Combining Images Across Multiple Subjects: A Study

of Direct Cortical Electrical Interference

Diana L. Miglioretti, Colin McCulloch, and Scott L. Zeger

This article introduces a Bayesian hierarchical model for combining information across multiple images. Our work was motivated by
an invasive functional brain mapping technique called direct cortical electrical interference that gives a sparse set of binary observations
of an underlying “true” region at multiple sites on the brain surface. To model region shapes that may vary widely across individuals,
we use mixtures of simple templates, for example, circles. These subject-speciŽ c templates are treated as random effects, governed by
a set of population templates that make up a population region. The numbers of subject-speciŽ c and population templates are treated
as unknown variables to be estimated from the data. Conditional on the subject-speciŽ c regions, the observed data are modeled using
logistic regression. To estimate the variability among images across patients, we develop a measure based on Baddeley’s error measure
for binary images. Because the dimension of the parameter space changes as the numbers of subject-speciŽ c and population templates
change, inference is made using reversible jump Markov chain Monte Carlo. Using a hierarchical approach, we may better estimate each
individual’s region by borrowing strength from other subjects’ data, we can estimate a population region by pooling information across
subjects, and we can use a collection of data from previous patients to predict the location of a future patient’s region of interest. The
approach is illustrated with DCE I data collected on 20 patients for two motor tasks: tongue and hand movements.
KEY WORDS: Binary image variation; Functional brain mapping; Hierarchical model; Reversible jump Markov chain Monte Carlo;
Template mixture model.

