Individual differences in causal structures inferred during feature negative learning

Abstract

Traditional associative learning theories predict that training with feature negative (A+/AB-) contingencies leads to the feature B acquiring negative associative strength and becoming a conditioned inhibitor (i.e., prevention learning). However, feature negative training can sometimes result in negative occasion setting, where B modulates the effect of A. Other studies suggest that participants learn about configurations of cues rather than their individual elements. In this study, we administered simultaneous feature negative training to participants in an allergist causal learning task and tested whether evidence for these three types of learning (prevention, modulation, configural) could be captured via self-report in the absence of any procedural manipulation. Across two experiments, we show that only a small subset of participants endorse the prevention option, suggesting that traditional associative models that predict conditioned inhibition do not completely capture how humans learn about negative contingencies. We also show that the degree of transfer in a summation test corresponds to the implied causal structure underlying conditioned inhibition, occasion-setting, and configural learning, and that participants are only partially sensitive to explicit hints about causal structure. We conclude that feature negative training is an ambiguous causal scenario that reveals individual differences in the representation of inhibitory associations, potentially explaining the modest group-level inhibitory effects often found in humans.

Full text

https://doi.org/10.1177/1747021820959286
Quarterly Journal of Experimental
Psychology
2021, Vol. 74(1) 150 –165
© Experimental Psychology Society 2020
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/1747021820959286
qjep.sagepub.com
The ability to detect contingent relationships in the world
is critical to make accurate predictions about the future.
The contingency between a given cue and an outcome may
be positive (the cue is associated with the outcome), or it
may be negative (the cue is associated with the absence of
the outcome). Although it is clearly important to learn
about positive (causal) relationships, it is just as important
to learn about negative (preventive) ones. Theoretical
accounts of learning often characterise prevention learning
as arising from the detection of negative associations
between cues and outcomes (e.g., Cheng, 1997; Cheng &
Novick, 1992; Rescorla & Wagner, 1972; Ward & Jenkins,
1965). Thus, at first glance, prevention learning can be
seen as simply the opposite to causal learning.
This view of causation and prevention as diametrically
opposed is formalised in traditional associative learning
models such as the Rescorla-Wagner model (Rescorla &
Wagner, 1972). These models, which were originally
developed to explain Pavlovian conditioning in animals,
have been cited as theoretical accounts of how humans
learn causal and preventive relationships between cues and
outcomes (Dickinson et al., 1984). The Rescorla-Wagner
model conceptualises predictive relationships between
cues and outcomes in terms of a single dimension of asso-
ciative strength. In associative terms, “excitors” (causal
cues) have a positive associative strength and activate the
representation of an outcome, while “inhibitors” (preven-
tive cues) have negative associative strength and inhibit
the representation of the outcome. In such models, associa-
tive strength changes on each trial in proportion to the
degree and direction of prediction error generated by the
cues present on that trial. Negative prediction error (expec-
tation of an outcome that does not occur) results in a
Individual differences in causal structures
inferred during feature negative learning
Jessica C Lee and Peter F Lovibond
Abstract
Traditional associative learning theories predict that training with feature negative (A+/AB-) contingencies leads to
the feature B acquiring negative associative strength and becoming a conditioned inhibitor (i.e., prevention learning).
However, feature negative training can sometimes result in negative occasion setting, where B modulates the effect of
A. Other studies suggest that participants learn about configurations of cues rather than their individual elements. In this
study, we administered simultaneous feature negative training to participants in an allergist causal learning task and tested
whether evidence for these three types of learning (prevention, modulation, configural) could be captured via self-report
in the absence of any procedural manipulation. Across two experiments, we show that only a small subset of participants
endorse the prevention option, suggesting that traditional associative models that predict conditioned inhibition do
not completely capture how humans learn about negative contingencies. We also show that the degree of transfer in
a summation test corresponds to the implied causal structure underlying conditioned inhibition, occasion-setting, and
configural learning, and that participants are only partially sensitive to explicit hints about causal structure. We conclude
that feature negative training is an ambiguous causal scenario that reveals individual differences in the representation of
inhibitory associations, potentially explaining the modest group-level inhibitory effects often found in humans.
Keywords
Feature negative; conditioned inhibition; occasion setting; associative learning; prevention; causal learning; causal
structure; configural
Received: 12 January 2020; revised: 12 July 2020; accepted: 20 July 2020
University of New South Wales Sydney, Sydney, NSW, Australia
Corresponding author:
Jessica C Lee, University of New South Wales Sydney, Sydney, NSW
2052, Australia.
Email: jessica.lee@unsw.edu.au
959286QJP0010.1177/1747021820959286Quarterly Journal of Experimental PsychologyLee and Lovibond
research-article2020
Original Article

