Content uploaded by Helena Matute
Author content
All content in this area was uploaded by Helena Matute
Content may be subject to copyright.
Memory & Cognition
2004, 32 (7), 1065-1075
Learning to predict future events from present events
is one of the most powerful adaptive tools, since it allows
an organism to find the necessary resources for survival
and to avoid dangerous situations. Given its importance,
this kind of predictive learning was the central focus of
animal behavior research throughout the twentieth cen-
tury. During the last decades, predictive learning has also
become important in the area of human cognition, where
it has given rise to a great amount of empirical and theo-
retical research.
The vast amount of evidence provided by this research
has sometimes turned out to be quite difficult to explain by
the available theoretical approaches. Many variables usu-
ally neglected by theoretical models influence the process
of human learning of predictive relations among events or
the way in which humans use the acquired information.
Among other things, it has been shown that the probe
question used to assess participants’ judgment (Matute,
Arcediano, & Miller, 1996; Matute, Vegas, & De Marez,
2002; but see also Cobos, Caño, López, Luque, & Al-
maraz, 2000), the perceived properties of the cue and the
outcome (De Houwer, Beckers, & Glautier, 2002; Mitchell
& Lovibond, 2002), and the frequency with which judg-
ments are emitted (Catena, Maldonado, & Cándido, 1998;
Collins & Shanks, 2002; Matute et al., 2002) can be re-
sponsible for the differential results observed.
Several experiments have shown that predictive and
causal judgments—that is, the subjective assessment of
the relationship between two events—can be altered de-
pending on the frequency with which judgments are pro-
duced (see, e.g., Catena et al., 1998; Collins & Shanks,
2002; Matute et al., 2002; but also see Wasserman, Kao,
Van Hamme, Katagari, & Young, 1996, who observed
the same result regardless of whether participants re-
sponded in every trial or only at the end of training).
Consider, for example, a situation in which participants
receive a series of cue–outcome trials followed by a se-
ries of trials in which the cue is no longer followed by the
outcome. If participants are required to make their judg-
ments while they are still receiving the information, their
judgments will be strongly influenced by the most recent
information received (i.e., recency effects will be ob-
served). In contrast, if participants must make their judg-
ments only after seeing all the information, their evalua-
tions will match the overall statistical relation between
the events quite well. That is, they will tend to integrate
(or average) the information received during the two
phases. In some studies (e.g., Dennis & Ahn, 2001), a
primacy bias (i.e., a tendency to give more importance to
the evidence presented in the first phase) has even been
observed in this situation.
This response mode effect is important because it can-
not be easily explained by most theoretical models of
causal and predictive learning. On the one hand, asso-
ciative models (e.g., Dickinson & Burke, 1996; Rescorla
& Wagner, 1972; Van Hamme & Wasserman, 1994) as-
sume that the learning process consists in the progressive
strengthening or weakening of the association between
the mental representation of the cue and the outcome.
According to these models, the predictive value of a cue
is updated after every trial. This updating process ac-
counts quite well for the recency effect observed in ex-
periments using a trial-by-trial response mode and, as-
1065 Copyright 2004 Psychonomic Society, Inc.
Support for this research was provided by Grant PI-2000-12 from De-
partamento de Educación, Universidades, e Investigación of the Basque
Government to H.M. M.A.V. and S.V. were supported by F.P.I. Fellow-
ships (BFI01.31 and BFI00.138, respectively) from the Basque Gov-
ernment. We thank Lorraine Allan, Leyre Castro, Jan De Houwer, and
Gabriel A. Radvansky for their helpful comments on an earlier draft, and
Fernando Blanco and Ann Meulders for their assistance in conducting
the third experiment. Correspondence concerning this article should be
addressed to H. Matute, Departamento de Psicología, Universidad de
Deusto, Apartado 1, 48080 Bilbao, Spain (e-mail: matute@fice.deusto.es).
Frequency of judgment as a context-like
determinant of predictive judgments
MIGUEL A. VADILLO, SONIA VEGAS, and HELENA MATUTE
Universidad de Deusto, Bilbao, Spain
Several studies have shown that predictive and causal judgments vary depending on whether the
question used to assess the relationship between events is presented after each piece of information
or only after all the available information has been observed. This effect could be understood by as-
suming that in the two cases people perceive that the test question requires that different sets of evi-
dence be taken into account. This hypothesis is tested in the present experiments through contextual
manipulations that take place at the time of training and at the time of test. Our results show that peo-
ple use this contextual information to infer which set of events should be considered when making their
subjective assessments. The results are at odds with current theoretical approaches, but it is possible
to develop mechanisms that would allow these models to account for the observed evidence.
1066 VADILLO, VEGAS, AND MATUTE
suming certain parameter values, it could also account
for the accurate judgments obtained when a global re-
sponse mode is used. But it is hard to see why these pa-
rameters should vary depending on the response mode.
On the other hand, statistical models view the process
of predictive learning as the computation of the contin-
gency between the cue and the outcome (Allan, 1980;
Cheng, 1997; Cheng & Novick, 1992; Jenkins & Ward,
1965). According to this perspective, people store all the
relevant information presented during the two training
phases and, when asked to rate the relationship between
the cue and the outcome, they simply use this stored in-
formation to compute the relevant statistical measure.
The predictions of these models match quite well the re-
sults obtained when a global response mode is used. How-
ever, they cannot account for the recency effects ob-
tained with a trial-by-trial procedure, unless it is assumed
that the evidence presented during the first phase of
training is neglected in this case (see Matute et al., 2002).
A potential explanation for the divergent results ob-
tained when the response mode is manipulated is pro-
vided by belief revision models, which state that the in-
formation available is processed and stored differently in
each case (Catena et al., 1998; Collins & Shanks, 2002;
Hogarth & Einhorn, 1992; Pennington & Hastie, 1992).
According to these models, each judgment is produced
taking into account the evidence presented between the
last judgment and the present judgment. This evidence is
used to compute the strength of the relationship between
the cue and the outcome and to update the assessment
made in the last judgment accordingly. Once a judgment
is made, all previous information is lost. When only one
judgment is required, this procedure leads to an integra-
tive judgment, since all the information will be consid-
ered. By contrast, when several judgments are required,
this procedure predicts that older information will be lost
as a function of the updating process and, thus, the judg-
ment will be highly influenced by the most recent infor-
mation received.