1. INTRODUCTION 1997; Lee 1997). Biostatistics now faces a new challenge to

When considering the results of a screening test, a physi-
cian rarely relies on an individual’s test results alone. He or
she also considers the sensitivity and speciŽ city of the test
and prevalence of the disease among similar people. Simi-
larly, when an experienced neurosurgeon considers resection
surgery to remove a seizure focus or brain tumor, he or she
considers his or her prior knowledge obtained from previous
patients along with any functional brain mapping results from
that speciŽ c patient. Thus, when analyzing functional brain
mapping data, it is natural to borrow strength across patients
to estimate an individual’s brain map. In addition, combining
information across patients is important for making popula-
tion inferences and predictions about future patients who have
not yet undergone testing. Hierarchical models are a natural
approach for meeting these goals.
The central idea in hierarchical modeling is to acknowl-
edge multiple levels of variability in data. In population and
biomedical research, such modeling can account for variation
among individuals in their true characteristics as well as varia-
tion due to imprecise measurement. Hierarchical models, also
called random-effects, mixed-effects, or multilevel models, are
well established for univariate continuous measures (e.g., Rao
1965; Lindley and Smith 1972; Laird and Ware 1982), binary
responses (e.g., Stiratelli, Laird, and Ware 1984; Zeger and
Karim 1991), and count data (e.g., Christiansen and Morris
Diana L. Miglioretti is Assistant ScientiŽ c Investigator, Center for Health
Studies, Group Health Cooperative of Puget Sound, Seattle, WA 98101
(E-mail: miglioretti.d@ghc.org). Colin McCulloch is Senior Statistician,
Applied Statistics Program, General Electric Corporate Research & Devel-
opment, Schenectady, NY 12301 (E-mail: mcculloc@crd.ge.com). Scott L.
Zeger is Professor and Chair, Department of Biostatistics, Johns Hop-
kins University, Baltimore, MD 21205 (E-mail: szeger@jhsph.edu). The
authors thank the associate editor for valuable comment s and Dana Boat-
man, John Hart, and Barry Gordon from the Department of Neurology,
Johns Hopkins University, for introducing us to this challenging prob-
lem and for providing the DCE I data along with helpful comments. This
work was partially supported by the National Institutes of Aging train-
ing grant T32 AG00247-01A 1 and the National Institutes of Health grant
R01-MH56639 . The program used for the example may be found at
http://www.centerforhealthstudies.org /perpages/migliore/index.html.
125
develop hierarchical models for images and other complex
measures of health that are beginning to supplant traditionally
used scalar measures in biomedical research.
In this article we introduce a hierarchical approach for ana-
lyzing multiple binary images that is an extension of the tem-
plate mixture model previously introduced for a single image
(Miglioretti, McCulloch, and Zeger 2000). Our work was
motivated by a speciŽ c type of functional brain mapping data
called direct cortical electrical interference (DCE I), reviewed
later; however, we seek a general approach that can be used
for other brain mapping and spatial data. The goal of DCE I
is to determine the cortical region in the brain that is nec-
essary for a speciŽ c task. The underlying brain map may be
thought of as a binary image, where each location in the brain
is either necessary for the task or not. DCE I results in a sparse
set of binary measurements of this brain map at multiple sites
throughout the brain. The inferential goal is to estimate each
individual’s image (i.e., brain map), along with a population
image with speciŽ c interest in the locations and sizes of the
region of interest (the area necessary for the task under study).
1.1 Direct Cortical Electrical Interference
DCE I, also known as cortical stimulation, is an invasive
functional brain mapping method used in surgical resection
candidates. Most patients receiving DCE I have medically
intractable epilepsy. DCE I is used to determine if the parts
of the brain causing the seizures can be safely resected with-
out damaging regions necessary for important daily func-
tions, such as motor control and language. For extraoperative
DCE I, arrays of electrodes are subdurally implanted into the
brain surface and left in the patient’s brain for several days
to a week or more (Lesser et al. 1987; Gordon, Hart, Boat-
man, and Lesser 1996). During DCE I testing, small temporary
functional “lesions” are created with a brief electrical current
© 2002 American Statistical Association
Journal of the American Statistical Association
March 2002, Vol. 97, No. 457, Applications and Case Studies
126
( 5 sec) passed between two electrodes about 1 cm apart
(Gordon et al. 1996). During this time, the patient is asked
to perform a task, such as name an object shown on a com-
puter screen or move his or her Ž ngers. By comparing the
patient’s performance with and without electrical interference,
the clinician makes a judgement as to whether the task has
been temporarily impaired by current applied to this electrode
pair. This testing results in binary data (impaired or intact) at
multiple sites throughout the brain. Examples of DCE I data
from two motor tasks are displayed later (see Figs. 4 and 5).
Because the electrical current during DCE I testing is dis-
tributed to a region beneath and around the electrode pair,
cortical processing is affected in an area rather than just at a
point location. It is important to take this into account during
analysis, because associating the observed value with a sim-
ple point location would result in overlocalizing the region of
interest—namely, underestimating the areas in the brain nec-
essary for the task. There is also evidence that the probability
of producing a deŽ cit during interference is a function of the
strength of the current. Under this assumption, DCE I would
be less likely to cause a task impairment if the region of inter-
est were located farther away from the electrode site where
the current strength is weaker.
Nathan, Sinha, Gordon, Lesser, and Thakor (1993) studied
the distribution of the electrical current using Ž nite element
analysis. They found that the current density peaks immedi-
ately beneath the electrodes and declines rapidly to almost
no current approximately .5 cm away from the electrode pair.
Just below the cortical surface, the current midway between
the two electrodes is about half the strength of the peak cur-
rent. We closely approximate the electrical current distribution
given by Nathan et al. (1993) using a mixture of three bivari-
ate Gaussian densities (Fig. 1). Two of the Gaussian densities
are circular and centered at each electrode in the site. The third
density is elliptical and centered at the midpoint between the
two electrodes with orientation equal to that of the electrode
pair.
Methods for analyzing DCE I data in multiple subjects have
been mostly limited to displaying the locations of tested sites
Figure 1. Approximation of the Current Density Beneath the Elec-
trode Site. The two electrodes are located at ( x, y ) D ( .5, 1) and
( 1.5, 1) .
Journal of the American Statistical Association, March 2002
and sites found to be associated with the task under study, typi-
cally after registering each brain to a common atlas (e.g., Boat-
man, Lesser, Hall, and Gordon 1994; Boatman, Lesser, and
Gordon 1995; Boatman, Hall, Goldstein, Lesser, and Gordon
1997; Gordon et al. 1996; Nii, Uematsu, Lesser, and Gordon
1996). In addition, the brain atlas may be divided into arbitrary
zones and the proportion of associated sites within each zone
calculated across subjects (e.g., Malow et al. 1996; Ojemann,
Ojemann, Lettich, and Berger 1989; Uematsu et al. 1992,
Urasaki, Uematsu, Gordon, and Lesser 1994). In this article,
we introduce a new approach for analyzing DCE I data that
overcomes the obvious limitations of these previously used
methods. We use a hierarchical approach that borrows strength
across subjects to estimate a brain map for each individual
and combines information across subjects to estimate a single
population brain map.
For DCE I, there is little prior information about the shape,
location, and size of the region of interest and limited spa-
tial resolution in the data. In addition, the region shapes may
vary widely across individuals. Therefore, we require a  exi-
ble method that can model a wide range of region shapes using
a small number of parameters that may be pooled across indi-
viduals. To meet these requirements, we model the individual
and population regions of interest using mixtures of subject-
speciŽ c and population templates of simple geometrical shapes
(e.g., circles). We treat the subject-speciŽ c templates as ran-
dom effects, governed by the population templates. We treat
the numbers of population and subject-speciŽ c templates as
unknown variables to be estimated from the data. Because the
dimension of the parameter space depends on the numbers of
templates, we make inference using reversible jump Markov
chain Monte Carlo (RJMCMC). Given the data, we summa-
rize our region analysis by producing, for each individual and
the population, a pixelwise map of the posterior probability
that each location in the image is contained in the region of
interest. These probabilities are implied by the posterior dis-
tribution on the template parameters, which is marginalized
over the numbers of templates. Under this marginalization, the
region estimates are not restricted to be a union of circles, but
instead can take on a wide range of shapes. We formulate the
model for the observed data using a generalized linear model
framework, so the method can be applied to varied types of
measurements.
In Section 2 we develop the template mixture model for
DCE I with multiple subjects. In Section 3 we describe the
algorithm used for model Ž tting and inference. We introduce
a measure of image variability in Section 4. In Section 5 we
present an example using data from DCE I studies conducted
at the Johns Hopkins Medical Institutions. We end with a dis-
cussion of our methodology.
2. THE TEMPLATE MIXTURE MODEL
The template mixture model for multiple subjects consists
of three main parts: (1) a set of prior assumptions about the
true image (brain map) for the population and (2) for each
individual, and (3) a speciŽ cation of the probability distribu-
tion of the observations given the true images. In the follow-
ing model formulation, we indicate subject-speciŽ c parameters
with subscript i and population parameters with superscript ü .
Miglioretti, McCulloch, and Zeger: Direct Cortical Electrical Interference