Lee and Lovibond 151
decrease in associative strength for any cues present, and
may result in a cue acquiring net negative associative
strength (i.e., becoming an inhibitor).
A procedure that generates negative prediction error is
feature negative training (A+/AB-, also known as condi-
tioned inhibition training). First demonstrated by Pavlov
(1927), this procedure consists of pairing a target (A) with
an outcome (+), as well as pairing a compound of the tar-
get and feature (AB) with no outcome (-). An intuitive
interpretation of these contingencies is that since the target
(A) predicts the outcome, the addition of the feature (B)
must therefore prevent the outcome from occurring. This
intuition is captured in the Rescorla-Wagner model’s cal-
culation of prediction error, which leads to the feature
accruing negative associative strength and becoming a
conditioned inhibitor.
A summation test is typically used to show that the fea-
ture B has acquired inhibitory properties (Rescorla, 1969).
This test consists of comparing conditioned responding to
a compound consisting of the feature (B) and a separately
trained target (e.g., C+) to a control compound where the
same target (C) is combined with a neutral control cue
(e.g., D, a cue paired with no outcome). If responding is
lower to the test compound (CB) than to the control com-
pound (CD), then the feature (B) is deemed inhibitory.
Conditioned inhibition has been reliably demonstrated
using the summation test in a variety of conditioning prep-
arations and animal species (see Savastano et al., 1999;
Sosa & Ramírez, 2019; Swartzentruber, 1995 for reviews).
Thus, there is good evidence in animals for the idea that
feature negative training produces inhibitory learning
about the feature (i.e., prevention learning) and that inhibi-
tors control behaviour opposite to excitation (Rescorla,
1969). Note that henceforth, we will refer to the degree to
which a negative feature suppresses responding to an exci-
tor in a summation test as transfer of inhibitory learning
(i.e., an empirical phenomenon), reserving the term condi-
tioned inhibition for “true” preventive learning where the
negative feature passes the summation test and possesses
the properties described by Rescorla (1969).
However in humans, evidence for transfer of inhibitory
learning at the group level is much weaker (e.g., Karazinov
& Boakes, 2007; Lee & Livesey, 2012; Williams, 1995;
Wilkinson et al., 1989). This suggests that inhibitory learn-
ing in humans may be more complex than predicted by
associative models that posit a single dimension of asso-
ciative strength (Rescorla & Wagner, 1972). In particular,
there might be individual differences in what participants
learn about the negative feature, with only a subset of par-
ticipants learning in the manner predicted by the Rescorla-
Wagner model. Glautier and Brudan (2019) have recently
provided evidence in support of this idea. They presented
participants with feature negative training where a context
served as the negative feature, and found that participants
varied in the degree to which they showed transfer of
inhibitory learning in a summation test. Critically, classi-
fying participants into “inhibitors” and “non-inhibitors”
predicted how participants responded to conditioning of a
different negative feature in a new task. There is growing
recognition of the existence of meaningful individual dif-
ferences in associative learning in humans (e.g., Byrom &
Murphy, 2014, 2016; Lee et al., 2018; Stegmann et al.,
2019) and even in animals (Jean-Richard-dit-Bressel et al.,
2019). In cases where the individual differences are quali-
tative, aggregate data may be misleading (e.g., see Lee
et al., 2018). Glautier and Brudan’s (2019) study provides
initial evidence that there might be stable individual differ-
ences in feature negative learning.
Individual differences are especially likely to arise in
feature negative designs because there are several alterna-
tive ways to solve the training task. Feature negative train-
ing can sometimes lead to negative occasion setting (see
Bonardi et al., 2017; Fraser & Holland, 2019, for reviews).
In contrast to learning that the feature B prevents or inhib-
its the outcome itself, negative occasion setting involves
learning that the feature B modulates (sets the occasion
for) the target’s (A’s) relationship with the outcome. In
other words, A by itself causes the outcome but when B is
also present, A does not cause the outcome. A popular view
of occasion setting is that it constitutes evidence of hierar-
chical control of associative behaviour, where the occasion
setter is associated not with the outcome but with an asso-
ciation between another cue and the outcome (Bonardi
et al., 2017; Fraser & Holland, 2019; see Konorski, 1948;
Rescorla, 1969 for an alternative threshold view). Indeed,
Glautier and Brudan (2019) attributed the difference
between their inhibitor and non-inhibitor subgroups to a
tendency to learn first-order (inhibition) or second-order
(occasion-setting) associations. In contrast to conditioned
inhibition, the defining feature of occasion-setters is that
they possess second-order (modulatory) links that are
independent of any direct associations with the outcome
(e.g., Holland, 1984, 1989).
Interestingly, negative occasion setting is typically
found with serial (i.e., B → A → no outcome) presentation
of the AB compound, while conditioned inhibition is typi-
cally found with simultaneous presentation (AB → no out-
come) in animals (e.g., Holland, 1984, 1989; Holland &
Lamarre, 1984; Holland & Morell, 1996) and humans
(Baeyens et al., 2004). Negative occasion setting (from
serial procedures) is distinguished from conditioned inhi-
bition (in simultaneous procedures) by multiple functional
properties. One of the key features of occasion setters is
that they show limited transfer in a summation test. They
tend not to transfer (pass the summation test) when pre-
sented serially with other targets, but may sometimes
transfer to specific targets that have themselves been part
of an occasion-setting discrimination (see Holland, 1989;
Lamarre & Holland, 1985; Wilson & Pearce, 1990).
Although the bulk of the occasion-setting literature comes