Even though this perspective provides a better account
than most statistical and associative models, it also has
some limitations. For example, Matute et al. (2002, Ex-
periments 1A–1C) showed that not only the response
mode but also the type of question used (e.g., predictive
vs. causal) could yield quite different results. If the re-
cency effect obtained with trial-by-trial procedures is
due to a particular information-processing and -storing
strategy, it is not clear why this effect should disappear
as a function of test question. In a different experiment,
Matute et al. (2002, Experiment 4) also observed that the
recency effect typically observed in trial-by-trial groups
can be counteracted if participants are exposed to an in-
structional screen just between the last training trial and
the test trial. After these instructions, an integrative
judgment, rather than a recency-based one, can be ob-
served. This result implies that even if participants have
been exposed to a trial-by-trial procedure, they have ac-
cess to all the information received, and not just to the in-
formation received between the last and the present judg-
ments, as suggested by belief revision models.
According to Matute et al. (2002), people do store all
the information received, and the divergent results ob-
tained using trial-by-trial and global response modes
occur because of the different interpretation of test de-
mands that participants make in each case. When partic-
ipants are asked only to make a single assessment at the
end of training, by default they interpret that this question
refers to all the information presented during training
and, thus, they make an integrative estimation. On the
other hand, when a judgment is required after each piece
of information is presented, people tend to perceive that
they are being asked to assess the current status of the
cue, which makes them focus on the most recent trials.
The advantage of this explanation is that it accounts
not only for the response mode effect, but also for the ef-
fect of other testing variables, such as the type of ques-
tion or the role of instructions, by the very same mecha-
nism (Matute et al., 2002). Different probe questions can
give rise to different results under a trial-by-trial re-
sponse mode because some questions (e.g., predictive
questions) make participants think that the most recent
information is most relevant, whereas others (e.g., causal
questions) suggest that all the information should be
taken into account. As an example, when contingencies
change over time and people have to make frequent pre-
dictions about the expected occurrence or nonoccur-
rence of the outcome, they tend to make predictions that
correspond to the current state of affairs; it would not
make much sense to make predictions on the basis of how
things used to be in the past. By contrast, when people
are asked to rate the causal relationship between the cue
and the outcome, it would not make much sense to use
only the most recent information, since causal relation-
ships are supposed to be relatively constant over time; in
this case people tend to give more integrative responses.
In the same framework, the effect of an instructional
screen presented just before testing becomes straightfor-
ward. Depending on what information people are re-
quested to consider at test, they will demonstrate one or
the other type of effect, thereby showing that they have
not lost the older information.
A testable prediction of this view is that any manipu-
lation that affects the perceived demands of the test ques-
tion will have an impact on the subjective assessments
reported. When asked to make a judgment about the pre-
dictive relationship between a cue and an outcome on the
basis of ambiguous evidence (e.g., when the relation be-
tween the cue and the outcome is not stable and changes
with time), participants must infer whether they are being
asked about the general relationship (in which case they
should take into account all the information available) or
about any of the momentary relationships (in which case
they should take into account only the information pro-
vided during the relevant block of trials). Therefore, if
participants have access to any type of information that
might help them decide which of the different sets of in-
PREDICTIVE JUDGMENTS 1067
formation should be considered at test, that information
will modulate their judgment.
An interesting way in which this hypothesis can be
tested is by studying how the manipulation of the re-
sponse mode interacts with contextual manipulations.
Contextual manipulations are often used in animal learn-
ing research (for a review, see Bouton, 1993). In order to
help nonhuman animals disambiguate the information
provided by changing contingencies, researchers in the
animal learning tradition have often introduced contex-
tual cues that signal different sets of trials containing
contradictory information. These contextual manipula-
tions have recently been shown to affect human causal
and predictive judgments as well (Matute & Pineño,
1998; Pineño & Matute, 2000; Vila & Rosas, 2001).
According to our hypothesis, participants responding
on a trial-by-trial basis will no longer make their default
recency-based estimation if the contextual cues pre-
sented at test suggest that earlier information is more rel-
evant than recent information. Similarly, the integrative
or intermediate judgments usually observed in global
procedures will not be observed when the contextual
cues presented at test indicate that one set of trials is
more relevant than the other. Thus, these experiments
will test whether the information provided by contextual
cues can counteract the response mode effect.
EXPERIMENT 1
Matute et al. (2002) suggested that, when a trial-by-
trial response mode is present, recency effects appear be-
cause people perceive that the final test question is just
one more instance of the same question, identical to that
presented in the preceding trials. Because nothing in this
question indicates that it is different from the previous
ones or that it refers to the evidence in its entirety, peo-
ple assume it refers to the most recent block of evidence.
If this is true, it follows that if something changed at the
time of test, people would notice that the test question
was different from the previous ones—in other words,
that it does not refer only to the information most re-
cently presented. In this case, trial-by-trial participants
could be made to make an integrative judgment instead
of their default recency-based response.
This prediction is tested in Experiment 1, in which three
groups of participants received the same cue–outcome
pairings but made their estimations in different contexts
and with different response modes. Two of these groups,
Global-111 and TT-111, were the standard global and
trial-by-trial groups, respectively. For these groups, the
context remained unchanged during two training phases
and a test phase. We use the term “111” to indicate that
the context was the same in these three phases. These
groups should replicate the basic response mode effect
(see, e.g., Matute et al., 2002). The third group, TT-112,
was a trial-by-trial group for which the test phase took
place in a novel context, different from that of the two
previous training phases. The term “112” is used to in-
dicate that the context is the same for the two training
phases but different for the test phase. If our hypothesis
is correct, this group should behave at test in a way sim-
ilar to that of the global group, rather than show the re-
cency bias that is usually observed in the standard trial-
by-trial group.
Method
Participants and Apparatus
. Fifty-four undergraduate students
from Deusto University volunteered for the experiment. Random as-
signment resulted in 17 participants in Group TT-111, 18 participants
in Group TT-112, and 19 participants in Group Global-111. All the
participants performed the experiment at the same time in a large com-
puter room. Adjacent participants were seated about 1.5 m apart from
each other and were exposed to different experimental conditions.
Design
. The design summary is shown in Table 1. During train-
ing, all groups of participants received two blocks of trials. During
the first block, they were exposed to 20 trials in which the cue was
always followed by the outcome (C–O). They then saw another 20
trials in which the cue was no longer followed by the outcome
(C–noO). The only differences between the three groups were the
response mode and the contextual manipulations at the time of test.
Two of the groups responded in a trial-by-trial mode. For one of
them, all three phases took place in Context 1 (Group TT-111); for
the other, Phases 1 and 2 took place in Context 1 and the test phase
took place in a second context (Group TT-112). The third group was
a global group for which the three phases took place in the same
context (Group Global-111).
