Illusory conjunctions in the time domain are errors made in binding stimulus features presented
In the same spatial position in Rapid Serial Visual Presentation (RSVP) conditions. Botella,
Barriopedro, and Suero (2001) devised a model to explain how the distribution of responses
originating from stimuli around the target in the series is generated. They proposed two routes
consisting of two sequential attempts to make a response. The second attempt (sophisticated
guessing) is only employed if the first one (focal attention) fails in producing an integrated
perception. This general outline enables specific predictions to be made and tested related to
the efficiency of focal attention in generating responses in the first attempt. Participants had to
report the single letter in an RSVP stream of letters that was presented in a previously specified
color (first target, T1) and then report whether an X (second target, T2) was or was not presented.
Performance on T2 showed the typical U-shaped function across the T1-T2 lag that reflects the
attentional blink phenomenon. However, as was predicted by Botella, Barriopedro, and Suero’s
model, the time-course of the interference was shorter for trials with a correct response to T1
than for trials with a T1 error. Furthermore, longer time-courses of interference associated with
pre-target and post-target errors to the first target were indistinguishable.
Keywords: illusory conjunctions, features binding, attention
Se llaman conjunciones ilusorias en el dominio del tiempo a los errores que se producen al
combinar rasgos estimulares presentados en la misma posición espacial en condiciones de
Presentación Rápida de Series Visuales (PRSV). Botella, Barriopedro y Suero (2001) han
formalizado un modelo para explicar cómo se genera la distribución de las respuestas de acuerdo
con su posición de origen respecto a la posición del blanco o diana en la serie. Propusieron
dos rutas que están constituidas por dos intentos secuenciales de alcanzar una respuesta. El
segundo intento (una forma de adivinación sofisticada) sólo se utiliza si fracasa el primero (basado
en la atención focal) en su intento por producir un percepto integrado. Esta estructura general
permite derivar y poner a prueba predicciones concretas relativas al rendimiento de la atención
focal en la generación de respuestas en el primer intento. Los participantes tienen que informar
de cuál es la única letra de una serie de letras mostrada en PRSV que apareció en un color
previamente especificado (primer blanco, T1) y después indicar si también se presentó una X
(segundo blanco, T2). El rendimiento con T2 a través de los diferentes desfases T1-T2 mostró
la típica función en forma de U que caracteriza el efecto de attentional blink (AB). Sin embargo,
tal y como se predice desde el modelo de Botella, Barriopedro y Suero, el curso temporal de
la interferencia fue más corto para los ensayos con una respuesta correcta a T1 que en los
ensayos con errores. Además, esos cursos temporales de la interferencia más largos, asociados
a los errores pre-blanco y post-blanco al primer blanco, fueron indistinguibles.
Palabras clave: Conjunciones ilusorias, pegado de rasgos, atención
Illusory Conjunctions in the Time Domain and the Resulting
Time-Course of the Attentional Blink
Juan Botella, Isabel Arend, and Manuel Suero
Autonomous University of Madrid
The Spanish Journal of Psychology Copyright 2004 by The Spanish Journal of Psychology
2004, Vol. 7, No. 1, 63-68 1138-7416
This research was supported by the Ministerio de Ciencia y Tecnología of Spain, project BSO2000-0112. We would like to thank
the two reviewers for their helpful comments.
Correspondence concerning this article should be addressed to Juan Botella. Facultad de Psicología, Universidad Autónoma de
Madrid, Campus de Cantoblanco, 28049 Madrid, Spain. E-mail: juan.botella@uam.es
63
The term illusory conjunctions in the time domain has
been proposed to describe the types of errors made in
binding stimulus features presented sequentially in the same
spatial location by means of the Rapid Serial Visual
Presentation (RSVP) technique (Botella, Barriopedro, &
Suero, 2001). In this procedure, participants must report
the feature from the to-be-reported dimension (response
dimension) of the only stimulus (the target) that has a
specific feature in the target-defining dimension (key
dimension). Errors are called pre-target and post-target
intrusions according to whether the feature which wasn’t
combined correctly with the target-defining feature was a
result of an item being presented before or after the target.
