Download Second-order conditioning of human causal learning

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Confirmation bias wikipedia , lookup

Educational psychology wikipedia , lookup

Milgram experiment wikipedia , lookup

Experimental psychology wikipedia , lookup

Structural equation modeling wikipedia , lookup

Vladimir J. Konečni wikipedia , lookup

Operant conditioning wikipedia , lookup

Learning theory (education) wikipedia , lookup

Psychological behaviorism wikipedia , lookup

Classical conditioning wikipedia , lookup

Transcript
Learning and Motivation 37 (2006) 230–246
www.elsevier.com/locate/l&m
Second-order conditioning of human causal
learning 夽
Elvia Jara a, Javier Vila a,¤, Antonio Maldonado b
a
Universidad Nacional Autónoma de México, F.E.S. Iztacala, México
b
University of Granada, Spain
Received 11 March 2005; received in revised form 30 November 2005
Available online 17 February 2006
Abstract
This article provides the Wrst demonstration of a reliable second-order conditioning (SOC) eVect
in human causal learning tasks. It demonstrates the human ability to infer relationships between a
cause and an eVect that were never paired together during training. Experiments 1a and 1b showed a
clear and reliable SOC eVect, while Experiments 2a and 2b demonstrated that Wrst-order extinction
did not aVect SOC. These results were similar to those found in animal and human conditioning and
suggested that a similar associative mechanism could explain these eVects. However, they can also be
used to look into the underlying causal mental model people build and store while they are learning
this task. From a cognitive view, overall results suggest that an independent rather than a chain
causal mental model is stored after second-order learning in human causal tasks.
© 2005 Elsevier Inc. All rights reserved.
Keywords: Pavlovian conditioning; Second-order conditioning; Extinction; Causal learning
Second-order conditioning (SOC) has been a widely studied topic of associative
learning, because it allows the study of an emergent relationship between two events
that were never presented together (Pavlov, 1927; Rizley & Rescorla, 1972). In the SOC
procedure, there are three phases. In the Wrst training phase, a conditioned stimulus
夽
This research was supported by founds from DGAPA-UNAM (IN302605) and CONACYT (34843-H) and
by BSO2003-03723 Spanish research project, granted by MCYT (Ministerio de Ciencia y Tecnología). We gratefully acknowledge the helpful comments of Bruce Overmier.
*
Corresponding author.
E-mail address: [email protected] (J. Vila).
0023-9690/$ - see front matter © 2005 Elsevier Inc. All rights reserved.
doi:10.1016/j.lmot.2005.12.001
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
231
(CS1) is followed by an unconditioned stimulus (US). In the second phase, a secondorder conditioned stimulus (CS2) is presented along with CS1. Finally, in the test phase,
CS2 is presented alone to the subjects while their responses are recorded. The results
have commonly shown that CS2 evokes the same conditioned response (CR) as does
CS1, even though CS2 was never directly paired with the US (Pavlov, 1927). SOC has
been studied in animals employing diVerent types of US and responses. For instance, it
has been demonstrated in aversive conditioning (Bevins, Delzer, & Bardo, 1996; Cheatle
& Rudy, 1979; Rescorla, 1976), in autoshaping of keypecking (Rashotte, GriYn, & Sisk,
1977; Rescorla & Cunningham, 1979) and in sexual behavior (Crawford & Domjan,
1995). Also, there have been demonstrations in humans using non-verbal stimuli in
aversive electrodermal conditioning (Davey & McKenna, 1983; Davey & Arulampalan,
1982).
Many theoretical models have been oVered, in an attempt to explain SOC; most of them
are derived from associative learning theories. These models can be grouped into four
diVerent types based on the diVerent possible kinds of associations that could be formed
during SOC, as shown in Fig. 1 (Barnet, Cole, & Miller, 1997). The most traditional model,
which is based upon the S-R model, as well as the Rescorla and Wagner (1972) model
(Rashotte, 1981; Rizley & Rescorla, 1972), proposed the acquisition of a direct connection
between CS2 and CR reinforced by the presence of CS1, due to temporal contiguity
between them during second-order training. A second type of model, the CS2–CS1–US–
CR model, suggests the existence of a chain of associative representations in such a way
that CS2 will evoke a representation of CS1 which will evoke the US representation that in
turn elicits CR (Hall, 1996; Mackintosh, 1974). A third type of model assumes the establishment of a direct association between CS2 and the US representation, because during
the second phase the US representation could be activated in the presence of CS2 by CS1;
due to this association, CS2 can evoke the CR (Barnet & Miller, 1996; Miller & Barnet,
1993). The fourth type of model considers the existence of an association between CS2 and
CS1 where CR emission is only due to the CS1 representation evoked by CS2 (Miller &
Escobar, 2002).
(1)
CS2
CS1
CR
(2)
CS2
CS1
US
CR
(3)
CS2
CS1
US
CR
(4)
CS2
CS1
US
CR
Fig. 1. DiVerent hypotheses about the associations established during SOC, derived from associative learning theories. Letters represent: CS1, Wrst-order stimulus; CS2, second-order stimulus; US, unconditioned stimulus; CR,
conditioned response.
232
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
To evaluate the nature of the learned associations during SOC, several post-conditioning revaluation techniques have been used, such as CS1 extinction (Rescorla, 1980; Rizley
& Rescorla, 1972), US devaluation (Crawford & Domjan, 1995; Holland & Rescorla,
1975), as well as instructions and habituation in humans (Davey & McKenna, 1983).