Procedure
. The experimental task was a computerized version
of the standard allergy task used in experiments on causal and pre-
dictive learning (see, e.g., Wasserman, 1990). The participants were
shown the medical records of fictitious patients. The record of each
patient consisted of two consecutive cards. The cue card was pre-
sented first in each trial and indicated whether or not that patient
had taken a fictitious medicine called Dugetil. The outcome card
then indicated whether or not the patient had developed an allergic
reaction to the medicine. Contextual cues were text labels that were
presented on each card and given symbolic contextual value
through instructions, as is usually the case when such manipula-
tions are performed with humans (see, e.g., Vila & Rosas, 2001).
The two contexts used in this experiment (i.e., “Fifteenth Century”
and “Sixteenth Century” text labels that appeared at the top of the
screen) were counterbalanced across participants.
1
The instructions
were presented on two consecutive screens. An English translation
of the Spanish instructions is offered in Appendix A.
In each trial, the participants from Groups TT-111 and TT-112
were asked to rate the degree to which they thought that the current
patient would develop the allergic reaction. This question was always
presented on the same card that indicated that the patient had taken
Table 1
Design Summary of Experiment 1
Training
Group Response Mode Phase1 Phase 2 Test
Global-111 Global (C–O)1 (C–noO)1 (C)1
TT-111 Trial by trial (C–O)1 (C–noO)1 (C)1
TT-112 Trial by trial (C–O)1 (C–noO)1 (C)2
Note—The cue (C) represents a fictitious medicine presented in every
trial; the outcome is an allergic reaction that is present during Phase 1
(O) but not during Phase 2 (noO). The first and second phases of the
study always occurred in the same context. The test trial took place in
the same context for Groups Global-111 and TT-111 but in a different
context for Group TT-112. The two contexts are represented by the
number outside the parentheses.
1068 VADILLO, VEGAS, AND MATUTE
Dugetil. This question can be translated into English as, “To what de-
gree do you think this patient will develop an allergic reaction?” To
register their judgments, the participants were asked to introduce a
number from 0 to 100 by using the
UP and DOWN arrow keys and the
ENTER key on the keyboard. The participants in Group Global-111
saw this question in the test trial only. For Groups TT-111 and Global-
111, the test trial took place in the same context as the training
phases. By contrast, for Group TT-112, this context changed at test.
Results and Discussion
The mean judgments at test are shown in Figure 1. As
can be seen in this figure, mean judgments at test were
higher in Group Global-111 than in Group TT-111. Judg-
ments at test were also higher in Group TT-112 than in
Group TT-111. These impressions were confirmed by a
one-way analysis of variance (ANOVA) performed on
the judgments at test, which revealed a main effect of
group [F(2,51) 10.27, MS
e
1,385.37, p .001], and
by planned comparisons with
α
divided by the number of
contrasts according to the Bonferroni correction procedure
(
α
.05/3 .017), which revealed a significant differ-
ence between Groups TT-111 and Global-111 [t(34)
5.01, p .001] and a significant difference between
Groups TT-111 and TT-112 [t(33) 3.04, p .01], but
no differences between Groups TT-112 and Global-111
[t(35) 1.36, p .18].
The difference observed between Groups Global-111
and TT-111 replicates the response mode effect observed
by Matute et al. (2002). Most important, the significant
difference found between the two trial-by-trial groups
suggests that, as was expected, a simple contextual ma-
nipulation at the time of test to make participants per-
ceive that the test question might be different from the
previous question gives rise to different predictive judg-
ments. Moreover, the absence of differences between
Groups Global-111 and TT-112 strongly suggests that,
regardless of the response mode, participants can make
a statistically accurate judgment whenever the test ques-
tion does not seem to refer only to the last set of trials.
This result is consistent with the idea that the recency ef-
fect usually observed with trial-by-trial procedures is at-
tributable to the participants’ perceiving that the test trial
is identical to the previous one and that the contingencies
determining the presence or the absence of the outcome
during the last block of trials remain unchanged.
EXPERIMENT 2
The results of Experiment 1 suggest that contextual
manipulations at the time of test can affect the results, and
that this effect might be due to the differential subjective
interpretation that the participants make of the test de-
mands. However, this manipulation was used only in order
to prevent the test question from being perceived as just
one more instance of the question presented during the
most recent training trials. Interesting results could also
be observed if contextual cues were manipulated not
only at the time of test, but also during the training stages.
These contextual cues would become more informative
if they were used to “mark” the different blocks of in-
formation. If during each of these blocks different con-
textual cues were used, participants could use the pres-
ence of these “marking” contextual cues during the test
phase to infer which block of trials should be considered
most relevant when dealing with the test question. One
purpose of Experiment 2 was to test this prediction.
In addition, in Experiment 1 contextual manipulations
were used only in a group that had been exposed to a
trial-by-trial response mode. Thus, a second aim of Ex-
periment 2 was to investigate whether this kind of con-
textual manipulation could also have an effect when the
response mode is global. Therefore, all the groups in Ex-
periment 2 were global groups—that is, they were asked
to make a single judgment after having received all the
information.
Figure 1. Mean predictive judgment at test for the three groups in Experiment 1.
Error bars represent the standard errors of the means.
PREDICTIVE JUDGMENTS 1069
Method
Participants and Apparatus
. Fifty-one undergraduate students
from Deusto University volunteered for the study. None of these
participants had taken part in the previous experiment or in any re-
lated experiment. Random assignment resulted in 13 participants
in Group Global-111, 12 participants in Group Global-121, 13 par-
ticipants in Group Global-112, and 13 participants in Group Global-
122. In this experiment, the participants were tested in small groups
in a room with individual cubicles. Except for this change, the ap-
paratus was the same as in the previous experiment.
Design and Procedure
. In this experiment, the four global
groups described in Table 2 were used. As in the previous experi-
ment, all groups of participants received a block of 20 trials in
which the cue was always followed by the outcome and a block of
20 trials in which the cue was never followed by the outcome. In this
case, the four groups of participants were asked to make a single
predictive judgment after having seen the two blocks of trials (i.e.,
global response mode). The groups differed in relation to the con-
textual manipulations during the training stages and at the time of
test. For Group Global-111, the two training phases and the test
phase took place in Context 1. For Group Global-112, the context
remained constant during both training phases but the test trial took
place in a second context. For Group Global-121, the first training
phase and the test phase were conducted in the same context, whereas
the second training phase took place in a second context. For Group
Global-122, the second training phase and the test phase were per-
formed in the same context, which was different from that of the
first training phase. All other procedural aspects were the same as
in Experiment 1.
