QEEG: Methodological Issues
Quantitative electroencephalography (EEG) has earned a
reputation of being noisy, unreliable, and imprecise in the minds
of psychologists, brain scientists, and laymen alike (Nuwer,
1988; Begley, 1992). A lack of methodological standards underlies
this inaccurate characterization. A researcher interested in
analyzing quantitative EEG during behavioral or mental processes
confronts a gauntlet of largely arbitrary methodological choices
about reference electrodes, recording electrodes, epoch
parameters, windowing function, bandwidths, spectral estimates,
and artifact control (see Table 2.1). Different methodological
configurations generate incompatible, or worse, conflicting
findings and conclusions (e.g., Davidson, Chapman, Chapman, &
Henriques, 1990). A separate family of alternatives is
encountered in non-spectral analysis of EEG (e.g., Gregson,
Britton, Campbell, & Gates, 1990). One of the goals of this work
is to survey methodological alternatives in quantitative EEG and
propose a set of standards for investigations of cognitive
processes. These choices will in turn be used to construct a
study to investigate EEG correlates of subjective interest in
cinematic narratives.
Recording references
Each EEG recording reference has its own set of
advantages and disadvantages. Linking reference electrodes from
two mastoids or earlobes provide a non-lateralized reference
(Miller, Lutzenberger, & Elbert, 1991). This common
method reduces the likelihood of artificially inflating activity
in one hemisphere. But the use of linked-ears references had been
criticized by Nunez (1991) because he claims that the "effective"
reference will drift away from the midline plane if the
electrical resistance at each electrode differs, the phenomenon
called "shunting". Maintaining small contact resistances compared
to amplifier input impedances (e.g., 5-10 K ohms) reduces
resistance variance to reasonable levels. Within-group
comparisons are also less affected by systematic reference
problems, such as shunting. Nevertheless, Nunez recommends
ipsilateral references or reference-free techniques in lieu of
linked references (Nunez, 1991; Nunez, Silberstein, Cadusch, &
Wijesinghe, 1993).
Table 2.1. Methodological issues in quantitative EEG
1. Data Acquisition
a. Number of electrodes
b. Reference method
c. Montage
d. Impedance
2. Data Collection
a. Epoch interval
b. Task duration
c. Window function
d. Frequency bands
e. Artifact control
f. Data transform
g. Spectral parameter
3. Data Analysis
a. Reliability
b. Baseline and control conditions
c. Statistical methods
d. Statistical corrections
e. Methodological and physiological validity
In response to Nunez's recommendation, linked-ear and
ipsilateral-ear references were compared for eyes closed (EC) and
eyes open (EO) conditions. For either condition alpha activity
differed slightly (2%) between reference types. Substantial
differences between reference types, however, were noted in the
beta range (12 Hz and above). This finding does not warrant the
rejection of the popular linked-ear reference, though more
subjects are required to further clarify differences between
reference types.
Linked-ears and ipsilateral-ear montages rely on
physical references. Electrical potentials may seep into a
physical reference and transform an otherwise electrically-
neutral area into an active site, producing topographic
distortion. This is evident in the topographic distribution
associated with the linked-ear montage. The presence of cortical
activity in each earlobe manifests less activity in lateral
electrodes while maximizing activity in the midsagittal plane
(Etevenon, 1986). Nuwer (1988) proposes that researchers run
several references in succession to identify active leads, an
impractical and costly approach. Active references, if stable
across recording conditions, are not problematic for within-
subject comparisons.
Reference-free techniques, such as common average
references or source derivation, do not suffer from problems
associated with an actual physical reference (Duffy, 1986). Local
common average references are especially accurate for small
localized regions (Pfurtscheller, 1988) but this reference is
poor for extensive topographic evaluations (Duffy & Maurer,
1989). Source-derived and Laplacian constructions require
multiple electrodes for each brain area and are inaccurate for
regions with few neighboring electrodes, such as lateral or
posterior areas (Nuwer, 1988; Duffy, 1986; Gevins, 1984).
Laplacian derivations also generate hard to interpret
complexities (Duffy & Maurer, 1989) with little standardization
between laboratories (e.g., Nunez, 1993; Gundel & Wilson, 1992;
Pfurtscheller, 1988).
