndrome. Thus, while variability existsas it does in virtually all areas of biologythe robust structure we uncovered in the data,
and its relation to clinical diagnostic tests of Brocas
aphasia, are clear.
Still, the variability debate is unlikely to stop here.
We would thus like to clarify what we think we did
by carrying out this analysis, and what we did not.
We begin on the negative side: rst, our analysis is
entirely empirical, and we have nothing to say regarding conceptual issues in neuropsychology. Second, our
analysis has nothing to say about the specicity of
decit in Brocas region (or lack thereof). That is, as
we focused on Brocas aphasics, we did not address
the question of whether there exist comprehension
tests that tease it apart from other syndromes. Hopefully, future work along the same lines will help determine this rather dicult issue. Finally, we do not
show that variation does not exist in this syndrome.
On the contrary: it is evident in the data, just like
in most results in neuroscience. What we did show,
however, is how our methods discover highly robust
structure in a large data set that was previously said
to be resistant to such structural analyses. It is this result and this analysis, and nothing else, that we would
now like to bring to the fore.

126

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

Appendix A. The data sources

English active/passive
Badecker, W., Nathan, P., & Caramazza, A. (1991).
Varieties of sentence comprehension decits: A case
study. Cortex, 27, 311321.
Berndt, R., Salasoo, A., Mitchum, C. C., & Blumstein,
S. E. (1988). The role of intonation cues in aphasic
patients performance of the grammaticality judgement
task. Brain and Language, 34, 6597.
Caplan, D., & Futter, C. (1986). Assignment of
thematic roles to nouns in sentence comprehension
by an agrammatic patient. Brain and Language, 27,
117134.
Grodzinsky, Y., Pierce, A., & Marakovitz, S. (1991).
Neuropsychological reasons for a transformational
derivation of syntactic passive. Natural Language and
Linguistic Theory, 9, 431453.
Hickok, G., & Avrutin, S. (1995). Representation,
referentiality, and processing in agrammatic
comprehension: Two case studies. Brain and Language,
50, 1026.
Linebarger, M., Schwartz, M., & Saran, E. M. (1983).
Sensitivity to grammatical structure in so-called
agrammatic aphasics. Cognition, 13, 361392.
Ostrin, R. K., & Schwartz, M. F. (1986). Reconstructing
from a degraded trace: A study of sentence repetition
in agrammatism. Brain and Language, 28(2),
328345.
Schwartz, M. F., Linebarger, M. C., Saran, E. M., &
Pate, D. S. (1987). Syntactic transparency and sentence
interpretation in aphasia. Language and Cognitive
Processes, 2(2), 85113.
Schwartz, M. F., Saran, E. M., & Marin, O. S. M.
(1980). The word order problem in agrammatism:
Comprehension. Brain and Language, 10,
249262.
Shankweiler, D., Crain, S., Gorrell, P., & Tuller, B.
(1989). Reception of language in Brocas aphasia.
Language and Cognitive Processes, 4, 133.
Sherman, J. C., & Schweickert, J. (1989). Syntactic and
semantic contributions to sentence comprehension in
agrammatism. Brain and Language, 37, 419439.
Non-english active/passive
Beretta, A., Harford, C., Patterson, J., & Pinango, M.
(1996). The derivation of post-verbal subjects: Evidence
from agrammatic aphasia. Natural Language and
Linguistic Theory, 14, (4), 725748.
Beretta, A., Schmitt, C., Halliwell, J., Munn, A., Cuetos,
F., & Kim, S. (2001). The eects of scrambling on
Spanish and Korean agrammatic interpretation: Why
linear models fail and structural models survive. Brain
and Language, 79, 407425.