We let u denote a two-dimensional location on the surface
of the brain. The true image for the ith individual fi 4u5,
i D 11 : : : 1 N , and the population image f ü 4u5 are assumed to
be binary such that
( parameters Œü1k and Œü2k , and Ž3k corresponds to the template
1 if u 2 Si
fi 4u5 D
0 otherwise
and
( formly over an a priori known reasonable range of values
ü 1 if u 2 S ü
f 4u5 D
0 otherwise1
where Si and S ü are the subject-speciŽ c and population regions
of interest. Thus each location on the cortex is assumed to be
either needed for the task under study or not, and the regions
of interest are the areas needed for the task.
We model the subject-speciŽ c regions of interest Si as mix-
tures of Ki simple geometrical shapes, or templates, Tik 2 Si D
SKi
kD1 Tik . We use circular templates Tik D 8u2 —u ƒ Ìik — < ik 9,
where Ìik is the two-dimensional location parameter indicat-
ing the center of the circle, ik is its radius, and —¢— repre-
sents Euclidean distance. The numbers of templates for each
individual K1 1 : : : 1 KN are taken to be unknown and possi-
bly different. Each template Tik is assumed to come from
a distribution centered near one of K ü population templates,
which are also taken to be circular with parameters Ìük and
ük , k D 11 : : : 1 K ü . These population templates make up the
S ü
population region of interest, S ü D KkD1 Tkü .
The full hierarchical model is given in Figure 2. The num-
ber of population templates K ü is assumed to follow a Pois-
son distribution with Ž xed mean parameter ‹ü . Each popula-
Figure 2. Full Hierarchical Model.
127
tion template has six associated parameters: a two-dimensional
location Ìük , a radius ük , and three precisions, Ž1k , Ž2k , and
Ž3k , used to deŽ ne individual distributions in the next level of
the hierarchy. Here Ž1k and Ž2k correspond to the two location
radii ük . The population template locations Ìü1 1 : : : 1 ÌüK ü are
independently distributed uniformly over the image space A,
and the radii ü1 1 : : : 1 üK ü are independently distributed uni-
4min 1 max 5, deŽ ned in Section 5. The distributions for the
population precisions 8Žp1 1 : : : 1 ŽpK ü 3 p D 11 21 39 are assumed
to be independent gammas. We do not constrain the precisions
for the two location parameters to be equal, facilitating extra
 exibility in the template parameter variability.
Given the population parameters, the model for the subject-
speciŽ c regions is deŽ ned as follows. We model the number of
individual templates K1 1 : : : 1 KN as independent Poisson4‹5
variables truncated to Ki µ K ü , where the mean number of
templates ‹ follows a gamma distribution. We make the con-
straint that each individual can have at most one template
associated with each of the population templates. For the ith
individual, the labels indicating to which population template
each individual template corresponds are considered latent
allocation variables zi1 1 : : : 1 ziKi , which are distributed uni-
formly on the permutations of Ki elements chosen from K ü
elements. The individual template two-dimensional locations
Ì i1 1 : : : 1 ÌiKi and radii i1 1 : : : 1 iKi are independently dis-
tributed as Gaussians with population means 4Ìü 1 Ðü 5 and
precisions 4Î5 determined by the allocation vector zi and
constrained to be in the same range as the population param-
eter space.
128
The third part of the model describes the observed data,
which we assume follow a generalized linear model. Let yij
represent the binary observation at the jth testing site of the
ith individual, j D 11 : : : 1 Ji 1 i D 11 : : : 1 N . Thus yij equals
1 if the task is impaired during DCE I and 0 if the task is
functional. Let xij D 4xij 1 1 xij2 5 be the pair of two-dimensional
coordinates for the two electrodes in the jth testing site for
individual i. Given the subject-speciŽ c parameters, we model
the observed values using logistic regression,
yij —fi Bernoulli4 ij 51
and
logit4 ij 5 D  0 C  1 H 4fi 1 xij 51
R the ith individual’s brain map, marginalizing over the num-
where H 4fi 1 xij 5 D u2²2 I4—u ƒ Ìik — < ik for any k D
11 : : : 1 Ki 5Dxij 4u5 du. Here I 4¢5 is the indicator function and
Dxij 4u5 is the electrical current distribution centered around
xij1 and xij 2 , assumed known. Thus H 4fi 1 xij 5 is a measure of
the amount of current at the jth electrode site that overlaps
the region of interest Si , constrained to be a value between
0 and 1. We estimate H 4fi 1 xij 5 using Monte Carlo integra-
tion by simulating J locations dij from the current distribution
Dxi 4u5 and setting H b4fi 1 xij 5 equal to the proportion of these
locations that fall within any of the circular templates.
The regression parameters  0 and  1 in (1) describe the
error rates for the task under study. SpeciŽ cally,  0 is the log
odds of observing an impairment during DCE I testing given
that the electrical current does not overlap the region needed
for the task. The slope parameter  1 describes the change in
log odds of an impairment as the amount of current overlap-
ping the region of interest increases. The prior distributions
for the regression coefŽ cients are taken to be Gaussian for  0
and  1 .
Observed values yij are assumed to be conditionally inde-
pendent given the true underlying image fi . It follows that the
likelihood is simply
N Ji
Y
N
Y Y
p4yi —Ìi 1 Ði 1 Ki 1 Á5 µ
y
 ijij 41 ƒ  ij 51ƒyij 1
iD1 iD1 jD1
where logit4 ij 5 D  0 C 1 H 4fi 1 xij 5. The posterior distribution
is written as
p4K ü 1 ‹1 K1 z1 Ìü 1 Ðü 1 Î1 Ì1 Ð1 Á—y5
D Z ƒ1 p4Á5p4K ü 5p4‹5p4Ìü 1 Ðü —K ü 5p4ΗK ü 5
" #
Y
N
€ ü ü ü ü
p4Ki —K 5p4zi —Ki 1 K 5p4Ìi 1 Ði —Ki 1 zi 1 Ì 1 Ð 1 Î5
iD1
N Ji
Y Y
€ p4yij —Ìi 1 Ði 1 Ki 1 Á51
iD1 jD1
where Z is an unknown normalizing constant. The Ž rst line of
(3) contains all hyperparameters and population distributions,
the second line gives the individual distributions, and the third
line is the likelihood component.
Journal of the American Statistical Association, March 2002
3. MODEL FITTING
If we assume that the numbers of individual and population
templates are unknown, then the dimension of the parameter
space changes as the number of templates changes. One  ex-
ible technique for estimating parameters in Bayesian models
of varying dimension is the RJMCMC methodology proposed
by Green (1995). One can explore the combined model space
by setting up a hybrid sampler (Tierney 1994) that proposes
to update the parameters within a Ž xed-dimensional subspace
in one type of move and proposes to jump between models of
different dimensions in another type of move. In this way, we
use RJMCMC to simulate from the joint posterior distribu-
tion deŽ ned in (3). An estimate of each subject-speciŽ c region
(1)
of interest Si may be calculated using the posterior mean of
ber of templates Ki and population parameters, p4u 2 Si —y5 ²
E4fi 4u5—y5. An estimate of the population region of interest
S ü may be calculated similarly as p4u 2 S ü —y5 ² E4f ü 4u5—y5.
We estimate E4fi 4u5—y5 by the usual Monte Carlo estimate,
" Kit
#
O 1 XT
[
fi 4u5 D I u2 Tik 4Ìikt 1 ikt 5 1 (4)
T ƒ t0 tDt0 C1 kD1
where Ti1 4Ìi1t 1 i1t 51 : : : 1 TiKit 4ÌiKit t 1 iKit t 5 are the templates
from the tth realization of the Markov chain, T is the total
number of iterations, and t0 is the number of burn-in iterations.
The population posterior mean E4f ü 4u5—y5 may be estimated
similarly, replacing the subject-speciŽ c parameters in (4) with
the corresponding population parameters. These estimates are
preferred over a measure of the region with maximum a pos-
teriori density, because they result in a smoothed estimate of
the region of interest, as opposed to a union of circles.
For the multiple-subject template mixture model, we set up
a hybrid sampler with six steps: (1) update the population
template parameters and precisions, (2) update the mean num-
ber of individual templates, (3) update the individual template
parameters, (4) update the regression parameters, (5) add or
remove a template for each individual, and (6) birth or death of
(2) a population template and associated individual templates. For
update steps 1–4, each parameter is updated in turn by condi-
tioning on all other parameters in the style of Gibbs sampling.
Because we cannot sample directly from the full conditional
distributions, we use Metropolis–Hastings steps. For step 1,
to update the population template locations and correspond-
ing precisions, we take the candidate distributions to be the
full Gaussian and gamma conditional distributions under the
assumption of no truncation in the subject-speciŽ c parameter
distributions. These distributions closely approximate the full
conditionals if the distributions of the subject-speciŽ c template
parameters are not highly skewed (i.e., if there is minimal
truncation). Because there is likely to be signiŽ cant trunca-
tion in the subject-speciŽ c radii distributions given the prior
(3)
range speciŽ ed in Section 5, we use a random walk Metropo-
lis sampler (Gilks, Richardson, and Spiegelhalter 1996) for
the population template radii and corresponding precisions.
Similarly, for step 2, we also use a random walk sampler to
update the mean number of individual templates. For steps 3
and 4, to update the subject-speciŽ c template parameters and
Miglioretti, McCulloch, and Zeger: Direct Cortical Electrical Interference
the regression parameters, we use grid-based candidate distri-
butions (Tierney 1994), which we found much more efŽ cient
than a generic proposal distribution.
To explore the combined parameter space, we include two
birth/death steps designed to allow jumps between models
with different numbers of individual and population templates.
For the individual birth/death step, performed separately for
each individual, we attempt to add or remove one subject-
speciŽ c template. For each individual, a random choice is
made between proposing the birth or death of a template with
probabilities bKi and dKi D 1 ƒ bKi . We take bKi D dKi D 05.
Appendix A gives a detailed exposition of the algorithm along
with derivations of the acceptance probabilities. Brie y, for
the ith individual’s birth step, a population template is selected
from those for which that individual does not have an associ-
ated template. A new set of individual template parameters is
drawn from their conditional distributions given the selected
population template. The new template is accepted with prob-
ability 4Ki 1 Ki C 15 D min811 A9, where
‹ p4yi —Ki C 11 Ìi 1 Ði 1 Ì0i 1 Ð0i 1 Á5
AD (5)
Ki C 1 p4yi —Ki 1 Ìi 1 Ði 1 Á5
and 4Ì0i Ð0i 5 is the vector of proposed new template parameters.
If the number of individual templates Ki equals the number of
population templates K ü at that sweep, then nothing is done.
For the reverse (death) step, an individual template to delete
is randomly chosen. If individual i has no templates at that