One hundred relevant EEG studies between 1965 and 1994
were surveyed by the author to asses the prevalent methodologies
in quantitative EEG. Seven different references were described in
the literature. Referencing to linked ears or vertex were
predominant. Eighty-four percent of all studies involved
referential recordings, in which scalp electrodes are connected
to a common neutral site or sites, such as the earlobe or base of
the neck. The remaining studies employed bipolar montages, which
involves connecting pairs of scalp electrodes together without
reference to a common inactive lead. Unfortunately, this second
configuration obscures localized functional activity (Gevins,
1986). Dissimilar reference configurations can produce strikingly
distinct topographies, reinforcing the need for reference
standardization (Pfurtscheller, 1988; Pfurtscheller & Klimesch,
1990; Lehmann, 1989; Davidson et al., 1990). As it
happens, no current reference technique is worthy of universal
application (Duffy, 1986). Accordingly, the present study
employed a linked-ears reference.
Recording electrodes
The international 10-20 system of electrode placement
standardized physical placement and designations of electrodes on
the scalp (Jasper, 1958). This coordinate system divides the head
into proportional distances from prominent skull landmarks
(nasion, preauricular points, inion) so as to provide adequate
coverage of all regions of the brain (see Appendix A). Electrode
placements are labelled for the adjacent brain area to facilitate
communication between researchers and laymen alike. The 10-20
system owes its endurance in part to its simplicity and the
fortuitous division of the head into functional regions that
remain relevant to human information processing research given
the present state of technology. According to CT-scan evidence,
F3 and F4 overlays Brodmann's area 46, C3 and C4 Brodmann's area
4, and P3 and P4 Brodmann's area 7 (Homan, Herman, & Purdy,
1987). Still, many researchers continue to position recording
electrodes arbitrarily, without regard to standardized coordinate
systems (e.g., Goodman & Mulholland, 1988; Pfurtscheller &
Klimesch, 1990). Others use a minimum number of electrodes,
obviating topographic differences (e.g., Davidson et al.,
1990).
A familiar argument against usefulness of topographic
EEG in research or clinical settings is the supposed poor spatial
resolution of scalp electrodes (Nunez et al., 1993; Lorig
& Schwartz, 1989). The problem arises not from the fact that
scalp electrodes record activity from a large pool of neurons
simultaneously (Nunez, 1981), but that the EEG signal can be
smeared or filtered as it passes through meninges, calvarium, and
unrelated cortical tissue. Although Gevins (1993) has attempted
128 scalp electrodes to increase spatial resolution, the
technique does not address the primary concern, which is one of
functional rather than physical resolution.
To address this problem, EEG was analyzed at 19 sites
for 20 subjects during two baseline resting conditions and one
viewing task. EEG magnitudes were averaged and normalized (see
below) to reduce inter-subject variability. Log mean magnitude at
each recording site was correlated to means at neighboring
electrodes so as to identify whether there is significant
smearing between proximal electrodes. If functional resolution
were poor, correlations between electrodes would depend solely on
physical proximity. Alpha activity at site P4 would resemble
activity at all neighboring sites (T6, C4, O2) as readily as it
mirrors activity in functionally-related areas (Pz, P3). This was
not the case, especially in the eyes closed condition (see Figure
2.1).
Discrepancies in the eyes open baseline condition and
film viewing task were ascribed to diverse processing demands.
Except for temporal regions, covariance between neighboring
electrodes across functional boundaries (e.g., parietal to
temporal) was much smaller than covariance within functional
regions (e.g., left parietal to midline parietal), indicating
that multiple distinct functional areas are assessed by
topographic EEG (Kooi, 1971; Bullock & McClune, 1989).
Figure 2.1. Functional resolution of alpha activity for
eyes closed condition (n=20). Lines indicate strong correlations
(r=.50) between proximal sites.
