ArticlePDF Available

Flexible Use of Recent Information in Causal and Predictive Judgments

Authors:

Abstract and Figures

Associative and statistical theories of causal and predictive learning make opposite predictions for situations in which the most recent information contradicts the information provided by older trials (e.g., acquisition followed by extinction). Associative theories predict that people will rely on the most recent information to best adapt their behavior to the changing environment. Statistical theories predict that people will integrate what they have learned in the two phases. The results of this study showed one or the other effect as a function of response mode (trial by trial vs. global), type of question (contiguity, causality, or predictiveness), and postacquisition instructions. That is, participants are able to give either an integrative judgment, or a judgment that relies on recent information as a function of test demands. The authors concluded that any model must allow for flexible use of information once it has been acquired.
Content may be subject to copyright.
Flexible Use of Recent Information in Causal and Predictive Judgments
Helena Matute and Sonia Vegas
Deusto University
Pieter-Jan De Marez
Leuven University
Associative and statistical theories of causal and predictive learning make opposite predictions for
situations in which the most recent information contradicts the information provided by older trials (e.g.,
acquisition followed by extinction). Associative theories predict that people will rely on the most recent
information to best adapt their behavior to the changing environment. Statistical theories predict that
people will integrate what they have learned in the two phases. The results of this study showed one or
the other effect as a function of response mode (trial by trial vs. global), type of question (contiguity,
causality, or predictiveness), and postacquisition instructions. That is, participants are able to give either
an integrative judgment, or a judgment that relies on recent information as a function of test demands.
The authors concluded that any model must allow for flexible use of information once it has been
acquired.
Learning to predict the events in our environment is critical for
survival. Both humans and other animals are known to learn
predictive and causal relations between the events in their envi-
ronment, and the question of how they do it has preoccupied
philosophers and psychologists for many years.
There are many different types of models that have tried to
answer this question, and some of them seem to contradict each
other. For example, according to statistical or computational mod-
els (generally developed in the area of human cognition), when
participants are asked to judge the degree to which a potential
predictor cue, C, predicts a given outcome, O, they will take into
account all of the information they have and will compute their
judgment using a statistical rule such as, for example, P (Allan,
1980). P is computed as the probability of the outcome occurring
when C is present—that is, p(O | C)—minus the probability of the
outcome occurring when C is absent—that is, p(O |noC).
However, according to associative models, generally developed
in the area of animal learning, predictive and causal judgments are
analogous to the conditioned response that an animal gives in a
conditioning experiment. On observing the occurrence of the con-
ditioned stimulus or cue, C (e.g., light or tone), the animal predicts
the immediate occurrence of the unconditioned stimulus or out-
come, O (e.g., food or foot shock), and hence, salivates or freezes
as a result of this prediction. According to associative theories, the
strength of the expectation of the outcome (whether assessed
through a conditioned response in rats or through a predictive or
causal judgment in humans), will be a function of the strength of
the association that is gradually established between C and O
during the acquisition trials (e.g., Allan, 1993; Rescorla & Wagner,
1972). Thus, according to these theories, the strength of the asso-
ciation will be updated on a trial-by-trial basis: Some types of trials
will strengthen the association between C and O, whereas others
will decrease it. For example, every time that a CO trial occurs,
the associative strength will increase, and any time that a C–noO
trial occurs, the associative strength will decrease.
Opposite Predictions of Associative and Statistical Models
Those two sets of models make opposite predictions concerning
recency effects (e.g., Chapman, 1991; Lo´pez, Shanks, Almaraz, &
Ferna´ndez, 1998). Associative models predict that the final re-
sponse will differ as a function of the most recent trials, whereas
statistical models predict that the order of trials should have no
effect on the final response. For example, according to associative
theories, a series of CO trials followed by a series of C–noO trials
(i.e., acquisition followed by extinction) should lead to a lower
judgment than a series in which the CO and C–noO trials are
presented in the reverse order. By contrast, statistical models
predict no difference between those two conditions. According to
these theories, participants will calculate their final judgment using
all the information received during the two stages, and thus, an
integrative judgment, rather than a recency-based judgment,
should be observed at test.
It is true that the predictions of associative models are parameter
dependent (see e.g., Baker, Mercier, Valle´e-Tourangeau, Frank, &
Pan, 1993; Wasserman, Kao, Van Hamme, Katagari, & Young,
1996) and that associative models that give less weight to recent
trials can easily be developed. Moreover, statistical models could
also be developed that compute only the latest trials so that they
could best fit the results of extinction experiments and other
recency effects (see Miller & Escobar, 2001; Shanks, Lo´pez,
Helena Matute and Sonia Vegas, Department of Psychology, Deusto
University, Bilbao, Spain; Pieter-Jan De Marez, Department of Psychol-
ogy, Leuven University, Leuven, Belgium.
Support for this research was provided by Department of Education,
Universities and Research of the Basque Government Grant PI-2000-12
awarded to Helena Matute. Sonia Vegas was supported by Beca de For-
macio´n Investigadores Fellowship BFI00.138 from the Basque Govern-
ment. Pieter-Jan De Marez was supported by the Socrates undergraduate
exchange program of the European Union. We thank Leyre Castro, Andre´s
Catena, Antonio Maldonado, Nuria Ortega, Oskar Pinen˜o, and Bram
Vervliet for insightful discussions concerning these experiments.
Correspondence concerning this article should be addressed to Helena
Matute, Departamento de Psicologı´a, Universidad de Deusto, Apartado 1,
Bilbao 48080, Spain. E-mail: matute@orion.deusto.es
Journal of Experimental Psychology: Copyright 2002 by the American Psychological Association, Inc.
Learning, Memory, and Cognition
2002, Vol. 28, No. 4, 714–725
0278-7393/02/$5.00 DOI: 10.1037//0278-7393.28.4.714
714
Darby, & Dickinson, 1996, for discussion). However, the cognitive
system may be much more flexible than any of those classes of
models predict. As predicted by associative theories, the predictive
or causal relations that we once learned may no longer be valid
(things change in our environment), and there must be a mecha-
nism that allows us to adapt to these changes. At the same time,
however, it might be quite maladaptive to discard our previous
knowledge and rely only on the most recent information. Quite
possibly, the cognitive system is able to use either the most recent
information or all of it as a function of the demands of the
environment.
However, the many studies on extinction available in the liter-
ature (e.g., Paredes-Olay & Rosas, 1999; Pavlov, 1927) provide
clear evidence in favor of the associative prediction and against the
statistical prediction. For this reason, our approach in these exper-
iments was to use an extinction design to test whether integrative
judgments could also be observed after extinction training as a
function of several testing manipulations.
Variables That Might Affect the Observation
of One or the Other Result
One of the variables that could affect the observation of trial-
order effects is the wording of the test question that is used to
request the participants judgments. These questions generally
vary from one published study to another, and the potential influ-
ence of this variable is generally overlooked. However, Matute,
Arcediano, and Miller (1996) presented evidence that the type of
question used to assess participants judgments affects the results
of judgmental experiments. More specifically, they used a cue
competition design and observed cue competition effects (i.e.,
judgments that become biased when there are several cues asso-
ciated to the same outcome) when participants were asked to give
predictive and causal judgments, but not when participants were
asked to judge the degree of contiguity (or co-occurrence) between
the cue and the outcome (but see also Cobos, Can˜o, Lo´pez, Luque,
& Almaraz, 2000). Although Matute et al. did not directly study
the potential differences between causal and predictive questions
in that study, there are reasons to believe that those two types of
questions could produce differential judgments (e.g., Cheng, 1997;
Miller & Matute, 1996). For example, a predictive test question
(e.g., To what degree do you expect the outcome to occur in this
particular trial?) seems to refer to a particular time and context
and could differ from one time to another (e.g., a light might
predict food in this time and context; the same light might predict
nothing at a different time or context). However, a causal question
(e.g., To what degree do you think that C is the cause of O?) seems
to imply a more general relationship between the cue and the
outcome that does not vary as much with time and context. If this
were true, it would be easier to observe extinction with predictive
than with causal questions. Finally, a contiguity (or co-occurrence)
question (e.g., To what degree would you say that C was followed
by O, even by mere chance?) should show that regardless of the
degree of extinction shown with the other two questions, partici-
pants are aware that C was followed by O in 50% of the trials. In
other words, the knowledge that C was followed by O during the
first phase should not be lost during the second phase in which C
is never followed by O even though extinction might be observed
as a function of some testing conditions. This was tested in
Experiments 1A1C.
Another variable that might affect the observation of trial-order
effects is whether participants are requested to make their judg-
ments as they are learning or only after the training session is
finished. Several researchers have found that trial-by-trial re-
sponses are more sensitive to recency effects, whereas global
responses given at the end of a sequence tend to integrate the
information received throughout the session. The influence of this
variable has generally been studied in the area of belief revision
(e.g., Hastie & Park, 1986; Hastie & Pennington, 1995; Hogarth &
Einhorn, 1992), but it has been applied to the area of causal
judgments as well (e.g., Catena, Maldonado, & Ca´ndido, 1998;
Wasserman et al., 1996). In general, the results of those studies
suggest that in our acquisitionextinction design, extinction
should be more readily observed if judgments are requested in a
trial-by-trial basis than if they are requested only at the end of
training. This was tested in Experiments 1A1C.