The distribution of intrusions for a given participant, or
group of participants, in a given experimental condition is
labeled as a pre-target or post-target pattern, according to
whether there is a predominance of intrusions from pre- or
post-target positions, or as a symmetrical pattern if there is
no predominance of either type (Botella et al., 2001). Most
research regarding this phenomenon is directed towards
explaining the behavior that determines the distribution of
intrusions around the target and why the empirical
distribution changes for different combinations of key and
response dimensions.
Botella et al. (2001) proposed a two-stage model to
account for patterns of illusory conjunctions over time, in
which responses are based on two consecutive attempts to
retrieve information from an RSVP stream. Two independent
but sequentially applied mechanisms are responsible for
both attempts. Although most components of the model have
received strong empirical support, there is still a lack of
empirical evidence for one of these components, namely,
the general architecture involved in the two sequential
response processes. In the present paper, we present new
evidence to support this architecture. First, we outline the
general architecture of the model, but include only those
aspects necessary to follow the rationale of the present
research (for a more detailed description, see Botella et al,
2001). Second, we explain how the general architecture of
the model involves a specific prediction related to the
detection accuracy of a second target. Then we describe the
experiment and discuss the results.
According to the model, two pre-attentive modules
routinely, and in parallel, extract the relevant dimensions of
all stimuli in the series; Module K searches for the key
target-defining dimension, and Module R records the
response features. When Module K detects the target-defining
feature, a third mechanism, focal attention, is triggered in
order to “capture” the target and integrate its features into
a single object or integrated perception. This is the first
attempt to generate a response, and a response is generated
if focal attention is successful. Given the short SOA (span
of attention) in most RSVP tasks, the focusing process is
assumed to be frequently interrupted by the presentation of
the items which follow in the series. However, the time
taken to develop an integrated perception by means of focal
attention can be characterized as a random variable. If the
integration process is completed rapidly, before the items
which follow interrupt the process, it produces a correct
feature combination. However, on trials in which focal
attention is not able to produce a response before it is
interrupted by the items which follow in the series, a second
attempt is made to trigger a response. Two main differences
separate both attempts. The first difference is that, whereas
the attempt based on successful focal attention is made on-
line, the second attempt can be considered an off-line
process, although the temporal delay is very short. The
second difference is that the second attempt is generated by
a fourth mechanism that takes into account the levels of
activation of the representations of the response features
from the items around the target and the target itself; it
employs that information to form a “sophisticated guess”
based on the application of Luce’s rule to those levels of
activation. For the purpose of the present research, it is not
important how the sophisticated guessing mechanism works.
The only important points are that this mechanism is
triggered only after focal attention has failed, and that it is
the process that can produce the erroneous combinations
we call illusory conjunctions.
An implication of the model, and its proposed general
architecture, is that trials in which the response is generated
in the first attempt (focal attention is successfully completed)
always produce correct responses. Trials in which the
response is generated by the second attempt can produce
either correct responses or intrusions. As a consequence,
any empirical distribution of intrusions can only partially
be assigned to the two response routes. Whereas all
intrusions are produced by the sophisticated guessing
mechanism (second attempt), correct responses are a mixture
of first and second attempt responses.
The main focus of the present research is to test a
prediction derived from the general architecture of Botella
et al.’s (2001) model. We used a combination of two effects
discovered by way of modern temporal methods for
studying attention (Shapiro, 2001). From our point of view,
a main difference between the two ways in which a
response is produced is the efficiency of the focal attention
process (“efficiency” here refers to the speed at which the
process is completed). Let’s suppose that we can order
trials according to the time invested in focusing attention.
If we divide this distribution arbitrarily in half, the lower
part of the distribution should include trials with shorter
values, and those trials are more likely to be solved in the
first attempt. In the upper part of the distribution, we have
the trials with longer values that should expose target items
to more interference from the items which follow in the
series. They are not likely to be identified in the first
attempt, so responses in those trials more likely result from
the sophisticated guessing mechanism on the second
attempt. This covariation makes the classification of trials
BOTELLA, AREND, AND SUERO
64
into correct responses and errors a credible reflection of
classification into two categories: fast and slow focusing
of attention.