152 Quarterly Journal of Experimental Psychology 74(1)
from studies with non-human animals, there is some evi-
dence that humans also show transfer to another target that
had previously been modulated (trained as an occasion-
setter), but not to a simple target paired with the outcome
in a serial design (Experiment 4, Baeyens et al., 2004).
Given how little research there is on occasion-setting in
humans, it is unknown whether participants can learn that
the feature is modulatory when the target and feature are
presented simultaneously.
Finally, there is a third way (in addition to conditioned
inhibition and occasion setting) by which participants may
solve the feature negative discrimination. Configural theo-
ries such as Pearce’s (1987, 1994) state that compounds
are represented as unique configurations rather than as the
sum of their elements, and that associative links form
between this configural representation and the outcome. In
contrast, elemental theories (e.g., Rescorla & Wagner,
1972) posit separate associative links for each element in a
presented compound (e.g., for AB, A and B have separate
associations with the outcome). In order for configural
theories to predict responding to novel configurations of
stimuli, they must incorporate some generalisation mecha-
nism. For example, Pearce’s (1987, 1994) model assumes
that associative strength generalises between configura-
tions based on the proportion of common elements they
share.
Evidence from human causal learning studies suggests
that participants do learn about stimuli in this configural
way. However, interestingly, they do not always general-
ise between configurations to the extent predicted by
Pearce (1987, 1994). Shanks, Charles, et al., (1998)
showed that reversal learning (i.e., conditioning) of the
negative feature B following feature negative training
(i.e., A+/AB-, then B+) left performance to the original
feature negative contingencies (A and AB) largely unaf-
fected. In other words, learning that B was causal did not
inflate ratings to the AB compound in a subsequent test,
contrary to the predictions of elemental theories of learn-
ing (e.g., Rescorla & Wagner, 1972) but also, to a lesser
degree, contrary to Pearce’s configural theory which pre-
dicts some degree of generalisation between B+ and AB-
configurations. Shanks et al. interpreted their findings of
intact feature negative performance as evidence for con-
figural learning, but noted that participants were more
configural than predicted by Pearce’s generalisation
mechanism (see also Shanks, Darby, et al., 1998; Williams,
1995).
This reluctance to generalise suggests a memorisation
strategy where participants attempted to remember the
associations between cue configurations and outcomes
during training with little attempt to infer the effects of
individual cues (see also Williams et al., 1994). One limi-
tation of this study is that Shanks, Charles, et al. (1998) did
not include a summation test and thus it is unclear what
participants learned about the feature B, and how they
would transfer that learning to another excitor. In the
absence of this test, an equally compatible conclusion
from their results is that participants learned that B was a
negative occasion-setter. Note, however, that the only
study of negative occasion setting in humans showed that
unlike animal studies, reversal training of the feature com-
pletely disrupted subsequent performance to the original
feature negative contingencies (Experiment 3, Baeyens
et al., 2004). Thus in the human literature, it is unknown
whether a separate “configural” strategy exists that can be
differentiated from negative occasion setting.
In summary, separate lines of evidence in both humans
and non-human animals suggest that feature negative
training, under different conditions, can lead to different
types of learning about the feature. The overall goal of the
study was to test whether participants learn different
causal structures in a feature negative discrimination cor-
responding to conditioned inhibition, occasion-setting,
and configural learning. These three types of learning dif-
fer in their underlying causal (or associative) structure (see
Figure 1), which has implications for how much the fea-
ture should transfer in a summation test. Conventionally,
conditioned inhibitors are assumed to directly inhibit
Figure 1. Associative/causal structures underlying conditioned inhibition (prevention), occasion-setting (modulation) and configural
learning arising from training with feature negative (A+AB-) contingencies. A is the target and B is the negative feature. Solid black
arrows indicate excitatory/causal associations, rounded-head, red dotted arrows indicate inhibitory/preventive associations.

Lee and Lovibond 153
activation of the outcome through an inhibitory associative
link with the outcome. If participants have learned that B
is inhibitory/preventive, they should show strong transfer
of the inhibitory properties of B to a new target since B
directly inhibits the representation of the outcome. Note
that an alternative view proposed earlier by Konorski
(1967) is that inhibition involves an association between
the negative feature B and a “no-outcome” unit, which
itself is in an antagonistic relationship to the “outcome”
unit. We decided against including Konorski’s view of
inhibition as it does not refer to causal relationships
between environmental events and is more difficult to
communicate than the simpler preventive option.
In contrast, Ross and Holland (1981) suggested that
occasion-setters act on the A-outcome association, modu-
lating the effect of the target. If participants have learned
that B is modulatory, they may or may not assume that the
modulatory effects of B will extend to a novel target that
had not previously appeared with B, producing an interme-
diate level of transfer. Finally, configural theories posit
associative links between each unique configuration of
cues and the outcome. For feature negative contingencies,
this means that B is represented as part of the AB configu-
ration that acquires a direct inhibitory link with the out-
come (see Figure 1), leaving the status of B ambiguous.
The results of Shanks and colleagues (Shanks, Charles,
et al., 1998; Shanks, Darby, et al, 1998) suggest that par-
ticipants memorise configurations of cues and their respec-
tive outcomes and do not generalise between configurations.
Under this view, participants should be agnostic for all
novel compounds, and therefore respond at baseline,
showing the least amount of transfer.
There were three specific aims for the current study.
The first aim was to test what proportion of participants
learn about a negative feature in the manner predicted by
the Rescorla-Wagner model (Rescorla & Wagner, 1972),
by reporting a prevention causal structure (corresponding
to conditioned inhibition). The second aim was to test
whether there are individual differences in participants’
self-reported causal structure for the negative feature, even
in the absence of any procedural manipulation. Specifically,
we wanted to know whether subsets of participants
reported a modulation (corresponding to occasion-setting)
or configural causal structure using the typical feature
negative procedure where elements of the compound are
presented simultaneously. The last aim was to test whether
the self-report measures of causal structure were consist-
ent with the degree of transfer in a summation test.
Critically, in our study, we used explicit self-report
measures to capture individual differences rather than
data-driven methods (e.g., Glautier & Brudan, 2019). We
have successfully used self-report in our previous work on
generalisation to identify subgroups of participants who
used different learning strategies, and found that these
measures were consistent with other behavioural measures
(Lee et al., 2018; Lovibond, Lee, & Hayes (2020); Wong
& Lovibond, 2017). Although conditioned inhibition and
occasion-setting have been distinguished in terms of their
underlying associative structure in previous work (see
Fraser & Holland, 2019), there has been little attempt to
connect these effects to explicit causal structures in
humans. The causal learning literature shows that partici-
pants interpret associations differently according to the
causal model implied by the cover story (Waldmann et al.,
2006; Waldmann & Holyoak, 1992), but there have been
few attempts to investigate whether different causal mod-
els (structures) arise from associative learning procedures
with more neutral cover stories. Such a connection would
further strengthen the view that associative learning theo-
ries can account for how humans learn about causal rela-
tionships (Dickinson et al., 1984) and help clarify the role
of higher-order cognitive processes in human associative
learning (Lee, Lovibond, & Hayes, 2019; Lee, Lovibond,
Hayes, & Navarro, 2019; McLaren et al., 2018; Mitchell,
De Houwer, & Lovibond, 2009).
This study is novel in being the first to attempt to meas-
ure individual differences in feature negative learning via
explicit self-report and to obtain occasion setting (and con-
ditioned inhibition) in the absence of any procedural
manipulation of the training conditions (e.g., serial presen-
tation). In both experiments, we use the allergist task
where participants learn which foods (cues) are causing
allergic reactions (the outcome) in a fictitious patient, Mr
X. Our choice of the allergist task was motivated by the
fact that it reliably shows other cue competition effects
like blocking. Although in real life, food allergies are not
often offset by eating other (inhibitory) foods, we note that
if anything, this works against us finding any evidence of
inhibitory learning. Thus, the allergist task can be seen as
a neutral cover story with respect to causal structure.
Experiments 1 and 2 measured the causal structures that
participants inferred following simultaneous feature nega-
tive training. Experiment 2 also attempted to manipulate
which causal structure participants inferred by providing
verbal hints prior to training.
Experiment 1
The aim of Experiment 1 was to test whether individual
differences exist for feature negative learning, and whether
these differences can be captured by self-report. The top
row of Table 1 shows the design of Experiment 1. The
training phase consisted of 6 trial types. A+ and AB- were
the feature negative contingencies; A was the target causal
cue and B was the negative feature, presented in a simulta-
neous compound. C+ was the transfer causal cue used in
the summation test. DE- and F- provided control cues for
the summation test; F- was trained as an unambiguous
non-causal cue, while D was trained as a non-causal cue
that appeared in combination with another cue. Note that