Among these, CS Wrst-order extinction had been the most used one, in which, after SOC
had been established, CS1 is presented without being followed by the US during a third
phase. The most frequent results, both in animals (Fujii, 1981; Rescorla, 1980; Rizley &
Rescorla, 1972) and humans (Davey & McKenna, 1983; Davey & Arulampalan, 1982),
have been that the CR evoked by CS2 persists more or less at the same magnitude, in spite
of the fact that the CR evoked by CS1 has been extinguished. The Wnding that CS2 still
evokes CR even after extinction of CS1 appears to suggest that the associations and the
contents of learning during a second-order conditioning could be diVerent from those of
Wrst-order learning. Although Wrst-order learning is aVected by CS1 extinction, by US
devaluation, by US habituation or instructions, SOC is not subsequently inXuenced by any
of these factors (Davey, 1983; Holland & Rescorla, 1975; Rizley & Rescorla, 1972). These
results suggest that the types of associations formed during second-order conditioning are
not mediated by CS1. However, the debate still continues whether associations are between
CS2 and CR or are mediated by the US representation.
Recent research into human learning using causal learning tasks could clarify this issue.
The interest in studying associative learning phenomena in humans led Dickinson, Shanks,
and Evenden (1984) to suggest that causal or contingency learning tasks could be considered similar to animal conditioning procedures. Contingency learning tasks consist in presenting a series of trials in which a cue (CS) and an outcome (US) may or may not be
presented together. After several trials, the participants are asked to give a judgment about
the degree of association between cues and outcomes (Dickinson et al., 1984). In a causal
learning task, causes are considered formally equivalent to CSs and, in consequence, the
eVects equivalent to USs in such a way that causal judgment can reXect the degree of association between them. Contingency, as well as causal learning tasks, have allowed the replication of diverse eVects of associative learning in humans, which were originally studied in
animals, enlarging the Weld of knowledge on explanatory models of learning (Shanks,
1995).
Although second-order conditioning has not yet been directly studied in humans using
causal learning tasks, Perales, Catena, and Maldonado (2004) have recently shown how
people are able to infer a relationship between two diVerent cues (C1, C2), even when
causes and eVects were never presented together, based on two diVerent and independent
experiences where each cue was paired with the same consequence. They also showed that
the most frequent assumption during the causal inference process was an independent
causal mental model. In their experiment, when the two cues were causes, (C1, C2) almost
all participants used a common eVect (E) mental model in which each cause could inXuence
independently the same eVect, i.e., a C1–E and C2–E mental model.
In second-order causal learning, however, the Wrst phase C1 produces a given eVect and
during the second phase another cause, C2, produces C1 as an eVect, this arrangement
could give rise to an inference process in which people learn that the second-order cause is
the cause of the Wrst-order one, which in turn produces the common eVect (i.e., C2–C1–E).
In this type of procedure, the normative causal mental model should be a chained one,
being more compatible with a CS2–CS1–US associative model. This chained mental model
could easily explain second-order conditioning in learning tasks as a process of transitive
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
233
inference, but it has been shown that after extinction of the Wrst component in animal and
human second-order conditioning tasks, CS2 still elicits CR. Such results argue against a
chained causal mental model because a post-conditioning extinction procedure, in which
people learn that the C1 does not produce the eVect, should also produce a disruption of
the causal relationships between C1 and the same eVect. Therefore, the use of such a procedure might reveal whether the results are or not the same in a causal learning and a conditioning task and might also clarify whether a chained mental model underlies causal
attributions during SOC.
In summary, Experiments 1a and 1b aimed to show the existence of SOC in humans
using causal learning tasks, in a similar way to that in which SOC was shown in classical
conditioning studies. The experimental task consisted of a series of trials presented to the
participants. In the Wrst phase, a cause C1 was followed by an eVect E1. In the second
phase, a diVerent cause C2 produced the cause C1. In the testing phase, the participants
were asked to judge to what degree C2 will cause the appearance of the eVect E1. The inference of a causal relationship between a second-order cause (C2) and an eVect (E) that were
never presented together would also bear out those associative models proposing a direct
connection between causes (CS) and eVects (US), independently of any CR elicitation that
could also be inXuenced by other factors, such as context, memory, and emotion. The second and more important objective was to evaluate the associative structure of SOC in
causal learning tasks by the subsequent extinction of the Wrst-order relationship (C1–E1).
Consequently, Experiments 2a and 2b used the same experimental task, but after SOC
training the participants experienced a third phase of extinction where C1 no longer produced E1. Then, they were asked about C2–E1 to Wnd out if SOC was or was not extinguished as well. The use of this extinction technique had two goals. First of all, the results
will show whether or not the same eVects happen in conditioning and causal learning tasks
and, second, the results can permit us to clarify whether or not a chained causal mental
model underlies SOC in human learning tasks.
Experiment 1a
The main objective of Experiment 1a was to demonstrate a second-order conditioning
phenomenon within human causal learning. To this end, a similar experimental design as
the one used by Rizley and Rescorla (1972) in animals was employed. In a Wctitious situation, all participants learned during the Wrst phase that a disease (C1) caused the appearance of a substance in the blood (E1) in several patients. In the second phase, a chemical
substance (C2) consumed by the patients caused the disease presented in the Wrst phase
(C1). At the end of both training phases, the participants were asked to make a causal judgment about the relationship between the second-order cause (C2, the chemical substance)
and the eVect (E1, appearance of a substance in the blood). As a control, another cause
(C3) and a second eVect (E2) were also presented during the Wrst phase and another new
cause (C4) and C3 were presented during the second-order phase. For half of the participants, C3 and the eVect (E2) were paired during the Wrst phase, but C3 and C4 were never
paired during the second phase. For the other half of the participants, C3 and E2 were not
paired during the Wrst phase, while C4 and C3 were presented together during the second
phase. In both cases, participants should infer the existence of causal relationships between
the second-order cause (C2) and the Wrst eVect (E1), while they should infer the non-existence of causal relationships between the other cause (C4) and the second eVect (E2).
234
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
Method
Participants
The participants were 24 undergraduate students of psychology between 18- and 25years-old who volunteered to take part in the experiment as part of their course requirements. The participants were randomly divided into two groups (12 participants in each
condition).
Apparatus and stimuli
An IBM compatible computer and a projector were used in the experiment. The Super
Lab Pro v 2.1 software for Windows (Cedrus Co.) was used to show the cues.
