Electroencephalography and clinical Neurophysiology , 1990, 77:295-308 295

Elsevier Scientific Publishers Ireland, Ltd.

EVOPOT 89042

Cross-correlation and latency compensation analysis of click-evoked

and frequency-following brain-stem responses in man
Gary C. Galbraith and Warren S. Brown
University of California, Los Angeles, School of Medicine, Mental Retardation Research Group at Lanterman Developmental Center,
Pomona, CA 91769 (U.S.A.), and Graduate School of Psychology, Fuller Theological Seminary, Pasadena, CA 91182 (U. S.A.)

(Accepted for publication: 28 September 1989)

Sununary Cross-correlation (CC) and latency compensation (LC) analyses were applied to the human click-evoked brain-stem
auditory evoked response (BAER) and the brain-stem frequency-following response (FFR). FFRs were elicited by pure tone stimuli
(230 Hz and 460 Hz) or by complex tones derived from the sum of 3rd (920 Hz), 4th (1150 Hz), and 5th (1380 Hz) harmonics of the
missing 230 Hz fundamental. The lower and upper harmonics always began in sine phase, while the middle harmonic varied in
starting phase, resulting in harmonically complex stimuli with differing amplitude and phase patterns.
Cross-correlations were computed between individual trials and a wave form template (smoothed wave V for BAER, pure tone
stimulus sinusoids for FFR). Trials were included in the analysis only if values of r 2 exceeded 0.5 (negative values of r were thus
included, whiah controlled for the chance occurrence of positive correlations). Although brain-stem recordings are noisy, requiring as
many as 1000 stimuli/average, correlation analysis consistently identified more positive than negative trials (approximately 2:1 ratio).
Trials were also deleted if the lag associated with the selected r 2 was at the maximum shift position ('extreme lag').
Averaging trials that satisfied the correlation and lag criteria led to sizeable enhancement of BAER (mean = 114%) and FFR
(mean ~ 68% for 230 Hz stimulus) amplitudes. LC analysis resulted in additional, albeit smaller, increases in amplitude (approxi-
mately 10%). FFRs to harmonically complex stimuli were characterized by a clear periodicity at the missing fundamental frequency
(230 Hz). However, amplitudes varied according to the modulation depth of the stimulus and, in certain cases, actually exceeded that
of the FFR response to a 230 Hz pure tone.
The results demonstrate the effectiveness of cross-correlation and, to a lesser degree, latency compensation analysis, applied to
two classes of brain-stem potentials. It is anticipated that such techniques will prove useful in the study of auditory signal processing
at the level of the brain-stem.

Key words: Brain-stem auditory evoked response; Frequency-following responses; Cross-correlation; Latency compensation;
(Human)

A number of investigators have employed
cross-correlation (CC) techniques to detect and
enhance long-latency evoked response compo-
nents (Palmer et al. 1966; Derbyshire et al. 1967;
Woody 1967; Weinberg and Cooper 1972a, b;
PfurtscheUer and Cooper 1975). Such techniques
typically utilize a correlation template, usually
Correspondence to: G.C. Galbraith, Lanterman Develop-
mental Center, University of California, Los Angeles, School
of Medicine, P.O. Box 100-R, Pomona, CA 91769 (U.S.A.).
some approximation to the known signal (gener-
ally the average wave form). Woody (1967) pro-
posed a latency compensation (LC) algorithm in
which the initial evoked response template is mod-
ified in successive iterations as the effects of tem-
poral jitter are systematically removed. However,
such iterative procedures have been criticized
(WasteU 1977; Aunon and Sencaj 1978); and
powerful new techniques are being proposed to
deal with the general problem of latency jitter
(e.g., MScks et al. 1988).
In contrast to the substantial body of research

0168-5597/90/$03.50 1990 Elsevier Scientific Publishers Ireland, Ltd.