154 Quarterly Journal of Experimental Psychology 74(1)
although F is less ambiguous than D in its predictive
power, D is arguably a more conservative control as it also
appeared in a compound. In contrast, F may suffer gener-
alisation decrement at test due to it never appearing in a
compound prior to test. Note that D and E are equivalent
and thus we only used one cue from the DE compound as
a control in the summation test. Finally, GH+ was a filler
compound included to prevent participants from learning
that all compounds predicted no outcome.
Following training, participants completed an outcome
prediction test where they continued to make the same pre-
dictions but in the absence of feedback. CB was the critical
summation test compound as it was composed of the fea-
ture B combined with the transfer causal cue C (Rescorla,
1969). Therefore, ratings for CB indexed how much par-
ticipants transferred their learning about B to a new target
(C). We included 3 different control compounds (CD, CF,
CI) to compare against CB. The control cues consisted of
the non-causal cues D (trained in compound) and F (trained
alone), as well as a novel cue I that participants saw for the
first time at test. We included these three different control
compounds as previous human studies (e.g., Karazinov &
Boakes, 2004; Shanks, Darby, et al, 1998) had used them
and we wished to explore whether there were systematic
differences between them.
We assessed causal structure in two ways. The first was
via open-ended self-report and the second was with a
three-alternative forced-choice (3AFC) question with
options describing a preventive, modulatory, and configu-
ral causal structure for B. We also included a standard
causal ratings test (e.g., Dickinson et al., 1984) to assess
what participants learned about the causal status of each
individual cue on a bipolar cause-prevent scale. Note,
however, that our main hypothesis involved testing
whether causal structure predicted differences in the trans-
fer of inhibitory learning in the outcome prediction test.
Method
Participants. 100 Mechanical Turk workers (27 female, M
age = 34.3, SD age = 9.6) participated in this experiment in
exchange for payment (12 min at US$10/hr). To be
eligible, workers had to have an approval rate > 95% and
have completed 500 Human Intelligence Tasks (HITs).
Apparatus and stimuli. The experiment was programmed
using the jspsych library (de Leeuw, 2015) with stimuli
and outcome pictures created as separate files (see osf.io/
esfm3/). Each stimulus consisted of a picture of a food
along with its verbal label on a white background and was
300 pixels wide and 400 pixels high. The experimental
cues (A-I, 9 in total) were selected randomly from a total
pool of 16 food pictures. The outcomes (allergic reaction
and no reaction) were displayed as image files (609 x 380
pixels).
Procedure. All experiments were approved by the Univer-
sity of New South Wales Human Research Ethics Advisory
Panel and participants provided written online consent.
Table 1 shows the sequence of phases in Experiment 1.
Training phase. Participants were told that an allergist
was trying to work out which foods cause allergic reac-
tions in “Mr X.” On each trial, participants were first pre-
sented with the text “Mr X eats the following meal”: at the
top of the screen, along with the contents of a meal that Mr
X had eaten (pictures of either one or two foods with their
verbal labels) below the text. Then, 500 ms after presenta-
tion of the cues, the prompt “Please rate the likelihood that
Mr X will show an allergic reaction after eating this meal”
appeared along with a visual analogue scale ranging from
“Definitely NO ALLERGIC REACTION” to “Definitely
ALLERGIC REACTION.” Participants made their predic-
tion by moving the slider to a point on the scale. Once
participants clicked on the scale (i.e., made any rating) a
“Continue” button appeared below the scale. Participants
had unlimited time to complete their rating in this, and all
subsequent phases. After clicking “Continue,” the predic-
tion scale and prompt disappeared and was replaced by
feedback. This feedback consisted of text (“Allergic Reac-
tion!” or “No allergic reaction”). If the outcome was an
allergic reaction, a sad face accompanied the text. After 2
s, the stimuli and feedback disappeared and the participant
entered the 1-s blank inter-trial-interval (ITI) period.
Table 1. Design of Experiments 1 and 2.
Experiment Training phase Outcome prediction test Open-ended questions Causal ratings test Causal structure
assessment
1 A+AB-
C+
DE-F-
GH+
A AB
C
DE D E F
CB CD CF CI
CB vs. CD
CB vs. CF
A B
C
D E F
I
3AFC causal
structure question
2 A+AB-
C+
DE- F-
GH+
A AB B
C
DE D E F
CB CD CF CI I
B A B
C
D E F
I
3AFC causal
structure question