The rationale of the experiment described here is based
on the assumption that the model is a good description of
how responses are produced in identifying feature-
conjunction targets in an RSVP stream. The differential
efficiency of the process of focusing associated with correct
responses and errors should show up if an appropriate
experimental task is designed. We have attempted to do
this by employing the Attentional Blink (AB) paradigm, in
which the identification of a first target (T1) impairs
performance in detecting a second target (T2) presented in
positions temporally close to it (Raymond, Shapiro, &
Arnell, 1992; Shapiro, Raymond, & Arnell, 1994). It is
assumed that the AB reflects the interference that focal
attention devoted to T1 produces on processing T2 (Isaak,
Shapiro, & Martin, 1999; Shapiro, Arnell, & Raymond,
1997) and that the AB deficit is a graded function of T1
processing difficulty (Chun & Potter, 1995; Seiffert &
DiLollo, 1997). In fact, the time-course of the attentional
blink reflects the extended time-course of T1’s demands
on focal attention (Chun & Potter, 1995; Ward, Duncan, &
Shapiro, 1996).
Our procedure consists of adding a second target (the
letter X), to the target search task in an RSVP stream which
the participant must detect after giving the response to the
first target (“identify the letter in a given color and then
report whether an X has been presented after the first
target”). Our prediction (derived from Botella, Barriopedro
& Suero’s model, 2001) is that, if we classify responses to
the first target as correct responses or intrusions we are, in
fact, classifying them also into trials that have associated
fast and slow focusing processes, respectively. As a
consequence, performance in detecting the second target
should reflect this fact. Specifically, if the shape of the AB
function reflects the time-course of the processes associated
with the application of Focal Attention to the first target,
then correct responses should be produced by a faster
attentional time-course than errors. This prediction agrees
with experimental results obtained with several variants of
the AB paradigm (Jolicoeur & Dell’Acqua, 1998; Ward,
Duncan, & Shapiro, 1996).
An associated, but not-so-obvious prediction, is that
intrusions coming from pre- and post-target positions should
show the same time-course, as they are all produced in the
second response-generating attempt via the sophisticated
guessing mechanism. This prediction is different from what
would be expected by other views. More traditional
explanations of conjunction errors (e.g., Lawrence, 1971;
McLean, Broadbent, & Broadbent, 1983) establish that errors
are produced because focal attention is applied at the wrong
moment (too early for pre-target errors, too late for post-
target errors). However, if this is true, the time course of
the mechanism generating errors should be different
according to the type of errors. Pre-target and post-target
intrusions should show, respectively, shorter and longer ABs
than correct responses, as earlier and later attentional
focusing, in an attempt to identify T1, should free attentional
focus for T2 earlier and later, respectively.
Notice that we use the AB phenomenon as a tool to
study how illusory conjunctions are produced. This is the
same logic (but in the opposite order) that Chun (1997) used
when he plotted the distribution of errors in perceiving the
second target as a function of the time course of the AB
(T1-T2 lag). He used the pattern of illusory conjunctions to
study how the AB is produced. We used the AB to study
how illusory conjunctions are produced.
The correct/error dichotomy in the first response is not
a pure classification for the efficiency of focusing in trials
(correct responses are composed of fast and slow trials,
whereas intrusions are all slow trials). However, if this
methodological flaw has any effect when trying to test
predictions associated to the efficiency of focusing, it must
consist in obscuring the difference. If even in these more
stringent circumstances an effect is still observed, a strong
argument could be made that the classification can be
used to separate successful and unsuccessful attentional
focusing.
Method
Participants
Eighteen staff members and graduate students participated
as volunteers. All of them had normal or corrected-to-normal
vision.
Apparatus
The experiment was run on an IBM compatible PC.
The experimental program was written with the MEL
program (Schneider, 1988).
Stimuli and materials
Four hundred and twenty series of 16 letters were
constructed. We will call the “critical set” the five items
composed of the target stimulus plus the two items before
and after it. The critical set occupied positions 4-8, 5-9 or
6-10. The letters were presented in five different colors (red,
green, white, yellow, and blue), and the background of the
screen remained gray throughout the experiment. All five
colors were used for the five stimuli in the critical set,
assigned randomly for each series. For the rest of the series,
the four colors not used for the target in that particular series
were employed, randomly, with the only restriction being
that there were never two consecutive stimuli with the same
color. Positions 6 to 8 were selected for the target in order
ILLUSORY CONJUNCTIONS AND ATTENTIONAL BLINK
65
to separate the critical set from the ends by several items.