Results and Discussion
The results of this experiment are shown in Figure 2.
As the figure shows, the contextual manipulations had a
large effect on the predictive judgments. As was ex-
pected, when the contextual cues presented at the time of
test did not provide information about the relative rele-
vance of each block of trials (i.e., when the contextual
cues were kept constant across the three phases [Group
Global-111] or when testing took place in a novel context
[Group Global-112]), the subjective estimations were
close to the middle of the scale—that is, the participants
tended to integrate (or average) the contradictory infor-
mation received during the two training phases. More-
over, when the contextual cue presented at the time of test
was related to one specific block of trials (i.e., Groups
Global-121 and Global-122), the estimations were high
or low depending on the block of trials that was being sig-
naled by the test context. These impressions were con-
firmed by a one-way ANOVA performed on the judg-
ments at test, which showed a main effect of group
[F(3,47) 15.27, MS
e
794.51, p .0001]. As was ex-
pected, planned comparisons with
α
adjusted according
to the Bonferroni correction procedure (
α
.05/2
.025) yielded no differences between Groups Global-111
and Global-112 [t(24) 0.74, p .46], but the observed
difference between Groups Global-121 and Global-122
proved to be significant [t(23) 8.3, p .001].
These results suggest that the tendency to integrate the
information usually observed under a global response
mode arises only when the cues available at the time of
test do not provide information that could be used to
infer that one set of trials is more relevant than the other.
However, if contextual (or other) cues signal each dif-
ferent block of trials, then the presentation of these cues
at the time of test can be used to infer that some trials are
more relevant. Thus, even under a global response mode
predictive judgments can be based on the most recent or
on most distant information, rather than on an integrative
strategy, whenever participants are induced to think that
this will be the most appropriate response. An interest-
ing result in this experiment is the lack of a significant
difference between Groups Global-111 and Global-112.
This result contrasts with the clear differences observed
in Experiment 1 between Groups TT-111 and TT-112 and
suggests that the high response observed in the 112 groups
is not merely due to the fact that testing in a new context
elevates the level of responding that would be observed
in the corresponding 111 group. If this were the case,
then responding in Group Global-112 should also have
been higher than responding in Group Global-111. How-
ever, in order to properly address the comparison be-
tween the global and TT groups of Experiments 1 and 2
when there is and when there is not a context shift, it
would be necessary to perform an additional experiment
in which all these groups are included.
EXPERIMENT 3
The results presented up to this point show that contex-
tual cues can play a key role when participants deal with a
predictive task based on conflicting evidence. Moreover,
the effect of these contextual cues seems to overcome the
influence of the response mode effect. Participants work-
ing under trial-by-trial procedures no longer make judg-
ments based only on recent information if the test context
suggests that the present trial does not belong to the last set
of trials (Experiment 1). Similarly, participants required to
make a single (i.e., global) judgment at the end of training
no longer make their default integrative estimations if con-
textual cues suggest that a partial estimation is more ade-
Table 2
Design Summary of Experiments 2 and 3
Training
Group Response Mode Phase 1 Phase 2 Test
Global-111 Global (C–O)1 (C–noO)1 (C)1
Global-121 Global (C–O)1 (C–noO)2 (C)1
Global-112 Global (C–O)1 (C–noO)1 (C)2
Global-122 Global (C–O)1 (C–noO)2 (C)2
TT-111 Trial by trial (C–O)1 (C–noO)1 (C)1
TT-121 Trial by trial (C–O)1 (C–noO)2 (C)1
TT-112 Trial by trial (C–O)1 (C–noO)1 (C)2
TT-122 Trial by trial (C–O)1 (C–noO)2 (C)2
Note—In Experiment 2, only the global groups were tested. In Exper-
iment 3, all eight groups were tested. The cue (C) represents a fictitious
medicine presented in every trial; the outcome is an allergic reaction
present during Phase 1 (O) but not during Phase 2 (noO). In these exper-
iments, the different phases of the study could occur either in Context 1
or in Context 2 (represented by the numbers outside the parentheses).
The groups differed with regard to whether the second training phase
and/or the test phase took place in Context 1 or in Context 2.
1070 VADILLO, VEGAS, AND MATUTE
quate (Experiment 2). Thus, our next step was to compare
within a single experiment the way these manipulations af-
fect both trial-by-trial and global procedures. In order to
allow for such comparisons, Experiment 3 comprises all
the groups from Experiment 2 together with analogous
trial-by-trial groups. This will allow us to explore other
predicted similarities between certain effects of the re-
sponse mode and certain contextual manipulations that
have not yet been tested. In addition, contexts were ma-
nipulated in a different way so as to make sure that the re-
sults of Experiments 1 and 2 can be generalized when dif-
ferent contextual cues are used to retrieve the information
related to one or the other set of trials.
Method
Participants and Apparatus
. One hundred ninety-one under-
graduate students from Deusto University volunteered for the study.
None of the participants had taken part in any related experiment.
Half of the participants were tested in a large computer room as in
Experiment 1, whereas the other half were tested in individual cu-
bicles as in Experiment 2. In both cases, the participants were ran-
domly distributed across the eight experimental groups.
Design and Procedure
. The design of this experiment is sum-
marized in Table 2. Four groups were exposed to a global response
mode, whereas four other groups were exposed to a trial-by-trial re-
sponse mode. Except for the changes in the contextual manipula-
tions mentioned below, all procedural details were the same as in
Experiments 1 and 2.
The same allergy task used in the previous experiments was used
in this experiment, but a few modifications were introduced in
order to improve the contextual manipulations. In Experiments 1
and 2, the contexts were two different centuries in which the ficti-
tious patients were supposed to have lived. This kind of context
could be quite strange or unnatural for participants, since it is not
clear why the centuries should have an influence on the medicine’s
effect. Because of this, in Experiment 3 we decided to use a more
natural set of fictitious contexts, in a way similar to that used by
other human learning researchers (e.g., Vila & Rosas, 2001). In Ex-
periment 3, the participants were informed that the medicines could
have been developed in different laboratories. The laboratory that
had produced the medicine was always indicated at the top of the
screen. They were named “Laboratory A” and “Laboratory B,” and
the distinction between these two laboratories was accentuated by
the use of different colors (yellow and blue, respectively) for the
screen. Both the names and the colors of the two laboratories were
counterbalanced. The participants were encouraged to pay attention
not only to the medicines and allergies, but also to the laboratories.