296
applying CC and LC techniques to long-latency
potentials, very little has been done to apply these
techniques to short-latency potentials. The prob-
lem is complicated by reduced signal-to-noise
ratios, often requiring a thousand or more stimuli
to define an average. Additionally, it might seem
that LC analysis is not necessary with short-latency
potentials since they are intrinsically more stable.
Thus, components of the brain-stem auditory
evoked response (BAER) are thought to depend
upon synaptically 'secure' neurons which dis-
charge with brief and stable firing latencies
(Buchwald 1983). In spite of such presumed tem-
poral stability, however, significant enhancement
of BAER components by LC analysis has been
demonstrated (Galbraith 1984). Dobie and col-
leagues (Dobie and Clopton 1980; Dobie and Wil-
son 1984; Wilson and Dobie 1987) have obtained
short-latency auditory responses by cross-cor-
relation using continuous (but periodic) broad-
spectrum, pseudorandom, noise stimuli. However,
they did not address the issue of possible single
trial temporal variability in their data.
In most correlational analyses there is no a
priori or independent template, other than some
representation of the response itself. Certainly in
the case of click-evoked responses the stimulus
wave form is unrelated to evoked response mor-
phology. The frequency-following response (FFR),
however, exhibits more favorable properties in this
regard. Thus, while the scalp-recorded FFR is
delayed (according to the site of origin within the
auditory pathway, presumably the inferior col-
liculi) and persists beyond the stimulus by an
equivalent duration, the response closely repro-
duces the stimulus wave form morphology (Smith
et al. 1975, 1978; Greenberg et al. 1987). Such
FFR properties are thus well suited to correla-
tional analysis since the stimulus wave form con-
stitutes the appropriate template.
In the present paper we use CC and LC meth-
ods to study both the scalp-recorded human BAER
and FFR. We chose to replicate the 'missing
fundamental' FFR paradigm of Greenberg et al.
(1987). This paradigm was selected because the
use of complex higher harmonic stimuli requires
greater neural integration and can thus test the
integrity and efficiency of the system. The missing
G.C. GALBRAITH,W.S. BROWN
fundamental paradigm thereby provides a useful
context within which to examine temporal vari-
ability of FFR signals.
Methods and materials
Subjects
Five male and 6 female subjects (Ss) par-
ticipated in the study. There was a heterogeneous
age range (8-43 years, mean = 14.3). Hearing was
within normal limits (assessed with Beltone Model
9D audiometer) and there were no known neuro-
logical disorders; all children were at age-ap-
propriate grade levels in school.
Stimulation and recording
BAERs were elicited by 80 dB (SPL) rarefac-
tion clicks, 100 #sec in duration. Tone stimuli for
eliciting FFRs (Fig. 1) were digitally synthesized
and presented via digital-to-analog converter at a
rate of 27,500 samples/sec. Stimulus duration was
17.4 msec, intensity was 72 dB. Stimuli were
damped sine waves consisting of (a) 230 Hz pure
tone (4 complete cycles; first and last half cycle
attenuated by linear ramp function), (b) 460 Hz
pure tone (8 complete cycles; first and last full
cycle attenuated), or (c) complex tones derived
from the sum of 3rd (920 Hz), 4th (1150 Hz), and
5th (1380 Hz) harmonics of the 230 Hz fundamen-
tal (such stimuli produce a sensation of pitch
equal in frequency to the missing fundamental;
Greenberg 1980; Greenberg et al. 1987). The lower
and upper harmonics always began in sine phase.
The middle harmonic, however, varied in starting
phase from 0 to 180 , in 45 increments. This
produced 5 harmonically complex stimuli with
differing amplitude and phase patterns (cf.,
Greenberg et al. 1987). Tones or clicks were pre-
sented at a rate of approximately 15/sec in blocks
of 1000. Order of stimulus presentation was coun-
terbalanced across Ss.
Monaural stimuli were presented to the right
ear by means of a TDH-39 headphone shielded in
a mu-metal enclosure which significantly reduced
stimulus artifacts. Stimulus intensity and head-
phone response properties were calibrated by
means of a Briiel and Kjaer type 2801 condenser
CROSS-CORRELATION OF BRAIN-STEM RESPONSES
230Hz
460Hz
450
//
900
I
I
135
180 ~ ' quent off-line processing.
Fig. 1. Digitally synthesized wave forms for pure tone (230 and
460 Hz) and harmonically complex stimuli. Complex stimuli
are the sum of 3rd (920 Hz), 4th (1150 Hz), and 5th (1380 Hz)
harmonics of a 230 Hz fundamental. Third and 5th harmonics
always begin in sine phase, however, 4th harmonic varies from
0 to 180 in 45 steps. Circles and dashed line superim-
posed on the complex stimuli approximate the variable onset
latencies of the largest initial stimulus component created by
phase shifting the middle harmonic. Stimulus duration is 17.4
msec (initial and final 2.18 msec attenuated by a linear ramp
function). Horizontal calibration is 5 msec.
microphone (2 ml coupler) and Hewlett-Packard
8052A impulse sound level meter. In the case of
complex stimuli the individual component fre-
quencies were calibrated separately. The final sig-
nal amplitude was controlled by a Hewlett-Packard
350D decibel attenuator. A control condition free
of audible stimulation was achieved by a 90 dB
attenuation of the signal.
Electroencephalographic (EEG) signals were re-
corded from an active electrode (10 m m gold disk)
placed on the vertex (Cz), referenced to the con-
tralateral (left) mastoid. The forehead served as
ground. Electrodes were applied with the aid of
skin prep paste, which typically resulted in inter-
electrode impedances below 2000 I2. The E E G
297
was amplified by a factor of 200,000, with an
analog bandpass of 30-3000 Hz + 3 dB.
Coherence between stimulus and response was
assured by a fast assembly language program that
alternated between digital-to-analog output of
stimulus wave form points and the time-correlated
analog-to-digital input of E E G data values. In the
case of FFRs, however, the initial 3 msec of E E G
input was discarded in favor of an additional 3
msec sample at the end of the response. No useful
F F R data was lost as a result of this maneuver
since the lower stimulus rise-times resulted in re-
sponse latencies even longer than the wave V
latency to click stimuli. An on-line artifact rejec-
tion routine repeated any stimulus for which ex-
treme E E G amplitudes (e.g., blocked amplifier,
movement) were sampled. Acceptable single trial
samples for each stimulus condition were stored
(1000 trials x 640 data points/trial) for subse-
Analysis
In the analysis of BAERs, evoked either by
click or as the initial response to tone stimuli, the
correlation template consisted of smoothed data
points of the BAER itself, centered on the peak of
wave V (cf., Galbraith 1984). For the FFR, how-
ever, the template consisted of the actual sine
wave stimulus for pure tones, and the 230 Hz
wave form for harmonically complex stimuli. Since
in the F F R analysis it was necessary to avoid the
initial BAER to sinusoidal stimuli, the template
was positioned beyond the B A E R and aligned
only with F F R waves. The center, or zero, lag
position to be used in LC was determined by
computing a lagged cross-correlation function be-
tween template and average FFR. Zero-lag align-
ment was achieved by shifting the template to the
position yielding the largest positive inter-correla-
tion between template and F F R (see Fig. 6).
All 1000 individual trials of an average re-
sponse were evaluated. For click-evoked BAERs
only wave V was analyzed. For FFRs, however,
separate analyses were performed for the initial
BAER, the entire FFR, as well as individual F F R
waves (3 waves for 230 Hz, 5 waves for 460 Hz).
Templates were incrementally shifted to lead and
lag the component being analyzed (36.4 #sec/lag;
298
G.C. GALBRAITH, W.S.BROWN
CROSS-CORRELATION OF BRAIN-STEM RESPONSES 299