Lee and Lovibond 155
Table 1 shows the six trial types presented in the train-
ing phase. Each of these six trial types (A+, AB-, C+,
DE-, F-, GH+) was presented twice in each block, with
the left-to-right presentation of the compound cues coun-
terbalanced within a block (e.g., AB and BA were pre-
sented in randomised order). Training consisted of three
blocks in total, meaning that each trial type was presented
six times. Trials were randomised such that the same trial
type could not be presented on successive trials.
Outcome prediction test. After training was completed,
participants were told that they would now be presented
with more meals, and that they should continue to make
the same predictions about the occurrence of an allergic
reaction. However, in this phase, they were told that they
would not receive any feedback. The trial types presented
in the outcome prediction test are presented in Table 1, and
consist of stimuli that were familiar (i.e., presented in train-
ing, A, AB, C, DE, D, E, F) as well as novel compounds for
the summation test (CB, CD, CF, CI). Each trial type was
presented twice at test, with the left-to-right presentation
of compound cues again counterbalanced. Ratings in the
training and outcome prediction test could range between
0 and 100.
Open-ended questions. Following the outcome predic-
tion test, participants were asked two open-ended ques-
tions designed to elicit explanations for their ordinal
pattern of ratings for the summation stimuli. The order
of the questions (CB vs. CD and CB vs. CF) was ran-
domised. On each trial, participants were presented with
the three relevant cues at the top of the screen (C, B, and
either D or F from left to right), and told that two of the
meals that Mr. X ate were C+B and C+D/F. They were
told that compared to CD/CF, they rated the likelihood
of allergic reaction as LOWER/HIGHER/SIMILAR for
CB. The displayed text was determined by their actual
ratings in the outcome prediction test. The threshold for
displaying “similar” was a ratings difference < 5 for CB
and CD/CF. Participants were asked to explain why they
rated the meals in this way by typing their response into
a field box.
Causal ratings test. Participants were then asked to
judge whether each food caused or prevented an allergic
reaction. Participants were presented with the cues (A, B,
C, D, E, F, and I) and asked “Please rate to what extent
this food tended to prevent or cause an allergic reac-
tion” on a visual analogue scale. The scale ranged from
“Strongly PREVENTED ALLERGIC REACTION” to
“Strongly CAUSED ALLERGIC REACTION,” with “No
effect” labelled in the middle of the scale. The ratings were
made in a similar way to the previous test phase with no
feedback. The causal ratings were transformed to range
between -100 and +100.
Causal structure assessment. In this phase, we asked
participants specifically what they learned about cue B
(the negative feature) using a 3AFC question. Participants
had to select which of the following options best described
what they thought about the food representing cue B. The
options for the three different causal structures (configu-
ral, modulation, prevention) were presented in randomised
order. The prevention option was phrased “It prevented
allergic reactions in general,” the modulation option was
phrased “It prevented allergic reactions caused by specific
foods,” and the configural option was phrased “It is hard
to know the exact role of individual foods such as this one.
I concentrated on remembering which combinations of
foods caused an allergic reaction and worked from there.”
Finally, the last question asked participants whether
they wrote anything down during the training phase.
Results
Exclusion criteria. Participants were excluded from analysis
if they failed to meet the training criterion. The training
criterion was an average rating > 75 for the stimuli that
predicted the outcome (A+, C+, GH+) and an average
rating < 25 for the stimuli that predicted no outcome (AB-,
DE-, F-) in the last block of training (i.e., last two presenta-
tions of each trial type). Participants were also excluded if
they reported writing things down during training, as the
instructions clearly stated for them not to. Of the 100 par-
ticipants, 18 reported writing during training (and all 18
also failed the training criterion). After exclusions, 72 par-
ticipants remained.
Causal structure subgroups. To examine the effect of
inferred causal structure, we divided participants into
subgroups based on their response to the 3AFC question.
We did not use the open-ended responses to classify par-
ticipants as many of the responses were not clearly inter-
pretable with respect to causal structure (e.g.,
re-descriptions of the training contingencies). Note,
however, that the open-ended responses did not reveal
any alternative causal structures not covered by our
existing forced choice options. Using the 3AFC ques-
tion, there were 25 participants in the configural sub-
group, 23 in the modulation subgroup, and 24 in the
prevention subgroup.
Data analysis. Our primary measure of transfer was the rat-
ings to the CB compound at test. In line with a traditional
summation test (Rescorla, 1969), we also computed differ-
ence scores in outcome prediction ratings between CB (the
transfer compound) and each of the three control com-
pounds (CD, CF, and CI, see Supplementary Materials for
these analyses). Note that we regarded CD as the most
conservative control compound and therefore the best con-
trol for assessing the degree of inhibitory transfer.