Cues C1 and C3 were names of two diVerent diseases (Midiasis or Xeritis), the eVects E1
and E2 were blood substances (Alpha or Beta), and Wnally C2 and C4 were chemical substances (Neocina or Licaina). All of them were used in this experiment as stimuli, and they
were counterbalanced across the participants.
Procedure
The experiment was conducted in a dark room, with half of the participants tested at
each of two sessions (because of counterbalance). They were seated in a chair facing a
screen, to which slides of the stimuli were to be projected. Paper and pencil were provided
to register their prediction and their judgment. The experiment was conducted in Spanish;
instructions translated into English were as follows:
Two new diVerent diseases were found in ‘Gersy Laboratories’ in The Netherlands,
and these diseases are known as Midiasis and Xeritis. These diseases are transmitted
quickly. You should identify whether these diseases are related to the substances
Alpha and Beta that are found in the blood of patients. Also, you will see that
patients ate sausages which contain the substances Neocina and Licaina, and you
should identify if theses chemical substances, used in food preservation, are related to
these diseases. By pressing keys, you will Wnd out whether they have some relation.
At the beginning, your answer will be uncertain, but little by little you will become an
expert. If you detect some changes in the experiment you should continue. Remember
that you can take as much time as you need to complete the experiment.
Press any key to continue.
After the participants read the instructions, the training phase began. The training
screen was present as follows (relations could change because of the counterbalance). At
the top of the screen was centered the laboratory name “Gersy Laboratories,” in pink capital letters. Below the laboratory name, and depending on the counterbalancing, a sentence
such as the following appeared in purple color: “The patient has Midiasis. Does this disease
cause the appearance of the blood substance Alpha?” Under the question, the name of the
two blood substances appeared in blue color each one associated with a key number,
Alpha with the number three and Beta with the number Wve. On each trial, the participants
had to give a prediction; for instance, if participants considered that Midiasis would cause
Alpha, they had to write the number three on a sheet of paper; on the contrary, if they considered that the disease would cause Beta, they had to write the number Wve. Once the prediction was made, the participants were given the correct feedback (1200 ms duration).
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
235
Experiment 1a employed both between and within design variables. Table 1 summarizes
the design of this experiment. For all participants, the experiment had two diVerent phases
of training; each phase consisted of two blocks of 12 diVerent trials, Paired (P) or nonpaired (N) presented randomly and intermixed. In each one of the two groups (PP/PN and
PP/NP), the causes C1 and C2 were always followed by their eVects (E1 or C1) in 12 trials
during each phase (the SOC or paired PP condition). In the non-paired PN or NP second
condition, the causes (C3 and C4) were followed (P) or not (N) by their eVects (E2 or C3)
in another 12 trials during the Wrst and the second phases.
After both training phases, all participants were asked for causal judgments about the
relationship between both the second-order paired cause and its possible eVect (C2–E1) and
the non-paired cause and its possible eVect (C4–E2). The participants were asked to make
causal judgments according to the following instructions, “To what degree do you believe
Neocina causes the blood substance Alpha to appear? You must use a 0–100 scale, where 0
represents that it never produced the substance and 100 that it always produces the substance.” Below the instructions, the stimulus name was written on the left and a little square
was drawn on the right; Wnally, a numerical scale from 0 to 100 appeared with marks at 0,
25, 50, 75, and 100%. The participants had to give their judgments on a sheet of paper.
Results and discussion
The upper in Fig. 2 panel shows the mean judgments for the SOC test phase. Participants estimated a higher relationship in SOC (C2–E1) type of causes than in the nonpaired (C4–E2) ones. A 2 £ 2 analysis of variance (ANOVA), with the Wrst factor being the
two Groups (PP/NP vs. PP/PN) and the second factor the two Types of Cause (C2 and C4)
was performed on the participants’ judgments. The results revealed only a main eVect of on
Type of Cause, F (1, 22) D 11.43, p < .01, MSE D 1714.67, because in both groups, causal
judgments about the relationships between the second-order cause and the eVect, C2–E1
Table 1
Experimental designs
Experiment
Group
(n D 12)
Phase 1
1a
PP/NP
C1–E1
C3–No E2
C1–E1
C3–E2
PP/PN
1b
PP/NP
PP/PN
C1–E1
C3–No E2
C1–E1
C3–E2
2a
C1–E1
C3–E2
2b
C1–E1
C3–E2
Test
Phase 2
Test
Phase 3
Test
C2–C1
C4–C3
C2–C1
C4–No C3
C1–E1?
C3–E2?
C1–E1?
C3–E2?
C2–C1
C4–C3
C2–C1
C4–No C3
C2–E1?
C4–E2?
C2–E1?
C4–E2?
C2–C1?
C4–C3?
C2–C1?
C4–C3?
C2–C1
C4–C3
C1–E1?
C3–E2?
C2–C1
C4–C3
SOC Test
C2–E1?
C4–E2?
C2–E1?
C4–E2?
C1–No E1
C5–No E2
C2–C1?
C4–C3?
C1–No E1
C5–No E2
C2–E1?
C4–E2?
C1–E1?
C5–E2?
C2–E1?
C4–E2?
Note. PP represents Paired type of causes. NP and PN represent Non-paired. The Wrst letter represent Wrst phase,
and the second letter represents the second phase, with the slash dividing the type of causes. The rest of the letters
represents the diVerent stimuli presented counterbalanced along training. C1, C3, and C5: diseases Midiasis, Xeritis or Zatuba. E1 and E2: blood substances Alpha or Beta. C2 and C4: chemical substances Neocina or Licaina.
236
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
100
EXPERIMENT 1a
80
60
40
Mean Judgment
20
0
100
EXPERIMENT 1b
80
60
40
20
C2-E1
C4-E2
0
PP / NP
GROUP
PP / PN GROUP
Fig. 2. Mean judgments for the SOC test. Open bars represent SOC Type of Cause, and hatched bars represent
Non-paired Type of Cause. Error bars denote standard errors of the mean.
(the paired PP conditions, i.e., SOC) were signiWcantly higher than those about C4–E2 (the
non-paired NP or PN conditions).