_ 12 lags for wave V and 230 Hz, 5:6 lags for 460
Hz). The maximum shift in each direction was
equivalent to 10% of the template period. Cross-
correlations were computed at each shift. Thus,
for each single trial the largest correlation-squared
and associated lag were determined. Any trial
component was excluded from further analysis if
the selected lag was in an extreme position (i.e.,
first or last lag position). Such extremes are am-
biguous since the correlation function may still be
increasing. This may be considered a form of
off-line artifact rejection, since low correlation
values and extreme lags are more frequent in noise
trials (Galbraith 1984).
Revised components were obtained by (a) sim-
ply reaveraging only single trial responses free of
extreme lags in which r 2 exceeded 0.5 (CC only),
and (b) reaveraging these same responses after
adjusting for lag differences, i.e., realigning not on
the stimulus but on the lag associated with the
largest value of r 2 ( L C ) . It should be noted that
use of largest r 2, rather than r, is a conservative
approach that may select negatively correlated
trials. Such trials are presumably a result of E E G
background activity, or noise. Inclusion of nega-
tively correlated trials controls against the selec-
tion of only positively correlated trials, some of
which may also result from ongoing E E G noise
(summing only positively correlated trials would
guarantee larger amplitudes). Yet, when a BAER
or F F R signal is present there are in fact more
positive than negative trials, resulting in an ap-
propriate amplitude increase. Moreover, by limit-
ing lag shifts to + 10% of the wave period, centered
on the zero lag, the data should not shift out of
phase with the template, resulting in a spurious
negative correlation.
Mean differences in peak-to-peak amplitudes
between successive stages of the analysis were
evaluated by t test for correlated data (Guilford
1965). Significant departure of regression slopes
from a slope of 1.0 (indicating no change in ampli-
tude between stages of analysis) was also evaluated
by means of t test (Cohen and Cohen 1983). Since
multiple t tests were computed, the Bonferroni
correction ( P / N ) was applied; in the present case
the significance level is P < 0.01/6 = 0.0017.
Results
Fig. 2 illustrates the effects of selecting single
trial data that range in correlatio~a with the tem-
plate. The results show increasing amplitudes and
improved definition of components with increas-
ing correlation (Fig. 2A-D). Fig. 3 graphically
illustrates, for the BAER and a single F F R com-
ponent, the additional effects of latency com-
pensation adjustment beyond the signal enhancing
properties of cross-correlation analysis. The re-
suits following LC show improved clustering of
the evoked signals, i.e., for both the BAER and
F F R there is a more homogeneous overlapping of
individual responses. Compared with the CC anal-
ysis this will lead to an additional, albeit propor-
tionally smaller, gain in average amplitude.
Fig. 4 illustrates responses to the 3 different
types of auditory stimulus: clicks, fundamental
sinusoids, and complex sinusoids. The click-evoked
BAER (Fig. 4A) shows a sequence of waves which
appear normal both as to morphology and latency.
In response to sinusoid stimulation the initial
components resemble those of the click-evoked
BAER. Thus, waves III and V can be clearly
identified in Fig. 4B, and wave V is present in Fig.

Fig. 2. Superimposedsingle trial displays of individual wave components (left panel) and averageresponse (right panel) according to
magnitude of correlation with wave form template (smoothed wave V for BAER, sine wave cycle for each FFR wave). Correlations
are based on individual wave components rather than entire sweep following stimulus. Components are selected according to the
indicated ranges of r 2 values, but are graphed (left panel) separately for negative (NEG) and positive (POS) values of r. Phase
reversals between NEG and POS components are apparent; these greatly influencethe overall average. Generally darker densities for
POS components reflect a preponderance of positively correlated data (in this sample more BAER than FFR trials were selected,and
no negativelycorrelated BAER components were observed for the highest NEG r values; D, left). The FFR(1), FFR(2), and FFR(3)
labels associated with the average response (D, right) indicate successive waves of the FFR, each derived from a unique subset of
trials. Calibrations are 5 msec and 5/zV (right column; this same calibration is 10/~Vfor left column).
300 G.C. GALBRAITH, W.S. BROWN

enlla
III A
N~0 ml
.~.",~.., A.

CC
N.

%01m
i
B.

-
',~, e'~,~,,~: , ~,
l.C ., " ,,

i i, ~i I

Fig. 3. Illustration of wave V BAER (left) and first wave of the

FFR (right) selected by cross-correlation (CC) criteria, and the
same data following latency compensation (LC) adjustment.
V
C