156 Quarterly Journal of Experimental Psychology 74(1)
To test the effect of causal structure, we planned 3
pairwise comparisons to examine the differences between
subgroups (configural vs. modulation, configural vs. pre-
vention, modulation vs. prevention). Since these con-
trasts were non-orthogonal, we applied Holm-Bonferroni
correction to control the family-wise type I error rate at
.05. For ease of interpretation, we report uncorrected
p-values and note whether each contrast is significant
after Holm-Bonferroni correction. The data were ana-
lysed in R (R Core Team, 2020) using the “afex”
(Singmann et al., 2018) and “emmeans” (Lenth, 2019)
packages.
Training. Figure 2 shows mean outcome predictions from
the training phase for each subgroup. It is clear from Fig-
ure 2 that participants showed differential responding to
stimuli predicting the outcome versus those that predict no
outcome. From the second presentation, there are high
overall predictive ratings of the outcome for the A+, C+,
and GH+ trials, and low overall predictive ratings of the
outcome for the AB-, DE-, and F-trials. We tested for dif-
ferences in causal structure subgroups on terminal perfor-
mance by calculating the difference between the average
of the predictive (A+, C+, GH+) trial types and the aver-
age of the non-predictive (AB-, DE-, F-) trial types on the
last two trials of training. There was no effect of causal
structure subgroup for this difference score, F < 1.
Outcome prediction test. For analysis of the outcome pre-
diction test data, each participant’s ratings were averaged
across the two test trials for each test stimulus, producing
a single rating for each participant for each test stimulus.
Figure 3 shows mean prediction ratings for the novel sum-
mation compounds split by participants’ reported causal
structure (prevention, modulation, or configural). For
brevity, we report only the analyses for the summation
compounds (see Supplementary Materials for analyses for
the other test stimuli).
As previously stated, our primary test of the effect of
inferred causal structure involved outcome predictions for
the CB compound. Examining Figure 3, participants who
inferred a preventive causal structure gave the lowest out-
come predictions for CB, participants who reported a con-
figural strategy gave the highest outcome predictions for
CB, and the participants who inferred a modulation causal
structure gave ratings in between. In partial support of this
observation, planned pairwise comparisons showed a sig-
nificant difference in ratings to CB between the configural
and prevention group, F(1, 69) = 12.2, p = .001, ηp
2 = .15,
but no significant difference between the configural and
modulation groups, F(1, 69) = 1.79, p = .185, ηp
2 = .03, nor
the modulation and prevention groups, F(1, 69) = 3.4,
p = .071, ηp
2 = .06. Thus, there was some evidence that
causal structure predicted the degree of transfer of learning
about the feature B. In contrast, there were no significant
pairwise comparisons between causal structure subgroups
for any of the control compounds (CD, CF, or CI, see
Supplementary Materials).
Causal ratings test. Figure 4 shows mean causal ratings for
each test cue split by inferred causal structure. These
Figure 2. Mean outcome prediction ratings (+/– 1 SE of the mean) for each causal structure subgroup in the training phase in
Experiment 1.
Figure 3. Mean outcome predictions (+/– 1 SE of the mean)
for the inhibition (CB) and control compounds (CD, CF, CI) in
Experiment 1 split by causal structure subgroup.

Lee and Lovibond 157
results differ from the outcome prediction test in that all
test cues were presented individually, and participants
were asked to provide ratings on the prevent-cause scale,
rather than simply predicting the likelihood of the out-
come. The figure shows that participants in all causal
structure subgroups gave ratings at ceiling for the causal
cues A and C. It also shows that participants gave negative
(preventive) causal ratings for the feature B, but also to
some extent for the non-causal cues (D, E, and F) and the
novel cue I. There were no significant differences between
causal structure subgroups when examining A, C, D, E, F,
or I (see Supplementary Materials). However for the fea-
ture B, causal ratings were significantly lower for the pre-
vention subgroup than for the configural subgroup,
F(1, 69) = 10.4, p = .002, ηp
2 = .13. There was again no sig-
nificant difference between configural and modulation,
F(1, 69) = 3.77, p = .056, ηp
2 = .05, and no significant differ-
ence between modulation and prevention subgroups, F(1,
69) = 1.51, p = .223, ηp
2 = .02. Thus, the results from the
causal ratings test mirrored the results from the outcome
prediction test, with the clearest difference in causal struc-
ture being between the configural and prevention sub-
groups for the feature B.
Discussion
In summary, participants learned the feature negative dis-
crimination and showed evidence of transfer of inhibitory
learning, producing lower outcome prediction ratings
when the feature B was combined with a new target (C),
compared to when the same target was combined with a
non-causal control cue trained in compound (CD). The
novel finding was that only a third of the participants actu-
ally reported a causal structure in line with conditioned
inhibition. Participants were evenly split over the other
two causal structures, demonstrating substantial individual
differences in inferred causal structure in standard simulta-
neous feature negative training. Participants’ self-reported
causal structure predicted the degree of transfer of inhibi-
tory learning of the feature to the new target C, with the
configural subgroup showing the highest predictive ratings
(least transfer), the prevention subgroup showing the low-
est predictive ratings (most transfer), and the modulation
subgroup falling in between.
The differences between causal structure subgroups for
CB on the outcome prediction test were mirrored in the
causal ratings test for cue B. The ordinal pattern of ratings
for B and the significant subgroup comparisons (preven-
tion < configural) for this cue were consistent with ratings
to CB on the outcome prediction test. Although typically
included in human causal learning studies, the main limita-
tion of the standard causal rating test for our purposes is
that it assumes a prevention structure whereby inhibition is
opposite to excitation, and hence may not differentiate
between modulatory and preventive causal structures.
Participants who inferred a modulation structure may have
given similarly negative ratings to the prevention subgroup
for the inhibitor B if they interpreted the question as asking
about the effect of B on A. It is interesting to note that
although cues D, E, and F were only ever shown to lead to
no outcome (i.e., zero associative strength according to the
Rescorla-Wagner model; Rescorla & Wagner, 1972), par-
ticipants on the whole rated these cues as negative on the
cause-prevent scale. It is possible that some participants
may have interpreted the scale in a similar fashion to the
outcome prediction scale, using the left (prevention) side
of the scale for stimuli that predicted no reaction rather
than prevention of the reaction.
In summary, Experiment 1 provided initial evidence of
individual differences in how participants interpret the
causal scenario following feature negative training, and
showed that these self-reported causal structures predicted
the degree of transfer of inhibitory learning. An important
follow-up question is whether these individual differences
are stable or amenable to manipulation. Experiment 2
sought to test this possibility.
Experiment 2
The primary aim of Experiment 2 was to test whether ver-
bal hints could influence the inferred causal structure
adopted by participants. If feature negative training is an
ambiguous scenario where participants learn different
causal structures, then participants may be sensitive to
explicit hints about how the cues exert their effects.
Alternatively, there may be stable individual differences
that lead participants to interpret the contingencies using
their preferred structure irrespective of any hints.
Experiment 2 used a between-subjects design with 3 hint
groups (configural hint, modulation hint, prevention hint).
Figure 4. Mean causal ratings (+/– 1 SE of the mean) for
individual test cues in Experiment 1. Positive ratings indicate
causation, negative ratings indicate prevention, zero indicates
no effect.