Experiments 1A and 1B
In Experiments 1A and 1B, three groups of participants received
the identical acquisitionextinction training but differed on
whether they received a predictive, causal, or contiguity question.
Presumably, extinction (i.e., recency) judgments should be ob-
served in response to the predictive question. However, a judgment
integrating the information from the two training phases should be
observed in response to the contiguity (co-occurrence) and causal
questions. The only difference between Experiments 1A and 1B
was that in Experiment 1A judgments were required only at the
end of the training session (i.e., global response mode) whereas in
Experiment 1B judgments were required on a trial-by-trial basis.
As noted in the introduction, in addition to the type of question
used at test, the response mode (global vs. trial-by-trial) could also
affect the participants sensitivity to recent trials.
Method
Participants and apparatus. Fifty-seven undergraduate students from
Deusto University volunteered for Experiment 1A. Random assignment of
participants resulted in 17 participants in Group G.contiguity, 20 partici-
pants in Group G.causal, and 20 participants in Group G.predictive.
Forty-seven undergraduate students from Deusto University volunteered
for Experiment 1B. Random assignment of participants resulted in 17
participants in Group T.contiguity, 14 participants in Group T.causal,
and 15 participants in Group T.predictive (in Experiment 1A, G refers to
global; in Experiment 1B, T refers to trial by trial).
The experiments were run using personal computers. For each experi-
ment, participants were run simultaneously and seated about 1.5 m apart
from each other. Each participant was exposed to a different experimental
condition than the two adjacent participants. The experiments were per-
formed with the allergy task (e.g., Wasserman, 1990).
Procedure. Table 1 depicts the design summary of Experiments 1A
1C. The G groups in Table 1 correspond to Experiment 1A, the T groups
correspond to Experiment 1B. In each of the two experiments, three groups
of participants received identical training consisting of acquisition fol-
lowed by extinction. Groups differed in that they received different types
of questions (i.e., contiguity, causality, or predictiveness). In Experiment
1A, these questions were presented to all three groups once all the infor-
mation had been presented (i.e., global mode). In Experiment 1B, the
questions were presented on a trial-by-trial basis.
In both experiments, the computer showed the records of fictitious
patients, one patient per trial. Participants saw two cards per patient (i.e.,
per trial). The first card indicated that the patient had taken a fictitious
medicine, Dugetil. The second card indicated whether the patient had
715
RECENCY VS. INTEGRATIVE JUDGMENTS
developed an allergic reaction to the medicine. That is, the first card
corresponded to the cue, C; the second card corresponded to the out-
come, O.
In all cases, Phase 1 consisted of 20 trials of C followed by O, and
Phase 2 consisted of 20 trials of C followed by noO. A screen showing the
sentence Press any key to see the next patient separated the records for
different patients (i.e., trials). A translation from Spanish of the instructions
used in these experiments reads as follows:
Imagine you are an allergist who wants to study to what degree the
consumption of a medicine called Dugetil causes, as a secondary
effect, an allergic reaction. The medical records of a series of patients
will be presented. Based on them, you will 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 the patient did or did not develop 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 that a particular patient is going to develop the
allergic reaction.
In Experiment 1A, after the 40 training trials were completed, the test
trial showed one more card of a patient that had taken Dugetil. This card
was identical to the ones used during training, but now, a test question and
a rating scale appeared at the bottom of the screen. This test question was
the only variable that we manipulated. These questions were as follows
(translated from Spanish): To what degree would you say that the patients
that have taken Dugetil have developed, even by mere chance, the allergic
reaction? (contiguity question); To what degree do you think that Dugetil
is the cause of the allergic reaction? (causal question); and To what degree
do you think that this patient will develop the allergic reaction? (predictive
question). Below the question, the following phrase was displayed:
Please, give a number between 0 and 100, with 0 being absolutely not, and
100 being absolutely. Participants introduced a number by using the up
and down arrow keys and the Enter key on the computer keyboard.
Experiment 1B was identical to Experiment 1A except that judgments
were required not only in the test trial but also in each of the 40 training
trials. The questions and the rating scale used in every trial were identical
to those used in the test trial of Experiment 1A.
Results and Discussion
The critical judgments at test are given in Figure 1 for Experi-
ment 1A and in Figure 2 (right panel) for Experiment 1B. For
reference, Figure 2 also shows the mean judgments during training
in Experiment 1B (left panel).
As can be seen in Figure 1, no differences were observed as a
function of test question in Experiment 1A, which used a global
response mode. This was confirmed by a one-way analysis of
variance (ANOVA), which showed no main effect of type of
question in Experiment 1A, F(2, 54) 0.91, p .05. All three
groups in this experiment gave an intermediate judgment, suggest-
ing an absence of extinction when the response mode was global.
The results of Experiment 1B, however, showed differences
between the different types of questions at the time of testing, and
good evidence of extinction was now observed in the group that
received the predictive question (see Figure 2). This was con-
firmed by a one-way ANOVA, which showed a main effect of type
of question in Experiment 1B, F(2, 43) 4.27, p .05, and by
planned comparisons, which showed that, as expected, extinction
was evident in Group T.predictive as compared with Group T.con-
tiguity, F(1, 43) 8.53, p .01. That is, participants in Experi-
ment 1B were able to make the right prediction for the test patient
according to the temporal context in which they were being tested
(i.e., no one patient was developing the allergic reaction by the end
of Phase 2). At the same time, they were also able to integrate the
information provided during the two stages when they were asked
not to make a prediction for a particular patient but to rate the
degree with which C and O had occurred together during training
(i.e., contiguity question).
Figure 2 also shows a greater tendency to integrate the infor-
mation when the question is causal than when it is predictive.
However, the difference between Group T.causal and the other two
groups did not reach statistical significance (both ps .05). That
is, the causal question in this experiment yielded a result that was
somehow in between the extinction shown in Group T.predictive
and the integrative judgment shown in Group T.contiguity. Thus,
the next experiment will try to further clarify this result.
Table 1
Design Summary of Experiments 1A–1C
Group
Response
mode Phase 1 Phase 2 Test
Experiment 1A
G.predictive G C–O CnoO Predictive
G.causal G C–O CnoO Causal
G.contiguity G C–O CnoO Contiguity
Experiment 1B
T.predictive T C–O CnoO Predictive
T.causal T C–O CnoO Causal
T.contiguity T C–O CnoO Contiguity
Note. The cue, C, is a fictitious medicine; O/noO indicate the presence or
absence of the outcome, an allergic reaction. Twenty trials of each phase
were presented. Groups differ on the type of question that they received
and on whether the response mode was global (G) or trial by trial (T).
Experiment 1C replicated the critical findings in Experiments 1A and 1B
using a 2 (response mode: global vs. trial by trial) 2 (type of question:
causal vs. predictive) design.
Figure 1. Mean judgment at test for the three groups in Experiment 1A.
Error bars represent the standard error of the mean. G refers to global
response mode.
716
MATUTE, VEGAS, AND DE MAREZ
Experiment 1C
The results of Experiments 1A and 1B, taken together, suggest
that a trial-by-trial response mode and a predictive question are
both important to observe the recency effects predicted by asso-
ciative theories as opposed to the integrative effects predicted by
statistical theories. However, because these conclusions are drawn
from a comparison between two different experiments, our next
step was to assess the influence of those two variables within the
same experiment.
Experiment 1C used a 2 (response mode: global vs. trial by
trial) 2 (type of question: predictive vs. causal) design. We did
not include the contiguity question in this experiment because the
results of Experiments 1A and 1B consistently showed that, re-
gardless of whether the response mode is global or trial by trial,
participants are able to give an accurate judgment of the degree to
which the two events have occurred together during training.
Although the results of Experiment 1B replicated those of
Experiment 1A with respect to the contiguity question and ex-
tended them with respect to the predictive question, they were not
clear cut with respect to the causal question. This question showed
a tendency toward a more integrative judgment than that of the
predictive question, but it was not clear how reliable this tendency
was. It might be worth noting that in Experiments 1A and 1B the
causal question was presented, just as the other questions, in the
same screen in which the cue (i.e., This patient has taken Dugetil)
was presented. Although this makes sense with respect to the
predictive question (i.e., To what degree do you think that this
patient will develop the allergic reaction?), the causal question
was not intended to refer to one particular patient but to the causal
relationship between Dugetil and the allergic reaction in general.
By asking this question in the same screen in which the informa-
tion that a particular patient had taken Dugetil was presented, we
might have confused participants: This screen might have sug-
gested to some of them that we were asking whether Dugetil would
be followed by the allergic reaction in this particular patient. If this
were the case, the causal question might have been interpreted as
predictive by some participants and as causal by other participants.
Therefore, to avoid this possible misinterpretation and to make it
more clear to participants what each question meant, the causal
group in Experiment 1C will first see the cue card, then the
outcome card, and only after they have seen the two cards will they
be asked about the causal relationship between the cue and the
outcome.