These results are the Wrst demonstration of SOC in human causal learning tasks, and
they also show that SOC in causal learning tasks, as well as in conditioning tasks, depends
on paired presentations in both training phases. These Wndings revealed a SOC eVect similar to that commonly observed in animals using appetitive and aversive conditioning
(Crawford & Domjan, 1995, Rizley & Rescorla, 1972; Rescorla, 1980) as well as in humans
using electrodermal associative conditioning (Davey & McKenna, 1983; Davey & Arulampalan, 1982). The results showed that people are able to infer a causal relationship between
a second-order cause (C2) and an eVect (E1) that were never presented together. This inference comes from independent experiences, Wrst, from whether the eVect was related to a
primary cause (C1) and, second, whether this cause was also an “eVect” of the second-
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
237
order cause (C2). The control for the type of causes showed the need for a positive correlation between elements in both phases to get such a second-order eVect, just as is the case in
associative conditioning procedures. The next experiment aimed to replicate these Wndings
while providing evidence for the existence of causal learning during each phase.
Experiment 1b
Although Experiment 1a showed SOC occurs in causal learning tasks, it had only a Wnal
test at the end of all training phases. This fact allows for the possibility that the results were
not based on learned relations in the training phases. To prove that phases 1 and 2 did produce causal learning, the Experiment 1b objective was to replicate the SOC eVect, while
investigating whether the participants accurately detected the contingent relationships
established among causes and eVects in each experimental phase. In this experiment, participants were asked to make a causal judgment not only at the end of training, but also
after each experimental training phase.
Method
Participants
Twenty-four undergraduate students of psychology participated in this experiment for
course credits. Their ages varied between 18 and 24, and they were randomly divided into
two groups (12 participants in each condition).
Procedure
The procedure and design were identical to those described in Experiment 1a, the only
diVerence being that the participants were requested to estimate the relationships between
causes and eVects at the end of each training phase. Finally, to assess the SOC association,
the participants were asked to judge the relationships between C2–E1 and C4–E2 as in the
previous experiment.
Results and discussion
Table 2 shows the mean judgment for each relationship after each experimental phase.
Participants were able to accurately detect both the relationship between causes and their
contingent eVects and the lack of relationships between non-paired causes and eVects.
Table 2
Mean judgment after each experimental phase for Experiments 1b and 2b
Experiment
Group
Type of Cause
First phase
Second phase
1b
PP/NP
SOC
UNP
SOC
UNP
91.66
4.16
86.25
83.75
97.91
95.83
84.16
20.83
EXT
SOC
89.58
89.58
91.25
87.91
PP/PN
2b
Third phase
SOC Test
79.16
18.33
72.91
33.33
4.16
8.3
66.16
67.91
Note. Types of causes are represented by: SOC Second-order, UNP Non-paired, and EXT Extinction. Types of
causes in Experiment 1b are represented by PP: Paired and NP: Non-paired.
238
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
Higher causal judgments were obtained when causes were followed by their eVects than
when they were not.
The 2 £ 2 £ 2 ANOVA (Groups £ Type of Cause £ Training phase) performed upon
causal judgments yielded a signiWcant interaction between the three factors,
F (1, 22) D 53.12, p < .01, MSE D 456.4, as well as the interaction between Type of
Cause £ Training phase, F (1, 22) D 28.30, p < .01, MSE D 456.4. A posteriori analyses
revealed that between-judgments diVerences were only signiWcant when the causes were not
followed by their eVects, as in the Wrst phase of the PP/NP group (paired C1–E1 versus
non-paired C3–E2) and the second phase of the PP/PN group (paired C2–C1 vs nonpaired C4–C3). These results always showed accurate judgments, higher when the causes
were followed by their eVects and lower when the causes were not followed by them.
The lower panel of Fig. 2 shows the mean judgment for C2–E1 and C4–E2 relationships.
The 2 £ 2 ANOVA (Groups x Type of Cause) showed only a main eVect for Type of Cause,
F (1, 22) D 13.99, p D .01, MSE D 2161.7, as in Experiment 1. Participants’ causal judgments
were higher about SOC or paired type of causes (C2) than about non-paired ones (C4).
These results replicated the novel Wnding of SOC in Experiment 1a.
In general, the results of Experiments 1a and 1b have shown that the procedure was
appropriate to study SOC in causal learning tasks and that the development of SOC
depends on the relationships between causes and eVects in both training phases. These
results are similar to those found in animal conditioning.
Although Experiments 1a and 1b results were expected, they are important because they
are the Wrst demonstrations of SOC in human causal learning tasks. Moreover, they also
showed that participants were able to infer the absence of relationships between a secondorder cause and the Wrst-order eVect, when causes and eVects were not paired in either of
the two phases. Experiments 2a and 2b extend our analysis of SOC in causal learning by
showing the eVect of the Wrst-order extinction relationships upon SOC.
Experiment 2a
The previous experiments identiWed necessary and suYcient conditions for secondorder conditioning (SOC) in a causal learning task. However, they left open the question of
the nature of the learned associations underlying SOC. The main goal of Experiments 2a
and 2b was to evaluate, Wrst, the nature of the second-order association in human causal
learning using post-conditioning extinction and, second, the generality of such eVects by
showing whether SOC in causal learning is aVected by the same variables that aVect associative conditioning tasks. The experimental design was similar to the previous ones, but
with a third phase that included extinguishing the Wrst-order relationship learned during
the Wrst phase.
The common Wnding in animal and human learning associative conditioning tasks has
been that SOC is not inXuenced by extinction of Wrst-order conditioning (Davey & McKenna, 1983; Davey, 1987; Holland & Rescorla, 1975; Rescorla, 1980; Rizley & Rescorla,
1972). The conditioned response continues to be elicited by the second-order CS after
extinction of the Wrst-order relationship, indicating that the SOC response is not mediated
by some association between CS2 and CS1. These results suggested the formation of a
direct link between CS2 and CR (Rizley & Rescorla, 1972), or possibly between CS2 and
US (see Barnet, Arnold, & Miller, 1991; Miller & Barnet, 1993). The use of a causal learning task could clarify this issue, while, at the same time, the eVect of extinction upon
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
239
second-order causal learning could also reveal the causal mental model people build when
they learn such causal relationships.
