PSVCHOPHYHOLOGY
© 1980 by Hie Society for Psychophysiological Research, Inc.
Vol. 17. N o . 2
Printed in U.S.A.
P300 and Stimulus Categorization: Two Plus
One is not so Different from One Plus One
RAY JOHNSON, JR. AND EMANUEL DONCHIN
Cognitive Psychophysiology Laboratory, Department of Psychology, University of Illinois, Champaign
ABSTRACT
Event related brain potentials (ERPs) were recorded from subjects who were instructed to count
one of three, equally probable tones presented in a random sequence. In another condition, the
subjects had to count one of two stimuli, one of which was presented with a probability of .33. The
data support the view that the pattern of variation of P300 amplitude with the sequential structure
of the series depends on the category to which events are assigned, rather than on the individual
stimuli eliciting the P300. Furthermore, the data support the idea that the amplitude of P300
elicited by task-relevant stimuli is determined by the subjective probability associated with the
eliciting event.
DESCRIPTIONS: ERP, P300, Subjective probability. Stimulus categorization.
The assessment of subjective probabilities is of
considerable importance in the study of the P300
component ofthe human event-related brain potential (ERP). When subjects are presented with a
series of Bernoulli events, the magnitude of the
P300 elicited by each stimulus is inversely related to
its prior probability (Duncan-Johnson & Donchin,
1977; Roth, Ford, Lewis, & Kopell, 1976; Tueting,
Sutton, & Zubin, 1970). In these studies, the variable controlling the magnitude of P300 amplitude
was assumed to be the prior probabilities of the
stimuli. Yet, the prior probability of a stimulus
accounts for only part of the variance of P300. In a
series of reports (Duncan-Johnson, 1978; DuncanJohnson & Donchin, 1977; Johnson & Donchin,
This research was supported by the Advanced Research Projects Agency's Cybernetic Technology Office, under ONR Contract #N-000-i4-76-C-0002 to E. Donchin.
The authors wish to thank Christopher D. Wickens, Connie C.
Duncan-Johnson, Gregory L. Chesney and Jack Isreal for their
helpful comments on the manuscript. A preliminary report of this
study was presented at the Seventeenth Annual Meeting of the
toteTi977'' ^'y*^^°P'^y'^°^*'S^'=^ ^"'^^*^' Philadelphia, OcT^- '
,,
.,. o
u
c r j,T
inepresent address ofthe first author is: Stanford University
1979; K. Squires, Petuchowski, Wickens, & Donchin, 1977; K. Squires, Wickens, N. Squires, &
Donchin, 1976), it has been demonstrated that the
P3(X) elicited by a sequence of Bernoulli events
varies from trial to trial despite constant prior probability. These studies further showed that this variability can be attributed to the specific sequence of
stimuli presented on the trials immediately preceding each stimulus.
To account for this variability in P300 amplitude,
K. Squires et al. (1976) proposed a model which
assumes that the subjective probability (or "expectancy'') associated with the outeome of each event
varies from trial to trial. According to this model,
the effect of prior probability on P300 is modulated
by an exponentially decaying memory trace for past
presentations of each Stimulus. Thus, a Stimulus (A)
induces an expectation, which decays over SUCCessive trials, that it will be repeated (AA). The subject's expectancy is further modulated, although tO
a lesser extent, by the occurrence of sequences of
stimulus alternations (i.e., ABABA). In this formulation, these two factors, which are related tO the
preceding Sequence of stimuli, combine with the
prior probabilities o f each stimulus to determine
,
,•
•
, i-,•
•
J
i
School of Medicine, Department of Psychiatry and Behavioral ^ ^ subjective probablhty assigned to each outcome
Sciences, Stanfoixl, CA 94305.
of a trial. There appears to be an inverse monotonic
Address requests for reprints to: Dr. Emanuel Donchin, Cog- relationship between the subjective probability of a
nitive Psychophysiology Laboratory, Department of Psychol- Stimulus and the amplitude of the P300 which it
ogy. University of Illinois, Champaign, Illinois 61820.
elicits.
167
0048-5772/80/010167-12$l .20/0
© 1980 The Society for Psychophysiological Research, Inc.
168
JOHNSON AND DONCHIN
Whereas the expectancy model formulated by K.
Squires et al. (1976) assumes that the ^ p l i t u d e of
P300 vanes with the subjective probablhty of HernouUi outcomes, it might be argued that a smaller
P300 is elicited by the last in a series of identical
stimuli because neural responses to repeating
stimuli are often diminished (Harris, 1943; Sharpless & Jasper, 1956). Explanations invoking noncognitive processes can be refuted by demonstrating
that physically identical sequences of stimuU will
,
j-iT
I
a
^
n'inn
i...- j
r ..t.
have diflFerent effects on P300 amplittide if the
instructions to the subjects are changed, or if the
information available to the subject is vaned. This
issue can be further clarified by considering an
experiment in which one of three, rather than one of
two, equiprobable stimuli can occur on any trial.
when the subject is told to count only one stimulus,
If the variance in P300 amplitude depends primarily
on the interaction between successive physical
stimuli, then such a series should be treated as a
^^,
•
i_ LI
^
jn-^nn
VoL 17, No. 2
Procedun
g^^jects counted the occurrences of 1000-Hz tones
^hich, in the two-stimulus session, were presented in
series with 1400-Hz tones. During the three-stimulus
session, the uncounted tones were 1400 and 1800 Hz. All
tones were 50 m^c in duration (10 msec rise/fall time),
and were presented at an intensity of 80 dB SPL (re .0002
dynes/cm^). The stimuli were mixed with a continuous
background of wide-band white noise at 55 dB SPL and
S^if??^ MnauraUy through Telephonies earphones
(TDH-39). Subjects were seated in a reclining lounge
^^^ ^ a dimly lighted room. A cross-hairfixationpoint
^ ^ continuously illuminated in one of the fourfieldsof
an Iconix tachistoscope.