The instructions for the experiment were modified accordingly by
a change to the second screen of instructions. A translation of the
second screen of instructions used in this experiment is offered in
Appendix B.
Results and Discussion
The mean judgments at test are shown in Figure 3. As
Figure 3 shows, both contextual manipulation and re-
sponse mode had an effect on predictive judgments at
test. A 2 (response mode: TT vs. global) 4 (contextual
manipulation: 111 vs. 121 vs. 112 vs. 122) ANOVA on the
final judgments during test yielded a main effect of re-
sponse mode [F(1,183) 20.42, MS
e
584.03, p .001]
and a main effect of contextual manipulation [F(3,183)
85.56, MS
e
584.03, p .001]. Moreover, the interaction
between them was also significant [F(3,183) 3.35,
MS
e
584.03, p .05]. Five subsequent contrasts with
α
adjusted according to the Bonferroni correction pro-
cedure (
α
.05/5 .01) were performed to look for the
source of this interaction.
The mean judgments made at test by the participants
in Groups Global-111 and TT-111 indicate that when the
contextual cues were kept constant during both training
and testing the results varied depending on the response
mode (see Figure 3). Contrasts showed that judgments at
test were lower for Group TT-111 than for Group Global-
111 [t(46) 4.74, p .001], which replicates the basic
response mode effect that we also observed in Experi-
ment 1 of the present work.
Figure 2. Mean predictive judgment at test for the four groups in Experiment 2. Error
bars represent the standard errors of the means.
PREDICTIVE JUDGMENTS 1071
Figure 3 also shows that when the contextual cue pre-
sented at the time of test signaled one and only one block
of trials the final judgment was coherent with the infor-
mation provided within that block. This occurred re-
gardless of the response mode with which the partici-
pants emitted their judgments. Planned comparisons
revealed significant differences between Groups Global-
121 and Global-122 [t(46) 15.76, p .001] and also
between Groups TT-121 and TT-122 [t(46) 8.84, p
.001]. This suggests that, regardless of their response
mode, the participants used the contextual cues pre-
sented at the time of test to make their judgments if these
cues had had a disambiguating role during training.
Moreover, when the context did not change between
the first and the second training phases, making the test
judgment in a novel context had an effect on judgments
under a trial-by-trial response mode (TT-112) but not
under a global response mode (Global-112). Planned
comparisons revealed significant differences between
TT-111 and TT-112 [t(47) 3.04, p .005], whereas no
differences were observed between Groups Global-111
and Global-112 [t(44) 0.65, p .52]. As was ex-
pected, the global participants maintained their default
integrative judgments, but trial-by-trial participants,
whose default response would be a recency response,
opted for an integrative response when tested in a novel
context. In both cases, the result was an integrative judg-
ment when the test context was new.
GENERAL DISCUSSION
Although some of the results presented here can be
partially explained by different theoretical models, in their
present state none of the main theoretical approaches can
provide a fully satisfactory explanation for all the results
observed. As was mentioned in the introduction, the two
main theoretical frameworks—namely, the associative
framework and the statistical framework—cannot explain
the differences obtained when response mode is manip-
ulated. The mechanisms proposed by the associative
view can account for the recency effects observed by de-
fault under trial-by-trial procedures but not for the inte-
grative judgments generally observed with global proce-
dures. On the other hand, the kind of probabilistic rules
proposed by the statistical models closely matches the
results generally obtained with global procedures, but it
does not explain the tendency to recency observed under
a trial-by-trial response mode. Similarly, it is not clear
how these models would explain the effects of some of
the contextual manipulations used here.
Some models that we have not yet considered could
provide partially satisfactory explanations for some of
these results. For example, the finding that predictive
judgments at test when a specific context is present tend
to match the particular cue–outcome contingency sig-
naled by that context (e.g., high ratings in the 121 groups
and low ratings in the 122 groups) can be easily ex-
plained by any associative model that assumes that what
is associated with the outcome is not the particular cue
but the cue–context configuration (Pearce, 1987, 1994).
According to this model, the participants in the 121 groups
should give high ratings because the Cue Context 1
node, which is activated at test, is associated with the
outcome. Similarly, a low rating in the 122 groups is ex-
pected because the Cue Context 2 presented at test is
associated with the absence of the outcome. However,
Groups 111
Groups 121
Groups 112
Groups 122
100
80
60
40
20
0
Global
TT
Res
p
onse Mode
Mean Predictive Judgment
Figure 3. Mean predictive judgment at test for the eight groups in Experiment 3. Error
bars represent the standard errors of the means.
1072 VADILLO, VEGAS, AND MATUTE
this model cannot account for the basic difference found
between the trial-by-trial groups and the global groups
(e.g., TT-111 vs. Global-111), which is one of the basic
phenomena observed in these experiments.
In a similar fashion, certain versions of statistical
models could be used to explain some of our results. For
example, one of our major findings is that participants
responding on a trial-by-trial basis stop giving low rat-
ings at test if the test takes place in a different and new
context (e.g., TT-112). According to the probabilistic
contrast model (Cheng & Holyoak, 1995), this higher
rating might be due to the uncertainty people might ex-
perience when they are asked to give a rating referring to
an unknown context. During training, the participants in
Group TT-112 had received information about the prob-
ability of the outcome given the cue in the presence of
one context. When the new context is presented, they
cannot be sure that the cue will be followed by the out-
come in this new context, and so they make an interme-
diate judgment. However, this model remains unable to
account for the recency-based judgments observed in
Group TT-111.
In order to account for these phenomena, a suitable
theoretical model would have to allow for the indepen-
dent storage and retrieval of the information presented
during the different stages. Most current associative
models fail to meet this requirement, since they postulate
that after a participant has seen all the evidence a single
associative strength is stored. Nor can current statistical
models meet this criterion, given that according to them
all the information available will be used to compute the
appropriate contingency measure; whether this informa-
tion comes from stable or unstable evidence lacks im-
portance. Nevertheless, to overcome this limitation both
associative and statistical models can be extended with a
few assumptions.
Extending Associative Models
One important limitation of the most popular associa-
tive models (see, e.g., Rescorla & Wagner, 1972) is that
they allow for the storage of only a single associative
strength, which summarizes the perceived relationship
between the cue and the outcome. When the associative
mechanism faces the problem that the contingencies are
unstable, its updating rule will adjust the connections ac-
cording to the present contingencies and, in this process,
the associative strength developed for previous contin-
gencies will be lost. This effect is known as catastrophic
forgetting (Lewandowsky, 1991; McCloskey & Cohen,
1989; Ratcliff, 1990). Although this forgetting mecha-
nism accurately predicts the recency effects obtained in
some of the groups in the present experiments, it also
prevents associative models from accounting for any re-
sult in which the evidence received first is used at test.