In previous experiments, Perales et al. (2004) employed a mediated learning procedure,
where two cues were subsequently linked to the same eVect in two diVerent phases. After
the training, they found that participants were able to establish an emergent relationship
between the two cues that were never presented together, while, at the same time, they
showed that participants used an independent cause mental model rather than a chained
cause mental model. Assuming that during a second-order causal learning task, people
form an association between “causes and eVects” (i.e., the CS and the US in a conditioning
procedure), the normative and rational causal model should be chained (C2–C1–E), as
they have learned that the Wrst-order cause (C1) produces the eVect (E) and, subsequently,
that the second-order cause (C2) produces the Wrst one (C1). Accordingly, the extinction of
the relationship between the Wrst-order cause and the eVect should aVect the perceived relationship between the second-order cause and the same eVect, if participants have formed
such a chained mental model, i.e., if there is a mediation of the Wrst cause (the CS1 in conditioning procedures). However, if the extinction of the relationship between the Wrst-order
cause and the eVect did not have any inXuence on the second-order relationship, this would
imply the formation of a diVerent mental model in which each cause could be directly and
independently linked with the same eVect, as seems to happen in animal and human associative conditioning procedures.
Method
Participants
Twelve undergraduate students of psychology participated in this experiment for course
credits. Their ages varied from 19 to 24.
Procedure
The experimental design is summarized in Table 1. The presentation of trials on the
screen was identical to that described in previous experiments. This experiment had two
diVerent types of causes, SOC and EXT, and they only diVered during the third extinction phase. First and second-order training phases were similar to those described in previous experiments with the diVerence that in both phases the contingency established
among the causes and the eVects was always 1.0. The third phase consisted in two intermixed counterbalanced blocks of 12 trials. In the EXT type of cause condition, the Wrstorder cause (C1) was presented without being followed by the eVect (E1). Depending on
the counterbalancing, the information given in these trials was for example: “The patient
has Midiasis.” Under this sentence, the next question that appeared was “Does this disease cause the appearance of the substance Alpha in the blood? Again, the participants had
to give an answer indicating “Yes” with the number three and,“No”with the number Wve.
After their response, feedback was presented in the sentence “Midiasis does not cause
Alpha” (1200 ms duration). In the SOC type of cause condition, the only diVerence was
that a novel cause, for example Zatuba (C5) and a second eVect (Beta, E2) were
presented. After the participant responded, the feedback indicated that “Zatuba does not
cause Beta.” In the test phase, the participants were required to judge the relationships
between the second-order causes and eVects, C2–E1 and C4–E2, using a scale from
0 to 100.
240
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
Results and discussion
The left panel of Fig. 3 displays the mean judgment given during the test phase. No
diVerences appear between the SOC and EXT types of causes, showing that the extinction
of C1–E1 relationships did not aVect the second-order causal judgments. A one-way
ANOVA performed on the participant judgments did not reveal any signiWcant eVect for
Type of cause F (1, 11) D .09, p D .759, MSE D 187.0.
These results suggest that although the C1–E1 relationships are a necessary condition to
form a SOC association, once learned, the C2–E1 association seems independent of the
C1–E1 association. Revaluation of C1–E1 association had no eVect on responding to C2.
These results replicated those obtained in animal conditioning experiments (Fujii, 1981;
Holland & Rescorla, 1975; Rizley & Rescorla, 1972) and those observed in electrodermal
associative conditioning with human participants (Davey & McKenna, 1983). However,
before further discussing this issue, it seems important to be sure that the participants had
actually extinguished the Wrst-order relationship.
Experiment 2b
In Experiment 2b, the participants were required to estimate the causal relationships
after each training phase, to determine whether they in fact learned the contingencies presented in each phase. In the previous Experiment 2a, the Wrst-order relationship was extinguished after second-order learning. The results showed that such extinction did not have
any eVect on the second-order causal learning, as was also found in animal conditioning
(Rizley & Rescorla, 1972) and human aversive conditioning (Davey & McKenna, 1983).
Experiment 2b attempted to replicate the same eVect and, in addition, assessed whether the
participants had learned the relations presented in each phase, especially in extinction.
100
Mean Judgment
80
60
40
20
C2-E1
C4-E2
0
EXPERIMENT 2a
EXPERIMENT 2b
Fig. 3. Mean judgments for SOC test. Open bars represent Extinction Type of Cause, and hatched bars represent
SOC Type of Cause. Error bars denote standard errors of the mean.
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
241
Consequently, the participants were asked to estimate the causal relationships between
each cause and its eVect after each experimental phase.
Method
Participants
Twelve undergraduate students of psychology, 19- to 24-year-old, participated in this
experiment for course credits.
Procedure
The procedure and experimental design were identical to those one used in Experiment
2a and are summarized in Table 1; the only diVerence was that the participants were
requested to estimate the relationships between causes and eVects at the end of each training phase. Finally, at the end of training phases, the participants had to give a judgment
about C2–E1 and C4–E2 relationships to show the level of SOC.
Results and discussion
The participants accurately detected the contingencies between causes and eVects
programmed in each experimental phase, as can be observed in Table 2. Participant
causal judgments were similarly higher after the Wrst and second-order phases and
lower after the extinction phase. These impressions were corroborated by a 3 £ 2
(Phases £ Type of Cause) ANOVA which yielded a signiWcant eVect of Phases,
F (2, 22) D 70.96, p < .001, MSE D 782.83. LSD Post Hoc tests did not reveal signiWcant
diVerences between the Wrst and the second phase judgments (p > .05), but diVerences
were signiWcant between each of these phases and the third phase (p < .05). This pattern
of results is important because it permits us to conWrm the extinction of the Wrst-order
relationship.