Burchert, F., De Bleser, R. & K. Sonntag. (2001). Does

case make the dierence?, Cortex, 37, 700703.
Friederici, A. & Graetz, P. A. M. (1987). Processing
Passive Sentences in Aphasia: Decits and Strategies.
Brain and Language, 30, 93105.
Friedmann, N., & Shapiro, L. P. (2003). Agrammatic
comprehension of simple active sentences with moved
constituents: Hebrew OSV and OVS structures. Journal
of Speech Language and Hearing Research, 46, 288297.
Hagiwara, H. (1993). The breakdown of Japanese
passives and theta role assignment principle by Brocas
aphasics. Brain and Language, 45, 318339.
Kolk, H., & van Grunsven, M. (1985). Agrammatism as
a Variable Phenomenon. Cognitive Neuropsychology,
2(4), 347384.
Kolk, H., van Grunsven, M. & Keyser, A. (1985). On
parallelism in agrammatism. In M. -L. Kean (Ed.),
Agrammatism. New York: Academic Press.
Relatives
Burchert, F., De Bleser, R. & K. Sonntag. (2001). Does
case make the dierence? Cortex, 3, 700703.
Beretta, A., Harford, C., Patterson, J., & Pi~
n ango, M.
1996. The derivation of post-verbal subjects: Evidence
from agrammatic aphasia. Natural Language and
Linguistic Theory, 14(4), 725748.
Beretta, A., Schmitt, C., Halliwell, J., Munn, A., Cuetos,
F., & Kim, S. 2001. The eects of scrambling on Spanish
and Korean agrammatic interpretation: Why linear
models fail and structural models survive. Brain and
Language, 79, 407425.
Friedmann, (2000). Moving verbs in agrammatic
production. In R. Bastiaanse & Y. Grodzinsky (Eds.)
Grammatical disorders in aphasia: A neurolinguistic
perspective.
Grodzinsky, Y. (1989). Agrammatic comprehension of
relative clauses. Brain and Language, 37, 480499.
Hickok, G., & Avrutin, S. (1995). Representation,
referentiality, and processing in agrammatic
comprehension: Two case studies. Brain and Language,
50, 1026.
Sherman, J. C., & Schweickert, J. (1989). Syntactic and
semantic contributions to sentence comprehension in
agrammatism. Brain and Language, 37, 419439.
Questions
Hickok, G., & Avrutin, S. (1996). Comprehension of
Wh-questions in two Brocas aphasics. Brain and
Language, 52, 314327.
Thompson, C. K., Tait, M. E., Ballard, K. J., & Fix, S.
C. (1999). Agrammatic aphasic subjects comprehension
of subject and object extracted wh-questions. Brain and
Language, 67, 169187.

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

Clefts
Caplan, D., & Futter, C. (1986). Assignment of thematic
roles to nouns in sentence comprehension by an
agrammatic patient. Brain and Language, 27, 117134.
Appendix B. The methodology
The performance of a subject X on a set of sentences
S is described by a couple (nX,S, mX,S) where
nX,S = number of trials and mX,S = number of correct
answers. We assume independence between trials, therefore interpreting the whole experiment, for a given patient, as repeated bernoulli trials with probability
parameter pX,S. We then proceed to evaluate pX,S in light
of the observed performances of patient X on the sentences S.
A 99% condence interval [pLX,S, pUX,S] for pX,S is
obtained by solving the following equations:

mX;S

X
nX;S
n
pUXk 1
pUX X;S
k .01=2;
k
k0
X
nX;S 
pLXk 1
pLXnX;S
k 1
.01=2.
k
k0

mX;S
1

This approach leads to a rst way of assessing the

pertinence of a contrast: we simply display condence
intervals for pX,S for each subject as well as the condence intervals for pX,S (where S denotes the trial
sentences that are not in S). As can be seen in Figs.
2A and B such a simple visual inspection gives a quite
striking impression of the relevance (or lack thereof)
of a contrast. Thus, a contrast in group behavior between the pX,S and the pX,S becomes visually obvious, but a rigorous hypothesis testing framework is
preferable.
One could try to summarize each subjects performance by the percentage of success and then analyze
the distribution of these numbers, but since the number
of trials varies from study to study, serious distortion of
the data would likely occur. Indeed, a deviation of 10%
from chance in 10 trials does not have the same statistical meaning as a deviation of 10% from chance in 100
trials. The conventional tools for testing the hypothesis
that two samples come from the same distribution
(e.g., MannWhitney or Wilcoxon) are therefore not
applicable here and we need a framework which suits
the binomial nature of the data, while leaving room
for the possibility that the probability coecient in the
corresponding Binomial distribution may vary from
subject to subject. Such a framework is provided by
the b binomial distribution, a natural generalization of
the binomial distribution, in which the probability
parameter (xed in the binomial) is allowed to vary