sweep, then nothing is done. The individual death step mov-
ing from Ki C 1 to Ki templates is accepted with probability
4Ki C 11 Ki 5 D min811 Aƒ1 9, where A is deŽ ned in (5).
For the population birth/death step, we propose to add or
delete a population template and associated individual tem-
plates. A random choice is made between the birth and death
steps with probabilities bK ü and dK ü D 1 ƒ bK ü , where we take
bK ü D dK ü D 05. Appendix B outlines the algorithm in detail.
For the birth step, a new set of population template parameters
is selected from their prior distributions. A random number
of individuals are chosen for which new subject-speciŽ c tem-
plates associated with the new population template are pro-
posed (with the possibility of no individuals being selected).
New template parameters for these individuals are drawn
from their conditional distribution given the new population
template. The new population and individual templates are
accepted with probability 4K ü 1 K ü C 15 D min811 LR € B9,
where LR is the likelihood ratio,
" #" #
‹ü Y ‹ YN
p4ki µ K ü —K ü 5
BD ü
K C 1 i2I Ki C 1 iD1 p4ki µ K ü C 1—K ü C 15
Q
iyI 4K ƒ K i C 15
ü the patient’s cortical surface so they cover the expected loca-
€ Q 1 (6)
4K ü C 15N i2I 4Ki C 15
and I is the set of individuals for which a new template is
proposed. For the reverse (death) step, a population template
and associated individual templates to delete are randomly
chosen. The acceptance probability for the death step mov-
ing from K ü C 1 to K ü population templates 4K ü C 11 K ü 5 is
min811 LR € B ƒ1 9, where B is deŽ ned in (6).
129
4. IMAGE VARIATION
In functional brain mapping, the cortical regions associated
with some tasks are more variable than the regions associ-
ated with other tasks. A quantitative measure of this vari-
ability would be useful for describing how the locations of
regions needed for different functions vary within populations.
For example, we might expect language regions to be more
variable in epilepsy patients with seizure foci in the temporal
lobe if language functions are displaced to areas outside this
region due to the disease (see, e.g., Ojemann and Whitaker
1978). Measuring variability in this case is not straightforward,
because each individual’s dataset is an image as opposed to
one or even a few numbers. In this section we propose a plau-
sible measure of variability in a set of binary images.
Following Baddeley (1992), we deŽ ne d4x1 S5 as the short-
est distance between a pixel in an image x 2 X and a region
in the image A 2 X2 d4x1 S5 D min84x1 a52 a 2 A9, where
4x1 a5 is some measure of distance, such as the Euclidean
distance, between the pixel and the point a. We propose the
following estimate of variation among N images:
V 4S1 1 : : : 1 SN 5
1 X
D var6d4x1 S1 51 : : : 1 d4x1 SN 57
M x2X
³ ´ N
1 X 1 N ƒ1 X X
D 6d4x1 Si 5 ƒ d4x1 Sj 572 0
M x2X 2 2 iD1 j>i
The term inside the large braces is motivated by an unbiased
estimate of the variance of d4x1 Si 5, because the expectation
of the squared difference between two independent and iden-
tically distributed random variables y1 and y2 is an estimate
of twice the variance of yi . Thus V 4Si 5 is a measure of the
average variance of the distance of a pixel x to Si , averaged
over all M pixels in the image X. This estimator is closely
related to Baddeley’s ã2 (Baddeley 1992), which is deŽ ned as
the square root of the average squared difference between the
shortest distance from a pixel to a region in one image and the
distance from the same pixel to a region in a second image,
averaging over all pixels in the image.
5. EXAMPLE
At the Johns Hopkins Medical Institutions, DCE I is used
for functional brain mapping in surgical resection candidates
to identify cortical areas critical for motor and language func-
tions to avoid postresection deŽ cits (Gordon et al. 1996). Elec-
trodes are arranged in rectangular arrays composed of 2.3-mm
diameter exposed platinum iridium disks spaced 1 cm apart
(Lesser et al. 1987). These arrays are surgically placed into
tion of the seizure focus to be resected and surrounding areas.
(For details on the clinical DCE I testing protocol used, see
Lesser et al. 1987 and Uematsu et al. 1990.)
To demonstrate our methodology, we use the template mix-
ture model to estimate the regions associated with movement
of the tongue and hands during DCE I testing in 20 patients.
The raw data are displayed in Figures 4 and 5. Black dots
with lines represent electrodes associated with a positive motor
130
response during DCE I testing (i.e., yij D 15. Gray lines rep-
resent tested electrode sites with no motor response (i.e.,
yij D 05. During DCE I testing, a patient is asked to stick out
and wiggle the tongue or to hold out the hands and wiggle
the Ž ngers. A positive motor response is deŽ ned as either
(1) an induced motor response, in which the tongue or hand
jerks or twitches during DCE I, or (2) motor arrest, in which
movement of the tongue or Ž ngers stops during DCE I. Before
analysis, each individual’s electrode locations were registered
to the Talairach atlas (Talairach and Tournoux 1988). Regis-
tration was performed manually by comparing the electrode
locations to structural features of the brain using photographs
taken at surgery, skull x-rays, and a three-dimensional com-
puted tomography (CT) scan of the electrodes superimposed
onto a three-dimensional brain reconstruction from volumetric
magnetic resonance imaging (MR I) scans.
Because there is only one observation per electrode site per
patient, there is little information to estimate the regression
coefŽ cients (see Miglioretti et al. 2000). However, there is
strong prior information about these parameters. The data con-
sist of the binary clinical judgement as to whether there was
a positive motor response at each electrode site, which appear
to have very low error rates, because any uncertainty about the
results at a particular site simply prompts continued testing
until a conŽ dent decision can be made. In addition, it is known
that a small overlap of the electrical current with the corti-
cal area responsible for motor control should cause a positive
motor response. Thus the probability of a positive response
should be very close to 0 when H 4fi 1 xij 5 ’ 0 (i.e., the elec-
trical current does not overlap with the region of interest) and
very close to 1 when even a small portion of the current den-
sity is over the region of interest, say when H 4fi 1 xij 5 > 025.
In addition, the probability should increase quickly from 0
to 1. There is some uncertainty about the regression coefŽ -
cients, because it is not known how much current overlap is
required to cause an impairment, although it is likely that the
amount of overlap needed is small. To take this uncertainty
into account, we chose the priors for  0 and  1 to include a
range of coefŽ cients that meet the aforementioned conditions.
We set the prior for  0 to Gaussian with mean ƒ12 and stan-
dard deviation .5 and the prior for  1 to Gaussian with mean
70 and standard deviation 15. Figure 3 shows the associated
uncertainty in the probability of impairment as a function of
the amount of current overlapping the region.
The prior for K ü was taken to be Poisson(3), and the prior
for ‹ was gamma 41051 055 to encourage a small number of
templates. We set the prior range for Ìük to be the range of
the electrode grids. The prior for kü was taken to be uniform
from .1 to 1 cm, with the maximum being the typical spac-
ing between two electrodes. The limits were chosen to keep
the circles reasonably sized to better approximate the regions.
Given the goals of analysis, circles more than 1 or 2 cm apart
should come from different distributions. Taking this into con-
sideration, we chose the prior parameters for Î as follows.
We took the precisions for Ο1 and Ο2 to be gamma 451 35,
to restrict circles from the same population to be reasonably
close to each other, and we took the precisions for ü to
be gamma 451 15. The sampler was programmed in SAS/ IML
(SAS Institute, Inc. 1990). We implemented each model for
Journal of the American Statistical Association, March 2002
Figure 3. Two Hundred Fifty Samples from  i j Induced by the Prior
Distribution on (  0 ,  1 ) . The prior for  0 is Gaussian with mean - 12
and standard deviation .5, and the prior for  1 is Gaussian with mean
70 and standard deviation 15.
40,000 MCMC sweeps, discarded the Ž rst 20,000 sweeps, and
saved every Ž fth sweep for the results.
To assess convergence, we ran multiple chains starting at
different initial values and noted no signiŽ cant difference in
the results. Acceptance rates were between 22% and 65% for
the individual birth/death step and between 7% and 9% for
population birth/death step. The mean number of population
templates is 8.0 (range 4–16) for tongue movements and 6.4
(range 4–14) for hand movements. The mode of the mean
number of individual templates ‹ is 7.9 for tongue movements
[95% highest posterior density interval (HPD) D 406–14.3] and
4.9 for hand movements (95% HPD D 2.6–11.2).
We examined model Ž t by comparing the observed numbers
of electrode sites associated with impairments for each indi-
vidual to the expected frequencies simulated from the poste-
rior predictive distributions for the two motor tasks (Gelman,
Carlin, Stern, and Rubin 1995). In addition, we overlaid spa-
tial plots of the observed data with the posterior probability
contours for each individual (Figs. 4 and 5). Both methods
indicate no lack of Ž t of the model.
Figures 4 and 5 show the individual posterior probability
contours for the regions associated with tongue movements
and hand movements. Positive tongue responses associated
with DCE I testing were found in 19 of the 20 patients. In
the patient without an observed tongue response (patient 20),
it appears that the electrode grid was likely implanted below
the region associated with tongue movement. Hand responses
were observed in 14 of the 20 patients during DCE I test-
ing, and it appears again that the grids were likely implanted
below the areas associated with hand movement in the other
patients. This is consistent with the literature, because the
cortical regions associated with hand movements are consis-
tently shown to be located superior and slightly posterior to
those for tongue movements (Crone, Miglioretti, Gordon, and
Lesser 1998a; Crone et al. 1998b; PenŽ eld and Boldrey 1937;
Pfurtscheller, Flotzinger, and Neuper 1994; Uematsu et al.
1992). The results also agree with recent studies showing that
the areas associated with the hand and the tongue overlap and
are diffusely spread throughout the sensorimotor cortex (Crone
Miglioretti, McCulloch, and Zeger: Direct Cortical Electrical Interference 131