The number of electrodes used in EEG research has
steadily increased during the past 30 years to an average of 12
electrodes in recent years. During the same period,
the number of subjects per study has gradually
increased (see Figure 2.2). Number and position of electrodes
often varies as a function of research requirements, such as
mapping of hemispheric dynamics or investigating focal activity
(Etevenon, 1986). Notwithstanding, a proposed minimum of eight
electrodes, over specific cortical areas, is recommended for
studying higher cognitive functions and subsequent acceptance in
psychophysiological and related journals. The recommended areas
from which EEG is to be acquired include left and right regions
of frontal, central, temporal, and parietal cortex (Kooi, 1971;
Etevenon, 1986). More recording electrodes may be necessary for
detecting complex patterns of cortical activity, but more than 20
electrodes are unwarranted for most research goals (Duffy, 1986;
Duffy & Maurer, 1989). This modest requirement followed standardization
in scientific research and
publishing and will immediately enhance study compatibility as
well as improve communication between EEG laboratories on the
whole. In the present study, EEG was recorded from 19 electrodes
using the international 10-20 coordinate system for electrode
placement.
Figure 2.2. Average number of subjects and recording sites
in EEG studies for 5 or 10 year periods.
Frequency Analysis
EEG can be characterized as either non-periodic (spikes,
random noise), non-sinusoidal and periodic (mu), or sinusoidal
(alpha, delta) signals (Nunez, 1981). Spectral analysis minimizes
the inherent redundancy in a periodic signal by evaluating
frequency components of the EEG signals (Gevins, 1986; Etevenon,
1986; Kunkel, 1978). Using spectral analysis, also called
frequency analysis, an EEG signal can be decomposed into
constituent periodic components, usually by means of a set of
filters or mathematical functions (e.g., Fourier analysis).
Frequency analysis provides an efficient means of summarizing
relevant information from longer records of EEG generated in
clinical and research settings.
Using spectral analysis, researchers have formulated
numerous indices of cerebral activity such as coherence, power,
peak frequency, and alpha periodicity (Remond & Lairy, 1972;
Nuwer, 1988; Sterman, Mann, Kaiser, & Suyenobu, 1992). Although a
few scientists report and attempt to interpret as many as 10
spectral indices or parameters for each recording session (e.g.,
de Rijke & Visser, 1989; Etevenon, 1986), the majority of
published EEG analysis concerns a single spectral parameter or
two (e.g., absolute and relative power). In addition to spectral
parameters, researchers can define a wide range of frequencies to
analyze. Some studies include a laundry list of spectral
parameters for a dozen or more frequency bands, impressively
diminishing the quality of the information communicated to the
reader (e.g., Etevenon, Bertaut, Mitermite, & Eustache, 1989;
Etevenon, 1986). In this author's experience, studies which
concentrate on single spectral parameters for multiple frequency
bands (e.g., Grillon & Buchsbaum, 1986) or multiple spectral
parameters for a single frequency band are optimal for both
experimenter and the audience. With this in mind, the present
study has been limited to three topographic comparisons (sites,
pooled sites, lateral sites differences) and four spectral
parameters (see below) of a single frequency band in the alpha
range (8-12 Hz).
Spectral bands and parameters
Conventional frequency bands in EEG research are the
following: 0-4 Hz (delta), 5-7 Hz (theta), 8-13 Hz (alpha) and 14
or more Hz (beta). A common and accepted practice in quantitative
EEG analysis involves designating adjacent bands of 4 Hz
intervals (e.g., 0-4 Hz, 4-8 Hz, 8-12 Hz, 12-16 Hz, 16-20 Hz). It
is well known that wide frequency bands (more than 1 to 2 Hz
intervals) encompass a variety of physiological processes (Lorig
& Schwartz, 1989; Grillon & Buchsbaum, 1986) and many functions
are better identified in narrower (e.g., 8-10 Hz, 11-13 Hz)
frequency bands (Pfurtscheller & Klimesch, 1990; Gale & Edwards,
1983; Sterman et al., 1994). Yet no definitive division of
the human EEG frequency range has been found. More than 20
arbitrary frequency boundaries have been specified in the
literature for studying the alpha rhythm (e.g., 7.81-14.06 Hz,
7.03-12.89, 8-15 Hz; Etevenon, Eustache, Mitermite, Lepaisant,
Lechevalier, & Zarifian, 1990; de Toffel & Autret, 1991; Ray &
Cole, 1985, respectively). Lack of standardization in frequency
bands fosters confusion between laboratory findings, but may be
required due to the range of variables addressed by quantitative
EEG (Remond & Lairy, 1972). Nevertheless, the use of traditional
frequency bands is most appropriate for preliminary
investigations and hypothesis testing.