127

according to a parametric law B (a, b). This distribution

is capable of a variety of shapes as its parameters vary,
so that no severe limitation is thus imposed on the way
the probability of success on a given type of sentence
uctuates from subject to subject [see (Skellam 1948)],
moreover it has been widely used on biological data
[see (Williams, 1975)].
Formally, under the B (a, b) model the probability of
m successes in n trials is
P Success mjTrials n


n

Ba m; n b
m=Ba; b;
m
with
Ba; b Ca  Cb=Ca b; C
being the usual Gamma function:
In the context of such a model we can, for each set of
subjects performance on a set of sentence, determine by
maximum likelihood computation the best tting value
for the parameters a, b. Furthermore for the performances of a set of subjects on two set of sentences
(e.g., passive/active, or +Movement/
Movement) we
can test the null hypothesis that the whole data set
comes from one parametrized distribution against the
hypothesis that they come from two distinct distribution. Indeed Ignoring constants involving only the
observations, the log-likelihood for a set of results{(n1,
m1). . .(nk, mk)} is given by
Ll; hjfn1 ; m1 . . . nk ; mk g

m
j
1
k
X
X
j1

Logl r  h

r0

nj
m
Xj
1

Log1
l r  h

r0

m
j
1
X

!
Log1 r  h :

r0

Note that this likelihood depends on the full information concerning the number of trials and of successes of
each subject so that no concern over unduly pooling
incomparable percentages arises. If {(n1, m1). . .(nk,
mk)}, represents the performances of subjects on one
type of sentences and {{(nk + 1, mk + 1). . .(nk + q,
mk + q)}} their performances on an other type, we can
proceed to test the hypothesis that the performances
are similar using the fact that if L0 is the maximal value
of
Ll; hjfn1 ; m1 . . . nk ; mk ; nk1 ; mkq . . . nk1 ; mk1 qg
and L1 is the maximal value of
Ll; hjfn1 ; m1 . . . nk ; mk g
Ll; hjfnk1 ; mk1 . . . nkq ; mkq g;
then an asymptotically valid test of the similarity of performance for the two types of sentences is given by com-

128

D. Drai, Y. Grodzinsky / Brain and Language 96 (2006) 117128

paring 2(L1
L0) with upper percentage points of the c2

distribution with two degrees of freedom (see Williams,
1975).
See Fig. 2 for the density curves of the probability of
success corresponding to various sentence types, and
Table 1 for a summary of the signicance of the various
contrasts we have tested. This procedure was implemented in Mathematica, using the FindMinimum function on the negative log-likelihood in order to nd the
parameters of the b-distribution. We validated our
implementation by reproducing the numerical results
presented in Williams (1975).

References
Avrutin, S. (2001). Linguistics and agrammatism. GLOT International,
5, 311.
Ben-Shachar, M., Hendler, T., Kahn, I., Ben-Bashat, D., & Grodzinsky, Y. (2003). The neural reality of syntactic transformations:
Evidence from fMRI. Psychological Science, 14.5, 433440.
Berndt, R. S., Mitchum, C. C., & Haedinges, A. N. (1996). Comprehension of reversible sentences in agrammatism: A meta-analysis.
Cognition, 58, 289308.
Caplan, D., & Waters, G. (1999). Verbal working memory and
sentence comprehension. Behavioral and Brain Science, 22, 77126.
Caramazza, A. (1986). On drawing inferences about the structure of
normal cognitive processes from patterns of impaired performance:
The case for single-patient studies. Brain and Cognition, 5, 4166.
Caramazza, A., Capitani, E., Rey, A., & Berndt, R. (2000). Agrammatic Brocas aphasia is not associated with a single pattern of
comprehension performance. Brain and Language, 76, 158184.
Chomsky, N. (1995). The minimalist program. Cambridge, MA: MIT
Press.
Drai, D. (in press). Evaluating decit patterns of Broca aphasics in the
presence of high inter subject variability. In: Y. Grodzinsky,
K. Amunts (Eds.), Brocas Region. New York: Oxford University
Press, forthcoming.
Fanselow, G. (2001). Features, h-roles, and free constituent order.
Linguistic Inquiry, 32(3), 405437.
Friederici, A., & Graetz, P. (1987). Processing passive sentences in
aphasia: Decits and strategies. Brain and Language, 30,
93105.
Goodglass, H. (1968). Studies in the grammar of aphasics. In S.
Rosenberg & J. Koplin (Eds.), Developments in applied psycholinguistics research. New York: Macmillan.
Goodglass, H. (1993). Understanding Aphasia. San Diego: Academic
Press.
Grodzinsky, Y. (1984a). The syntactic characterization of agrammatism. Cognition, 16, 99120.