Figure 4. Individual Posterior Probability Contours for the Regions Associated With Tongue Movement in 20 Patients. Black dots with lines
represent electrodes associated with a positive motor response during DCEI testing. Gray lines represent tested sites with no motor response.
From lightest to darkest, posterior contours correspond to .05 µ p < .10, .10 µ p < .20, and p > .20.

et al. 1998a; Pascual-Leone, Cohen, Brasil-Neto, and Hallett
1994; Schlaug, Knorr, and Seitz 1994).
There are several interesting items to note about the
individual posterior contours. First, several patients had no
observed positive motor responses during DCE I testing. By
borrowing strength across patients, the template mixture
model gives a coherent estimate of the areas associated with
hand and tongue movements. If information from the other
subjects had been ignored, then the data would indicate that
these individuals had no area associated with tongue and hand
movements. In addition, most subjects have areas estimated to
be associated with both hand and tongue movements outside
the testing area. If the information from the other patients had
been ignored, then the area associated with hand and tongue
movements would be smaller in these patients. Finally, in
many patients, there is shrinkage of the area estimates toward
the population mean. For example, for patients 1, 2, 6, and 7
for the tongue task and patients 1, 3, 5, and 7 for the hand
task, there are electrode pairs where the posterior probability
is more concentrated near one electrode–the one closest to the
population region.
The next set of Ž gures highlights the richness of results
that we can obtain from this type of analysis, because we can
calculate posterior distributions on any function of the model