Method
Participants and apparatus. Sixty-eight undergraduate students from
Deusto University volunteered for the study. None of the participants had
taken part in any related experiment. Random assignment of participants
resulted in 16 participants in Group T.predictive, 18 participants in Group
T.causal, 17 participants in Group G.predictive, and 17 participants in
Group G.causal. The participants were run in individual cubicles.
Procedure and design. The experiment was run again with the allergy
task, and the design was a combination of Experiments 1A and 1B. The
causal and predictive groups in Table 1 are the ones that were replicated in
this experiment. Participants received again 20 CO trials followed by 20
CnoO trials, as in the previous experiments. However, there were two
groups trained in the trial-by-trial response mode (T groups) and two other
groups trained in the global response mode (G groups). In both response-
mode conditions, half of the participants received predictive questions, the
other half received causal questions.
The main procedural difference with respect to Experiments 1A and 1B
concerns the causal groups. In these groups, participants were now re-
quested to introduce their judgment only after they had seen both the cue
and the outcome screen, so that they could not assume that they were
supposed to make a predictive judgment for a particular patient. In the
same vein, the second paragraph of the instructions that participants re-
ceived at the beginning of the experiment was also modified as follows for
the causal groups: At some points during the experiment, you will have to
indicate to what degree you think that Dugetil is the cause of the allergic
reaction.
Our presenting the predictive question before the outcome and the causal
question after the outcome implies that participants in the causal groups
would receive one more training trial than the predictive groups before
making their final test judgment. For this reason, the causal groups in this
experiment were tested in the 40th trial rather than in the 41st trial.
Figure 2. Mean judgment during training (left panel) and testing (right panel) for the three groups in
Experiment 1B. Error bars represent the standard error of the mean. T refers to trial-by-trial response mode.
717
RECENCY VS. INTEGRATIVE JUDGMENTS
Results and Discussion
The left panel of Figure 3 shows the training data for the two
groups that responded during training (i.e., the T groups). The right
panel shows the critical data, that is, the mean judgment at test for
all groups.
The results of this experiment replicated the findings of Exper-
iments 1A and 1B that extinction is most clearly observed at test
when a predictive question and a trial-by-trial response mode is
used, and that, by contrast, global and causal questions tend to
favor more integrative judgments. A 2 (response mode) 2 (type
of question) ANOVA on the final judgments during testing yielded
a main effect of type of question, F(1, 64) 8.23, p .01, a main
effect of response mode, F(1, 64) 14.79, p .01, and no
interaction, F(1, 64) 0.74, p .05. Planned comparisons be-
tween the two G groups replicated the results of Experiment 1A, in
that no differences were observed between predictive and causal
judgments when the response mode was global, F(1, 64) 2.01,
p .05. In that case, participants tended to integrate the informa-
tion received during the two training phases regardless of whether
they were asked a predictive or a causal question. However,
planned comparisons between the two T groups replicated and
clarified the results of Experiment 1B, in that they show that the
predictive question produced lower judgments than the causal
question when the response mode was trial by trial, F(1,
64) 6.95, p .05. Thus, questions that are trial by trial and
predictive seem to favor the observation of recency effects,
whereas a global response mode or a causal question seem to favor
judgments that integrate the two training phases.
Experiment 2
The results of Experiments 1A1C, taken together, suggest that
participants making causal or global judgments tend to integrate
the information received through the training session, whereas
those making predictive judgments on a trial-by-trial basis tend to
be more sensitive to the information received during recent trials.
Thus, both the response mode and the question type seem to be
critical factors that modulate the observation of one or the other
type of judgment.
Of importance, it seems that recency effects, which are often
used to discriminate between statistical and associative predictions
(see, e.g., Shanks et al., 1996), tend to occur only in predictive
trial-by-trial situations. Before drawing such a general conclusion,
however, it is important to make sure that our results are not
restricted to the acquisitionextinction design that we have been
using in Experiments 1A1C. For this reason, in Experiment 2 half
of the participants received the two training phases in the same
order as in the previous experiments, whereas the other half
received the two training phases in the reverse order (i.e., the
CnoO trials presented before the CO trials). Orthogonally, the
response mode was either global or trial by trial for half of the
participants within each group. The question used was always
predictive. If the above conclusions are correct, trial order should
affect predictive judgments in the T groups (with the group re-
ceiving the CO trials during the second phase giving a higher
judgment at test than the group receiving CnoO trials during the
second phase), but not in the G groups.
Method
Participants and apparatus. Eighty undergraduate students from Deu-
sto University volunteered for the study. None of the participants had taken
part in any related experiment. Sixty-five participants were run in a large
computer room, as in Experiment 1A. The remaining participants were run
in individual cubicles, as in Experiment 1C. In both testing conditions, the
participants were randomly distributed across the four experimental
groups. This resulted in 20 participants per group.
Procedure and design. Table 2 depicts the design summary of this
experiment. Two groups received trial-by-trial training; the two other
groups received global training. Half of the participants in each condition
received the same acquisitionextinction training of the previous experi-
ments, that is, C was followed by O in all trials during Phase 1 and never
during Phase 2 (Groups G.1.0 and T.1.0). For the other half of participants,
the order of phases was reversed, that is, CnoO trials were followed by
Figure 3. The left panel shows the mean judgment during training for the two trial-by-trial groups in
Experiment 1C. The right panel shows the mean judgment at test for all four groups as a function of response
mode (G vs. T) and type of question (predictive vs. causal). Error bars represent the standard error of the mean.
G refers to global response mode; T refers to trial-by-trial response mode.
718
MATUTE, VEGAS, AND DE MAREZ
CO trials (Groups G.0.1 and T.0.1). In all cases, 20 trials of each type
were presented, and the type of question was always predictive.
Results and Discussion
The left panel of Figure 4 shows the training data for the two
groups that responded during training (i.e., the T groups). The right
panel shows the critical data, that is, the mean judgment at test for
all groups. As expected, trial order affected predictive judgments
at test in the two trial-by-trial groups but not in the global groups.
These impressions were confirmed by a 2 2 ANOVA, which
showed a main effect of trial order, F(1, 76) 20.21, p .01, no
main effect of response mode, F(1, 76) 0.02, p .05, and a
Trial Order Response Mode interaction, F(1, 76) 10.11, p
.01. As shown in Figure 4, the results at test depended strongly on
trial order when the response mode was trial by trial but not when
it was global. Planned comparisons showed a significant difference
between the two T groups, F(1, 76) 29.45, p .01, but no
significant difference between the two G groups was observed,
F(1, 76) 0.87, p .05. Thus, as expected, the two G groups
integrated the information received during the two training phases,
regardless of trial order, whereas the two T groups tended to
respond during testing according to the information received dur-
ing the latest trials. This suggests that the results of Experiments
1A1C are not restricted to the acquisitionextinction design that
we used, and can be generalized to other situations in which the
order of trials is not randomly distributed through the training
session.
Experiment 3
As expected, the differences that we have observed among the
several conditions tested in Experiments 1A1C and 2 suggest that
participants can respond differently as a function of how they
interpret the test question. Both the trial-by-trial response mode
and the predictive wording seem to be interpreted as requiring a
specific judgment for a particular case in a particular time and
context; however, the global response mode and the causal and
contiguity questions seem to be interpreted as demanding a re-
sponse that integrates the two training phases.
However, there is also a possibility that the differences observed
between the global and the trial-by-trial conditions are merely due
to a lack of attention in the global conditions. These participants
have nothing to do during training but reading the information on
the screen and pressing the Enter key to go on to the next trial.
Thus, it is possible that they simply keep pressing the Enter key
without even reading the information presented in each trial. The
present experiment tested this possibility.
Method
Participants and apparatus. Forty undergraduate students from Deu-
sto University volunteered for the study. None of the participants had taken
part in any related experiment. The participants were run in a large
computer room, as in Experiment 1A, and were randomly distributed
across the four experimental groups. This resulted in 11 participants in
Group G, 11 participants in Group T, 9 participants in Group G.predict,
and 9 participants in Group G.read.
Table 2
Design Summary of Experiment 2
Group
Response
mode Phase 1 Phase 2 Test
G.1.0 G COCnoO Predictive
G.0.1 G CnoOCO Predictive
T.1.0 T COCnoO Predictive
T.0.1 T CnoOCO Predictive
Note. The cue, C, is a fictitious medicine; O/noO indicate the presence or
absence of the outcome, an allergic reaction. Groups differ in their re-
sponse mode (G vs. T), as well as in the order of the two training phases
(e.g., 1.0 means that the outcome always occurred during Phase 1 and
never during Phase 2; 0.1 means the reverse order of trials). G global;
T trial by trial.
Figure 4. The left panel shows the mean judgment during training for the two trial-by-trial groups in
Experiment 2. The right panel shows the mean judgment at test for all four groups as a function of response mode
(G vs. T) and trial order (1.0 vs. 0.1). Error bars represent the standard error of the mean. G refers to global
response mode; T refers to trial-by-trial response mode.