158 Quarterly Journal of Experimental Psychology 74(1)
We used the same 3AFC question as in Experiment 1 to
assess to what extent the hint manipulation influenced the
causal structure participants inferred from the feature neg-
ative contingencies.
Method
Participants. Three hundred Mechanical Turk workers
(105 female, M age = 34.6, SD age = 10.4) took part in
Experiment 2. The eligibility criteria were the same for
Experiment 1 with the addition that participants could not
have participated in Experiment 1 before. Due to a techni-
cal issue, the data file for one participant was not saved,
leaving 299 data files.
Procedure. The method was identical to Experiment 1
except for the following changes. Participants were ran-
domly allocated to one of three hint groups (configural hint,
modulation hint, prevention hint). The hint was presented
prior to the training phase, with the following wording:
Prevention Hint
“The doctor suspects:
There is at least one food that causes an allergic reaction
in Mr X
and also
there may be another food that suppresses allergic
reactions in Mr X.”
Modulation Hint
“The doctor suspects:
There is at least one food that causes an allergic reaction
in Mr X
and also
there may be another food that blocks that food from
causing an allergic reaction in Mr X.”
Configural Hint
“The doctor suspects:
There are specific meals that lead to allergic reactions in
Mr X.”
B and I were added as test cues to the outcome prediction
test, and the open-ended question was changed with the aim
of eliciting clearer responses. Instead of prompting partici-
pants to explain why they had rated CB in a particular way
compared to CD and CF, the open-ended question in
Experiment 2 asked participants to report what they had
learned about B. We also changed the instructions for the
causal ratings test slightly to help differentiate the scale from
the outcome prediction scale used in the previous phase.
Results
Exclusion criteria. The same exclusion criteria were used as
in Experiment 1. Eighty-seven participants failed the train-
ing criterion, and 49 reported writing things down during
training (42 of whom also failed the training criterion).
After exclusions, 205 participants remained.
Causal structure subgroups. Table 2 shows the division of
participants into hint group and reported causal structure
after exclusions. The open-ended answers were more
informative than in Experiment 1, and aligned moderately
well with the answers to the forced-choice question (see
Supplementary Materials), but for consistency we again
used the forced-choice question to classify participants
into causal structure subgroups for analysis. Compared to
Experiment 1, where the division of participants into
causal structure subgroups was much more even, the
majority of participants in Experiment 2 (55.1%) reported
a modulatory causal structure.
Effectiveness of hints. To test the effectiveness of the hints
on reported causal structures we conducted log-linear con-
trasts on the frequencies displayed in Table 2. Comparing
the cell frequencies where the reported causal structure
matched the hint (i.e., the diagonal entries) against the
other cells produced a significant difference, χ2(1,
N = 205) = 6.5, p = .011. The effect of each hint was exam-
ined by testing a contrast comparing the proportion of par-
ticipants in each hint group who reported the matching
causal structure relative to the proportion who reported
that structure in the other two hint groups. This contrast
was significant for the prevention hint, χ2(1, N = 205) = 8.1,
p = .004, but not for the configural, χ2(1, N = 205) = 3.1,
p = .078, or modulation hints, χ2(1, N = 205) = .193,
p = .660. Thus, there was statistical evidence that the hints
influenced participants’ causal structures overall, but when
examining the hints individually, only the prevention hint
produced a significant increase in the proportion of partici-
pants adopting that causal structure.
Data analysis. Since the hints appeared to influence the
causal structures participants adopted, we did not include
Table 2. Number of participants reporting each causal
structure by hint group in Experiment 2.
Causal structure Total
Configural Modulation Prevention
Hint group
Configural hint 14 48 15 77
Modulation hint 8 38 20 66
Prevention hint 6 27 29 62
Total 28 113 64 205

Lee and Lovibond 159
the interaction between hint and causal structure in our
analysis of the test data. Instead, we analysed the effects of
hint group and causal structure separately using the same
pairwise contrasts as in Experiment 1.
Training. Figure 5 shows the training data for each causal
structure subgroup averaged over hint groups (see Supple-
mentary Materials for the breakdown by hint group). The
acquisition curves look very similar in all cells despite the
greater amount of variability for the configural subgroup
due to the smaller n. Using the same measure for terminal
performance as Experiment 1, there was a non-significant
effect for hint group, F < 1, marginally non-significant
effect for causal structure subgroup, F(2, 196) = 2.89,
p = .058, ηp
2 = .029, and a non-significant interaction, F(4,
196) = 1.84, p = .123, ηp
2 = .036.
Outcome prediction test. Figure 6 shows the mean ratings
to the critical test compounds split by causal structure (a)
and by hint group (b). Similar to Experiment 1, there
appear to be no differences between hint groups or causal
structures for the control compounds (CD, CF, and CI), but
there do appear to be differences for the test compound
CB. Figure 6a shows a similar pattern between causal
structure subgroups to Experiment 1, with the lowest rat-
ings to CB in the prevention subgroup, the highest ratings
in the configural subgroup, and the modulation subgroup
in between. The ordinal pattern of ratings to CB is similar
when examining the hint groups (Figure 6b), except that
the modulation hint group gave similarly low ratings to the
prevention subgroup.
The tests of causal structure differences were much
stronger in Experiment 2 than in Experiment 1 (Figure 6a).
For absolute ratings to the CB compound, the prevention
subgroup gave significantly lower predictive ratings com-
pared to the configural subgroup, F(1, 204) = 25.2,
p < .001, ηp
2 = .11, and also compared to the modulation
subgroup, F(1, 204) = 21.7, p < .001, ηp
2 = .10, but the
difference between configural and modulation subgroups
was marginally non-significant, F(1, 204) = 3.76, p = .054,
ηp
2 = .02. For the CB-CD difference score, clear statistical
differences were found in Experiment 2 between all three
causal structure subgroups (see Supplementary Materials
for full details).
Examining the hint groups, the prevention hint group
rated CB significantly lower than the configural subgroup,
F(1, 204) = 22.5, p < .001, ηp
2 = .10, and the modulation
hint group also rated CB lower than the configural hint
group, F(1, 204) = 21.1, p < .001, ηp
2 = .09, but there was
no difference between the prevention and modulation sub-
groups, F < 1 (see Figure 6b). This pattern of results for
absolute ratings to CB was mirrored in the results examin-
ing the CB-CD difference scores (see Supplementary
Figure 5. Mean outcome prediction ratings (+/– 1 SE of the mean) for each causal structure subgroup in the training phase in
Experiment 2.
Figure 6. Mean outcome predictions (+/– 1 SE of the mean)
for inhibition (CB) and control compounds (CD, CF, CI) in
Experiment 2 split by (a) causal structure subgroup and (b) hint
group.