Figure 5. Individual Posterior Probability Contours for the Regions Associated With Hand Movement in 20 patients. Black dots with lines
represent electrodes associated with a positive motor response during DCEI testing. Gray lines represent tested sites with no motor response.
From lightest to darkest, posterior contours correspond to .05 µ p < .10, .10 µ p < .20, and p > .20.
132
Figure 6. Posterior Densities for the Individual Region Sizes for
Hand (Black) and Tongue (Gray) Movements. The solid lines corre-
spond to posterior modes, and the dashed lines correspond to the 95%
highest posterior density intervals.
parameters. For example, the region size can be estimated at
each sweep by calculating the size of the area covered by at
least one template. The variability can be calculated using our
measure deŽ ned in Section 4.
Figure 6 shows the posterior densities for the average
individual region sizes for hand (black) and tongue (gray)
movements. The solid lines correspond to the modes, and
the dashed lines correspond to the 95% HPD intervals. The
modal region size for hand movements in the patients stud-
ied is 4.4 cm2 (95% HPD D 301–7.0 cm2 ). The region size
for tongue movements is 50% larger (mode D 607 cm2 , 95%
HPD D 5.2–8.8 cm2 ).
There is some evidence that the region associated with
tongue movement is larger than the region associated with
hand movement in these patients (posterior probability D 092,
posterior mode for difference D 200 cm2 , 95% HPD D ƒ083–
4.7 cm2 ). However, the difference could be due to the fact that
the cortical area associated with tongue movements is inferior
to that for hand movements, and the electrode grids in these
patients tended to cover more of the inferior portion of the
sensorimotor cortex.
The posterior densities for the intersubject image variation
for the two tasks is displayed in Figure 7. In this study popu-
lation, the region associated with hand movements is slightly
more variable than the region associated with tongue move-
ments (posterior probability D 074). The image variability for
the regions associated with the hand is 1.5 (95% HPD D 071–
3.1), whereas that for the tongue is 1.3 (95% HPD D 055–2.0).
The posterior contours for the population regions associated
with tongue and hand movements are displayed in Figure 8.
Across the population, there is considerable overlap between
the regions associated with hand and tongue movements; how-
ever, the cortical area associated with hand movements tends
to be superior and posterior to tongue movements.
Figure 9 shows the posterior densities for the popula-
tion region sizes for the hand (black) and tongue (gray)
movements. The population region associated with tongue
movements (mode D 905 cm2 , 95% HPD D 502–13.5 cm2 ) is
slightly larger than the region associated with hand movements
(mode D 701 cm2 , 95% HPD D 302–12.3 cm2 ); however, the
Journal of the American Statistical Association, March 2002
Figure 7. Posterior Densities for the Intersubject Image Variation for
Tongue (Gray) and Hand (Black) Movements. The solid lines corre-
spond to the posterior mode, and the dashed lines correspond to the
95% HPD intervals.
population posterior distributions overlap more than the dis-
tributions for the individual region sizes. Although the indi-
vidual hand regions are smaller than the tongue regions, there
is some evidence of greater variability in the regions associ-
ated with hand movements in this set of patients (see Fig. 7).
This combination results in population regions that are more
similar in size.
6. DISCUSSION
In this article, we have extended the template mixture model
proposed in an earlier article (Miglioretti et al. 2000) to mul-
tiple subjects by embedding the single-subject model into a
larger hierarchical model with a population component. We
model the individual and population regions of interest using
a mixture of subject-speciŽ c and population templates. The
subject-speciŽ c templates are treated as random effects, gov-
erned by the population templates. To model the observed
data, we use a generalized linear model. SpeciŽ cally, for mod-
eling binary DCE I data, we use logistic regression. Because
the dimension of the parameter space depends on the num-
bers of templates, we Ž t the model using the reversible jump
methodology of Green (1995).
Figure 8. Population Posterior Probability Contours for the Regions
Associated With Tongue Movement and Hand Movement in the Study
Population. Black dots with lines represent electrodes associated with
a positive motor response during DCEI testing in 20 patients. Gray lines
represent electrodes with no motor response. From lightest to dark-
est, posterior contours correspond to .05 µ p < .10, .10 µ p < .20, and
p > .20.
Miglioretti, McCulloch, and Zeger: Direct Cortical Electrical Interference 133

DCE I analysis, because the length of time an electrode grid

Figure 9. Posterior Densities for the Population Region Sizes for
Hand (Black) and Tongue (Gray) Movements in 20 Patients. The solid
lines correspond to posterior modes, and the dashed lines correspond
to the 95% HPD intervals.
To estimate each individual image and the population
image, we use an estimate of the posterior mean of the region.
Because these estimates are marginalized over the number of
templates, the region shapes are not restricted to be a union of
circles. Thus this method is useful when the analytic goal is to
model binary regions using a small number of parameters, but
there is minimal information about the shapes and locations of
the areas of interest in the underlying image. The model’s par-
simony is essential for combining information across subjects.
By using mixtures of subject-speciŽ c templates, this approach
can model a wide range of shapes across subjects.
For DCE I and other functional brain mapping techniques,
including multiple subjects in a single hierarchical model has
several advantages. First, we may better estimate an individ-
ual’s brain map by borrowing strength from other subjects’
data. DCE I testing for a single subject is limited by both the
location of the electrode grid and the fact that all electrode
pairs are not tested. By using information from other patients,
we can make reasonable estimates about the areas associated
with the task under study outside the testing area in that spe-
ciŽ c patient. This was demonstrated in our example, where
the regions associated with hand and tongue movements were
at least partially outside the testing area for many subjects.
A second advantage of a hierarchical approach is that infor-
mation from past patients can be used to predict where to
test a new patient. This is becoming an important objective of
can be left in a patient for testing is limited by cost concerns,
greatly limiting the amount of research testing that is possible.
The RJMCMC method allows us to avoid prespecifying the
numbers of templates used in the model. More templates are
automatically Ž t for individuals with larger regions, and fewer
templates Ž t for those with smaller regions. Without this  ex-
ibility, if the number of templates were Ž xed to be too small,
then regions of interest would not be modeled adequately, and
if the number of templates were set too large, then the addi-
tional templates not needed to model the region either would
be distributed throughout the region where there is no infor-
mation (i.e., where there are no electrodes) or would overlap
with other templates. By modeling the joint posterior distri-
bution for the numbers of individual and population templates
and the parameters, and by using a Poisson prior for the num-
bers of templates with a low mean, RJMCMC discourages the
addition of templates that do not increase the likelihood. Thus
a parsimonious model is automatically encouraged.
When implementing the RJMCMC methodology, develop-
ment of the dimension-changing move types requires some
care to ensure that the model mixes well across the differ-
ent number of templates. In their article on modeling mix-
tures with an unknown number of components, Richardson
and Green (1997) implemented a birth/death step with empty
clusters only. In our model, this would correspond to popula-
tion templates with no associated individual templates. Origi-
nally, we tried using this birth/death step instead of the more
complicated population birth/death step described in Section 3.
However, we found that there were rarely any empty popula-
tion templates to propose deleting, especially if the number of
population templates, K ü , was close to the number of individ-
ual templates, Ki . As a result, the algorithm would not mix
well for K ü close to any Ki . Similar behavior was noted by
Castelloe (1998, 1999) in his model for spatial Poisson cluster
processes.
APPENDIX A: INDIVIDUAL BIRTH/DEATH STEP
ALGORITHM AND ACCEPTANCE PROBABILITIES
To simplify notation, let Èi D 4Ìi 1 Ði 5 and Èü D 4Ìü 1 Ðü 5 represent
the subject-speciŽ c and population template parameters. The indi-
vidual birth and death step is detailed in Table A.1. The acceptance

Table A.1. Individual Birth/Death Step

For i D 11 : : : 1 N, with probability .5, go to the birth step. Otherwise, go to the death step.
Birth step
1. Select a population template v U({ 11 : : : 1 K ü } n{ z i 1 1 : : : 1 z i Ki } ).
2. Draw new individual templates parameters u p(Èi —Èüv 1 Îv ).
3. Draw an index for the new individual template l U({ 11 : : : 1 Ki C 1} ).
4. With probability (Ki 1 Ki C 1), update the parameters for individual i as follows:
For k < l , keep Èi k and z i k .
For k D l , set Èi k D u and z i k D v.
For k > l , set Èi k C1 D Èi k and z i k C1 D z i k .
Otherwise, keep the current values.
Death step
1. Randomly select an individual template l U({ 11 : : : 1 Ki } ).
2. With probability (Ki 1 Ki ƒ 1), delete template Èi l and allocation variable z i l .
Otherwise, keep the current values.
134 Journal of the American Statistical Association, March 2002

probability for the individual birth step is tion birth and death steps are detailed in Table B.1. For the popula-
tion birth step, the prior ratio is given by
4Ki 1 Ki C 15 D min411 prior ratio € likelihood ratio µ ¶
p4K ü C 15 Y
N
p4Ki0 —K ü C 15 p4zi0 —Ki0 1 K ü C 15
prior ratio D
€ proposal ratio € Jacobian50 p4K ü 5 iD1 p4Ki —K ü 5 p4zi —Ki 1 K ü 5