719
RECENCY VS. INTEGRATIVE JUDGMENTS
Procedure and design. Groups G and T replicated the global and the
trial-by-trial conditions of Experiments 1A1C, with the type of question
being predictive in all cases. To test whether a lack of attention could be
responsible for the results of the global conditions, two other global groups
were used in this experiment that emitted a judgment only at test but that
had to type a yes/no response in each training trial. Group G.predict was
required to give a yes/no prediction for the outcome when the cue was
presented in each trial. The wording of this question was Will this patient
develop the allergic reaction? (y/n). Group G.read was simply forced to
read the screen: While the information on the allergic reaction was still
visible, these participants were required to give a yes/no response to the
question Did this patient develop the allergic reaction? (y/n). In this group,
the program did not go on to the next trial until the information was read
and the correct yes/no response was typed.
Results and Discussion
The left panel of Figure 5 shows the training data from the group
that emitted judgments during training (Group T). The insert in this
figure shows the percentage of yes predictions during training
for Group G.predict. This insert shows that, like participants in
Group T, participants in Group G.predict were also paying atten-
tion to the information presented in the screen: they predicted the
occurrence of the outcome during Phase 1 and its absence during
Phase 2. Finally, the percentage of yes responses in Group
G.read is not shown in the figure, but these participants were also
paying attention to the screen and their responses were 100%
correct (recall that in this group the program did not continue to the
next trial until the correct response had been typed).
The right panel of Figure 5 shows the critical data in this
experiment, that is, the mean judgment at test for all groups. As
can be seen in this panel, the three global groups behaved homo-
geneously and differed from the trial-by-trial group. This was
confirmed by a one-way ANOVA, which showed a main effect of
group, F(3, 36) 10.76, p .01, and by planned comparisons,
which showed that, as expected, extinction was evident in Group
T as compared with Group G, F(1, 36) 18.96, p .01, Group
G.predict, F(1, 36) 20.81, p .01, and Group G.read, F(1,
36) 22.51, p .01. This shows that the results of the previous
experiments cannot be attributed to a lack of attention in the global
conditions.
Experiment 4
Experiments 1A3 have shown that both the type of question
and the response mode affect the way in which participants make
their judgments at the end of training. With regard to the type of
question, it is not that surprising that participants receiving a
predictive question tend to rely on recent information to make their
prediction because this is probably an efficient way to make
accurate predictions in a changing environment. The large amount
of data existent on extinction in animal research can also be
regarded as a predictive trial-by-trial situation in which, as in the
present experiments, participants tend to make their predictions on
the basis of the most recent information received. At the same
time, it is not that strange that participants that were asked to make
a causal rather than a predictive judgment tended to consider a
greater amount of information that included also the early trials. As
suggested by the results to the contiguity questions in Experiments
1A and 1B, participants were aware of the probability of the
outcome occurring in the presence of C being .5 (i.e., 1 during the
early trials, 0 during the latest trials). However, they also were able
to predict a probability of 0, or close to 0, for the next trial when
the question was predictive, or to infer a causal relation of an
intermediate level between the two events when they were asked to
make causal judgments. That is, participants seemed to be able to
use the acquired information in a flexible way as a function of the
test question that they received (see also Matute et al., 1996, for a
similar finding).
More puzzling is the response-mode effect. Why should partic-
ipants respond differently as a function of whether they are re-
sponding in every trial or at the end of training? The idea that
Figure 5. The left panel shows the mean judgment during training for Group T in Experiment 3. The insert
shows the percentage of yes predictions during training for Group G.predict. The right panel shows the mean
judgment at test for all four groups. Error bars represent the standard error of the mean. G refers to global
response mode; T refers to trial-by-trial response mode.
720
MATUTE, VEGAS, AND DE MAREZ
participants giving global judgments store all trials whereas those
giving trial-by-trial judgments do not (Catena et al., 1998; Hastie
& Park, 1986; Hastie & Pennington, 1995; Hogarth & Einhorn,
1992) is an appealing one. However, this cannot explain the
difference we observed among trial-by-trial participants that re-
ceived different test questions: if they had stored only the recent
trials, then they should have given judgments close to 0 in all cases
(except for Group T.0.1 in Experiment 2).
Our observation that participants receiving contiguity or causal
questions tended to integrate the information of the two phases at
test even when their response mode was trial by trial, suggests that
the locus of their integration was at the response stage rather than
at the acquisition stage. That is, like the global participants, trial-
by-trial participants must also have had all the information at the
end of training to respond at test one way or another as a function
of how they interpreted the test question. If this is true, the
information could probably be integrated at the end of training, not
only as a function of the response mode and the type of question,
but also as a function of many other testing demands. Indeed,
anything that during testing would suggest to participants that an
integrative judgment would be more appropriate than the default
(i.e., recency) judgment in a trial-by-trial situation should produce
an integrative response. This view was tested in the present
experiment.
In this experiment, three groups of participants received the
conditions that should produce the strongest recency effect, that is,
predictive questions in a trial-by-trial response mode. Our primary
manipulation consisted of inserting an instructional screen just
before testing. The purpose of this screen was to suggest that the
test trial should not be interpreted as an additional Phase 2 trial (for
which a recency response would be appropriate), but as a whole
new phase. In this way, all they had learned during the two training
phases, rather than only the information from Phase 2, should
become relevant at test. If the effects we have been observing in
the previous experiments are response effects, participants trained
with the trial-by-trial mode and predictive questions should be able
to produce integrative judgments if they interpret the test phase in
this way. By contrast, if these effects are due to the trial-by-trial
participants failing to store the information received through the
training session, these participants will be unable to give an inte-
grative judgment even if the test instructions suggest that it will be
appropriate to do so.
Method
Participants and apparatus. Sixty-eight undergraduate students from
Deusto University volunteered for the study. None of the participants had
taken part in any related experiment. They were randomly assigned to the
various conditions, which resulted in 17 participants per group. The par-
ticipants were run in individual cubicles, as in Experiment 1C.
Procedure and design. Table 3 shows the design summary for this
experiment. Three groups were exposed to a trial-by-trial response mode
and one group was exposed to a global response mode. Groups G.20 and
T.20 replicated the global and trial-by-trial conditions of the previous
experiments, with the question being predictive in all cases. That is, they
received 20 CO trials followed by 20 CnoO trials. Two other groups
(Groups T.20.i and T.33.i) were exposed to a trial-by-trial mode, but at the
end of the training sequence, just before testing, they received the follow-
ing instruction (translated from Spanish):
You now should have enough information about the relationship
between Dugetil and the allergic reaction. Now imagine that a patient
of yours has taken Dugetil and she asks you if she is going to develop
the allergic reaction. You will have to tell this person to what degree
you think she is going to develop the allergic reaction so that she can
decide whether or not she should give up the treatment with Dugetil.
One of the instructed groups (Group T.20.i) received the same 20 CO
and 20 CnoO trials as the T.20 group. If the instruction that they received
before testing was enough for them to integrate what they had learned
during the two training phases, they should give a judgment close to 50 in
our 0100 scale, that is, significantly higher than that of Group T.20 and
similar to that of Group G.20. However, there was a possibility that
participants in Group T.20.i gave a judgment of 50 not because they
integrated the information of the two phases, but simply because they
wanted to say, I dont know; so fiftyfifty. To avoid this possible
interpretation of the expected judgment of 50 in this group, the other
instructed group (Group T.33.i) received 33, rather than 20, extinction
trials. Thus, if the instructed groups were integrating the information that
they had received and were giving an accurate judgment rather than a mere
I do not know response, the judgments of these two instructed groups
should differ from each other.
Results and Discussion
The left panel of Figure 6 shows the training data for the three
groups that responded during training (i.e., the T groups). The
critical results of this experiment are depicted in the right panel for
all four groups. As expected, Groups T.20 and G.20 replicated
during testing the basic response-mode effect, but more important,
Table 3
Design Summary of Experiment 4
Group
Response
mode Phase 1 Phase 2 Instruction Test
G.20 G 20 CO 20 CnoO No Predictive
T.20 T 20 CO 20 CnoO No Predictive
T.20.i T 20 CO 20 CnoO Yes Predictive
T.33.i T 7 CO 33 CnoO Yes Predictive
Note. The cue, C, is a fictitious medicine; O/noO indicate the presence or absence of the outcome, an allergic
reaction. Groups G.20 and T.20 are intended to replicate the basic response-mode effect of Experiments 13.
Group T.20.i was included to assess the prediction that an instruction received just before testing would be
sufficient to produce an integrative judgment in participants that had received trial-by-trial training with
predictive questions. Group T.33.i was a control group that was identical to Group T.20.i except that it received
a larger number of extinction trials (i.e., 33 rather than 20). G global; T trial by trial.
721
RECENCY VS. INTEGRATIVE JUDGMENTS
Group T.20.i behaved as the global group rather than as a trial-
by-trial group. Participants in Group T.20.i were able to integrate
the information from the two training phases when giving their
final judgment. Moreover, the judgment of Group T.20.i cannot be
interpreted as an I dont knowresponse, because the judgment of
this group differed from that of Group T.33.i, which also showed
an integrative, but lower, judgment.
These impressions were confirmed by a one-way ANOVA on
the mean judgments at test, which showed a main effect of group,
F(3, 64) 11.94, p .01. Planned comparison between Groups
T.20 and G.20 showed a basic response-mode effect, with judg-
ments in Group T.20 being significantly lower than those in Group
G.20, F(1, 64) 30.36, p .01. Moreover, when an instruction to
integrate the information was inserted in the T.20 group before
testing (i.e., Group T.20.i), the judgment of the T.20 group became
significantly increased in Group T.20.i, F(1, 64) 16.91, p .01.