This way, the data were not contaminated by the well-known
trend for participants to report the first and last items from
the series (Lawrence, 1971). Nevertheless, although the
participants were able to use the serial position as a cue to
select the response item, it cannot explain any eventual
difference in the comparisons to be made. The first target
(T1) was a letter that was defined by a color cued at the
beginning of each trial. The second target (T2) was the letter
X that could vary from position +1 to +7 after the first target.
To report the first target, participants received a menu with
six alternatives including the response features from the
target, the two items before and after the target, and the
option “don’t know.” The items were presented in the menu
in a random order (the use of the menu makes the
experimental paradigm more efficient, avoiding the tendency
for participants to be too conservative by selecting the “don’t
know” alternative frequently; see Botella et al, 2001; Botella
& Eriksen, 1992). To report if the X was present, participants
were required to press the number 1 on the keyboard for
“yes” and the number 0 for “no.” There were 48 trials for
each T1-T2 condition and 84 (20%) for the no-X condition.
In those trials in which the letter X was presented in
positions +1 or +2, it was replaced, in the menu, by the
letter from position +3, as participants were aware that T1
will never be an X.
Procedure
Participants sat with their eyes about 40 cm from the
screen. At the beginning of each trial a horizontal string of
three “&” signs appeared at the center of the screen, in the
position where the letters would appear and in the target
color for that trial. Participants began the series of letters
by pressing the space bar. Each colored letter remained on
the screen for 83 ms, then was immediately replaced by the
following one, and so on, until the end of the trial. Upon
completion of the series, a response menu for the first target
appeared, containing the letters that pertained to the critical
set, but presented in a random order along with the option
“don’t know” (except for +1 and +2 trials; see above).
Participants were required to report which letter appeared
in the color specified at the beginning of the trial. After
giving their response, the second target question appeared:
“Was an X present?” Participants were asked to press the
number 1 or 0 on the keyboard for yes or no answers,
respectively.
Results
Table 1 presents the average percentages of responses
to the item from each part of the critical set (the average
percentage of “don’t know” responses was 6.7%). The
distribution shows a pattern of predominance of post-target
intrusions, a result previously found in the “Report the letter
in this color” version of the task (Botella et al., 2001). As
a consequence, the average number of trials on which the
rest of the calculations are based, was 57, 231, and 86, for
the pre-target, correct, and post-target responses, respectively.
In order to test our main prediction, for each participant
we calculated the percentage of correct T2 detections
conditioned on the type of response to T1 (correct, pre-target
or post-target intrusion) for each T1-T2 lag. These data were
submitted to a 7x3 (lag by T1 response) ANOVA. There
was a statistically significant main effect of lag, F(6, 84) =
12.58, p < .001, and a marginally significant main effect of
response type F(2, 28) = 3.17, p = .058. More importantly,
however, is the presence of a significant interaction between
lag and response type, F(12, 168) = 2.66, p = .003. Figure
1 shows the nature of the interaction. As expected, the time
course of the attentional blink is different when participants
make a correct response to T1 in comparison to pre- and
post-target intrusions, whereas these two last types of errors
for T1 produce the same time course for the attentional
blink. In fact, when repeating the ANOVA with only the
pre- and post-target conditions, the interaction is not
significant, F (6, 84) = 0.606, p = .269.
BOTELLA, AREND, AND SUERO
66
Figure 1. Mean percentages of correct detections of T2 conditionalized
to the response to T1 (correct responses, pre-target intrusions or post-
target intrusions) as a function of T1-T2 lag.
Table 1
Mean Percentages of the Three Types of Responses to T1
Pre Correct Post
15.5 61.5 23.0
Pre
Post
Correct
0,6
0,5
0,4
0,3
0,2
0,1
0
1 2 3 4 5 6 7
T1-T2 LAG
P(T2/T1)
ILLUSORY CONJUNCTIONS AND ATTENTIONAL BLINK
67
In order to check whether performance in lag 1 is
different as a function of the type of response to T1 (different
size of lag 1 sparing) we submitted those data to a one-way
within-subjects ANOVA, which showed a significant effect,
F(2, 28) = 4.532, p = .02; post hoc comparisons showed
significant differences between hits and both types of
intrusions (p = .02 in both cases) but not between the two
types of intrusions (p = .45).
We also obtained separate cubic functions for each T1
response condition. In all three the fit was very good (R
2
:
.968, .996 and .964 respectively for pre-, correct, and post-
target conditions).