However, this feature does not affect all associative mod-
els in the same way.
The design used in the experiments presented here
(i.e., a cue always followed by an outcome in the first
phase and never followed by that outcome in the second
phase) is an acquisition–extinction design and has a long
tradition in animal conditioning research. The results of
many experiments in which this design has been used
have provided interesting data that show that associa-
tions between a cue and an outcome are not destroyed
when the cue no longer predicts the outcome. For exam-
ple, many studies have shown that the conditioned re-
sponse can be recovered after extinction if the test trial
takes place in a context different from that of the extinction
phase (an effect known as renewal; Bouton & Bolles,
1979) or if there is a time interval between the end of the
extinction phase and the test phase (spontaneous recovery;
Pavlov, 1927). Moreover, these effects have also been
replicated in human contingency learning experiments
(see, e.g., Pineño & Matute, 2000; Rosas, Vila, Lugo, &
López, 2001). The evidence provided by these studies
has motivated the development of new associative models
that are less sensitive to catastrophic forgetting.
According to some associative models (e.g., Bouton,
1993, 1997; Pavlov, 1927), the information relative to pos-
itive and negative contingencies is stored in independent
associations (excitatory and inhibitory, respectively).
Thus, when the contingency between the cue and the out-
come changes from positive to negative, a new inhibitory
association is developed which does not require the de-
struction of the previous excitatory one. According to
Bouton (1993, 1997), the problem of storing these two as-
sociations is that the cue becomes an ambivalent predic-
tor of the outcome. In this situation, the factor that deter-
mines which of the two associations is expressed is the
relative activation of each of them during testing, which
depends on the physical or temporal contextual cues that
are available at test. If the test phase is presented immedi-
ately after the extinction stage (i.e., in the same temporal
context) or in the same physical context, then the associ-
ation developed second (i.e., the inhibitory one) will be
expressed; but if the test phase is conducted at a different
time or in a context different than that of the second stage,
then the association developed first will be expressed.
2
Within this framework, the different results of the ex-
periments presented here can be understood as involving
the kind of context-based retrieval of the associations
proposed by Bouton (1993, 1997) to account for the re-
newal effect and other animal conditioning phenomena.
When different contexts are present during each training
phase (e.g., Groups 121 and 122 in Experiment 3), the
presence of one or the other context at the time of test
will determine which association will be expressed. The
only extension needed in this type of model is the as-
sumption that when the context is constant the second
association will have a higher activation in memory at
the time of test (Group TT-111), unless there is anything
in the test question that makes people think that the test
trial is different from the most recent trials. That is, not
only contexts but also other types of variables can be
used to separate the test trial from the most recent trials
and, in consequence, reactivate older information in mem-
PREDICTIVE JUDGMENTS 1073
ory. Instances of variables that produce this rupture be-
tween the second phase and the test trial are the presen-
tation of a new question at the time of test (Group Global-
111) and the presentation of the same question in a new
context (Group TT-112). In these cases, both the first and
the second acquired associations will be partially acti-
vated by the novel test context or question.
Extending Statistical Models
Similarly, current statistical models also face impor-
tant limitations in accounting for the present results.
They would need to allow for the separation of the dif-
ferent sets of evidence and for a relatively independent
storage of the information contained in them. Although
current statistical models do not accomplish this re-
quirement directly, some recent normative considera-
tions regarding contextual and conjunctive causal power
(Cheng, 2000) could supply the theoretical framework to
overcome this problem.
3
According to Cheng’s (2000)
normative analysis, the power of a given cause can change
from one context to another and, when this is the case,
the assessments of the relation between the cause and the
effect cannot be generalized from one context to another.
Thus, in the groups in which we used contextual switches,
our participants may have regarded the different contex-
tual cues as reflecting different contextual causal pow-
ers. From this perspective, the independent use of the in-
formation contained at each stage would be perfectly
coherent. In this case, the judgment made at the time of
test will depend on the contextual cues then presented.
However, the results observed when the contextual cues
are kept unchanged during the two training phases (i.e.,
when the contexts were only temporal, such as in Group
TT-111) are more difficult to explain. Although the in-
formation participants receive can also be regarded as re-
flecting different contextual causal powers, the relevant
contextual factors are unobservable and their presence or
absence has to be inferred.
Thus, as was the case of most associative models, the
main limitation of statistical models, in their present
state, is that they cannot compute contingency values
separately for each block of evidence. In this case, when
asked to make an assessment the participant would just
use all the evidence available, regardless of its distribu-
tion over time. When a different contextual cue is pre-
sented at different phases (Groups 121 and 122 in these
experiments), this contextual cue might be used to sepa-
rate the information contained in each phase. Neverthe-
less, this separation cannot be directly performed when
these contextual cues are absent (e.g., Groups TT-111,
TT-112, Global-111, and Global-112 in these experi-
ments). Of course, a partition of the available informa-
tion can be established a posteriori by researchers, but
no a priori mechanisms have yet been suggested to ac-
complish this task. However, the development of such
mechanisms would not be difficult.
In order to account for such results, a statistical model
would need the following extension: a mechanism to de-
tect, in the absence of contextual cues, whether the evi-
dence presented reflects a single homogeneous contin-
gency or, by contrast, a variable interevent relation. In
the second case, the model should also be able to iden-
tify the different blocks in which the available informa-
tion could be optimally separated. Both these tasks could
be accomplished by an algorithm that tests the proba-
bilistic coherence of the sequence of events under analy-
sis. In our experiments, for example, the evidence pre-
sented taken as a whole suggests that the probability of
the outcome given the cue is .50. However, under this
general conditional probability it is highly improbable that
in a sequence of 20 consecutive trials all of them would
consist in the cue followed by the outcome (Phase 1 in
our experiments) or in the cue followed by no outcome
(Phase 2 in our experiments). This suggests that a mo-
mentary probability of the outcome given the cue equal
to 1, rather than a general conditional probability of .50,
might be responsible for the sequence of events in Phase 1,
and that a momentary probability of 0 might be respon-
sible for the Phase 2 sequence. Accordingly, the evidence
can be better seen as reflecting two different phases with
different interevent relations. The sequences of evidence
that are maximally improbable given the conditional
probabilities suggested by the information in its entirety
might be regarded as independent blocks. The partici-
pants’ judgments could be based on either of the differ-
ent blocks or on the conjunction of all the information pre-
sented, depending on the perceived similarity between
the situation of the test trial and that of each of the train-
ing phases.