Indeed, the judgment of Group T.20.i is almost as high as that
reached by Group G.20, F(1, 64) 1.95, p .05. This suggests
that a global response mode is not necessary for participants to
integrate what they had learned in the two phases. Finally, the
difference between Group T.20.i and Group T.33.i, F(1,
64) 5.36, p .05, indicates that these two groups are accurately
integrating the information rather than giving an I dont know
response.
Thus, the response of Group T.20.i replicates the results of the
global groups in the previous experiments. In those cases, partic-
ipants did not show recency effects. Instead, they regarded the two
training phases as relevant when making their final judgment. The
same result was now observed in Group T.20.i through a different
manipulation. Therefore, it seems that it is not the frequency with
which judgments are requested, but how the demands of the test
phase are interpreted, that produces one or the other result.
General Discussion
The experiments presented here show that there are some testing
conditions that favor the recency effects predicted by associative
theories, whereas there are other testing conditions that favor the
absence of trial-order effects predicted by statistical theories.
There are probably many other conditions, in addition to the ones
we have tested here, that could also favor the observation of one or
the other type of judgment. Nevertheless, summarizing our results,
these are the variables that we have found to be critical:
First, the type of question used to assess participantsjudgments
was a critical variable. Among the different types of questions that
we assessed, the contiguity question showed that participants ac-
curately store all the information and can give accurate judgments
on the degree with which C and O occurred together through the
session, regardless of whether the response mode is global or trial
by trial (Experiments 1A and 1B). However, participants tended to
use only the most recent information when the test question was
predictive, particularly when the response mode was trial by trial
(Experiments 1B4). The causal questions, however, tended to
yield more integrative judgments than the predictive questions.
Causal judgments do not normally refer to a particular case in a
particular situation and time, as predictive questions do, but to a
more general relationship between the cue and the outcome (Ex-
periments 1B and 1C). Thus, in general, extinction and trial-order
effects tend to be observed more readily in predictive situations.
Second, we observed a very strong response-mode effect that
was evidenced in the trial-by-trial mode being generally more
sensitive to recent information than the global mode, particularly
when predictive questions were used (Experiments 1A4). More-
over, Experiment 3 showed that the response-mode effect was not
due to a lack of attention in the global conditions, and Experi-
ment 2 showed that this effect is not restricted to the acquisition
extinction design that we used in all the other experiments: In
Experiment 2 we tested the reverse order of trials, and the results
were consistent with the results of the other experiments. That is,
trial order affected participants judgments in the trial-by-trial
response mode but not in the global response mode.
Third, informing participants just before testing that the test trial
should not be interpreted as an additional Phase 2 trial was suffi-
cient for the trial-by-trial participants that were receiving predic-
tive questions to make an integrative judgment rather than their
Figure 6. The left panel shows the mean judgment during training for the three trial-by-trial groups in
Experiment 4. The right panel shows the mean judgment at test for all four groups. Error bars represent the
standard error of the mean. G refers to global response mode; T refers to trial-by-trial response mode.
722
MATUTE, VEGAS, AND DE MAREZ
default recency-based judgment (Experiment 4). Thus, as robust as
the recency effect might be in the trial-by-trial conditions, it seems
that it can be counteracted not only by the test question (Experi-
ments 1B and 1C) but also by simply separating the test trial from
the Phase 2 trials, therefore making all previous phases equally
relevant at the time of testing (Experiment 4). This suggests that
the response-mode effect has less to do with the frequency with
which judgments are requested than with how participants inter-
pret the demands of the environment at the time of testing.
At first glance, our results might seem inconsistent with previ-
ous results reported by Wasserman et al. (1996, Experiments 2 and
3). In some conditions, their participants received a positive con-
tingency during the first half of the study and a negative contin-
gency during the second half. In some other conditions, the order
of trials was reversed. Moreover, some participants responded on
a trial-by-trial basis, whereas some responded on a global mode.
Wasserman et al. did not report trial-order effects. However, the
purpose of their experiments was different from ours, and they
used an overall P of 0 and a rating scale that went from 10
to 10. Thus, all judgments were close to 0 at asymptote, and the
differences we observed could not be observed in their study (note
that if the overall P is 0, there is no reason to expect our results:
the predictions of associative and statistical models are coincident
at asymptote in that case). However, their training data showed
sensitivity to the order of trials, and the trial-by-trial pattern that
they reported was very similar to the one observed in the present
series of studies: participants having the positive contingency first
gave higher judgments during the first half of the study, and this
was reversed for participants having the negative contingency first.
Translated to their design and scale, our results would probably be
very similar to theirs.
Potential Explanations
Trial-order effects are frequently regarded as one of those phe-
nomena that can best be used to discriminate between the statis-
tical and associative predictions (e.g., Allan, 1993; Lo´pez et al.,
1998): Statistical theories predict response effects that should
result in integrative judgments, whereas associative theories, such
as that of Rescorla and Wagner (1972), predict acquisition effects
that should result in recency effects (i.e., extinction). However,
because we observed both trial-order effects and an absence of
trial-order effects as a function of several testing variables, none of
those two sets of theories can directly provide a unified account of
the observed results.
One possibility would be to consider the belief revision models.
These models predict that information is processed and stored
differently as a function of the response mode, with only partici-
pants exposed to the global mode storing all the information
(Catena et al., 1998; Hogarth & Einhorn, 1992; Pennington &
Hastie, 1992). These models could at first glance explain the
response-mode effect that we observed in Experiments 13. How-
ever, as shown above, our results suggest that the response-mode
effect is not due to differential processing or storing of the infor-
mation but to differential responding at test. In Experiment 4, if
participants in Groups T.20.i and T.33.i had not stored the infor-
mation received in all trials, they would not have been able to give
an integrative judgment at test. However, an instructional screen
presented just before testing was sufficient for these participants to
be able to integrate the information. This suggests that the infor-
mation storing strategies of the trial-by-trial response mode are not
different from those of the global response mode.
Indeed, our results show that participants respond one or the
other way as a function of test demands, which suggests that any
model must allow for flexible use of information once it has been
acquired: Sometimes it might be beneficial to make a prediction
based on recent information, sometimes it might be more adaptive
to integrate the information received at different stages, sometimes
it might be beneficial to rely on information received during the
early trials. But, to be able to use either the most recent or the
oldest information, or all of it, and to respond appropriately to each
of the changing environmental demands, it also seems necessary to
store as much information as possible (e.g., CO and CnoO
information in the present case) in a simple and efficient way.
Perhaps the most obvious possibility would be to extend statis-
tical theories because they at least do store all the information. One
of the most popular current statistical models of causal learning is
Chengs (1997) Power PC model. However, Cheng has explicitly
argued that her model does not apply to predictive judgments, and
those are the only ones that we used in our Experiments 24.
Nevertheless, Chengs model was a revision of an older model by
Cheng and Novick (1992), which applied to predictive judgments.
This older model introduced the notion of focal sets: not all the
information is used when computing P, but only the information
in the relevant focal set. These focal sets are formed when there is
more than one potential predictor of the outcome and the two or
more predictors sometimes occur together. However, when there is
only one cue, as in the present experiments, the predictions of the
focal-set P rule are identical to those of the original P rule.
Thus, in the present experiments, the predictions of the focal-set
P rule and those of the original P rule are coincident. They
predict that participants will use all the CO and the CnoO trials
when computing their judgments at test. This allows these theories
to explain the integrative judgments observed in some conditions
in the present studies but prevents them to explain the trial-order
effects observed in the other conditions. Indeed, the lack of an
acquisition function prevents these theories to account for any
acquisition-related effects, such as trial order, learning curves, and
preasymptotic judgments (see also Allan, 1993; Baker et al., 1993;
Lo´pez et al., 1998; Shanks, 1993).
It should be possible, however, to develop statistical models
provided with an acquisition function as well as with time-
sensitive focal sets. This would probably require the addition of a
recency parameter to each focal set, so that at test, participants
could decide whether the most recent focal set, the oldest one, or
all of them should be taken into account when calculating their
judgments. This would also require one to specify how (and when)
such time-sensitive focal sets should be created. As others have
noted (e.g., Shanks, 1993), one of the main problems in trying to
apply the notion of focal sets to a particular experiment is how to
define in advance which focal sets will be created in each partic-
ular situation.
The alternative possibility is to consider associative theories. As
previously noted, the main difference between statistical and as-
sociative models is that statistical models focus on the response
process and associative models generally focus on the trial-by-trial
updating of the associations during acquisition. This updating
process allows them to explain the trial-by-trial results. At the
same time, however, this updating process prevents them from
accounting for the integrative judgments observed in the present
723
RECENCY VS. INTEGRATIVE JUDGMENTS
experiments because it results in the destruction of the CO
association by the end of the extinction phase (e.g., Rescorla &
Wagner, 1972). This problem has been called catastrophic for-
getting by some researchers (e.g., McCloskey & Cohen, 1989;
Ratcliff, 1990). However, there are some exceptions to this rule.