Grodzinsky, Y. (1984b). Language decits and linguistic theory.

Doctoral dissertation, Brandeis University.
Grodzinsky, Y. (1986). Language decits and the theory of syntax.
Brain & Language, 27, 135159.
Grodzinsky, Y. (2000). The neurology of syntax: Language use
without Brocas area. Behavioral and Brain Science, 23, 121.
Grodzinsky, Y. (2004). Variation in Brocas region: Preliminary crossmethodological comparisons. Chapter 8 in Lyle Jenkins, ed.,
Variation and Universals in Biolinguistics. Oxford: Elsevier.
Grodzinsky, Y. (2004a). Variation, canonicity and movement. Paper
presented at the EURESCO Science of Aphasia V Meeting,
Potsdam, September.
Grodzinsky, Y. (in press). A blueprint for a brain map of syntax. In
Grodzinsky and Amunts, eds.
Grodzinsky, Y., Pierce, A., & Marakovitz, S. (1991). Neuropsychological reasons for a transformational analysis of verbal passive.
Natural Language and Linguistic Theory, 9, 431453.
Just, M. A., Carpenter, P. A., Keller, T. A., Eddy, W. F., & Thulborn,
K. R. (1996). Brain activation modulated by sentence comprehension. Science, 274, 114116.
Kolk, H. (2004). Agrammatism. In R. Kent (Ed.), MIT encyclopedia of
communicative disorders. Cambridge, MA: MIT Press.
Kolk, H. H. J., & Heeschen, C. (1992). Agrammatism, Paragrammatism and the management of language. Language and Cognitive
Processes, 7, 82129.
Marantz, A. (1984). On the nature of grammatical relations. Cambridge, MA: MIT Press.
Mohr, J. P. (1978). Brocas area and Brocas aphasia. In H. Whitaker
(Ed.). Studies in Neurolinguistics (Vol. II, pp. 201235). San Diego:
Academic Press.
Roder, B., Stock, O., Neville, H., Bien, S., & Rosler, F. (2001). Brain
activation modulated by the comprehension of normal and
pseudo-word sentences of dierent processing demands: A functional magnetic resonance imaging study. NeuroImage, 15,
10031014.
Saito, M., & Fukui, N. (1998). Order in phrase structure and
movement. Linguistic Inquiry, 29, 439474.
Schwartz, M., Saran, E., & Marin, O. (1980). The word-order
problem in agrammatism: I. Comprehension. Brain & Language,
10, 249262.
Skellam, J. G. (1948). A probability distribution derived from the
binomial distribution by regarding the probability of success as
variable between the sets of trials. Journal of the Royal Statistical
Society Series B (Methodological), 10.2, 257261.
Wasow, T. (1977). Transformations and the Lexicon. In P. W.
Culicover, Thomas Wasow, & Adrian Akmajian (Eds.), Formal
Syntax (pp. 327360). New York: Academic Press.
Williams, D. A. (1975). The Analysis of binary responses from
toxicological experiments involving reproduction and teratogenicity. Biometrics, 31.4, 949952.
Zurif, E. B. (1995). Brain regions of relevance to syntactic processing.
In L. Gleitman & M. Liberman (Eds.) (2nd ed.). An invitation to
Cognitive Science (Vol. I.). Cambridge, MA: MIT Press.