The stimuli were presented in a random order at a fixed,
1350-msec interstimulus interval in blocks of 205 trials,
Subjectsreceived11 blocks of trials, each lasting apprt)ximately 5 min during both conditions. The sequence of
stimuli constituting each trial block was identical for all
subjects. Furthermore, the sequence of counted, relative
*? uncounted, stimuli was the same in both conditions.
Subiects Were given a bonus of 20 cents Der block of tiials
sequence of tiiree equiprobable outcomes, and P300 .^ ^^^^ they reported the correct number of 1000-Hz
amplitudes should be the same for all three stimuli. ^^^^^. ^^ JQ ^^^^ ^^^^ ji^^i^ ^^^^^ ^^ ^i^^i^ ^^^ ^f
Altemately, P300 amplitude may depend on the ^^e actual number. Four subjects received the two-tone
sequence of Bernoulli events with two possible cotidition during the first session, and 4 received it duritig
outcomes, a counted or an uncounted event, with the second. The two experimental conditions were prethe latter outcome comprising two distinct stimuli, sented in separate, 3-hrrecordingsessions.
In this case, the prior probabilities are .33 and .67
A
*
_ ,
,
,
,
Apparatus
for the counted and uncounted outcomes, resj^ctively. A larger P300 should be elicited by the
Burden Neurological Institute Ag-AgCl eiecti-odes
counted stimulus and a smaller P300 should be ^^""^ ^^®^ ^ ^^^ subject's scalp with collodion at F^,
elicited by both uncounted stimuli.
^ - ^"^ \}:^^t
"'^'^'** electrodes were used as a
r™.
.
J • J . ...1. reference. The subject was grounded with a forearm
Thepresentexpenmentwasdesignedtotestthese ^ , ^ ^ j ^ ^ ^ j ^ ^ electro-oculogjam (EOG) was recorded
two hypotheses by presenting subjects with a bet^^en supra- and sub-orbital electrodes. Beckman
stimulus series consisting of three, equiprobable biopotential electrodes were used for reference, ground,
tones, only one of which was to be counted. In this and EOG electrodes. The signals were amplified by Grass
way, it was possible to determine which categoriza- 7P122 amplifiers, set to a time constant of 2.5 sec (cf.
tion of the stimuli is the more potent determinant of Duncan-Johnson & Donchin, 1979), and upper halfP300 amplitude—the categorization established by amplitude cutoff of 35 Hz (3 dB/octave rolloff).
the subject's responses (count vs no count) or the
Experimental control and data acquisition were mancategorization established by the physical prop- aged with a PDP 11/40 computer (see Donchin & Heffley,
^r ^u *• 1
•
K^*u u A *
1975). The Signals from each of the four channels were
erties of the stimulus senes. Although data pre- ^. ^ ^ ^ ^ ^^^^ ^^ 200 samples/sec for a 1150-msec
sented by a number of investigators (Courchesne, ^^^ ^^^^ ^egan 150 msec prior to stimulus onset. The
Hillyard, «& Courchesne, 1977; Friedman, Sim- EEG, EOG, and ERP waveforms were monitored
son, Ritter, & Rapin, 1975; Kutas & Donchin, throughout the session.
1979) suggest that it is response-defined categoriza^
. ^ .
tions which determine die relationship between
P300 amplitude and prior probability, the relationFirst-order ERPs were computed by using data from all
ship between sequential expectancies and P300 ^^^ t"^*^ «" ^^^^^ ^ particular stimulus was presented.
»^«i;^,^<* ^^^A ^r^t u» *u^ oo«,z.
amplitude need not be the same.
^
first-order
Thus, in the two-stimulus condition, there were two
^ . j i-nr. i u i J .»*.»^ *u \nnn u *
A
ERPs, labeled A for tire 1000-Hz tone and
"a" for the 1400-Hz tone. Second-order ERPs were
Method
computed by sorting all trials according to the immediately preceding stimulus. There were four such seSubjects
quences (aA, AA, Aa, aa). Similarly, third-, fourth-, and
Eight University of Illinois students (5 females), aged fifth-order ERPs were computed. The three-stimulus
19-^25 yrs, were paid for their participation in the experi- condition yielded threefirst-orderERPs (A, B, and C for
ment. One ofthe subjects had previous experience in ERP the 1800-Hz tone), nine second-order ERPs (AA, BA,
experiments.
CA, BB, AB, CB, CC, AC, BC), and so on.
March, 1980
169
P300 AND STIMULUS CATEGORIZATION
Whereas all trials were considered in coding the
stimulus sequences, trials contaminated by EOG artifact
were not used in computing the averages. A trial was
rejected if at least five time-points were detected in which
the EOG signal exceeded a criterion value, determined
during preliminary work. One subject's data were eliminated from further analysis, since more than 40% of the
trials were rejected due to EOG contamination.