There are some models of animal learning that, unlike the most
prevalent approach (e.g., Rescorla & Wagner, 1972), regard ex-
tinction and other interference effects as retrieval effects rather
than as acquisition effects (e.g., Bouton, 1993). In these cases,
extinction does not imply the destruction of the CO association
acquired during Phase 1. Instead, the organism acquires two dif-
ferent associations: a CO association during Phase 1, and a
CnoO association during Phase 2. The problem here is that the
meaning of the cue becomes ambiguous, and when it is presented
during testing, C could predict either O or noO. Thus, it is
necessary to specify when extinction will be observed and when it
will not. According to Bouton (1993), the (physical or temporal)
context in which the test phase takes place plays the disambigu-
ating role. If the test phase is conducted immediately after Phase 2
and in the same physical context, as is usually the case, the most
recently acquired association (i.e., CnoO) is the one that will be
activated by the test context. Thus, extinction will be the default
result. However, if testing occurs in a novel (temporal or physical)
context, the most recently acquired CnoO association does no
longer prevail and good responding will be observed. The many
experiments showing spontaneous recovery when a time interval
elapses between extinction and testing and renewal of responding
when testing occurs in a novel context give support to this view
(see e.g., Bouton, 1993, for a comprehensive review).
Boutons (1993) theory was developed in the area of animal
conditioning and cannot directly account for the results observed
in the present experiments. However, it could be readily extended
to human learning if we assume that in humans the disambiguating
role does not need to be restricted to the (temporal or physical)
context in which testing takes place. In principle, not only con-
texts, but also verbal stimuli and other test demands such as
response mode or mere instructions can help participants disam-
biguate whether the CO or the CnoO association, or the two of
them, are relevant at a particular time and context. Apparently,
instructions and test questions can also activate one or the other
association just as contexts do. Moreover, according to Bouton
(1993), spontaneous recovery and renewal of responding occur
because the CnoO association acquired in a particular time and
context does not transfer to novel contexts as readily as the CO
association does. However, it is also possible that because the
novel context cannot be of great help in disambiguating whether O
or noO will occur after C, the two associations might become
partially activated in such cases by the test context. Thus, each of
them could interfere with the expression of the other one. In this
case, the combined excitatory and inhibitory strength of these two
associations could be responsible for the intermediate degree of
responding that is normally observed when testing occurs in a
novel (temporal or physical) context. This could also explain the
integrative judgments observed in the present research.
For example, the response-mode effect can be understood if we
assume that in the T groups the test trial must have been perceived
as an additional Phase 2 trial (there was nothing that allowed
participants to assume that it belonged to a different phase) and in
consequence, the default Phase 2 association (i.e., CnoO) pre-
vailed at test. The G groups, however, could not perceive the test
trial as an additional Phase 2 trial: It was clear that it was a whole
new (i.e., test) situation in which they were required to introduce
a judgment for the first time after the two training phases had been
completed. Therefore, the two associations could become equiva-
lently activated in such testing conditions, thus resulting in an
intermediate (or integrative) judgment.
Similarly, the causal and contiguity questions that we used were
explicitly worded to suggest that we were not referring exclusively
to the latest trials, and in consequence, participants integrated the
information from the two training phases (i.e., the two associations
become activated). However, the predictive test question explicitly
referred to a particular patient in a particular time and context (i.e.,
at the end of Phase 2). Thus, unless responding on a global mode
(Experiments 14) or receiving instructions to the contrary (Ex-
periment 4), participants receiving the predictive question tended
to respond as if the test trial belonged to Phase 2 (i.e., if patients
are no longer developing the allergic reaction and nothing has
changed for the last 20 trials, why should I predict that the next
patient should be different?) The observed type-of-question effect
provides a replication of the basic finding reported by Matute et al.
(1996) in that associations seem to be acquired on the basis of
mere contiguity, and these associations can then be flexibly used at
test. As a function of how the test phase is interpreted, one or more
associations can become activated. In consequence, interference
can be observed when more than one association becomes acti-
vated at test.
In the same vein, the instructional manipulation that we used in
Experiment 4 could readily be interpreted as serving the purpose of
separating the test trial from the Phase 2 trials. That is, these
instructions told participants in the T groups that the next trial (i.e.,
the test trial) was not to be regarded as an additional extinction
trial, but as a new type of trial. In consequence, both the CO and
the CnoO associations become similarly activated and interfered
with the expression of the other one.
Conditions That We Have Not Tested
There are many conditions that we have not tested. For example,
we have not tested how these variables apply to more complex
manipulations, such as those in which cues are presented in com-
pound (as in forward and backward blocking), or those in which
other trial types are used in addition to the ones we have used here
(e.g., noCO and noCnoO trials, instead, or in addition to, the
CO and CnoO trials that we used). These are interesting ques-
tions for future research because, for example, the observation of
both forward and backward blocking (i.e., evidence for an absence
of trial-order effects) in some studies (e.g., Shanks, 1985), along
with the observation that backward blocking is weaker and harder
to observe (i.e., evidence for a trial-order effect) in other studies
(e.g., Chapman, 1991), suggest that the conditions that yield to one
or the other type of result are not yet clear. It is possible that some
of the variables that we have tested here could also have an effect
on those compound conditions as well. Moreover, the results of
Catena et al. (1998), show that the response mode affects the
observation of trial-order effects in situations in which the four
trial types (CO, CnoO,noCO, and noCnoO) are manipulated.
Although they did not test the effect of the type of question in their
experiments, it is possible that both the type of question and the
instructions presented just before testing could also affect condi-
724
MATUTE, VEGAS, AND DE MAREZ
tions in which the four trial types, rather than the two used here,
are distributed throughout the experiment.
Finally, other response-mode manipulations, such as asking
participants to emit one judgment at the end of each phase, rather
than on every single trial, have been used in some studies of human
extinction (e.g., Vila, 2000). The results observed under those
conditions show sensitivity to the extinction trials. Thus, it seems
that as long as there is at least one judgment per phase, participants
tend to respond accordingly to the information provided during
that phase rather than to respond in an integrative manner. This,
again, suggests that trial-order effects depend on whether the test
question is interpreted as referring to the most recent phase or to
the whole training session.
References
Allan, L. G. (1980). A note on measurement of contingency between two
binary variables in judgment tasks. Bulletin of the Psychonomic Soci-
ety, 15, 147149.
Allan, L. G. (1993). Human contingency judgments: Rule based or asso-
ciative? Psychological Bulletin, 114, 435448.
Baker, A. G., Mercier, P., Valle´e-Tourangeau, F., Frank, R., & Pan, M.
(1993). Selective associations and causality judgments: Presence of a
strong causal factor may reduce judgments of a weaker one. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 19, 414
432.
Bouton, M. E. (1993). Context time and memory retrieval in the interfer-
ence paradigms of Pavlovian learning. Psychological Bulletin, 114,
8099.
Catena, A., Maldonado, A., & Ca´ndido, A. (1998). The effect of the
frequency of judgment and the type of trials on covariation learning.
Journal of Experimental Psychology: Human Perception and Perfor-
mance, 24, 481495.
Chapman, G. B. (1991). Trial order effects cue interaction in contingency
judgment. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 17, 837854.
Cheng, P. W. (1997). From covariation to causation: A causal power
theory. Psychological Review, 104, 367405.
Cheng, P. W., & Novick, L. R. (1992). Covariation in natural causal
induction. Psychological Review, 99, 365382.
Cobos, P. L., Can˜o, A., Lo´pez, F. J., Luque, J. L., & Almaraz, J. (2000).
Does the type of judgment required modulate cue competition? Quar-
terly Journal of Experimental Psychology: Comparative and Physiolog-
ical Psychology, 53(B), 193207.
Hastie, R., & Park, B. (1986). The relationship between memory and
judgment depends on whether the judgment task is memory-based or
on-line. Psychological Review, 93, 258268.
Hastie, R., & Pennington, N. (1995). Cognitive approaches to judgment
and decision making. In G. H. Bower (Ed.), The psychology of learning
and motivation (Vol. 32, pp. 131). San Diego, CA: Academic Press.
Hogarth, R. M., & Einhorn, H. J. (1992). Order effects in belief updating:
The belief-adjustment model. Cognitive Psychology, 24,155.
Lo´pez, F. J., Shanks, D. R., Almaraz, J., & Ferna´ndez, P. (1998). Effects
of trial order on contingency judgments: a comparison of associative and
probabilistic contrast accounts. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 24, 672694.
Matute, H., Arcediano, F., & Miller, R. R. (1996). Test question modulates
cue competition between causes and between effects. Journal of Exper-
imental Psychology: Learning, Memory, and Cognition, 22, 182196.
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.
109165). San Diego, CA: Academic Press.
Miller, R. R., & Escobar, M. (2001). Contrasting acquisition-focused and
performance-focused models of acquired behavior. Current Directions
in Psychological Science, 10, 141145.
Miller, R. R., & Matute, H. (1996). Animal analogues of causal judgment.
In D. R. Shanks, K. J. Holyoak, & D. L. Medin (Eds.), The psychology
of learning and motivation: Causal learning (Vol. 34, pp. 133166). San
Diego, CA: Academic Press.