4C a n d D. These B A E R c o m p o n e n t s occur several

m i l l i s e c o n d s later a n d are g e n e r a l l y less well de=
fined t h a n click-evoked c o m p o n e n t s due to the
m o r e g r a d u a l rise-time o f the d a m p e d sinusoid
stimulus. F o l l o w i n g the initial B A E R o n e observes
F F R waves that a r e d e l a y e d a p p r o x i m a t e l y 6 msec
relative to the stimulus a n d w h i c h persist a s i m i l a r
a m o u n t b e y o n d stimulus t e r m i n a t i o n (this con-
stitutes strong evidence that the F F R is n o t an
artifact). F o l l o w i n g at the f r e q u e n c y o f the stimu- D
lus results f r o m 230 H z a n d 460 H z p u r e t o n e
s t i m u l a t i o n (Fig. 4B a n d C), a n d a d o m i n a n t 230
H z r e s p o n s e results f r o m c o m p l e x s t i m u l a t i o n (Fig. I
4D). Fig. 4. Sample average responses from subject DA. A: click-
Fig. 5 p r e s e n t s a g r o u p - a v e r a g e d c l i c k - e v o k e d evoked BAER showing expected sequence of components (wave
B A E R a n d a n o - s t i m u l u s c o n d i t i o n . Fig. 5 A shows V latency = 5.8 msec). B: 230 Hz pure tone stimulus and FFR
response showing initial BAER components (wave V latency =
the original B A E R ( i n d i v i d u a l w a v e f o r m a n d
9.02 msec). Dashed line at left of response indicates that data
l a t e n c y differences slightly d i s t o r t the g r o u p - a v e r - sampling begins 3 msec after stimulus onset to allow additional
a g e d wave V m o r p h o l o g y ) . Fig. 5B shows results sampling of the end of response. Dots above individual waves
o f the c o r r e l a t i o n analysis a p p l i e d to the wave V illustrate delayed correspondence between peaks of stimulus
c o m p o n e n t . T h e C C stage o f analysis shows a and response. C: 460 Hz pure tone stimulus and response
(wave V latency = 8.11 msec). D: complex stimulus with 4th
p r o n o u n c e d a m p l i t u d e increase ( a p p r o x i m a t e l y
harmonic in sine phase (0 ) and response (wave V latency =
100%; Fig. 5B, d a s h e d line). T h e average n u m b e r 7.38 msec) showing dominant 230 Hz FFR. Vertical calibra-
of trials selected in the p r e s e n t analysis shows a tion is 0.25 gV, horizontal calibration is 5 msec.
CROSS-CORRELATION OF BRAIN-STEM RESPONSES
A and negative trials are approximately equal (46 vs.
+,/_
B 133/66
C yielding the largest positive correlation (maximum
D 46/52
Fig. 5. A: group-averaged ( N = l l ) BAER elicited by click
stimulus. B: group-averaged subset of data selected by cross-
correlation of wave V (dashed line) and following latency
compensation (solid line) analysis. Average numbers of posi-
tively ( + ) and negatively ( - ) correlated trials are shown. C:
group-averaged (N = 4) wave form of 'noise' trials in which
auditory stimulus is omitted. D: cross-correlation (dashed line)
and latency compensation (solid line) analysis of omitted
stimulus trials using wave V template when stimulus was
present (same template as in B). Average numbers of positiv,,iy
and negatively correlated trials are shown. All plots graphed on
same scale.
2 : 1 advantage (133 vs. 66) in favor of positively
correlated trials. Temporal realignment of the same
trials by LC (Fig. 5B, solid line) resulted in an
additional, but smaller, amplitude gain (approxi-
mately 5%).
Fig. 5C shows a group-averaged 'response'
when no stimulus was present (noise trials). As
expected, there is no demonstrable evoked activ-
ity. In order to carry out an analysis, since the
noise data do not yield a meaningful template, the
correlation template consisted of the subject's own
click-evoked BAER wave V. Obviously, some noise
trials will correlate with such a template by chance
alone. However, positive and negative correlations
should be equally distributed and thus effectively
cancel. This is evident in the results of Fig. 5D,
which shows very little coherent activity in the CC
301
and LC plots. Moreover, the numbers of positive
52), and within the range of the observed number
of negatively correlated trials when a signal is
present.
Fig. 6 illustrates the process by which the
sinusoidal stimulus template (Fig. 6B) is aligned
with FFR components of the original response
(Fig. 6C). A cross-correlation function (Fig. 6A) is
computed by systematically shifting the template
relative to the average FFR. The particular lag
alignment of template and FFR) defines zero lag
in the LC analysis. Fig. 6D shows individual F F R
components for the CC and LC analysis. As in the
1.0
A o.o
-0 5
o -11o I
I
I
I
I
C
I
3ms 8ms t
Fig. 6. Illustration of cross-correlation and latency compensa-
tion analysis for 230 Hz FFR. A: lagged correlation function
between 3 cycle sine wave template (Fig. 6B) and average FFR
response registered after the initial BAER (Fig. 6C). Vertical
dashed time line at maximum positive inter-correlation defines
zero lag alignment in subsequent latency analysis. B: aligned
230 Hz template (initial cycle that elicits BAER not included).
C: original FFR based on 1000 trials (BAER seen at the left of
the dashed line is analyzed separately). D: average of selected
FFR trials in which extreme lags are deleted and individual
component correlations (r 2) exceed 0.5 (dashed line), and
following latency compensation (solid line). Visible discontinu-
ities between successive waves emphasize that each wave is
analyzed separately and is based upon a different subset of
trials, Horizontal calibration is 5 msec (initial 3 msec of
response not sampled), vertical calibration is 0.25/~V.
302 G.C. GALBRAITH, W.S. BROWN

230 Hz ~ 112/57 ."x,.,~-r~'%. , 218/17a

460 Hz 1,,//,o3 ._ P , : - ,, , .... , /188

I
I
45

/
900
13,)/83 -f/'-~, >,-~,.....,.._"~. (~....,-..~ 198/182

135

I
1800
119171 .'/"'-,,,: -", ~."
~ ~. -""-.., 21o/168

No 5tim 17//21 ~ ~ .-~..,,..~..,.~,,..,.~.,.... " ~ , . . o . . . . , ~ . ' . , - . . , . , . . ~ . 144//143

Fig. 7. Group-averaged wave forms (I'4= 11, except for the no-stimulus condition in which N = 4). Left panel presents the original
FF1Ls based on 1000 stimuli aligned on the stimulus. Right panel presents the cross-correlation (dashed line) and latency
compensation (solid line) results. In this example BAERs are back-averaged from the first FFR component to maintain temporal
relationships between BAER and FFR. Vertical dashed line connects BAER wave V for harmonically complex stimulk Average
number of positively (+) and negatively ( - ) correlated trials that make up the respective wave forms are shown: left column of
entries refers to the BAER; right column refers to the FFR (averaged across subjects and FFR waves).

case of the click-evoked B A E R (Fig. 5), sizable
amplitude gains result in the CC stage, with ad-
ditional gains following LC.
Correlation analysis of the entire F F R wave
form yielded very few trials which exceeded the
selection criteria. Of course, the increased n u m b e r
of data points included in each correlation (N =
300 for the 460 Hz condition, N = 360 for all
other conditions) makes r 2 > 0.5 statistically less
likely. Indeed, in the whole wave analysis only 6
subjects yielded acceptable data, half of these in
only one stimulus condition. Due to relatively
incomplete whole wave data, only the individual
wave results will be considered further.
Fig. 7 (left panel) shows original group-aver-
aged F F R s for all sinusoid stimuli. Average F F R s
to the pure tone stimuli (top 2 traces) show well
defined following at the stimulus frequency. F F R s
to complex stimuli show a dominant 230 Hz peri-
odicity, with a visible trend towards reduced am-
plitudes for the 90 o stimulus. The b o t t o m trace
represents a no-stimulus control condition which,
as expected, is essentially flat.
Fig. 7 (right panel) shows the effects of re-
processing the data: CC amplitudes show large
increases over the original F F R , with LC showing
additional but smaller increases. The vertical
dashed line seen in the right panel connects initial
B A E R wave V peaks for the harmonically com-
plex stimuli (initial 3 msec of response not shown).
This pattern of latency shifts (longer for 0 and
45 o than for 135 and 180 o ) is virtually identical
to the pattern of initial stimulus maxima (see Fig.
1). The pairs of numbers in the right panel indi-
cate the average n u m b e r of positively/negatively
correlated trials. The left-hand column of num-
bers, referring to the initial BAER, are similar in
magnitude to the positive/negative ratio for the
click-evoked B A E R (Fig. 5). The number of trials
selected for waves of the F F R (right column) are
consistently larger than for the B A E R component.
Moreover, without exception there is a relative
CROSS-CORRELATION OF BRAIN-STEM RESPONSES 303