The alpha rhythm refers to the dominant or peak
frequency recorded from cortex, especially pronounced during non-
processing, relaxed conditions (Berger, 1930). At least
eighty-nine percent (if not 100 percent) of normal healthy adults
exhibit alpha activity, though magnitudes vary between
individuals (Remond & Lairy, 1972). Most adults generate between
20 to 60 æV of alpha activity during eyes closed conditions
(Kooi, 1971). Peak alpha frequency is fairly stable (10 ñ 0.5 Hz)
in most individuals during a single session and from day to day
(Nuwer, 1988) and is consistent for various subject populations
(Nunez, 1981; Kooi, 1971; Sterman et al., 1994). In this
study, the mean and mode of the peak frequency recorded from 20
subjects during an eyes-closed baseline condition (EC2) was
approximately 10 Hz (see Figure 2.3).
Figure 2.3. Distribution of peak frequency recorded
from 19 recording sites during an eyes closed baseline condition
(n=20). Dashed lines indicate mean and dotted lines indicate the
95 % confidence interval.
As seen in Figure 2.4, peak frequency exhibits
topographic variability, with higher peak frequencies in
posterior cortex (10.3-10.6 Hz) and lower peak frequencies in
anterior cortex (9.7-10.3 Hz; cf. Gratton, Villa, Fabiani, &
Colombis, Palin, Bolcioni, & Fiori, 1992, for similar results).
Nunez (1981) reported that 96% of peak frequencies fell between 8
and 12 Hz for 135 subjects. In this study, a frequency band of 8-
12 Hz was used to measure alpha activity. Integer frequency
boundaries were chosen in order to facilitate comparisons to
other studies.
Figure 2.4. Topography of peak frequency for an eyes
closed baseline (n=20). Bars indicate 95 % confidence interval.
Note the greater variance at anterior regions.
Numerical transformations of alpha amplitude
Physiological data distributions are typically skewed, a
fact which contradicts the normality assumption underlying the
Analysis of Variance test (Duffy & Maurer, 1989). Deviations from
normality in EEG data distributions arise from various origins
(Kramer, 1991; Nuwer, 1988), including: 1) scale examined (e.g.,
power), 2) intrinsic biological mechanisms, 3) inter-subject
heterogeneity, 4) artifacts, and 5) spectral parameters that are
arbitrary and not biologically meaningful. Transformations such
as log power (twice log magnitude) and the square root of power
(magnitude) result normal data distribution (Gasser, Bacher, &
Mocks, 1982), but power distributions are usually skewed. Despite
this information, the majority of studies analyze power means.
Figure 2.5 shows three mathematical transformations of
mean amplitude for high interest films. Squared transformation
(power) result in highly skewed distributions for high interest
conditions (Dmax= 0.17, p<.01; cf. Kolmogorov, 1941).
Linear transformation (magnitude) result in moderately skewed
distributions for high interest conditions (Dmax= 0.12,
p<.05). Finally, logarithmic transformation (natural log
of mean magnitude) result in relatively normal distributions
(Dmax= 0.056, ns) and it has been shown that many psychological
phenomena such as loudness correlate with the logarithm of
physiological data (e.g., Baird, Berglund, Berglund, Lindberg,
1991; Krueger, 1989). The present study analyzes the natural log
of mean magnitude.
Epoch parameters
As mentioned above, spectral analysis usually involves
Fourier analysis of the EEG signal. Fourier analysis, or Fast
Fourier Transformation (FFT), requires
that a signal be divided into segments, usually short in
duration, called epochs.
Figure 2.5. Distribution of EEG data for 19 sites
during a 2-min high interest film (n=20). Power and magnitude
distributions deviate from normality (Kolmogorov-Smirnov test).