In this case, the Jacobian is 1, because the generated random vari- p4Èü 0 1 Î0 —K ü C 15 Y
N
p4È0i —Ki0 1 zi0 1 Èü 0 1 Î0 5
€
ables are simply equal to the proposed new template parameters. The p4Èü 1 ΗK ü 5 iD1 p4Èi —K1 zi 1 Èü 1 Î0 5
likelihood is deŽ ned in (2). The prior ratio is given by µ ¶µ N ¶
‹ü Y ‹ Y p4ki µ K ü —K ü 5
D ü
p4Ki C 1—K ü 5 p4z0i —Ki C 11 K ü 5 K C 1 i2I Ki C 1 iD1
p4ki µ K ü C 1—K ü C 15
prior ratio D Q
p4Ki —K ü 5 p4zi —Ki 1 K ü 5 i62 I 4K
ü
ƒ Ki C 15
€ p4È0ülü 1 Î0l ü 5
p4È0i1 1 : : : 1 È0ikC1 —z0i 1 Ki C 11 Èü 5 4K ü C 15N
€ Y
p4Èi1 1 : : : 1 ÈiKi —zi 1 Ki 1 Èü 5 € p4È0ili —zili D lü 1 Èülü 0 1 Î0l ü 50
‹ 1 i2I
D p4È0il —zil D v1 Èüv 1 Îv 51
Ki C 1 K ü ƒ Ki The proposal ratio is given by

where z0i and È0ik are the candidate values. The proposal ratio is proposal ratio
given by
bkC1 hd 4lü —K ü C 15
D Q
dKi C1 hd 4l—Ki C 15 dk hb 4uü 1 vü 5hb 4lü —K ü 5 i2I 6hb 4li —K ü 1 Ki 5hb 4ui —uü 1 v ü 57
proposal ratio D Q
bKi h b 4v—zi 1 Kü 5hb 4l—Ki 5hb 4u—z0il D v1 Èüv 1 Îv 5 i2I 4Ki C 15
D Q 0
ü
K ƒ Ki p4Èül 0 1 Î0l 5 i2I p4È 0 0 ü 0 0
ili —K i 1 zili D l 1 Èlü 1 Îl ü 5
ü
D 1
p4È0il —zil D v1 Èüv 1 Îv 5
The Jacobian is equal to 1. Thus the acceptance probability for the
where hb and hd are the candidate distributions for the birth and death birth step 4K ü 1 K ü C 15 is min811 LR € B9, where
steps. Thus the acceptance probability simpliŽ es to 4Ki 1 Ki C 15 D
µ ¶µ N ¶
min811 A9, where A D 4‹=Ki C 154p4yi —Ki C 11 È0i Á5=p4yi —Ki 1 Èi 1 Á55. ‹ü Y ‹ Y p4ki µ K ü —K ü 5
BD
For the death step moving from Ki C 1 to Ki 1 the acceptance proba- K C 1 i2I Ki C 1
ü
iD1 p4ki µ K C 1—K C 15
ü ü

bility is 4Ki C 11 Ki 5 D min811 Aƒ1 9. Q ü

i62I 4K ƒ Ki C 15
€ Q
APPENDIX B: POPULATION BIRTH/DEATH STEP 4K ü C 15N i2I 4Ki C 15
ALGORITHM AND ACCEPTANCE PROBABILITIES
and LR is the likelihood ratio. The likelihood is deŽ ned in (2). Simi-
To simplify notation, let Èi D 4Ìi 1 Ði 5 and Èü D 4Ìü 1 Ðü 5 represent larly, the acceptance probability for the death step moving from K ü C
the subject-speciŽ c and population template parameters. The popula- 1 to K ü population templates 4K ü C 11 K ü 5 is min811 LR € B ƒ1 9.