The right panel of Fig. 3 displays the mean judgment given on the SOC test and reveals
that the judgments remained high despite extinction of the Wrst-order relationship. A oneway ANOVA, F (1, 11) D .02, p D .88, MSE D 46.5, indicated that there was no diVerence
between SOC and EXT type of causes during the test phase. However, a second 2 £ 2
ANOVA was carried out to compare Type of Cause during the third phase and the Wnal
test. This analysis revealed diVerences between phases but not between type of causes,
F (1, 33) D 39.60, p D .01, MSE D 584.86. This last result further demonstrated that SOC was
not aVected by extinction of the Wrst-order relationship, because judgments about the second-order causes were signiWcantly diVerent from judgments about the Wrst-order extinguished causes. In general, the results of Experiment 2b showed a similar eVect to the one
found in Experiment 2a, because no diVerences appeared between SOC and EXT conditions.
In summary, the main Wnding of both Experiment 2a and 2b was that extinction of the
Wrst-order relationship did not aVect the second-order relationship. This Wnding is similar
to those found in conditioning experiments carried out with animals and humans (Davey
& McKenna, 1983). Overall results suggest that the structure which mediates the secondorder learning is based on an independent causal relationship between C2 and the EVect
without any mediation by C1. This conclusion is parallel to the original Wnding of Rizley
and Rescorla (1972) with the classical conditioning paradigm.
242
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
General discussion
The primary novel Wnding of this research was a reliable second-order conditioning
eVect in human causal learning tasks in two independent experiments. It is important
because it suggests people are able to infer a causal relationship between causes and eVects
that were never experienced together. Experiments 1a and 1b also showed that the main
determinant of SOC is the prior correlation between cause and eVect. The additional and
most important Wnding of Experiments 2a and 2b was that extinction of the Wrst-order
causal relationship (C1–E) did not inXuence the acquired second-order relationship (C2–
E). This last result suggests that causal learning and conditioning tasks probably are mediated, at least in part, by similar learning mechanisms (Dickinson et al., 1984). A theoretical
beneWt of Experiments 2a and 2b was the information about association types that are
learned during SOC. At the same time, the results could clarify the type of causal mental
model participants may be using to establish the emergent relationship between C2 and the
EVect.
Until now, the only explanations of SOC have been derived from associative views
based upon the four types of possible learned associations (Fig. 1) during second-order
conditioning. The usual Wnding had been that extinction of Wrst-order conditioning did not
reduce the conditioned response to the second-order stimulus (CS2), suggesting the
absence of any CS1 mediation in the elicitation of the second-order response. However,
there still remained two possible associations responsible for the eVect. There could be a
direct connection between CS2 and the conditioned response itself, as suggested by S-R
theory (cf. Rizley & Rescorla, 1972), or the response could be dependent on a direct link
between CS2 and the US, as suggested by more recent theories (Barnet et al., 1991). The
results obtained using causal learning tasks suggested that there could be such a direct link
between the CS (the cause) and US (the eVect) because the participants were able to infer
the existence of a causal relationship between them, without any conditioned response,
unless such causal judgment is considered a conditioned response. In addition, these results
seem compatible with those found in animal and human sensory-preconditioning experiments, where if CS2 and CS1 are paired in a Wrst training phase and, subsequently, CS1 is
paired with the US, an emergent CS2–CR association is formed even though CS2 and the
US were never presented together (Brogden, 1939). When using such a sensory-preconditioning procedure in humans, White and Davey (1989) found that they still showed a
strong relationship between CS2 and CR, even after post-conditioning inXation of US.
These results bear out Rizley and Rescorla’s (1972) proposals that SOC and sensory-preconditioning may be products of diVerent underlying associative structures. Therefore,
whether the type of associations formed during sensory-preconditioning and SOC are the
same and similarly aVected by extinction of C1–US, remains an open question for future
research into human causal learning.
To summarize, the overall results showed that second-order causal learning, as well as
the perception of a relationship between C2 and E1, is dependent on the strength of the
contingent acquisition relationship between the Wrst-order cause and the eVect (C1–E1).
Accordingly, the successive cue-outcome pairings progressively strengthen the connection
between the mental representation of the cue and the outcome giving rise to the CR (Allan,
1993). In the absence of such contingent pairings, as happened in the non-paired conditions, SOC does not develop. But, once the second-order association is established, then it
seems no longer dependent on the Wrst-order association. This view is entirely consistent
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
243
with those associative theories that assume a direct CS2–US link formed during SOC, giving rise to the conditioned response without the need of any CS1 mediation (Barnet et al.,
1991).
However, the results might also be explained from more recent causal learning theories,
as the result of a process of causal inference (Johnson-Laird, 1983). From this perspective,
the Wnding that SOC happens in human causal learning is not at all surprising. If participants are taught by means of instructions or by direct experience that a cause (C1) produces an eVect (E1) and that another cause (C2) produces this Wrst cause, it seems very easy
to infer the existence of a causal relationship between all components as a causal chain,
C2–C1–E1. In agreement with these experimental results on human reasoning (Goldvarg
& Johnson-Laird, 2001), our results could be seen as a product of simple deductive reasoning in the form of “axioms or inference rules,” such as “If C1 causes E1, and C2 causes C1,
then C2 causes E1.” Moreover, participants were also able to infer the absence of relationships when one of the conditions was preventative, as also happened in human reasoning
experiments (Goldvarg & Johnson-Laird, 2001) and as Experiment 1b convincingly
showed. When learning was that “C3 causes E2, and C4 prevents C3” or “C3 prevents E2,
and C4 causes C3,” then participants logically inferred that C4 did not cause E2.
By contrast, the lack of inXuence of extinction of the Wrst-order relationships upon the
second-order causal relationship was an unexpected and important result. It was unexpected, because it is normatively inaccurate. According to logical inference rules and the
results obtained in human deductive reasoning (Goldvarg & Johnson-Laird, 2001), “if C2
causes C1, and C1 prevents E1, then C2 should also prevent E1.” However, Experiments 2a
and 2b showed that participants were unable to make such a logical inference, suggesting
that once the causal link between C2 and the eVect is formed, this new knowledge seems to
be rather independent of the previous knowledge about the relationship between C1 and
the same eVect. These results are important, because they could allow a reinterpretation of
the nature of the causal relationship learned during second-order conditioning tasks and
the type of causal mental models naïve participants built during such a learning experience.