In our experiments, when each set of training trials is
presented together with particular contextual cues (i.e.,
Groups 121 and 122), the contextual cues presented at
the time of test will provide the information required to
determine which of the partial blocks of information is
more applicable. On the other hand, when these contex-
tual cues do not provide such information (i.e., Groups
111 and 112), since it is not clear which of the partial
contingencies is more adequate, the judgment will be
made taking into account both blocks of information, un-
less the test question itself indicates which is more rele-
vant. For example, for the participants in Group Global-
111, the test question is completely new and, although
nothing else changes, this signals that this question does
not belong to the most recent set of trials. In this case, an
integrative judgment is emitted. However, for the partic-
ipants in Group TT-111 nothing changes between the last
training trial and the test trial, and this leads the partici-
pants to assume that the particular contingency of the
second phase still applies. As can be observed, once the
different blocks of trials have been separated, the factors
that could determine which of the partial contingencies
should be taken into account at the time of test would be
quite similar to those described in the discussion of the
extension of associative models above.
Some Concluding Comments
It is worth noting that the type of extensions we have
suggested for the main theoretical approaches will make
1074 VADILLO, VEGAS, AND MATUTE
them progressively more indistinguishable from each
other and their differential predictions harder to test.
However, they need not be seen as competing theoretical
alternatives but, rather, as different levels of explanation,
one of them (the statistical account) being a computa-
tional or general approach and the other (the associative
account) an algorithmic description of the low-level mech-
anisms involved in causal and predictive learning.
Of course, the kind of account that we have suggested
can also be extended to belief revision models (Catena
et al., 1998; Hogarth & Einhorn, 1992), provided they
allow for the storing of all the information regardless of
the response mode. These models should explicitly as-
sume that even under trial-by-trial responding some kind
of record of all the information presented needs to be
stored. Moreover, these models should assume some
mechanism by which the different blocks of trials could
be separated. These procedures would allow for the same
type of extensions as those that we have proposed for as-
sociative and statistical models.
These experiments, together with those of Matute et al.
(2002), clearly show that the most well-known theoretical
approaches for the study of causal and predictive induction
are, in their present state, unable to account for very sim-
ple experimental results. However, a model’s lack of fit-
ness need not necessarily mean that it is wrong; it could
also mean that it is just incomplete. As we have tried to
show, the major theoretical approaches can easily be ex-
tended in order to take into account many of the neglected
variables that obviously influence the predictions that
people make. Thinking about the limitations of several
theoretical models and about the means by which these
limitations might be overcome in most models could
lead us to a deeper understanding of the processes un-
derlying human causal and predictive judgment.
REFERENCES
Allan, L. G. (1980). A note on measurement of contingency between
two binary variables in judgment tasks. Bulletin of the Psychonomic
Society, 15, 147-149.
Bouton, M. E. (1993). Context, time, and memory retrieval in the in-
terference paradigms of Pavlovian conditioning. Psychological Bul-
letin, 114, 80-99.
Bouton, M. E. (1997). Signals for whether versus when an event will
occur. In M. E. Bouton & M. S. Fanselow (Eds.), Learning, motiva-
tion, and cognition: The functional behaviorism of Robert C. Bolles
(pp. 385-409). Washington, DC: American Psychological Association.
Bouton, M. E., & Bolles, R. C. (1979). Contextual control of the ex-
tinction of conditioned fear. Learning & Motivation, 10, 445-466.
Catena, A., Maldonado, A., & Cándido, A. (1998). The effect of the
frequency of judgment and the type of trials on covariation learning.
Journal of Experimental Psychology: Human Perception & Perfor-
mance, 24, 481-495.
Cheng, P. W. (1997). From covariation to causation: A causal power
theory. Psychological Review, 104, 367-405.
Cheng, P. W. (2000). Causality in the mind: Estimating contextual and
conjunctive causal power. In F. C. Keil & R. A. Wilson (Eds.), Ex-
planation and cognition (pp. 227-253). Cambridge, MA: MIT Press.
Cheng, P. W., & Holyoak, K. J. (1995). Complex adaptive systems as
intuitive statisticians: Causality, contingency and prediction. In H. L.
Roitblat & J.-A. Meyer (Eds.), Comparative approaches to cognitive
science (pp. 271-302). Cambridge, MA: MIT Press.
Cheng, P. W., & Novick, L. R. (1992). Covariation in natural causal in-
duction. Psychological Review, 99, 365-382.
Cobos, P. L., Caño, A., López, F. J., Luque, J. L., & Almaraz, J.
(2000). Does the type of judgement required modulate cue competi-
tion? Quarterly Journal of Experimental Psychology, 53B, 193-207.
Collins, D. J., & Shanks, D. R. (2002). Momentary and integrative re-
sponse strategies in causal judgment. Memory & Cognition, 30,
1138-1147.
De Houwer, J., Beckers, T., & Glautier, S. (2002). Outcome and cue
properties modulate blocking. Quarterly Journal of Experimental
Psychology, 55A, 965-985.
Dennis, M. J., & Ahn, W.-K. (2001). Primacy in causal strength judg-
ments: The effect of initial evidence for generative versus inhibitory
relationships. Memory & Cognition, 29, 152-164.
Dickinson, A., & Burke, J. (1996). Within-compound associations
mediate the retrospective revaluation of causality judgements. Quar-
terly Journal of Experimental Psychology, 49B, 60-80.
Goddard, M. J. (1999). Renewal to the signal value of an uncondi-
tioned stimulus. Learning & Motivation, 30, 15-34.
Hogarth, R. M., & Einhorn, H. J. (1992). Order effects in belief up-
dating: The belief-adjustment model. Cognitive Psychology, 24, 1-55.
Jenkins, H., & Ward, W. (1965). Judgment of contingency between re-
sponses and outcomes. Psychological Monographs, 7, 1-17.
Lewandowsky, S. (1991). Gradual unlearning and catastrophic inter-
ference: A comparison of distributed architectures. In W. E. Hockley
& S. Lewandowsky (Eds.), Relating theory and data: Essays on
human memory in honor of Bennet B. Murdock (pp. 445-476). Hills-
dale, NJ: Erlbaum.
Matute, H., Arcediano, F., & Miller, R. R. (1996). Test question
modulates cue competition between causes and between effects.
Journal of Experimental Psychology: Learning, Memory, & Cogni-
tion, 22, 182-196.