The magnitude ofthe P300 complex was expressed as a
discriminant score, computed by applying discriminant
functions (Donchin, 1969; K. Squires & Donchin, 1976).
See Donchin and HefBey (1979) for a discussion of the
rationale for using discriminant analysis in this context. A
step-wise discriminant function (Dixon, 1975) was developed for each electrode site, for each subject. The data
recorded in the two-stimulus condition served as a training set. Specifically, each function was based on the trials
associated with fifth-order "repetition-disconfirmation"
sequences (aaaaA) for the counted stimuli and
"repetition-confirmation" sequences (aaaaa) for the uncounted stimuli. The training sets were chosen to represent extremes of the expectancy continuum and, thereby,
to differ primarily in the magnitude of the P300 and slow
wave components. While the discriminant analyses
selected a few points from latencies outside of the P300slow wave region, the weightings of these points constituted a small fraction of those associated with the later
points. The distribution of points selected for all subjects
is shown in Fig. 1. This pattern is sitnilar to that reported
by Ehincan-Johnson and Donchin (1977), as well as K.
Squires and Donchin (1976).
The discriminant functions for each subject were
applied to the single-trial EEG activity in each condition.
The resultant discriminant scores were averaged over
electrodes according to preceding stimulus sequence.
These values are used in all tree diagrams displayed
below. In general, larger (more positive) scores indicate
greater amplitude in the P300 and slow wave components
of the waveform. Since the magnitudes of the P300 and
slow wave components were positively correlated, to
avoid circumlocution we refer to the discriminant score as
a measure of P300 amplitude. Statistical analyses were
applied to single-trial discriminant scores. A series of
paired comparisons was conducted in a nested factorial
design in which the factors were' 'subjects'' and' 'preceding stimulus." The individual trials served as a replication factor, which was nested in the other two factors
(Winer, 1971).
Results
Two-Stimulus Condition
O 6cr
^ 4-
The eflFect of the preceding sequence of tones on
the auditory ERP during the two-stimulus condition
is clearly shown in Fig. 2a, in which the grand-mean
(across subjects) ERPs at Cz are presented. The
ERPs elicited by the counted, target stimuli (A) are
superimposed on the ERPs elicited by the uncounted, non-target stimuli (a). The averages for
successively longer sequences of stimuli (higher
orders) branch from their respective lower-order
nodes. The corresponding grand-mean "discriminant score trees'' are shown in Fig. 2b. In all important details both the ERP waveforms and the discriminant score data replicated previous results from
this laboratory (Duncan-Johnson & Donchin, 1977;
K. Squires et al., 1976). The difiFerence between the
two first-order mean discriminant scores was significant (F = 127.68, MSe = 7.68) and refiects the
expected difference between P3OOs elicited by
stimuli with different prior probabilities. In addition, part of this effect reflects the fact that counted
' 'target" stimuli elicit a somewhat larger P3(X) than
equally probable non-target stimuli (cf. DuncanJohnson & Donchin, 1977). It is evident that when
either stimulus was preceded by a run of like stimuli
(e.g., AAAAA and aaaaa), P300 amplitude was
significantly smaller (Table 1) than that elicited by a
stimulus preceded by a run of unlike stimuli (e.g.,
aaaaA and AAAAa).^ Between these two extremes,
the ERP varied in a relatively systematic manner,
with the size of the P300 elicited by a particular
sequence varying as a function of the degree of its
similarity to either the repetition-confirmation or
disconfirmation sequences.
1 2
Three-Stimulus Condition
w 10
—
—
_
z
5 8
o
12 3 4
p
^z
BIN
1
2
3
A\
mi
12 3 4
12 3 4
r
Fz
^z
LATENCY RANGE
8 0 - 150 msec
151 -285msec
286 - 435 msec
4 3 6 - 690 msec
Fig. 1. Latency histograms of the first three points extracted
from the discriminant analyses for 7 subjects.
The data from the three-stimulus condition were
analyzed in two ways. First, the trials were coded by
the required response, treating the two uncounted
stimuli as members of a single category. The events
important to note that the relMiosships being tested in the
paired comparisons are specific to each sequential order and the
tests were only perfonned on all seqiKtntial orders to demonstr^e
that the results axe similarregardlessof the length of the preceding seqmnce.
170
JOHNSON AND DONCHIN
VoL 17, No. 2
aaaaa
Fig. 2a. Grand-mean (averaged over subjects) vertex ERP waveforms for each sequence in the two-stimulus condition. Solid
lines indicate ERPs to counted stimuli, dashed lines indicate uncounted stimuli. Positive voltages are represented as downward
deflections in this and all subsequent figures. The stimulus presentation is indicated by the block rectangle on the time scale.
thus constituted a Bernoulli series, with uncounted
and counted outcomes occuning at probabilities
of .67 and .33, respectively. Second, the trials were
sorted and averaged according to which of the three
stimuli was presented on each trial.
The grand-mean ERP waveforms and discriminant score trees for the data from the three-stimulus
session, treated as a Bernoulli series, are presented
in Figs. 3a and 3b. It is apparent that the amplitude
of P300 depended on the category to which a
stimulus was assigned regardless of the physical
parameters ofthe stimulus. Even though all stimuli
in this condition occurred equally probably, the
ERPs elicited by the non-target stimuli suggest that
the subject treated each uncounted stimulus as if its
prior probability were .67—^that is, twice its actual
value. Visual inspection of these data reveals a
marked similarity between the ERPs obtained in the
two- and three-stimulus conditions.