Recent developments in the Weld of human causal learning (Cheng, 1997; Waldmann &
Hagmayer, 2001) emphasize that causal relationships can not be reduced to covariation
between the events or to mere associations, because they are a product of top–down higher
cognitive processes. In this way, people are able to accurately detect true causal relationships and diVerentiate them from spurious correlations. Accordingly, it is assumed that
causal relationships are learned within a “causal mental model” by encoding assumptions
about the causal status of the events (causes and eVects, as shown by the eVects of causal
directionality in cue competition, see Waldmann, 2000, 2001), or in mediated learning
eVects, see also Perales et al. (2004) and about hypothetical direct and indirect causal relationships. Non-observed relationships between a cause and an eVect can then be inferred
from such mental models built either by the integration of the information provided via
instructions or by the knowledge acquired during the task. Assuming that a causal mental
model is a working memory representation of real (or imaginary) related events (JohnsonLaird, 1983), its structure being isomorphic to that provided by the situation, at least two
diVerent causal models could be built during SOC training. The most obvious one should
be a chained cause mental model in which C2 produce C1 which in turn cause E1 (i.e., C2–
C1–E1), allowing the inference that C2 also causes E1. If this was the model built by the
participants, we should expect that extinction of the Wrst link, i.e., C1 no longer produces
E1, should lead to disruption of the causal link between C2 and the same eVect. The
244
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
absence of inXuence of this new knowledge upon the relationship between C2 and E1 suggests that participants build or store a diVerent causal model during such training.
In accordance with related research on mediated learning and emergent causal relationships between events never experienced together (Perales et al., 2004), there are at least two
diVerent possibilities. The Wrst one proposed by Perales et al. (2004) is that once the causal
relationship between C2 and the eVect is formed, both C1–E1 and C2–E1 relationships are
stored as an independent causal mental model. It seems that once the between-events relationship is stored, the causal chain is no longer necessary, as the data of Experiment 1b and
2b convincingly show. It is also important to acknowledge that it is not necessary to
express a causal judgment of each relationship, as Experiment 2a clearly suggests; participants were aware of all the relationships, as Experiment 2b demonstrated. A second possibility, however, also reviewed in Perales et al. (2004) and proposed by an anonymous
reviewer, could be that participants adopted a “common-cause mental model” in which
they saw C2 as a common cause of C1 and E1. In other words, they stored a “causal mental
model” in which they saw C1 and E1 as independent eVects of C2. The results of experiment 2b are not clearly compatible with this view, because participants had indicated that
they considered C1 a cause of E1 rather than an independent eVect of C2. The exact nature
of the mental model people build and store during SOC remains a question for future
research.
According to the actual results, it is proposed that during the second-order conditioning
task naïve people built an “independent causal mental model” of the situation. According
to this model, each cause can produce the eVect in an independent way, explaining the
emergence of the second-order relationship between the second-order cause and the eVect,
as well as the absence of inXuence of extinction of the Wrst causal component upon the second one. Whether other factors, such as instructions, additivity or cause directionality,
could also inXuence extinction and the building and storing of the underlying causal mental model in human second-order causal learning, as has happened with other well established associative eVects like cue competition (Beckers, De Houwer, Pineño, & Miller, 2005;
Waldmann, 2000, 2001), remains an open question for further research.
In summary, this work provides the Wrst demonstration of a reliable second-order conditioning eVect in human causal learning and shows that this second-order learning is independent of the degree of relationships between the Wrst-order cause and the eVect. These
Wndings match Wndings from animal and human conditioning tasks and emphasize that
similar learning mechanisms are probably at work. From an associative point of view, the
learning mechanism in causal learning tasks would be the formation of a direct association
between the mental representations of CS2 (Cause) and the US (EVect), giving rise to the
conditioned response. From a cognitive causal theory point of view, this association
reXects the inference of a direct causal relationship between C2 and the EVect, as a product
of an underlying stored independent cause mental model. Moreover, the results of this
research bear out previous Wndings using a mediated learning task (Perales et al., 2004) and
agree with those causal theories (Waldmann & Martignon, 1998) which assume that in a
causal scenario where diVerent causes can produce the same eVect, people use by default an
independent model rather than a cause chained one, even though the latter would be more
rational. This independent causal mental model means that the perceived causal power of
each cause to inXuence the eVect seems to be independent of the causal power of the other
causes. It is important to acknowledge that an independent and additive causal model
seems to be a necessary condition to Wnd and explain others eVects, similarly obtained both
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
245
in causal and conditioning tasks, such as blocking (Beckers et al., 2005). In conclusion,
although this research emphasizes the similarity between conditioning and causal judgment tasks, the question of how to diVerentiate between associative and cognitive explorations of human learning still remains open for future research. This work shows a new way
to look for an answer to this long debate.
References
Allan, L. (1993). Human contingency judgments: Rule based or associative. Psychological Bulletin., 114, 435–438.
Barnet, R. C., & Miller, R. R. (1996). Temporal encoding as a determinant of inhibitory control. Learning and
Motivation, 27, 73–91.
Barnet, R. C., Arnold, M., & Miller, R. (1991). Simultaneous conditioning demonstrated in second-order conditioning: Evidence for similar associative structure in forward and simultaneous conditioning. Learning and
Motivation, 22, 253–528.
Barnet, R. C., Cole, R. P., & Miller, R. R. (1997). Temporal integration in second-order conditioning and sensory
preconditioning. Animal Learning & Behavior, 25, 221–233.
Beckers, T., De Houwer, J., Pineño, O., & Miller, R. R. (2005). Outcome additivity and outcome maximality inXuence cue competition in human causal learning. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 31, 238–249.
Bevins, R. A., Delzer, T. A., & Bardo, M. T. (1996). Second-order conditioning detects unexpressed morphineinduced salt aversion. Animal Learning & Behavior, 24, 221–229.