1.750 1.750
A. Wove V ,," / D. Wove V -"
//
1.500 ~' D~"/
1.250-

1.250. ,' /
0.750. (t=9.25) Md=,O8
1.000. (t=2.42)
// b=l O0 b=1.25"*
/ (t=~l) (t=5.32)
0.250 t t 0.750 r I t
0.250 0.750 1.250 1.750 0.750 1.000 1.250 1.500 1.750

> 0.800 B. 230 Hze,,,'" / 0.900 E. 2.30 Hz -_,~/

>
<1)
-1o
/,/ / 0.650- "O
'O'~/~/'
:'

0.550 O','","~MM
d
(3. =.17"* %_
E 0.400- ei / Mj=.08""
.< o.500 ;-./ !,=52!I. E '~ (t=4.85)
C) C) b=0.96
0 (t=5.52) _J ,, (t= 1.29)
0.050 i t 0.150 i
0.050 0.,300 0.5,50 0.800 0.050 0.500 0.550 0.800

C. 460 Hz ,,"/
0.800- ,,""
0.550800"F. 4 6 0 Z ~ , , H z=. ,,,e'"
///
0.550- ; ~
e,/"
0.500- .,~/ (t~1.98) 0.300 . ~ (t=1.19)
""7 b=1.11 -- ,~ b = 1.18"*
/ (t=0.67) 0.050 "~'/ t t(t=8"67) i
0.05C I i
0.050 0.300 0.550 0.800 0.050 0.,300 0.550 0.800
Initial A m p l i t u d e (IJV) CC Amplitude (pV)
Fig. 8. Scatter diagrams showing amplitude increases (in #V) from initial response to that following cross-correlation (CC) analysis
(left panel: A-C), and from CC to that following latency compensation (LC) analysis (right panel: D-F). Each plot shows a
unity-gain diagonal (solid line) wherein any data point above the diagonal indicates increased amplitude in the next stage of analysis;
a least-squares regression function (broken line) is fit to the data points. Summary statistics presented in each graph include the mean
amplitude difference (M d, in/W) and associated t test, and the regression slope (b) and associated t test of a departure from a slope
of 1.0 (Cohen and Cohen 1983). Since the Bonferroni correction is applied the significance level (* *) is P < 0.01/6 = 0.0017. BAER
measurement of wave V (A and D) is the average voltage of the rising (trough following wave III to peak) and failing (peak to
following trough) components. (One BAER subject is deleted due to unreliable data in which only 9 positively correlated trials passed
the cross-correlation criteria.) Measurement of the 230 Hz response (B and E) is the average peak-to-trough amplitude of 3 successive
waves; for the 460 Hz response (C and F) it is based on the average of 5 successive waves.

dominance of positively correlated trials, although
the differential between positive and negative tri-
Ms varies somewhat depending upon stimulus con-
ditions (greatest for 460 Hz pure tone, least for
90 complex). The no-stimulus response shows
low amplitude random waves, with fewer trials
selected than when a stimulus is present, and
virtual equality between the number of positive
and negative trials (144 vs. 143).
Bivariate scatter plots of the data are presented
in Fig. 8. The left panel compares initial ampli-
tudes (x-axis) of click-evoked BAER (Fig. 8A) or
304 G.C. GALBRA1TH, W.S. BROWN

pure tone average F F I s (Figs. 8B, C) with ampli-

tudes derived from t~: CC stage of the analysis
A 230 Hz SPECTRUM
(y-axis). The effect of this comparison is to quan-