The discriminant score tiees for this condition are
March, 1980
1.40
I .20
1.00
171
P300 AND STIMULUS CATEGORIZATION
AAAaA
a Aaa A
aaA
oA
AaaaA
AAaaA
aaAoA
.80
UJ
cc
.60
.40
.20
0.0
E
aAAaA
aAaAA
AaAAA
aaAAA
AaAaA
AAAAA
aAAAA
AAaAA
AaaAA
aaaAA
AA
AAAa
- .20
O -.40
-.60
- .80
AaoAa
aaaAa
aAaAa
AAaAa
AAAaa
-1.00
-I 20
aAAao
aAaaa
oaAaa
Aaaoa
aaaaa
AAaaa
-I 40
-L60
aaa
3
3
ORDER
ORDER
Fig. 2b. Tree diagrams of grand-mean discriminant scores for each stimulus sequence in the two-stimulus condition. Data
for the counted stimuli (A) are on the left, and those for uncounted stimuli (a) on the right. Larger discriminant scores are indicative
of larger P3OOs.
TABLE 1
Paired comparisons on stimulus alternations versus repetitions: two-stimulus condition
Sequential Order
Second
Source
MS e
BA/AA
AB/BB
df
407. 02
81. 22
F
57 .08**
11.54*
(1/1372)
MS,
573.54
430.83
Fifth
Fourth
Third
F
MS.
86 .58**
65 .23**
(1/1372)
363.64
257.67
F
56 .19**
44 .20**
(1/602)
MS,
176.97
122.14
F
24 .21**
18 70**
(1/280)
*p < .001.
**p < .0001.
consistent with this interpretation. Again, as in the
average waveforms, the three-stimulus trees are
remarkably similar (except for the flattening of the
lower limbs of the uncounted tree) to the trees
obtained in the two-stimulus condition. To quantify
the apparent similarity of the data from the two
conditions, correlational analyses were perfonned.
The mean discriminant scores associated with the 31
possible sequences in each tree (from first- through
fifth-order) for the two-stimulus condition were
correlated with the corresponding values obtained
during the three-stimulus condition. Tlie correla-
tions for the counted and uncounted stimuli are
presented for each subject in Table 2. The high
correlations between the mean discriminant scores
from the two conditions further confirms that the
two uncounted stimuli were treated by the subjects
as members of a single category. It is, however,
necessary to (ktermine if the subjects ignored altogether the distinction between the two stimuli in
the uncounted stimulus category. This cannot be
determined from the above analysis, in which the
distinctiveness of the two stimuli was ignored.
The grand-mean waveforms, as well as the as-
172
JOHNSON AND DONCHIN
VoL 17, No. 2
Fig. 3a. Grand-mean vertex waveforms for the three-stimulus condition when the two uncounted stimuli were sequentially
coded as a single stimulus. Solid lines indicate ERPs elicited by counted stimuli, dashed lines indicate uncounted stimuli.
sociated discriminant score trees, for the three
stimuli are presented in Fig. 4. Since the total
number of nodes in these trees would be 363, a
prohibitively large value, only ERPs associated
with the first- through third-order sequences are
displayed in Fig. 4a. In addition, only the outer
limbs of the discriminant score trees are shown in
Fig. 4b. In this condition, the difference between
the first-order discriminant score means for the
counted and uncounted (stimulus B) stimuli, was
significant (F = 86.23, MSe = 10.91). The dif-
ference between the first-order means for the two
uncounted stimuli, however, was not significant (F
< 1, MSe = 10.59). When the upper limbs of the
tree are examined, it can be seen that runs of either
of the two uncounted stimuli prior to the counted
stimulus elicited nearly identical P3OOs. The P3OOs
elicited by the A stimulus were similar regardless of
whether it was preceded by a run of Bs or by a run of
Cs, and both were significantly different from the
P300 elicited by an A preceded by a run of As.
Finally, a run of As enhanced the P300s elicited by
173
P3(K) AND STIMULUS CATEGORIZATION
March,
1.60
1.40
aaaaA
1.20
aAaaA
1.00
aaaA
aaA
AaaaA
aA
aAAaA
.80
S .60
O
U) 40
I •"'
Z
0.0
I-
AAAaA
AAaaA
AaAAA
aaAaA
aAaAA
aaaAA
AaAaA'
AAAAA
AAOAA;
aMAA
AagAA
aaAAA '
AAA
AAAA
aAAA
AAAAo
AAAa
aAAAa
o
*5 -.40
o
AaAAa
aaAAa
AaaAo
- .60
oaaAa
AAaAa
AaAaa
AAAaa
aaAaa
aAAaa
AAaaa
Aaaaa
aaaaa
aAaaa
aAaAa
- .80
-1.00
-1.20
AoAa
Aaa a
aaaa
aaa
-L40
-L60
3
3
ORDER
ORDER
Fig. 3b. Grand-mean discriminant scores for each sequence in the three^stimulus condition. Data for the counted stimuli (A)
are on the left, and those for uncounted stimuli (a) on the right.