In order to rule out any explanation of the observed
effect based on bias shifts in detecting T2 across the
conditions, the false alarm rates were also analyzed. The
average rates were 18.6%, 17.1% and 19.3% for pre-, hits,
and post-target responses respectively; the differences were
not statistically significant, F(2, 34) = 0.273, p = .762.
Discussion
The present study was designed to test a prediction
derived from the model proposed by Botella et al. (2001) to
account for what they called illusory conjunctions in the time
domain. According to the model, participants can make two
sequential attempts to generate a target identification response
for a single target embedded in an RSVP stream of distracters
(the second one is activated only when the first one fails).
The first attempt is produced by focal attention and always
produces correct responses when attention is focused on the
target. The second attempt is generated by a sophisticated
guessing mechanism that samples the levels of activation of
the response features of the target, and the stimuli around it,
in the series. It can produce correct responses or the type of
erroneous combinations called illusory conjunctions.
According to the model, the classification of trials in the
correct/error dichotomy covariates with the efficiency of
focal attention. We are then able to make the prediction that
the interference (AB) produced upon detecting a second target
will show a different time course that has a faster recovery
when the response to the first target is generated from focal
attention than when it is based on sophisticated guessing.
Our results match this prediction. The AB shows a longer
recovery in trials where the T2 detection was conditioned to
an intrusion error on T1, than when it was conditioned to a
correct response. This is reflected in the significant interaction
between the T1-T2 lag and the type of response to T1. The
time-course of the attentional blink is different in that: (a)
lag 1 sparing (Chun & Potter, 1995; Visser, Bischof, & Di
Lollo, 1999) is larger following errors (performance is higher
in the shortest lag when T1 intrusions are recorded); b)
performance recovers earlier following correct responses. In
short, assuming Botella et al’s model (2001) is correct, when
focal attention succeeds (most probably, when it works
efficiently), there is a smaller amount of lag 1 sparing but a
speedier recovery from the AB than when the sophisticated
guessing mechanism is employed to produce the response.
Trials solved by the attentional focusing mechanism
hypothesized in Botella et al.’s model require less time to
produce a response, exactly as is predicted by the model.
Even more important is the fact that the time course of
interference produced on T2 in trials with pre- and post-target
intrusions to T1 are indistinguishable. According to the model,
all intrusions have their origin in the same route: the
sophisticated guessing mechanism triggered after focal attention
fails. Therefore, the time course of the interference over the
second target should be the same for pre-target and for post-
target intrusions. Confirmation of this result provides specific
evidence for the general structure of the model, as it accounts
for the idea that hits and intrusions are typically generated by
two different mechanisms. Also, the results indicate that
processes resulting in pre- and post-target intrusions are probably
similar in some respects, as they generated the same type and
amount of interference on T2 processing. This result is very
different from the predictions derived from alternative
explanations of illusory conjunctions (Lawrence, 1971; McLean,
Broadbent & Broadbent, 1983). According to these explanations,
errors are produced because focal attention is applied to the
wrong item. However, in that case, pre-target errors should
show a faster recovery from AB than correct responses, whereas
post-target errors should show a slower AB recovery.
In short, our results support the predictions made from
Botella et al.’s (2001) model; specifically, the general
structure based on two sequential attempts to generate a
response. As reflected in the differential time course of the
AB on a second target, success of focal attention in
identifying the first target results in quick recovery from T1
processing and shorter AB. If focal attention fails, however,
the sophisticated guessing mechanism takes more time to
pick a response, with the results that the AB effect is
lengthened in time, although it does not matter whether the
eventual response is pre- or post-target intrusions.
It is worthy to highlight that the present results concur
with the view of the AB being a result of a central
processing bottleneck (e.g., Jolicoeur, Dell’Acqua, &
Crebolder, 2001). Jolicoeur (1999) found a modulation of
AB as a function of the duration of T1 processing, by
manipulating the number of alternatives in a speeded choice
reaction task. We have shown that the central processing
stage responsible for the bottleneck is also involved in
processes that Botella et al (2001) related to focal attention.
References
Botella, J., Barriopedro, M. I., & Suero, M. (2001). A model of
the formation of illusory conjunctions in the time domain.
Journal of Experimental Psychology: Human Perception and
Performance, 27, 1452-1467.