lt
tify the generally consistent improvement in signal 1200 1
definition when extreme-lag and poorly correlated
trials are removed. The greatest overall amplitude
increase (0.61 ffV) occurs for the BAER (Fig. 8A).
This represents a 114% increase ( r a n g e = 3 2 -
177%). Thus the data clearly manifest improved
wave V definition following removal of unaccept-
able trials. CC amplitudes to the 230 Hz stimulus
are also significantly greater (0.17/~V) than origi-
nal average F F R waves (Fig. 8B). This represents 30011
a 68% increase (range = 26-135%). Moreover, the
regression slope (dashed line) is significantly c:
greater than 1.0, indicating that larger amplitudes
B. 4 6 0 Hz SPECTRUM
in the initial F F R yield proportionally larger am- o
plitudes in the CC stage of analysis. Mean and .m 90O
...o
slope differences are non-significant for the 460
{3
Hz condition (Fig. 8C) due to greater variability in
the data (mean increase = 30%, range = - 5 0 to 600
96%). r"
I---
The right panel of Fig. 8 compares amplitudes
bJ
from the CC (x-axis) and LC (y-axis) stages of rl
bo 300
analysis. The BAER (Fig. 8D) shows a pattern in
which the regression slope is significantly greater
than 1.0, but the overall difference in means (CC
vs. LC) is non-significant (mean increase = 6%,
Ld
rl
I--
Z
4 hdadh
(._9
range = - 5 to 22%). The interpretation of such a <
920+1150+1380 Hz SPECTRA
slope function, however, suggests a pattern in 1200 1 C.
which Ss with larger CC amplitudes do yield sig- FFR
nificantly larger LC amplitudes. For the 230 Hz CC
900 -}- LC
response (Fig. 8E) the pattern is reversed. Thus,
the slope does not differ from 1.0, but the mean
increase (0.08 #V) in LC amplitudes is statistically 600
significant (note the more restricted data range in
Fig. 8E as compared with 8D, which partially
explains why the same mean difference is signifi- 300
cant for the F F R but not the BAER). This repre-
sents a 23% increase (range = - 1 0 to 67%). The
460 Hz response (Fig. 8F) shows a pattern similar
oFIl
230 460 0 45 90 't35 180
to that in Fig. 8D, viz., non-significant mean dif-
Pure Complex
ference but a significant slope (mean increase =
5%, range = - 2 3 to 30%). Fig. 9. Spectral magnitudes in the 230 (A), 460 (B), and
Fig. 9 summarizes a spectral analysis of group- 920 + 1150 + 1380 (C) Hz bands. Group-averaged (N = 11) data
showing relative spectral intensities of original FFR as well as
averaged FFR, CC, and LC data for all stimulus following CC and LC analyses. Ordinate is in arbitrary units
conditions. Results are presented for narrow spec- proportional to voltage 2. Abscissa represents the various pure
tral bands centered at the 230 Hz (Fig. 9A) and tone and complex stimulus conditions.
CROSS-CORRELATIONOF BRAIN-STEMRESPONSES
460 Hz (Fig. 9B) pure tone frequencies, as well as
the sum of spectral activity centered in the 3
harmonic frequency bands that combine to form
the complex stimuli (Fig. 9C). Such an analysis
quantifies the auditory frequency response to the
various stimuli. Fig. 9A shows spectral power in
the 230 Hz band. As expected, there is substantial
activity in this band to the 230 Hz stimulus, and
minimal activity to the 460 Hz stimulus. Interest-
ingly, 230 Hz power elicited by the complex stimuli
equals or exceeds that resulting from the 230 Hz
pure tone in every case except for the 90 o stimu-
lus. Fig. 9B shows spectral power in the 460 Hz
band. There is dominant 460 Hz activity to the
460 Hz stimulus, but relatively little 460 Hz activ-
ity to any other stimulus. Fig. 9C, which presents
the combined spectral power in the bands of the
complex stimuli, shows that there is little or no
activity in the actual frequency range of the
harmonics contributing to the complex stimuli.
The utility of CC and LC analysis is illustrated
by an interesting comparison seen in Fig. 9A.
Spectral intensities of the original FFRs to the 460
Hz pure tone and the 90 complex stimuli are
virtually identical. Subsequent CC and LC
analyses of the 460 Hz data show no change,
confirming the absence of 230 Hz activity. Analy-
sis of the complex data, on the other hand, shows
a sizable increase in 230 Hz activity. These results
confirm that significant signal enhancement oc-
curs when actual signals are present, even though
the original responses may be very poorly defined.
Discussion
The present results demonstrate that click-
evoked BAERs and FFRs to pure tone and
harmonically complex sinusoids can be signifi-
cantly enhanced by means of data selection and
latency compensation algorithms. The largest gains
result when a subset of trials is selected that is
devoid of (1) extreme lags (the largest correlation
occurs in the first or last position of the lagged
correlation function), and (2) waves that correlate
poorly (r 2 < 0.5) with the template (cross-correla-
tion (CC) stage of analysis). Additional significant
gains may result when the selected individual tri-
305
ads are temporally realigned on the lag associated
with the largest value of r 2 (latency compensation
(LC) stage of analysis).
Brain-stem-evoked potentials suffer from re-
duced signal-to-noise, typically requiring a mini-
mum of 1000 stimuli (cortical evoked potentials,
by comparison, can be defined with fewer than
100 stimuli). The present results demonstrate,
however, that reliable and improved definition of
brain-stem responses can be obtained by using a
subset of approximately 20% of the individual trial
data. Moreover, there is internal validation of the
subset because of the 2:1 predominance of trials
that positively correlate with the template when-
ever a stimulus is present. Such analysis tech-
niques may prove useful in the study of difficult
subject populations, such as severely and pro-
foundly retarded persons.
The CC phase of the analysis yielded BAER
wave V amplitudes more than double that of the
original response, as well hs markedly increased
FFR amplitudes. Further processing by LC re-
suited in additional, but smaller, gains. There were,
in all measures, noticeable individual differences.
Even small LC differences may, however, reflect
subtle but functionally important variations in the
temporal stability of information processing in
auditory brain-stem pathways. A model of LC
effects in the FFR was tested by using sine wave
calibration signals with fixed amplitude and vari-
able onset latencies. The average amplitude of
such temporally 'smeared' trials is reduced, but
perhaps less than one might expect (LC analysis
of course restores the correct amplitude). Thus,
variance as large as 32% of the sinusoid period
resulted in only a 10% amplitude reduction. Such
results suggest that FFR amplitude is robust in
spite of considerable temporal variability in the
data. Thus, even small LC amplitude enhance-
ments might reflect important signal variabilities