TABLE 2
stimuli (the lower limbs). If the physical distinction
Individual subject correlations between mean
between the two uncounted stimuli had no effect on
discriminant scores for all stimulus sequences in the the P300 associated with a B following a run of Cs
two- and three-stimulus conditions
(i.e., CCCB), then there should have been no disCorrelations
suDject
Counted
P.S.
J.B.
D.S.
J.D.
J.P.
S.L.
K.L.
.67***
-.23
.69***
.69***
.42*
.41*
.52**
Stimulus
Uncounted
.02
.73***
.54**
.68***
.71***
.62***
.53**
n = 31.
*p < .05.
**p < 005.
***p < .0005.
either a B or a C. The F values associated with the
paired comparisons on these data are shown in
Table 3.
A somewhat different pattern emerges upon examination of the limbs of the discriminant score
trees which are associated with the two uncounted
tinguishable difference between these P3OOs and
those elicited by a B following a run of Bs (i.e.,
BBBB). Identical results should have obtained for
the other non-target (C) stimulus. This, however,
was not the case.
When either of the two uncounted stimuli was
preceded by a run of like stimuli, it elicited a slightly
smaller P300 than was elicited when either stimulus
was preceded by a run of the other uncounted
stimulus. (This result accounts for the flattening of
the bottom of the discriminant score tree for the
uncounted stimulus as seen in Fig. 3b.) As Fig. 4b
shows, the mean discriminant score for the P3OOs
elicited by Bs which were preceded by a run of Cs
was larger than that computed for the P3OOs elicited
by those Bs which were preceded by a run of Bs;
albeit, both are substantially smaller than the discriminant score associated with the P3OOs elicited
by Bs preceded by a run of As. When the discriminant scores associated with the CB sequence are
compared to those found for the BB sequence,
significant differences are obtained at fourth-order
but not at third-order. When the BC ERPs are
174
VoL 17, No. 2
JOHNSON AND DONCHIN
BBX
COUNTED - 1000 Hz, X = A
UNCOUNTED - 1400 Hz, X ' 8
UNCOUNTED - 1800 Hz, X » C
Fig. 4a. Grand-mean vertex ERPs for the three-stimulus condition when all stimuli were coded individually. Solid lines indicate
ERPs to counted stimuli (A), dashed lines to uncounted, 1400-Hz stimuli (B), and dotted lines to uncounted, 18(X)-Hz stimuli.
The portrayal of three superimposed ERP averages necessitated the use of the " X " notation in the sequence labels. To determine
which averages are superimposed, each stimulus letter (A, B, C) is substituted in turn for " X . " For example, in the averages
denoted as "BX," the solid line represents the sequential average for the sequence BA; the dashed line represents the sequence
BB; and the dotted line represents the sequence BC.
compared to the CC ERPs, significant effects are
observed at third-order but not at fourth-order although in the latter case, the lack of a significant
difference appears to be due to the temporary increase in the P3OOs elicited by the CCCC sequence .^
It thus appears that the subject's response to uncounted stimuli is affected to some extent by preceding uncounted stimuli. The effects are, however,
rattier small, as can be seen from an inspection of
the waveforms in Fig. 3a.
Discussion
The data we present provide further evidence that
the amplitude of P3(X) is inversely proportional
to the subjective probability assigned to a taskrelevant stimulus. Furthermore, it is demonstrated
that even though the objective, prior, probability
remains constant over a series of trials, the subjective probabilities vary from trial to trial, depending
^These sequences cannot be expected to be significantly different from one another at second-order due to the fact that there
can be little difference between any two sequences at this early
stage.
on the specific sequence of stimuli preceding
each event (Duncan-Johnson & Donchin, 1977;
Duncan-Johnson & Donchin, 1978; Johnson &
Donchin, 1979; K. Squires et al., 1976,1977; Tueting et al., 1970). This study was undertaken to
determine if the subjective probabilities which
modulate P300 amplitude are those associated with
the specific physical stimuli or, rather, with the
classes into which stimuli are categorized by the
subject's task. To this end, we presented sequences
of three stimuli, only one of" which was to be
counted. The P3OOs elicited by the two uncounted
stimuli (/7 = . 33) were identical to the ERPs elicited
by a single uncounted stimulus which was twice as
probable. Thus, this trial-to-trial variability in P300
amplitude appears to be determined by the structure
of the task.
It is important to note that the subjects apparently
discriminated between the two uncounted stimuli.
This we infer from the f^t that, if either of the two
uncounted stimuli were preceded by a run of the
other, uncounted stimulus, it elicited a P300 which
was somewhat larger than when either of these
events was preceded by a run of itself. This effect on
March, 19m
1.60
P3(X} amiplitude, while small, is nevertheless reliable and consistent. Campbell, Courchesne, Picton,
and K. Squires (1979) have also found that physically different stimuli which carried the same
' 'feedback " to the subject were differentiated by the
subjects.
The data from two recent experiments supplement those of the present experiment (DuncanJohnson, 1978; Johnson, 1979) in demonstrating
that the effects of stimulus sequences on P300
amplitude reflect the subject's cognitions rather
than the activity of peripheral, stimulus-bound,
ads^tation or habituation processes. Our data constitute evidence that the brainresponseto seemingly
identical stimuli, presented in a random series,
governed by consistent and unchanging sequencegenerating rules, can vary with the immediately
preceding sequence of events. In other words, the
response to a constant environment appears to be
modulated by short-term, local perturbations of the
environment.