Paredes-Olay, M. C., & Rosas, J. M. (1999). Within-subjects extinction
and renewal in predictive judgments. Psicologica, 20, 195210.
Pavlov, I. P. (1927). Conditioned reflexes. London: Clarendon Press.
Pennington, N., & Hastie, R. (1992). Explaining the evidence: Tests of
story model for juror decision making. Journal of Personality and Social
Psychology, 62, 189206.
Ratcliff, R. (1990). Connectionist models of recognition memory: Con-
straints imposed by learning and forgetting functions. Psychological
Review, 97, 285308.
Rescorla, R. A., & Wagner, A. R. (1972). A theory of Pavlovian condi-
tioning: Variations in the effectiveness of reinforcement and nonrein-
forcement. In A. H. Black & W. F. Prokasy (Eds.), Classical condition-
ing II: Current theory and research (pp. 6499). New York: Appleton-
Century-Crofts.
Shanks, D. R. (1985). Forward and backward blocking in human contin-
gency judgment. Quarterly Journal of Experimental Psychology: Com-
parative and Physiological Psychology, 37(B), 121.
Shanks, D. R. (1993). Human instrumental learning: A critical review of
data and theory. British Journal of Psychology, 84, 319354.
Shanks, D. R., Lo´pez, F. J., Darby, R. J., & Dickinson, A. (1996).
Distinguishing associative and probabilistic contrast theories of human
contingency judgment. In D. R. Shanks, K. J. Holyoak, & D. L. Medin
(Eds.), The psychology of learning and motivation: Causal learning
(Vol. 34, pp. 265311). San Diego, CA: Academic Press.
Vila, N. J. (2000). Extincio´n e inhibicio´n en juicios de causalidad [Extinc-
tion and inhibition in causality judgments]. Psicologica, 21, 257273.
Wasserman, E. A. (1990). Attribution of causality to common and distinc-
tive elements of compound stimuli. Psychological Science, 1, 298302.
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:
Causal learning (Vol. 34, pp. 207264). San Diego, CA: Academic
Press.
Received January 4, 2001
Revision received February 1, 2002
Accepted February 1, 2002
725
RECENCY VS. INTEGRATIVE JUDGMENTS
... Bouton's model can accommodate these findings by assuming that they do also involve a change in the (subjective) context. More recently, human contingency learning experiments have shown that procedural details like the instructions provided to participants, the frequency and type of responses requested, and the presence of filler trials can also modulate the retrieval of first-or secondlearned information in interference paradigms (Matute et al., 2002Matute et al., , 2011 Vadillo et al., 2004; Rosas et al., 2006a). The results of our experiments and those conducted by Vervliet et al. (2004Vervliet et al. ( , 2005) suggests that the configuration of the target cues might also have the functional properties of a context change (see also, Bouton and Swartzentruber, 1991; Rowe and Craske, 1998; Lovibond et al., 2000). ...
Article
Full-text available
Decades of research in extinction and interference show that contexts can play a critical role at disambiguating the meaning of cues that have been paired with different outcomes at different times. For instance, if a cue x is followed by outcome 1 in the first phase of an experiment and by outcome 2 in a second phase, responses to cue x tend to be consistent with outcome 2 when tested in a context similar to that of the second phase of the experiment. However, if participants are taken back to the original context of the first phase (i.e., ABA renewal) or to a completely new context (i.e., ABC or AAB renewal), their responses to x tend to be more consistent with outcome 1. Although the role of physical and temporal contexts has been well studied, other factors that can also modulate the selective retrieval of information after interference have received less attention. The present series of experiments shows how changes in cue configuration can modulate responding in a similar manner. Across five experiments using a human predictive learning task, we found that adding, removing or replacing elements from a compound cue that had undergone an interference treatment gave rise to a recovery of responding akin to that observed after context changes in AAB renewal. These results are consistent with those of previous studies exploring the effect of changes of cue configuration on interference. Taken together, these studies suggest that a change in cue configuration can have the functional properties of a context change, a finding with important implications for formal models of configural learning and for classical accounts of interference and information retrieval.
... We do not question the idea that there are manipulations that might affect one dependent variable without affecting others. In fact, part of our own research has been directed at showing that judgments of contingency can vary considerably depending on seemingly minor procedural details like the wording of the test question (Matute, Vegas, & De Marez, 2002; Vadillo et al., 2005; see also Crocker, 1982; De Houwer et al., 2007; Perales & Shanks, 2008). Moreover, part of our discussion of Perales et al. (2005) relies on the idea that participants sometimes adopt response strategies that might mask their true perception of contingency. ...
Article
Decades of research in causal and contingency learning show that people’s estimations of the degree of contingency between two events are easily biased by the relative probabilities of those two events. If two events co-occur frequently, then people tend to overestimate the strength of the contingency between them. Traditionally, these biases have been explained in terms of relatively simple single-process models of learning and reasoning. However, more recently some authors have found that these biases do not appear in all dependent variables and have proposed dual-process models to explain these dissociations between variables. In the present paper we review the evidence for dissociations supporting dual-process models and we point out important shortcomings of this literature. Some dissociations seem to be difficult to replicate or poorly generalizable and others can be attributed to methodological artefacts. Overall, we conclude that support for dual-process models of biased contingency detection is scarce and inconclusive.
... In addition to active vs. passive roles of participants, there are many other variants that can be introduced in this task and that have been shown to affect the participants' estimations of causality. Examples include changing the wording of questions asked at the end of the experiment about the causal relationship (Crocker, 1982; Vadillo et al., 2005 Vadillo et al., , 2011 Collins and Shanks, 2006; De Houwer et al., 2007; Blanco et al., 2010; Shou and Smithson, 2015), the order in which the different trial types are presented (Langer and Roth, 1975; López et al., 1998), the frequency with which judgments are requested (Collins and Shanks, 2002; Matute et al., 2002), the description of the relevant events as causes, predictors, or effects (Waldmann and Holyoak, 1992; Cobos et al., 2002; Pineño et al., 2005), the temporal contiguity between the two events (e.g., Shanks et al., 1989; Wasserman, 1990; Lagnado and Sloman, 2006; Lagnado et al., 2007), and many other variables that fortunately are becoming well known. In the following sections, we will focus on the variables that seem to affect the illusion most critically in cases of null contingency. ...
Article
Full-text available
Illusions of causality occur when people develop the belief that there is a causal connection between two events that are actually unrelated. Such illusions have been proposed to underlie pseudoscience and superstitious thinking, sometimes leading to disastrous consequences in relation to critical life areas, such as health, finances, and wellbeing. Like optical illusions, they can occur for anyone under well-known conditions. Scientific thinking is the best possible safeguard against them, but it does not come intuitively and needs to be taught. Teaching how to think scientifically should benefit from better understanding of the illusion of causality. In this article, we review experiments that our group has conducted on the illusion of causality during the last 20 years. We discuss how research on the illusion of causality can contribute to the teaching of scientific thinking and how scientific thinking can reduce illusion.
... Algunos autores han señalado que esta actuación puede reflejar una integración de la información acerca de las dos fases del experimento. Esto es, el cambio de contexto podría llevar a los participantes a tomar en cuenta toda la información disponible en el experimento, comportándose de un modo estadísticamente apropiado a la relación real entre la clave y las dos consecuencias (v.g., Alvarado, Vila, Jara y Rosas, 2003; Matute, Vegas y De Marez, 2002). De acuerdo con esta idea, delta P para la relación A+ iguala a delta P para la relación A* en nuestros experimentos. ...
Article
Full-text available
Resultados recientes en la literatura ponen en cuestión si la atenuación de la interferencia retroactiva que se produce como consecuencia del cambio de contexto tras la fase de interferencia se deben al abandono de este contexto de interferencia (Bouton, 1993) o al regreso al contexto de adquisición. Se realizaron dos experimentos para comprobar si era posible obtener renovación en un paradigma de aprendizaje causal. Se utilizó una tarea de interferencia retroactiva que consistía en emparejar una clave y una consecuencia (A+) y posteriormente esa misma clave con una consecuencia incompatible (A*). El cambio de contexto tras la fase de interferencia produjo una atenuación de la interferencia retroactiva independientemente de si este cambio supuso un regreso al contexto de adquisición (Experimento 1) o simplemente el abandono del contexto de interferencia (Experimento 2). Los resultados obtenidos se ajustan a lo predicho desde la teoría de la recuperación de Bouton (1993).
Article
Full-text available
Temporal binding occurs when an action and an outcome that follows it after a short period of time are judged as occurring closer to each other in time than they actually are. This effect has often been studied using Libet’s clock methodology. Garaizar et al. (2016) presented Labclock Web, a free HTML5 open source software that allows researchers to conduct temporal binding and other experiments using Libet’s clock through the Internet. The purpose of the three experiments presented here was to test how certain methodological modifications in the Labclock Web task could impact the temporal binding effect. In comparison with the original study, we aimed to: (a) reduce the interval between action and outcome in the delayed condition to 100 ms, instead of 500, (b) present the two types of trials, immediate and delayed, in two separate consecutive blocks, instead of intermixed, (c) use a visual, rather than auditory, outcome following the action, and (d) reduce the number of trials. In addition to its potential theoretical implications, the results confirm that Labclock Web is a useful and reliable tool for conducting temporal binding experiments and that it is well suited to measure temporal binding effects in a broad range of situations.