Brogden, W. J. (1939). Sensory pre-conditioning. Journal of Experimental Psychology, 21, 55–58.
Cheatle, M., & Rudy, J. (1979). Analysis of second-order odor-aversion conditioning in neonatal rats: Implications for Kamin’s blocking eVect. Journal of Experimental Psychology: Animal Behavior Processes, 4, 237–249.
Cheng, P. W. (1997). From covariation to causation: A causal power theory. Psychological Review, 104, 367–405.
Crawford, L., & Domjan, M. (1995). Second-order sexual conditioning in males Japanese quail (Coturnix japonica). Animal Learning & Behavior, 23(4), 327–334.
Davey, G. C. L. (1983). An associative view of human classical. In G. C. L. Davey (Ed.), Animal models of human
behavior (pp. 95–114). Chichester: John Wiley & Sons.
Davey, G. C. L. (1987). An integration of human and animal models of Pavlovian conditioning: Associations,
cognitions, and attributions. In G. C. L. Davey (Ed.), Cognitive processes and Pavlovian conditioning in humans
(pp. 83–114). Chichester, England: Wiley.
Davey, G. C. L., & McKenna, I. (1983). The eVects of post-conditioning revaluation of CS1 and UCS following
Pavlovian second-order electrodermal conditioning in humans. Quarterly Journal of Experimental Psychology, 35B, 125–133.
Davey, G. C. L., & Arulampalan, T. (1982). Second-order “fear” conditioning in humans. Persistence of CR2 following extinction of CR1. Behavior Research and Therapy, 20, 391–396.
Dickinson, A., Shanks, D., & Evenden, J. L. (1984). Judgment of act-outcome contingency: The role of selective
attribution. Quarterly Journal of Experimental Psychology, 36A, 29–50.
Fujii, M. (1981). Second-order conditioning in licking-based conditioned suppression situation. Japanese Psychological Research, 23, 149–159.
Goldvarg, E., & Johnson-Laird, P. N. (2001). Naïve causality: A mental model theory of causal meaning and reasoning. Cognitive Science, 25, 565–610.
Hall, G. (1996). Learning about associatively activated stimulus representations: Implications for acquired equivalence and perceptual learning. Animal Learning & Behavior, 24(3), 233–255.
Holland, D. C., & Rescorla, R. A. (1975). The eVects of two ways of devaluing the unconditioned stimulus after
Wrst- and second-order appetitive conditioning. Journal of Experimental Psychology, 88, 459–467.
Johnson-Laird, P. N. (1983). Mental models: Towards a cognitive science of language, inference, and consciousness.
Cambridge MA: Cambridge University Press.
Mackintosh, N. J. (1974). The psychology of animal learning. London: Academic Press.
Miller, R., & Barnet, R. C. (1993). The role of time in elementary associations. Current Directions in Psychological
Science, 2(4), 101–111.
Miller, R., & Escobar, M. (2002). Learning: Laws and models of basic conditioning. In C.R. Gallistel (Ed.), Learning, motivation and emotion (Vol. 3, pp. 47–102). In Stevens’ handbook of experimental psychology, 3rd ed. (H.
Pashler, Ed-in-chief). New York: John Wiley & Sons. Invited Chapter.
246
E. Jara et al. / Learning and Motivation 37 (2006) 230–246
Pavlov, I. (1927). Conditioned reXexes. England: Oxford University Press.
Perales, J. C., Catena, A., & Maldonado, A. (2004). Outcome mediated contingency learning is sensitive to causal
directionality. Learning and Motivation, 35, 115–135.
Rashotte, M. (1981). Second-order autoshaping: Contributions to the research and theory of Pavlovian reinforcement by conditioned stimuli. In C. M. Locurto, H. S. Terrace, & J. Gibbon (Eds.), Autoshaping and conditioning theory (pp. 139–180). New York: Academic Press.
Rashotte, M., GriYn, R. W., & Sisk, C. L. (1977). Second-order conditioning of the pigeon’s keypeck. Animal
Learning & Behavior, 5, 25–38.
Rescorla, R. (1976). Second-order conditioning of Pavlovian conditioned inhibition. Learning and Motivation, 7,
161–172.
Rescorla, R. (1980). Pavlovian second-order conditioning: Studies in associative learning. New Jersey: Lawrence
Erlbaum Associates.
Rescorla, R. A., & Cunningham, C. L. (1979). Spatial contiguity facilitates Pavlovian second-order conditioning.
Journal of Experimental Psychology: Animal Behavior Processes, 5, 152–161.
Rescorla, R. A., & Wagner, A. (1972). A theory of Pavlovian conditioning: Variations in the eVectiveness of reinforcement & no reinforcement. In A. H. Black & W. K. Prokasy (Eds.), Classical conditioning II. Current
research and theory (pp. 64–99). New York: Appleton-Century-Crofts.
Rizley, R. C., & Rescorla, R. A. (1972). Associations in second-order conditioning and sensory preconditioning.
Journal of Comparative and Physiological Psychology, 81, 1–11.
Shanks, D. R. (1995). The psychology of associative learning. Oxford UK: Oxford University Press.
Waldmann, M. R. (2000). Competition among causes but not eVects in predictive and diagnostic learning. Journal
of Experimental Psychology: Learning, Memory, and Cognition, 26, 53–76.
Waldmann, M. R. (2001). Predictive versus diagnostic causal learning: Evidence from an overshadowing paradigm. Psychonomic Bulletin & Review, 8, 600–608.
Waldmann, M. R., & Hagmayer, Y. (2001). Estimating causal strength: The role of structural knowledge and processing eVort. Cognition, 82, 27–58.
Waldmann, M. R., & Martignon, L. (1998). A Bayesian network model of causal learning. Proceedings of the 20th
annual conference of the cognitive science society. Mahwah, NJ: LEA.
White, K., & Davey, G. (1989). Sensory preconditioning and UCS inXation in human “fear” conditioning. Behavior Research and Therapy, 27, 161–166.