It is crucial to understand that the "sequential
effect" is the subject's response to the randomness
ofthe sequence, and that the effect will be abolished
in a non-random sequence. The model presented by
K. Squires et al. (1976), and supported in the papers
cited above, specifically suggests that the subject's
memory of the sequential structure of the series
affects P300 amplitude because the subject estimates the probability of stimuli on the basis of the
contents of this memory. In these terms, the sequen-
.40
BBA
BA
1 .20
BBBA
1 .00
.80
CCCCA
UJ
.60
to
.40
cc
o
o
.20
0.0
AAAAA
AAAAC
AAAAB
--.20
o
5^- 40
- .60
- .80
CCCCB
BBBBC
-1.00
-1.20
-I .40
175
P300 AND STIMULUS CATEGORIZATION
CCCCC
-I .60
ORDER
Fig. 4b. Grand-mean discriminant scores for selected sequences of the three-stimulus condition. Data for the counted
stimuli (A) are above the superimposed trees for the uncounted,
1400-Hz (B) and 1800-Hz stimuli (C).
TABLE 3
Paired comparisons on stimulus alternations versus repetitions: three-stimulus condition
Sequential Order
Third
Second
Fourth
Source
MS,
F
MS,
F
F
MS,
BA/AA
CA/AA
BA/CA
153.89
75.75
13.70
14.40***
7.50**
1.31
178.44
114.49
7.06
17.57****
11.93***
<1
47.70
11.54
12.31
5.32*
1.43
1.29
AB/BB
CB/BB
AB/CB
89.78
2.36
63.02
9.78**
< 1
6.20*
211.73
33.12
77.39
19.44****
3.52
7.24**
70.05
63.82
< 1
5.21*
6.41*
<1
AC/CC
BC/CC
AC/BC
34.50
0.22
29.23
3.14
< 1
2.64
253.84
68.14
58.95
25.58****
7.93**
6.04**
45.59
20.59
27.45
10.12***
2.35
2.91
(1/1036)
df
*p < .05.
**p < .01.
***p < .001.
****p < .0001.
(1/686)
(1/168)
176
JOHNSON AND DONCHIN
tial effects are merely extensions ofthe well established inverse relationship between the amplitude of
P300 and the probability of the eliciting stimulus,
The model asserts that the amplitude of P300 depends on the subjective probability (expectancy)
which is assigned to each event. The prior probability of events can be modulated by a variety of f«;tors
to yield the subjective probability. When events
arrive in a Bernoulli sequence, subjective probability is estimated on the basis of the immediate past
history ofthe sequence.
Courchesne (1978) recently reported results
which, in his opinion, "cast some considerable
doubt about the generality of their [K. Squires etal.,
1976] conclusion." Courchesne is careful to point
out tiiat his data are "not contradictory to the
specific findings in the K. Squires et al. (1976)
r e p o r t . ' ' ^ fact, he states in his discussion that P300
waves are affected by a variety of factors among
which a r e . . . sequential event structures. Yet, it is
instructive to examine the basis of Courchesne's
assertion. His subjects were shown sequences in
which either a letter or various colored pattems were
presented. The subjects were instructed to count one
of the letters. The probability of these targets
was. 12. Courchesne measured the amplitude of
P3OOs elicited by counted stimuli which were
preceded by 2, 5, 8, or 11 non-targets. These
amplitudes were in turn compared to those elicited
by the first target stimulus in the series, which was
preceded by an unspecified number of non-target
stimuli. Courchesne thus attempted to assess the
effect of the preceding sequence on P300 by studying die difference between the ERP elicited by the
first target in a series and the ERPs elicited by
targets preceded by a run of non-targets, i.e., he
analyzed the ' 'upper-limb'' of the expectancy tree
(i.e., the 3rd, 6th-, 9th-, and 12th-order repetitiondisconfirmation sequences). Since these analyses
failed to yield differences among the amplitudes
of the P3OOs, he concluded that there was no relationship between P300 amplitude and stimulus
sequence in his data.
This method of analysis is subject to two criticisms. First, Courchesne did not use a Bernoulli
series. His stimulus series was constructed such that
two targets were never presented in succession,
Thus, his series was similar to the series used by
Duncan-Johnson and Donchin (1978) to demonstrate that the sequential effects can be eliminated,
Second, an evaluation of the sequential effects requires a comparison between ERPs elicited by
repetition-confirmations and disconfirmations at
the same order. That is, one must compare corresponding nodes from the upper and lower limbs of
the tree. Only in this way can one assume an
equivalent memory decay factor for the two ERP
VoL 17, No. 2
waveforms being compared. Using Courchesne's
n^thod, a P300 elicited by a stimulus with an
unknown number of preceding non-targets (i.e..
memory ^ c a y ) is compared to a P300 elicited by
sequences containing stimuli which occurred up to
15 sec prior to the eliciting event. When P3OOs are
compared across sequential orders, what is being
assessed is, perhaps, the subject's meinory for past
events rather than the effect of preceding events on
the P300 elicited by the current event. It seems,
therefore, that Courchesne's conclusions can be
discounted. On the strength of the available data, it
is reasonable to conclude that the amplitude of
P300, elicited by a task-relevant event, reflects
trial-to-trial fluctuations in the subjective probability of events.