Table B.1. Population Birth/Death Step

With probability, .5, go to the birth step. Otherwise, go to the death step.
Birth step
1. Draw new population template parameters uü p(Èü ) and v ü p(Î).
2. Draw a population template index l ü U({ 11 : : : 1 K ü C 1} ).
3. Draw an integer M U({ 01 : : : 1N } ).
4. If M > 0,
Draw a set of individual indices I D { i 1 1 : : : 1 i M } without replacement from { 11 : : : 1 N } .
For i 2 I:
Draw new template parameters ui p(Èi —Èüv 1 Îv ).
Draw an index for the new template l i U({ 11 : : : 1 Ki C 1} ).
5. With probability  (K ü 1 K ü C 1):
Update the population parameters as follows:
For k < l ü , keep Èük and Îk .
For k D l ü , set Èük D uü and Îk D vü .
For k > l ü , set ÈükC1 D Èük and Îk C1 D Îk .
For i 2 I, update the individual parameters as follows:
For k < l i , keep Èi k and z i k .
For k D l i , set Èi k D u and z i k D l ü .
For k > l i , set Èi k C1 D Èi k and z i kC1 D z i k .
For i 62 I, keep the current values.
Otherwise, keep the current values.
Death step
1. Randomly select a population template l U({ 11 : : : 1 Ki } ).
2. With a probability  (K1 K ƒ 1), delete population template Èül , the associated precisions Îl ,
and all individual template parameters Èi k and allocation variables z i k such that z i k D l .
Otherwise, keep the current values.
Miglioretti, McCulloch, and Zeger: Direct Cortical Electrical Interference
[Received April 2000. Revised September 2001.]
REFERENCES
Baddeley, A. J. (1992), “An Error Metric for Binary Images,” in Robust Com-
puter Vision, Quality of Vision Algorithms, in Proceedings of the Interna-
tional Workshop on Robust Computer Vision, pp. 59–78.
Boatman, D., Hall, C., Goldstein, M., Lesser, R., and Gordon, B. (1997),
“Neuroperceptua l Differences in Consonant and Vowel Discrimination
as Revealed by Direct Cortical Electrical Interference,” Cortex, 33,
83–98.
Boatman, D., Lesser, R., and Gordon, B. (1995), “Auditory Speech Processing
in the Left Temporal Lobe: An Electrical Interference Study,” Brain and
Language, 51, 269–290.
Boatman, D., Lesser, R., Hall, C., and Gordon, B. (1994), “Auditory Percep-
tion of Segmental Features: A Functional Feuroanatomi c Study,” Journal
of Neurolinguistics, 8, 225–234.
Castelloe, J. M. (1998), “ Issues in Reversible Jump Markov Chain Monte
Carlo and Composite EM Analysis, Applied to Spatial Poisson Clus-
ter Processes,” Ph.D. dissertation, University of Iowa, Department of
Statistics.
(1999), “Reversible Jump Markov Chain Monte Carlo Analysis of
Spatial Poisson Cluster Processes With Bivariate Normal Displacement,”
Computing Science and Statistics, 31, 306–315.
Christiansen, C. L., and Morris, C. N. (1997), “Hierarchical Poisson Regres-
sion Modeling,” Journal of the American Statistical Association, 92, 618–
632.
Crone, N. E., Miglioretti, D. L., Gordon, B., and Lesser, R. P. (1998a), “Func-
tional Mapping of Human Sensorimotor Cortex With Electrocorticographi c
Spectral Analysis II. Event-Related Synchronizatio n in the Gamma Band,”
Brain, 121, 2301–2315.
Crone, N. E., Miglioretti, D. L., Gordon, B., Sieracki, J. M., Wilson, M. T.,
Uematsu, S., and Lesser, R. P. (1998b), “Functional Mapping of Human
Sensorimotor Cortex With Electrocorticographi c Spectral Analysis I. Alpha
and Beta Event-Related Desynchronization,” Brain, 121, 2271–2299.
Gelman, A., Carlin, J. B., Stern, H. S., and Rubin, D. B. (1995), Bayesian
Data Analysis, London: Chapman and Hall.
Gilks, W. R., Richardson, S., and Spiegelhalter, D. J. (1996), Markov Chain
Monte Carlo in Practice. London: Chapman and Hall.
Gordon, B., Hart, J. J., Boatman, D., and Lesser, R. P. (1996), “Cortical
Stimulation ( Interference) During Behavior,” in Behavioral Neurology and
Neuropsychology, eds. T. E. Feinberg and M. Farah, Chicago: McGraw-Hill.
Green, P. J. (1995), “Reversible Jump Markov Chain Monte Carlo Computa-
tion and Bayesian Model Determination,” Biometrika, 82, 711–32.
Laird, N. M., and Ware, J. H. (1982), “Random-Effects Models for Longitu-
dinal Data,” Biometrics, 38, 963–974.
Lee, H. S. (1997), “Random Effects Models for Count Data,” Communications
in Statistics, Part A—Theory and Methods, 26, 1893–1904.
Lesser, R. P., Lüders, H., Klem, G., Dinner, D. S., Morris, H. H., Hahn,
J. F., and Wyllie, E. (1987), “Extraoperative Cortical Functional Localiza-
tion in Patients With Epilepsy,” Journal of Clinical Neurophysiolog y, 4,
27–53.
Lindley, D. V., and Smith, A. F. M. (1972), “Bayes Estimates for the Linear
Model,” Journal of the Royal Statistical Society, Ser. B, 34, 1–41.
Malow, B., Blaxton, T., Sato, S., Bookheimer, S., Kufta, C., Figlozzi, C., and
Theodore, W. (1996), “Cortical Stimulation Elicits Regional Distinctions in
Auditory and Visual Naming,” Epilepsia, 37, 245–252.
Miglioretti, D. L., McCulloch, C., and Zeger, S. L. (2000), “Template Mixture
135
Models for Direct Cortical Electrical Interference Data,” Biostatistics, 1,
403–421.
Nathan, S. S., Sinha, S. R., Gordon, B., Lesser, R. P., and Thakor, N. V.
(1993), “Determination of Current Density Distributions Generated by
Electrical Stimulation of the Human Cerebral Cortex,” Electroencephalog-
raphy and Clinical Neurophysiolog y, 88, 183–192.
Nii, Y., Uematsu, S., Lesser, R. P. and Gordon, B. (1996), “Does the
Central Sulcus Divide Motor and Sensory Functions?,” Neurology, 46,
360–367.
Ojemann, G. A. (1979), “ Individual Variability in Cortical Localization of
Language,” Journal of Neurosurger y, 50, 164–169.
(1981), “ Intrahemispheric Localization of Language, Memory, and
Motor Mechanisms in Human Cortex and Thalamus,” in New Perspec-
tives in Cerebral Localization, ed. R. Thompson, New York: Raven Press,
pp. 157–175.
Pascual-Leone, A., Cohen, L. G., Brasil-Neto, J. P., and Hallett, M. (1994),
“Non-invasive Differentiation of Motor Cortical Representation of Hand
Muscles by Mapping of Optimal Current Directions,” Electroencephalog-
raphy and Clinical Neurophysiolog y, 93, 42–48.
PenŽ eld, W., and Boldrey, E. (1937), “Somatic Motor and Sensory Represen-
tation in the Cerebral Cortex of Man as Studied by Electrical Stimulation,”
Brain, 60, 389–443.
Pfurtscheller, G., Flotzinger, D., and Neuper, C. (1994), “Differentiation
Between Finger, Toe and Tongue Movement in Man Based on 40
Hz EEG,” Electroencephalograph y and Clinical Neurophysiolog y, 90,
456–460.
Rao, C. R. (1965), “The Theory of Least Squares When the Parameters
are Stochastic and Its Application to the Analysis of Growth Curves,”
Biometrika, 52, 447–458.
Richardson, S., and Green, P. J. (1997), “On Bayesian Analysis of Mixtures
With an Unknown Number of Components,” Journal of the Royal Statistical
Society, Ser. B, 59, 731–792.
SAS Institute, Inc. (1990), SAS/ IML Software, Usage and Reference, Version
6, First Edition, Cary, NC: Author.
Schlaug, G., Knorr, U., and Seitz, R. (1994), “ Inter-Subject Variability of
Cerebral Activations in Acquiring a Motor Skill, a Study With Positron
Emission Tomograph y,” Experimental Brain Research, 98, 523–534.
Stiratelli, R., Laird, N., and Ware, J. H. (1984), “Random-Effects Models for
Serial Observations With Binary Response,” Biometrics, 40, 961–971.
Talairach, J., and Tournoux, P. (1988), Co-Planar Stereotactic Atlas of the
Human Brain, New York: Thieme.
Tierney, L. (1994), “Markov Chains for Exploring Posterior Distributions”
(with discussion), The Annals of Statistics, 22, 1701–1762.
Uematsu, S., Lesser, R., Fisher, R. S., Gordon, B., Krauss, G., Vining,
E. P., and Webber, R. W. (1992), “Motor and Sensory Cortex in Humans:
Typograph y Studied With Chronic Subdural Stimulation,” Neurosurgery,
31, 59–72.
Uematsu, S., Lesser, R., Fisher, R., Krauss, G., Hart, J., Vining, E. P., Free-
man, J., and Gordon, B. (1990), “Resection of the Epileptogenic Area in
Critical Cortex With the Aid of a Subdural Electrode Grid,” Stereotactic
and Functional Neurosurger y, 54–55, 34–45.
Urasaki, E., Uematsu, S., Gordon, B., and Lesser, R. P. (1994), “Cortical
Tongue Area Studied by Chronically Implanted Subdural Electrodes—With
Special Reference to Parietal Motor and Frontal Sensory Responses,” Brain,
117, 117–132.
Zeger, S. L., and Karim, M. R. (1991), “Generalized Linear Models With
Random Effects: A Gibbs Sampling Approach,” Journal of the American
Statistical Association, 70, 138–144.