Matute, H., & Pineño, O. (1998). Cue competition in the absence of
compound training: Its relation to paradigms of interference between
outcomes. In D. L. Medin (Ed.), The psychology of learning and mo-
tivation (Vol. 38, pp. 45-81). San Diego: Academic Press.
Matute, H., Vegas, S., & De Marez, P. J. (2002). Flexible use of recent
information in causal and predictive judgments. Journal of Experi-
mental Psychology: Learning, Memory, & Cognition, 28, 714-725.
McCloskey, M., & Cohen, N. J. (1989). Catastrophic interference in
connectionist networks: The sequential learning problem. In G. H.
Bower (Ed.), The psychology of learning and motivation (Vol. 24,
pp. 109-165). San Diego: Academic Press.
Mitchell, C. J., & Lovibond, P. F. (2002). Backward and forward
blocking in human electrodermal conditioning: Blocking requires an
assumption of outcome additivity. Quarterly Journal of Experimen-
tal Psychology, 55B, 311-329.
Nakajima, S., Tanaka, S., Urushihara, K., & Imada, H. (2000). Re-
newal of extinguished lever-press responses upon return to the train-
ing context. Learning & Motivation, 31, 416-431.
Pavlov, I. P. (1927). Conditioned reflexes (G. V. Anrep, Trans.). Lon-
don: Oxford University Press.
Pearce, J. M. (1987). A model for stimulus generalization in Pavlov-
ian conditioning. Psychological Review, 94, 61-73.
Pearce, J. M. (1994). Similarity and discrimination: A selective review
and a connectionist model. Psychological Review, 101, 587-607.
Pennington, N., & Hastie, R. (1992). Explaining the evidence: Tests
of story model for juror decision making. Journal of Personality &
Social Psychology, 62, 189-206.
Pineño, O., & Matute, H. (2000). Interference in human predictive
learning when associations share a common element. International
Journal of Comparative Psychology, 13, 16-33.
Ratcliff, R. (1990). Connectionist models of recognition memory:
Constraints imposed by learning and forgetting functions. Psycho-
logical Review, 97, 285-308.
Rescorla, R. A., & Wagner, A. R. (1972). A theory of Pavlovian con-
ditioning: Variations in the effectiveness of reinforcement and non-
PREDICTIVE JUDGMENTS 1075
reinforcement. In A. H. Black & W. F. Prokasy (Eds.), Classical con-
ditioning II: Current research and theory (pp. 64-99). New York: Ap-
pleton-Century-Crofts.
Rosas, J. M., Vila, N. J., Lugo, M., & López, L. (2001). Combined ef-
fect of context change and retention interval on interference in causal-
ity judgments. Journal of Experimental Psychology: Animal Behavior
Processes, 27, 153-164.
Tamai, N., & Nakajima, S. (2000). Renewal of formerly conditioned
fear in rats after extensive extinction training. International Journal
of Comparative Psychology, 13, 137-146.
Van Hamme, L. J., & Wasserman, E. A. (1994). Cue competition in
causality judgments: The role of nonpresentation of compound stim-
ulus elements. Learning & Motivation, 25, 127-151.
Vila, N. J., & Rosas, J. M. (2001). Reinstatement of acquisition per-
formance by the presentation of the outcome after extinction in
causality judgments. Behavioural Processes, 56, 147-154.
Wasserman, E. A. (1990). Attribution of causality to common and dis-
tinctive elements of compound stimuli. Psychological Science, 1,
298-302.
Wasserman, E. A., Kao, S.-F., Van Hamme, L. J., Katagari, M., &
Young, M. E. (1996). Causation and association. In D. R. Shanks,
K. J. Holyoak, & D. L. Medin (Eds.), The psychology of learning and
motivation: Vol. 34. Causal learning (pp. 207-264). San Diego: Aca-
demic Press.
NOTES
1. By using different centuries as our symbolic contexts, we do not
mean to extract any particular conclusions with respect to the role of
real temporal contexts. The only point we mean to make is that contex-
tual cues (of one type or another) associated with one or the other block
of trials can counteract the response mode effect. In Experiment 3, we
use (perhaps more natural) symbolic physical contexts.
2. It should be noted that in the present experiments (see also Matute
et al., 2002) as well as in several experiments in the animal literature
(see, e.g., Goddard, 1999; Nakajima, Tanaka, Urushihara, & Imada,
2000; Tamai & Nakajima, 2000), the results show that when testing
takes place in a novel context, the expression of the excitatory associa-
tion is not complete and is usually weaker than when it is tested in the
context in which it was acquired. This suggests that when the test con-
text is novel, the two associations, and not only the one acquired first,
are partially active in memory. This is a slight deviation from Bouton’s
(1993) predictions, but it does not weaken the extension of this theory
that we are suggesting to account for human predictive learning.
3. Although Cheng’s (2000) analysis applies theoretically only to
causal relations and our task requires just predictive judgments (whose
normative referents are different), the causal scenario used in these ex-
periments could induce participants to make their predictive judgments
on the basis of the underlying causal structure.
(Manuscript received July 17, 2003;
revision accepted for publication February 19, 2004.)
APPENDIX A
Translation of Instructions for Experiments 1 and 2
Screen 1
Imagine that you are an allergist who wants to study to what degree the
consumption of a medicine called Dugetil causes, as a side effect, an aller-
gic reaction. The medical records of a series of patients will be presented.
Based on them, you will have to make your predictions. For each patient,
you will first see a card that tells you whether that patient has taken Dugetil.
Once you have read it you will see, on a second card, whether or not the pa-
tient developed the allergic reaction. After that, you will see the cards for the
next patient, and so on.
At some points during the experiment, you will have to indicate to what
degree you think a particular patient is going to develop the allergic reaction.
Screen 2
But not everything will happen in the same period of time! You will have
the opportunity to travel through time with a newly developed time machine.
Thus, you will be able to see the cards of patients living in different cen-
turies. At the top of the screen you will see the century you are in. Pay at-
tention to this information, since it is important, but don’t forget that the
medical records are equally important.
At certain points, the time machine will automatically transport you to a cen-
tury that may or may not be the same as the one you are in. The machine will
warn you when this process is taking place.
APPENDIX B
Second Screen of Instructions for Experiment 3
But be careful, because the medicines taken by the patients might have
been developed in different laboratories. That is, you will see the medical
records of patients that have taken medicines from different laboratories. At
the top of the screen you will always see which laboratory is involved. Pay
attention to that message because it is important, but be aware that the med-
ical records are equally important.