Such sequential dependencies have been Observed in many contexts. Yet, the nature of the
cognitive processes underlying their generation has
been the subject of controversy for more than a
decade. Hyman's thorough study ofthe relationship
between reaction time (RT) and the information
value of a stimulus first revealed that the RT to a
stimulus depends on the immediately preceding
stimulus (Hyman, 1953). He noted that successive
presentations of a stimulus yield faster RTs than
trials on which the stimulus is different from that
presented on the preceding trial. Bertelson (1961)
confirmed that RTs to repeated stimuli are shorter,
Bertelson's so-called "repetition effect" on RT
attracted considerable interest. Consequently, there
have been numerous attempts to define and understand its nature in the two-choice situation as well as
in paradigms utilizing multiple stimuli, multiple
responses, and various combinations of each (see
Audley, 1973, for a review),
It is instructive to examine the extent to which the
modulation of P300 amplitude by preceding sequences is analogous to the modulation of RT by
preceding stimuli. The form of the two relationships
is strikingly similar. The expectancy trees in the
present study, as well as the trees described by K.
Squires et al. (1976, 1977) and Johnson and Donchin (1979), are remarkably similar to the RT trees
reported by Remington (1969). The effect of stimulus repetitions on RT, as described by Bertelson and
others, is consistent with our interpretation of the
effect of repetitions on the amplitude of P300. The
reduction in P300 amplitude with stimulus repetitions indicates that subjects expect a stimulus to
repeat. If so, it is reasonable that the subject would
also be more prepared to either perceptually identify
or to respond to a repeating stimulus—Whence the
reduction in RT. This view predicts that the pattern
of P3(X) amplitude as a function of prior sequence
would be identical to that exhibited by RT. And, in
fact, this is the case: sequences which diminish
March. 1980
P300 AND STIMULUS CATEGORIZATION
177
P3(X) amph.t\i6&tendto shorten RT. While in most affect the RT, and thus iK^count for the effects of
of the previous research, the effect of only the stimulus sequence on RT. Laming's model emimmediately preceding stimulus (second-order ef- phasizes the subject's appraisal of the relative firefects) was analyzed, a few investigators have re- quency of the stimuli themselves as well as the
ported the same trends for higher-order sequences, relative frequency of stimulus alternations and repthroughfifth-order(Kirby, 1972,1976; Remington, etitions. The model presented by K. Squires et al.
1969, 1971). Moreover, the P300 results of the (1976) to account for changes in P300 amplitude
present experiment directly parallel the RT data was explicitly patterned after Laming's model. To
from a similar paradigm reported by Bertelson the extent that P300 reflects the assignment of
(1965).
subjective probabilities, these data provide strong
The pattern presented by the RT data is quite support for Laming's model. (For a complete recomplex. Yet, to the extenttiiatdata are available, it view of the evidence on the relationship between
appears that the basic pattern is affected by neither P300 and subjective probability, see Donchin,
the interstimulus nor the intertrial interval. This 1979, Duncan-Johnson, 1978, and Johnson, 1979.)
seems to be the case for P300 as well. McCarthy,
In summary, our data support the notion that the
Kutas, and Donchin (Note 1) varied the in- variability in P3(X) amplitude as a function of
terstimulus interval across series firom 1000 to 3000 stimulus sequence is the result of cognitive promsec. The sequential effects appeared to hold at all cesses rather than adaptation or habituation. Moreinterstimulus intervals. Similarity, Chesney and over, the similarity of results in both P300 and RT
Donchin (1979) described an experiment in which studies of how individuals formulate their subjecthe interval between stimuli was 3 sec, and still the tive expectancies suggests that the P300 can serve as
repetition effect dominated.
a useful adjunct to the more traditional measures of
Laming (1969) hypothesized that subjects esti- information processing. This is particularly true
mate probabilities of stimuli from the immediate when the use of an overt response may obscure the
past history of the series. These estimates in turn processes which are being investigated.
REFERENCES
Audley, R. J. Some observations on theories of choice reaction
potentials in psychiatry. New York: Plenum Press, 1979. Pp.
time: Tutorial review. In S. Komblum (Ed.), Attention and
13-75.
performance IV. New York: Academic Press, 1973. Pp. Donchin, E., & Heffley, E. Minicomputers in the signal509-545.
averaging laboratory. American Psychologist, 1975, 30,
Bertelson, P. Sequential redundancy and speed in a serial two299-312.
choice responding task. Quarterly Journal of Experimental Donchin, E., & HefRey, E. Multivariate analysis of eventPsychology, 1%1, 13, 90-102.
related
potential data: A tutorial review. In D. Otto (Ed.),
Bertelson, P. Serial choice reaction time as a function of re- Multidisciplinary perspectives in event-related brain potential
sponse versus signal-and-response repetition. Nature, 1965,
research. EPA-600/9-77-043, Washington, D.C: U.S. Gov206(4980), 217-218.
_
emment Printing OflSce, 1979. Pp. 555-572.
Campbell, K. B., Courchesne, E., Picton, T. W., & Squires, K. Duncan-Johnson, C. C. The P300 component of the cortical
C. Evoked potential correlates of human information processevent-related potential as an index of subjective probability
ing. Biological Psychology, 1979, 8, 45-68.
and processing duration. Unpublished doctoral dissertation,
Chesney, G. L., & Donchin, E. Predictions, their confirmation
University of Illinois, 1978.
and the P300 component. Psychophysiology, 1979, 16, 174. Duncjui-Johnson, C.C.,& Donchin, E. On quantifying suiprise:
(Abstract)
The variation in event-related potentials with subjective probCourchesne, E. Changes in P3 waves with event repetition: ability. Psychophysiology, 1977, 14, 456-467.