Article
Full-text available
Illusory causation refers to a consistent error in human learning in which the learner develops a false belief that two unrelated events are causally associated. Laboratory studies usually demonstrate illusory causation by presenting two events—a cue (e.g., drug treatment) and a discrete outcome (e.g., patient has recovered from illness)—probabilistically across many trials such that the presence of the cue does not alter the probability of the outcome. Illusory causation in these studies is further augmented when the base rate of the outcome is high, a characteristic known as the outcome density effect. Illusory causation and the outcome density effect provide laboratory models of false beliefs that emerge in everyday life. However, unlike laboratory research, the real-world beliefs to which illusory causation is most applicable (e.g., ineffective health therapies) often involve consequences that are not readily classified in a discrete or binary manner. This study used a causal learning task framed as a medical trial to investigate whether similar outcome density effects emerged when using continuous outcomes. Across two experiments, participants observed outcomes that were either likely to be relatively low (low outcome density) or likely to be relatively high (high outcome density) along a numerical scale from 0 (no health improvement) to 100 (full recovery). In Experiment 1, a bimodal distribution of outcome magnitudes, incorporating variance around a high and low modal value, produced illusory causation and outcome density effects equivalent to a condition with two fixed outcome values. In Experiment 2, the outcome density effect was evident when using unimodal skewed distributions of outcomes that contained more ambiguous values around the midpoint of the scale. Together, these findings provide empirical support for the relevance of the outcome density bias to real-world situations in which outcomes are not binary but occur to differing degrees. This has implications for the way in which we apply our understanding of causal illusions in the laboratory to the development of false beliefs in everyday life. Electronic supplementary material The online version of this article (10.1186/s41235-018-0149-9) contains supplementary material, which is available to authorized users.
Article
Full-text available
Select literature regarding cue competition, the contents of learning, and retrieval processes is summarized to demonstrate parallels and differences between human and nonhuman associative learning. Competition phenomena such as blocking, overshadowing, and relative predictive validity are largely analogous in animal and human learning. In general, strong parallels are found in the associative structures established during learning, as well as in the basic phenomena associated with information retrieval. Some differences arise too, such as retrospective evaluation, which seems easier to observe in human than in nonhuman animals. However, the parallels are sufficient to indicate that the study of learning in animals continues to be relevant to human learning and memory.
Book
Full-text available
Este libro nace a partir de un symposium titulado "Extinción y recuperación de la información en aprendizaje causal: perspectivas teóricas" organizado en el marco del XV Congreso de la Sociedad Española de Psicología Comparada, celebrado en Barcelona en septiembre del año 2003 y en el que la mayoría de los grupos hispanos dedicados a este tema presentaron los últimos avances de sus investigaciones y reflexiones teóricas.
Article
Full-text available
This study showed that accuracy of the estimated relationship between a fictitious symptom and a disease depends on the interaction between the frequency of judgment and the last trial type. This effect appeared both in positive and zero contingencies (Experiment 1), and judgments were less accurate as frequency increased (Experiment 2). The effect can be explained neither by interference of previous judgments or memory demands (Experiment 3), nor by the perceptual characteristics of the stimuli (Experiments 4 and 5), and instructions intended to alter processing strategies do not produce any reliable effect. The interaction between frequency and trial type on covariation judgment is not predicted by any model (either statistical or associative) currently used to explain performance in covariation detection. The authors propose a belief-revision model to explain this effect as an important response mode variable on covariation learning.
Article
Ciiven the task of di the source of a patient's aUer^'ic reav-tion. college students jiuigcii the causal efficacy of common (A') and distinctive (A and Bj elements of compound stimuli: AX and BX. As the differential correlation of AX and BX with the occurrence and nonoccurrence ofthe allergic reaction rose from .00 to 1.00. ratings of ihe distinctive A and B elements diverged; most importantly, ratings ofthe common X element fell. These causal judgments of humans closely parallel the conditioned responses of animals in associa-tive learning studies, and clearly disclose that stimuli compete with one another for control over behavior.
Article
The study of the mechanism that detects the contingency between events. in both humans and non-human animals, is a matter of considerable research activity. Two broad categories of explanations of the acquisition of contingency information have received extensive evaluation: rule-based models and associative models. This article assesses the two categories of models for human contingency judgments. The data reveal systematic departures in contingency judgments from the predictions of rule-based models. Recent studies indicate that a contiguity model of Pavlovian conditioning is a useful heuristic for conceptualizing human contingency judgments.
Article
Four experiments examined trial sequencing effects in human contingency judgment. In Experiments 1–3, ratings of contingency between a target cue and outcome were affected by the presentation order of a series of trials distributed in 2 distinct blocks and showed a recency bias. Experiment 4 replicated this effect when the trials were partly intermixed. These recency effects are predicted by an associative learning model that computes associative strengths trial by trial and incorporates configural coding of cues but are problematic for probabilistic contrast accounts, which currently have no provision in the contingency computation for the differential weighting of trials as a function of their order of presentation. (PsycINFO Database Record (c) 2012 APA, all rights reserved)
Article
This chapter discusses that the cognitive approach to judgment and decision making behavior is gradually developing and expanding in scope and in number of adherents. It presents a collection of experimental methods and results from one laboratory studying several different judgment tasks to illustrate some major trends that are central in the cognitive approach: evidence representation in the form of a model or explanation of the decision situation, enhanced memory for decision-relevant information after a decision has been made, and adaptive flexibility in individual decision strategies in response to variations in task conditions. One obstacle for the development of cognitive theories of judgment and decision making behavior is that there are considerable differences among the theories that are called cognitive. This situation is probably good for the eventual development of cognitive theories, but it is troublesome for individual researchers, especially those who are currently attempting to spread the faith to new task domains. In the field of judgment and decision making, it is essential to think about the relationships between alternate theoretical developments and formulations based on traditional expected utility theory. The expectation is that the crucible of empirical evaluation forces all of the approaches to converge on a common theoretical framework. The chapter describes that cognitive precepts are core of the next generation of theories of judgment and decision making.
Article
What is encoded during a learning experience? In a learning situation, a training (acquisition) phase is often followed by a test phase. Acquisition-focused models (most associative models) emphasize information processing that occurs during training and assume that only summary statistics (associative values) are retained to influence behavior during testing. Performance-focused models (a k a “computational” models) emphasize information processing that occurs at test and often assume fairly comprehensive memory of prior experiences. In this article, we contrast these two families of models. We note that neither family can claim unique support from the various learning phenomena commonly cited as favoring one or the other position. Within each family, there are highly diverse models that preclude blanket generalizations. However, the acquisition-focused view that subjects retain only summary statistics suggests unrealistically impoverished memories relative to the fairly veridical memories that clearly are formed. But this oversimplifying assumption of acquisition-focused associative theories is exactly what has made them superior to performance-focused models to date in stimulating research.
Article
This chapter reviews that nonverbal behavioral assessment of causal judgment is apt to be more veridical than is verbal assessment, which is compromised by the demand characteristics and ambiguities of language. Organisms presumably evolved the ability to learn cause-effect relationships in order to prepare for and sometimes influence future events in the real world, not in order to verbally describe these causal relationships. The use of nonverbal behavioral assessment invites direct comparisons between human causal judgment behavior and animal behavior in similar situations. Cues of high biological relevance appear to be relatively invulnerable to cue competition compared to cues of low biological relevance, which are quite susceptible to cue competition. It discusses that this convergence of findings in the causal judgment and animal learning literatures suggests that the two fields can each benefit by attending to the findings of the other. Another likely finding from studies of cue competition in animals that is profitably examined in causal judgment situations with humans is the learning-performance distinction. There is also some discussion that causal judgments results from those associations that have a forward relationship from one event to another and that are not nor in competition with other associations that are active at the time the target association is tested.
Article
This chapter discusses theoretical issues concerning contingency judgment. One empirical result exists that appears straightaway to challenge the idea that contingency judgments can be modeled by the Rescorla-Wagner theory. This is the finding that judgments under noncontingent schedules do not always appear to converge across trials. The idea that stimuli are represented configurally allows the results of the experiments to be accommodated; it should be acknowledged that there are a number of problems facing this approach. Account of retrospective revaluation effects requires an elemental rather than a configural analysis: in an AB → 0, B → 0 design, subjects are assumed to relate what they learn in the second stage about element B to what they already know about compound AB, such that the balance of associative strengths of A and B is altered. It is difficult to see how a configural analysis, whereby the compound AB is represented quite independently of its elements, would allow this to happen. Some recent data raise the possibility that subjects behave configurally only under certain conditions. Many researchers agree that the appropriate normative theory is provided by the Δp metric: contingency judgments should then be evaluated for their objective accuracy against Δp and are assumed to be biased whenever they deviate from that statistic. Rather than proving that contingency judgment is nonnormative, however, results should be viewed in the same way as visual illusions: manifestations of an incorrect output from a system that fundamentally does provide a true picture of the world but that can be misled as a result of having to produce a response on the basis of insufficient evidence.