which, in turn, could result in significant percep-
tual/cognitive errors. There is suggestive evidence
that temporal variability of brain-stem responses
may be altered in certain forms of mental retarda-
tion. Thus, Down syndrome individuals differ from
controls in their patterns of BAER variability
(Galbraith 1984).
The frequency-specific portion of the FFR is
306
typically preceded by BAER wave V, with earlier
BAER waves often visible as well (Figs. 2 and 4).
The current methodology successfully separates
initial BAER and later F F R waves of the re-
sponses to pure tones and harmonically complex
stimuli. Prior to this report the BAER occurring
with the F F R was not differentially treated. The
fact that wave V latencies in the F F R are longer
than for click-evoked BAERs by several millisec-
onds is due to properties of the sinusoid stimuli
such as relatively slow rise time and variable wave
form envelope. Indeed, the pattern of wave V
latencies elicited by the complex stimuli (dashed
line in Fig. 7) is virtually identical to the pattern
defined by the latency of the initial large ampli-
tude component of the complex stimuli (dashed
line in Fig. 1).
The presence of characteristic BAER waves, the
range of F F R onset latencies, and the duration of
the F F R response beyond termination of the
stimulus, all support the hypothesis that the F F R
has an upper brain-stem origin. However, the ob-
served morphological and temporal differences be-
tween BAER and F F R indicate that these two
responses are functionally (and perhaps anatomi-
cally) distinct. For example, there is a shifting of
BAER wave V latency to the complex stimuli
relative to subsequent waves of the F F R (Fig. 7,
right panel). Thus, successive waves of the F F R
do not appear to be repeated bursts of BAER
wave V. Moiler and Jannetta (1982) and Moiler
and Burgess (1986) implicate the lateral lemniscus
as the likely source of BAER wave V, rather than
inferior colliculus. This could also be a factor in
the FFR, but the evidence favors a collicular
origin (Smith et al. 1975). However, even if the
BAER and F F R share a common upper brain-stem
site of origin, it is probable that different popula-
tions of neurons are involved. At least at the level
of the cochlear nucleus, phase locking auditory
neural units are found primarily in the ante-
roventral division, terminating in unique and par-
ticularly secure synaptic endings known as the
Calyces of Held (Rose 1980).
The most obvious finding in the present study
is the consistent increase in amplitude when indi-
vidual responses are deleted which show extreme
lags or low values of r 2. Extreme lags in brain-stem
G.C. GALBRAITH,W.S. BROWN
responses are most often associated with noisy
trials (Galbraith 1984), and low correlation values
represent trial wave forms that differ from the
analysis template (i.e., the expected wave form).
Thus, although based on typically no more than
10-25% of the original trials, CC averages
dramatically enhance the signal definition. This
was most evident for the BAER (Fig. 8A), but was
also pronounced for the 230 Hz fundamental (Figs.
8B, and 9A) and complex sinusoidal stimuli (Fig.
9A). It is possible, of course, that too few trials
exceed the r 2 criterion to yield a reliable average
(this was the case for one subject in the BAER
analysis and was typical of whole wave LC analy-
sis with F F R signals). However, in most instances
it appears that this form of signal processing re-
sults in a significant improvement in signal defini-
tion.
It should again be stressed that a conservative
approach was used to control for inflated ampli-
tudes, viz., all negatively correlated trials were
summed if the r 2 value of that correlation (which
ignores sign) was the maximum and exceeded 0.5
on a given trial. This practice guards against the
spurious enhancement of amplitudes based on
random occurrences of correlated noise if only
positive trials are summed. With exception of the
no-stimulus condition there were considerably
more positive than negative trials, suggesting a
demonstrable signal-to-noise ratio. However, even
a few large amplitude negative trials can be dis-
ruptive. This may account for some CC ampli-
tudes being smaller than the original response
(e.g., some data points fall below the unity gain
diagonal in Fig. 8).
The F F R wave form is unique in that it closely
matches frequency and wave form characteristics
of the stimulus. Even in the case of complex
stimuli, based on the sum of 3rd, 4th (with vari-
able phase) and 5th harmonics, the F F R closely
approximates the wave form of the 'missing
fundamental' frequency. Thus, the actual cross-
correlation template is known a priori, and it is
not necessary to perform successive iterations in
which the template is modified as a result of
preceding analyses (termed 'adaptive filtering';
Woody 1967). Wastell (1977) and MiScks et al.
(1988) have criticized the iterative procedure, but
CROSS-CORRELATIONOF BRAIN-STEMRESPONSES
this is not a factor in the present analysis.
Mtcks et al. (1988) have proposed an elegant
method for estimating the latency shift of single
potentials (iterative Fisher scoring) which per-
forms significantly better than the Woody al-
gorithm. The standard Woody algorithm over-
estimated potential amplitudes, especially when
signal-to-noise ratios were low, since some back-
ground noise always correlates positively with the
wave form template. McGiUem and Aunon (1977)
utilized a technique which aligns individual com-
ponents of the signal, which may vary separately;
and Aunon and Sencaj (1978) found this approach
to be more effective than a more global analysis
by means of the Woody algorithm. Clearly, these
techniques represent substantive contributions to
the quantitative study of single trial variability
and could profitably be applied to the study of the
FFR. Although we have not used the techniques
proposed by McGillem and Aunon (1977) and
MiScks et al. (1988), we have nevertheless studied
individual FFR waves and have benefited from
the a priori knowledge of FFR wave form char-
acteristics.
Greenberg (1980) and Greenberg et al. (1987)
have shown that the FFR to the 90 o phase-shifted
stimulus is equal to the phase-coherent (0 )
stimulus when the signal is near FFR threshold.
Thus, at relatively low intensity levels the FFR is
not based on the envelope modulation pattern of
the acoustic wave form. At higher intensities (10
dB or more above threshold), however, the FFR
does follow the wave form envelope of the stimu-
lus. The present stimuli were well above threshold,
which accounts for the observed amplitude varia-
tions due to stimulus phase. Indeed, FFR re-
sponses to harmonically complex stimuli (Figs. 7
and 9A) show a consistent amplitude pattern that
covaries with the depth of stimulus envelope mod-