Long-term effects on scalp distribution and amplitude. Elec- Duncan-Johnson, C. C , & Donchin, E. Series-ba^d vs. trialtroencephalography & Clinical Neurophysiology, 1978, 45,
based determinants of expectancy and P300 amplitude. Psy754-766.
chophysiology, 1978, 15, 262. (Abstract)
Courchesne, E., Hillyard, S. A., & Courchesne, R. Y. P3 waves Duncan-Johnson, C. C , & Donchin, E. The time constant in
to the discrimination of targets in homogeneous and
P300 recording. Psychophysiology, 1979, 16, 53-55.
heterogeneous stimulus sequences. Psychophysiology, 1977, Friedman, D., Simson, R., Ritter, W., & Rapin, I. Cortical
14, 590-597.
evoked potentials elicited by real speech words and human
Dixon, W. J. BMDP-biomedical computer programs. Los
sounds. Electroencephalography & Clinical NeurophysiolAngeles: University of Califomia Press, 1975.
ogy, 1975, 38, 13-19.
Donchin, E. Discriminant analysis in average evoked response Harris, J. D. Habituatory response decrement in the intact
studies: The study of single trial data. Electroencephalogorganism. Psychology Bulletin, 1943, 40, 385-422.
raphy & Clinical Neurophysiology, 1969, 27, 311-314.
Hyman, R. Stimulus information as a determinant of resK^tion
Donchin, E. Event-related brain potentials: A tool in the study of
time. Journal of Experimental Psychology, 1953,45, 188human information processing. In H. Begleiter (Ed.), Evoked
1%.
178
JOHNSON AND DONCHIN
Johnson, R., Jr. Etectrophysiological manifestations of decision
making in a changing environment. Unpublished doctoral
disseitittion. University of Illinois, 1979.
Johnson, R., Jr., & Donchin, E. Subjective probability and P300
amplitude in an unstable world. Psychophysiology, 1979,16,
174. (Abstract)
Kirby, N. H. Sequential effects in serialreactiontime. Journal of
Experimental Psychology, 1972, 96, 32-36.
Kirby, N. H. Sequential effects in two-choice reaction time:
Automatic facilitation or subjective expectancy? Journal of
Experimental Psychology, 1976, 2, 567-577.
Kutas, M., & Donchin, E. Variations in the latency of P300 as a
function of variations in semantic categorizations. In D. Otto
(Ed.), Multidisciplinary perspectives in event-related brain
potential research. EPA-600/9-77-043, Washington, D.C:
U.S. Government Printing Office, 1979. Pp. 198-201.
Laming, D. R. J. Subjective probability in choice-reaction
experiments. Journal of Mathematical Psychology, 1969, 6,
81-120.
Remington, R. J. Analysis of sequential effects in choice reaction times. Journal of Experimental Psychology, 1969, 82,
250-257.
Remington, R. J. Analysis of sequential effects for a four-choice
VoL 17, No. 2
reaction txtot experiment. Journal of Psychology, 1971, 77,
17-27.
Roth, W. T., Ford, J. M., Lewis, S. J., & Kopell, B. S. Effects
of stimulus probability and task-relevance on event-related
potentials. Psychophysiology, 1976, 13, 311-317.
Sharpless, S., & Jasper, H. H. Habituation of the arousal
reaction. Brain, 1956, 79, 655-680.
Squires, K. C , & Donchin, E. Beyond averaging: The use of
discriminant functions to recognize event related potentials
elicited by single auditory stimuli. Electroencephalography &
Clinical Neurophysiology, 1976, 41, 449-459.
Squires, K., Petuchowski, S., Wickens, C , & Donchin, E. The
effects of stimulus sequence on event related potentials: A
comparison of visual and auditory sequences. Perception &
Psychophysics, 1977, 22, 31-40.
Squires, K. C , Wickens, C , Squires, N. K., & Donchin, E. The
effect of stimulus sequence on the waveform of the cortical
event-related potential. Science, 1976, 193, 1142-1146.
Tueting, P., Sutton, S., & Zubin, J. Quantitative evoked potential correlates ofthe probability of events. Psychophysiology,
1970, 7, 385-394.
Winer, B. J. Statistical principles in experimental design (2nd
ed.). New York: McGraw-Hill, 1971.
REFERENCE NOTE
1. McCarthy, G., Kutas, M., & Donchin, E. The effects of
interstimulus interval on sequential dependencies and P300.
Manuscript in preparation, 1980.
(Manuscript received April 11, 1979; accepted for publication October 30, 1979)
Announcement
ONE-WEEK COURSE AT MIT:
DESIGN AND ANALYSIS OF SCIENTIHC EXPERIMENTS
From June 23rd through 28th, 1980, Massachusetts Institute of Technology will offer
a one-week elementary course in Design and Analysis of Scientific Experiments. Applications will be made to the physical, chemical, biological, medical, engineering, and
industrial sciences, and to experimentation in psychology and economics. The course
will be taught by Professors Harold Freeman and Paul Berger. Further particulars may
be obtained by writing to the Director of the Summer Session, Room E19-356, Massachusetts Institute of Technology, Cambridge, MA 02139.