Underlying its use is the assumption that the signal analyzed
will be stationarity or time-invariable, but human EEG varies in
character over time (Isaksson & Wennberg, 1976; Lopes da Silva,
1978; Praetorius, Bodenstein, & Creutzfeldt, 1977). Some
researchers attempt to compensate by tailoring epoch lengths to
signal properties (Lopes da Silva, 1978; Nuwer, 1988; Lehmann,
1989) or optimizing epoch duration for each psychological
property under investigation (Walter, 1978; Gevins, 1984). The
use of relatively brief epochs (1-2 s) is a general solution that
overcomes the stationarity problem (McEwen & Anderson, 1975) as
well as resolving other issues such as artifact control. As
discussed in detail below, data loss from artifact rejection
techniques is directly related to epoch length (Nuwer, 1988;
Duffy, 1986; Duffy & Maurer, 1989).
Aside from these considerations, epoch duration does not
appreciably affect study compatibility in the literature. In the
studies surveyed, epoch length ranged anywhere from 0.25 s to 30
s or longer, with a median length of 2.5 s. In this study, an
epoch duration of 2 s was used. Subjective interest, the
psychological phenomenon under investigation, changes once every
7.55 s, on averages (see Chapter 6). Two s epochs are short
enough to capture fluctuation of subjective interest as well as
being useful in estimating other spectral parameters of the alpha
rhythm such as variability (Duffy, 1986).
The duration of EEG data acquired for each task (i.e.,
number of epochs multiplied by epoch duration) should be
representative of task demands, resource allocation, and
strategic factors. Investigations of continuous tasks such as
mental imagery or problem-solving typically record EEG for 30 s
to 2 min. Studies which explored repeated events in EEG usually
recorded dozens or even hundreds of trials of very short
intervals (Pfurtscheller, Flotzinger, Mohl, & Peltoranta, 1992).
In the present study, EEG will be acquired for 50 to 75 epochs
(100 to 150 s) in each condition.
Artifact problem
Electrodes do not differentiate electrical activity
generated by cortex from that originating in extracerebral
sources (Gasser, Stroka, & Mocks, 1985). Non-cerebral potentials
generated by movements of the eye, tongue, face or neck muscles,
heartbeat, or changes in skin conductance, can contaminate
cortical activity (Barlow, 1986; Torello, 1989). Fortunately, low-
and high-pass filters minimize most artifacts generated by
muscles, changes in skin conductivity, and heart beat (Nuwer,
1988), but the problem of ocular artifact remains.
Eye blinks can last from 200 to 400 ms and produce
electrical magnitudes up to 800 æV, more than 10 times the
amplitude of cortical signals (Stern, Walrath, & Goldstein,
1984). The initial fast components of an eye blink can yield non-
cortical electrical potentials with frequency components up to
10.5 Hz, well within the alpha range (Gasser et al., 1985;
Brunia, Mocks, & Van den Berg Lenssen, 1989). Ocular artifacts
can contaminate any recording site, though they are largest in
frontal areas (Torello, 1989).
Artifact minimization and artifact rejection techniques
have developed to address the problem of eye movements and
blinks. Artifact minimization encompasses those techniques in
which contribution of suspected artifacts are estimated and
removed from the EEG signal, usually achieved in the time domain
(Kenemans, Molenaar, Verbaten, & Slangen, 1991; van den Berg
Lenssen, Brunia, & Blom, 1989). Artifact rejection involves
discarding epochs which are contaminated prior to averaging or
further analysis. Elimination of epochs that contain damaging
artifacts works well as long as epoch duration is relatively
short and the presence of artifacts does not co-vary with
experimental conditions (Berg, 1986). In this study, contaminated
epochs accounted for less than 10 percent of the data (cf. Berg,
1986, for higher values). Eye blinks did not interact with task
demands and was nearly identical across conditions. (The lone
exception was eyes closed conditions which consisted of almost no
contaminated epochs.) This situation is particularly suitable for
controlling artifacts by means of artifact rejection methods.
Other techniques to reduce physiological artifact
include instruction to the subject and provision of rest breaks.
Subjects can be instructed to make conscious attempts to reduce
eye blinks, especially during critical events (e.g., de Rijke &
Visser, 1989). This approach essentially introduces an additional
task component which may distract the individual and interfere
with the task under investigation (Semlitsch, Anderer, Schuster,
& Presslich, 1986). An experimenter may also provide comfortable
chairs and rest breaks for subjects to reduce ocular and muscular
artifacts (Torello, 1989). In this study, besides the use of a
padded lounge chair, brief rest breaks (5-10 s) were provided
after each film and longer breaks were provided (1 min) after
subjects watched the first set of films.