ulation, with the 90 amplitude at a minimum,
45 and 135 amplitudes intermediate, and 0
and 180 o at a maximum. The similarity of FFR
amplitudes to symmetrically phased stimuli
(0 /180 o, 45 /135 o ) also argues for the intrinsic
reliability of the present data. Finally, it should be
noted that response amplitudes to the 0 and
180 stimuli exceed that seen in response to the
230 Hz pure tone. These findings are in agreement
307
with those of Greenberg et al. (1987).
The complex harmonic stimuli were chosen in
part because they provide an extra challenge to
the signal extraction of responses having a brain-
stem origin. Our results, however, do not bear
upon the issue of whether frequency coherent ac-
tivity in the lower auditory system mediates per-
ception of the missing fundamental pitch. The
overall loudness of the stimuli (80 dB), as well as
the equal loudness of the various harmonics, means
that the FFR was determined by the shape of the
wave form envelope, driving units from broad
areas of the cochlear membrane rather than exclu-
sively those whose center frequency is in the area
of the fundamental (Greenberg 1980; Greenberg
et al. 1987). This relationship is easily seen com-
paring complex stimulus wave form envelopes (Fig.
1) with the evoked FFRs (Fig. 7) or FFR spectra
(Fig. 9A).
FFRs to broad-spectrum, pseudorandom noise
stimuli (consisting of as many as 127 harmonic
components, all with random phase) have been
effectively analyzed by means of cross-correlation
(Dobie and Clopton 1980; Dobie and Wilson 1984;
Wilson and Dobie 1987). The response in this case
is essentially an FFR to many pure tones pre-
sented simultaneously. Cross-correlating stimulus
and response wave forms clearly differentiated
artifact, cochlear microphonic, and neural compo-
nents at successively longer response latencies.
Moreover, it was possible to anal3/ze frequency
response characteristics by cross-correlating raw
responses obtained using broad-band stimuli with
band-limited stimulus subsets derived from the
'parent' stimulus. It should be noted, however,
that this is not analogous to the technique used in
the present study in which the missing fundamen-
tal wave form is cross-correlated with the FFR
evoked by complex stimuli (the missing funda-
mental is not a subset of the parent stimulus).
Also, Dobie and colleagues analyzed only the
averaged FFR wave form, rather than possible
trial-to-trial temporal variability in individual re-
sponses.
It appears that cross-correlation methods sig-
nificantly clarify and enhance the definition of
short latency potentials known to be of brain-stem
origin. It is felt that such improved signal defini-
308
t i o n will b e u s e f u l in c l i n i c a l a s s e s s m e n t s , e s p e -
c i a l l y in d i f f i c u l t r e c o r d i n g s i t u a t i o n s .
We express appreciation to Dr. James Smith and Vickie
McWain for technical assistance and to Dr. James Marsh for
helpful suggestions in preparation of the manuscript.
References
Aunon, J.I. and Sencaj, R.W. Comparison of different tech-
niques for processing evoked potentials. Med. Biol. Eng.
Comput., 1978, 16: 642-650.
Buchwald, J. Generators. In: E.J. Moore (Ed.), Bases of Audi-
tory Brain-Stem Evoked Responses. Grune and Stratton,
New York, 1983: 157-195.
Cohen, J. and Cohen, P. Applied Multiple Regression/Corre-
lation Analysis for the Behavioral Sciences (2nd Edn.).
Lawrence Erlbaum, Hillsdale, NJ, 1983: 54-60.
Derbyshire, A.J., Driessen, G.J. and Palmer, C.W. Technical
advances in the analysis of single, acoustically evoked
potentials. Electroenceph. clin. Neurophysiol., 1967, 22:
476-481.
Dobie, R.A. and Clopton, B.M. Auditory evoked responses
obtained by cross-correlation: a preliminary report.
Otolaryngol. Head Neck Surg., 1980, 88: 797-802.
Dobie, R.A. and Wilson, M.J. Short-latency auditory responses
obtained by cross-correlation. J. Acoust. Soc. Am., 1984,
76: 1411-1421.
Galbraith, G.C. Latency compensation analysis of the auditory
brain-stem evoked response. Electroenceph. clin. Neuro-
physiol., 1984, 58: 333-342.
Greenberg, S. Temporal neural coding of pitch and vowel
quality. UCLA Working Papers in Phonetics, 1980, 52:
1-183.
Greenberg, S., Marsh, J.T., Brown, W.S. and Smith, J.C. Neu-
ral temporal coding of low pitch. I. Human frequency-fol-
lowing responses to complex tones. Hear. Res., 1987, 25:
91-114.
Marsh, J.T., Brown, W.S. and Smith, J.C. Far-field recorded
frequency-following responses: correlates of low pitch audi-
tory perception in humans. Electroenceph. clin. Neurophys-
iol., 1975, 38: 113-119.
G.C. GALBRAITH, W.S. BROWN
McGillem, C.D. and Aunon, J.I. Measurement of signal com-
ponents in single visually evoked brain potentials. IEEE
Trans. Biomed. Eng., 1977, 24: 232-241.
MOcks, J., KShler, W., Gasser, T. and Pham, D.T. Novel
approaches to the problem of latency jitter. Psychophysi-
ology, 1988, 25: 217-226.
Moller, A.R. and Burgess, J. Neural generators of the brain-
stem auditory evoked potentials (BAEPs) in the rhesus
monkey. Electroenceph. clin. Neurophysiol., 1986, 65: 361-
372.
Mallet, A.R. and Jannetta, P.J. Evoked potentials from the
inferior colliculus in man. Electroenceph. clin. Neurophys-
iol., 1982, 53: 612-620.
Palmer, C.W., Derbyshire, A.J. and Lee, A.W. A method of
analyzing individual cortical responses to auditory stimuli.
Electroenceph. clin. Neurophysiol., 1966, 20: 204-206.
Pfurtscheller, G. and Cooper, R. Selective averaging of the
intracerebral click evoked responses in man: an improved
method of measuring latencies and amplitudes. Electroen-
ceph. clin. Neurophysiol., 1975, 38: 187-190.
Rose, J.E. Neural correlates of some psychoacoustic experi-
ences. In: D. McFadden (Ed.), Neural Mechanisms of
Behavior. Springer, New York, 1980: 1-37.
Smith, J.C., Marsh, J.T. and Brown, W.S. Far-field recorded
frequency-following responses: evidence for the locus of
brain-stem sources. Electroenceph. clin. Neurnphysiol.,
1975, 39: 465-472.
Smith, J.C., Marsh, J.T., Greenberg, S. and Brown, W.S. Hu-
man auditory frequency-following responses to a missing
fundamental. Science, 1978, 201: 639-641.
Wastell, D.G. Statistical detection of individual evoked re-
sponses: an evaluation of Woody's adaptive filter. Electro-
enceph, clin. Neurophysiol., 1977, 42: 835-839.
Weinberg, H. and Cooper, R. The use of correlational analysis
for pattern recognition. Nature, 1972a, 238: 292.
Weinberg, H. and Cooper, R. The recognition index: a pattern
recognition technique for noisy signals. Electroenceph. din.
Neurophysiol., 1972b, 33: 608-613.
Wilson, M.J. and Dobie, R.A. Human short-latency auditory
responses obtained by cross-correlation. Electroenceph. clin.
Neurophysiol., 1987, 66: 529-538.
Woody, C.D. Characteristics of an adaptive filter for the
analysis of variable latency neuroelectric signals. Med. Biol.
Eng., 1967, 5: 539-553.