Although various procedures have been developed to
reduce artifact in EEG recordings, no clear solution has emerged.
Artifact minimization generally requires additional electrodes
dedicated to detecting eye movements (Duffy, 1986; Coburn &
Moreno, 1988), construction of a propagation model (Lins, Picton,
Berg, & Scherg, 1993a; 1993b), estimation of variable time-
delays, and other assumptions and obstacles which can hinder
accuracy and effectiveness of this approach (Berg, 1986; Brunia
et al., 1989). Using discriminant analysis, MacCrimmon,
Durocher, Chan, Hay, and Saxena (1993) constructed a reliable,
highly accurate, non-subjective method of detecting artifacts for
use in artifact minimization methods; unfortunately, this
computer-intensive method requires 112 features to be calculated
off-line for each epoch. Artifact rejection has likewise been
criticized as subjective (MacCrimmon et al., 1993),
unreliable, and susceptible to significant data loss (Barlow,
1986). The problem of artifact will likely continue to plague EEG
science for some time to come.
Another common form of artifact can be attributed to
data analysis. As mentioned above, spectral analysis of the EEG
signal is usually achieved by means of an FFT of epoch values.
Each epoch is a truncated segment of the EEG signal which
consists of 2n data points. The truncation of an ongoing signal
results in sharp edges (non-zero initial or final values) at
beginning and end of the epoch. FFT of non-zero edges generates
spurious frequency information about a signal, also called
"leakage" (Jervis, Coelho, & Morgan, 1989). Mathematical
functions called "windows" or "frames" can be applied to epoch
values to taper data at epoch edges and reduce the effect of
leakage. A 4-term Blackman-Harris was used in this study (Harris,
1978). While a tapering function effectively eliminates leakage,
it results in artificial broadening of frequency peaks
(smearing), a reduction of signal power (Jervis et al.,
1989), and preferential sampling of the EEG signal. Use of
multiple overlapping windows can remedy the effect of
preferential sampling (Davidson et al., 1990; Sterman
et al., 1994), but this practice is rare. A design using
wide frequency bands, within-subject comparisons, and relatively
stable psychological conditions, will not be adversely affected
by a tapering function.
Appropriateness of parametric statistics
Although inferential statistics are required for
generalizing sample results to a subject population, Etevenon
et al. (1989) and others have proposed the use of
nonparametric tests in response to the skewness inherent in
physiological data. However, as mentioned above, the log
transformation of alpha magnitude results in normal
distributions. In addition, physiological data often do not
satisfy the independence assumption for statistical tests such as
Analysis of Variance. Statistical corrections can be applied to
treatment and error degrees of freedom in order to compensate for
nonsphericity of physiological data (Greenhouse & Geisser, 1959;
Huynh & Feldt, 1976; Keselman & Rogan, 1980). Because the
Greenhouse-Geisser correction is overly conservative and produces
Type II errors (Klimesch, Pfurtscheller, Mohl, & Schimke, 1990),
the Huynh-Feldt correction will be used in this study.
Additional statistical controls can be used to
counteract the effect of multiple Analysis of Variance tests on
overall à (probability) level. The Bonferroni procedure
partitions the overall test à level (p=.05) across all
statistical comparisons. A compensatory strategy based on the
Bonferroni procedure was adopted in the present study.
Probability (à) level for all site comparisons and site-pair
comparisons was reduced to p=.01 and findings will be
qualified or ignored unless they are confirmed in more than one
spectral parameter.
Conclusion
This review has compared various approaches taken by
investigators of spectral correlates of attention and higher
cortical functions. This information has been incorporated into
an experimental design. Decisions about reference electrodes,
recording electrodes, epoch parameters, windowing function,
frequency band, numerical transformation, and artifact control
have all been justified on grounds of reliability, practicality,
and relevance to the psychological phenomenon under
investigation, subjective interest. The methodological and
physiological validity of these choices (section 3.e. in Table
2.1) will be evaluated in part by the results described in the
following chapters.
|