160 Quarterly Journal of Experimental Psychology 74(1)
Materials for full details). Thus, it appears that the hints
were partially effective in influencing the degree of trans-
fer of inhibitory learning in the summation test.
Causal ratings test. Figure 7 shows the results from the
causal ratings test split by causal structure (a) and by hint
group (b). Examining the causal structure subgroups, Cue
B was rated significantly lower (i.e., more preventive) in
the modulation subgroup relative to the configural sub-
group, F(1, 202) = 39.7, p < .001, ηp
2 = .16, and also lower
in the prevention subgroup relative to the configural sub-
group, F(1, 202) = 40.0, p < .001, ηp
2 = .17 (see Figure 7a).
There was no significant difference comparing the modu-
lation to prevention subgroup, F < 1. The pattern of sub-
group causal ratings to B was similar to the pattern of
ratings to CB in the outcome prediction test, except that
there was no difference between causal ratings to B in the
modulation and prevention subgroups. In contrast, the
majority of pairwise comparisons between hint groups and
causal structures were non-significant (see Supplementary
Materials for full details). For the feature B, there were
also no significant pairwise comparisons between hint
groups, Fs <= 3.36. All 3 hint groups rated B as strongly
preventive on the scale (Figure 7b).
Discussion
In Experiment 2, the majority of participants reported a
modulation causal structure, again demonstrating that only
a small subset of participants learn that a negative feature
is preventive in the manner described by Rescorla and
Wagner (1972). In comparison to Experiment 1,
Experiment 2 provided stronger statistical evidence that
self-reported causal structure predicted the degree of trans-
fer of inhibitory learning. Participants who reported a pre-
ventive causal structure showed the strongest transfer of
inhibitory learning, participants who inferred a configural
strategy showed the least transfer, and participants who
inferred a modulatory causal structure showed an interme-
diate amount of transfer. Although one of the pairwise
comparisons in the outcome prediction test fell short of
significance (configural vs. modulation), Experiment 2
replicated the key qualitative pattern of transfer results in
Experiment 1, and the configural and modulation sub-
groups differed significantly on the causal ratings test.
Experiment 2 also extended Experiment 1 by showing
that causal structure could be manipulated to some extent
via verbal hints given prior to feature negative training.
Participants given the prevention and modulation hint
showed strong transfer, while participants given the con-
figural hint showed weaker transfer. However, a substan-
tial number of participants also showed resistance to the
hint, reporting different causal structures to that presented
in the hint. Our analyses revealed that although the hints
were effective overall, the prevention hint was the only
one that was effective in increasing the proportion of par-
ticipants choosing the corresponding causal structure,
explaining why the same pattern of results was not found
comparing hint group and self-reported causal structure.
One potential reason for this result may be that the hints
differed in their degree of clarity, with the prevention hint
offering the clearest indication of the correct underlying
causal structure. Most importantly however, the ordinal
pattern of transfer between causal structure subgroups was
consistent with Experiment 1.
General discussion
Across two experiments, we showed considerable indi-
vidual differences in the way participants learn the under-
lying causal structure of feature negative (A+AB-)
contingencies. We found that a subset of participants
reported learning that B prevented the outcome from
Figure 7. Mean causal ratings (+/– 1 SE of the mean) for
individual test cues in Experiment 1 split by (a) causal structure
subgroup and (b) hint group. Positive ratings indicate causation,
negative ratings indicate prevention, zero indicates no effect.

Lee and Lovibond 161
occurring, another subset reported that B modulated the
causal effect of A, and another subset were agnostic about
the effects of B alone as it was never presented outside of
the AB compound. The proportion of participants report-
ing each causal structure was approximately equal in
Experiment 1, and slightly biased towards the modulation
subgroup in Experiment 2. Thus, the structures we chose
seemed adequate in covering the range of possible causal
structures that participants learned (the open-ended
responses in Experiment 2 confirmed this conclusion; see
Supplementary Materials). Notably, only a modest propor-
tion of participants in each experiment reported a preven-
tive causal structure, suggesting that traditional associative
models (e.g., Rescorla & Wagner, 1972) do not adequately
capture what the majority of participants learn during fea-
ture negative training. These experiments demonstrate,
along with Glautier and Brudan (2019), that feature nega-
tive training with the typical simultaneous procedure leads
to a variety of learned causal structures in humans.
We also showed that self-reported causal structure was
consistent with the degree to which participants transferred
their learning about the feature (B) to a novel target (CB).
Across both studies, participants who reported a preven-
tive causal structure showed the greatest transfer of learn-
ing about the feature to a compound containing a new
target (CB), the configural subgroup showed the least
transfer, and the modulation subgroup showed an interme-
diate amount of transfer. Although we only assessed self-
reported causal structure for B and not the control cues,
participants also gave causal ratings to the feature B on the
prevent-cause scale in broad alignment with their predic-
tive ratings to CB, lending some support to the idea that
our findings were specific to the negative feature B.
However, for future work, it would be useful to assess par-
ticipants’ inferred causal structure for the control cues
(e.g., D) to ensure that participants’ reported knowledge is
specific to B. Finally, Experiment 2 showed that explicit
hints about the potential effect of the feature B were only
partially effective in determining the causal structure that
participants adopted, and there were less clear distinctions
in terms of transfer in the summation test compared to self-
reported causal structure. Even with the effect of the hint,
the proportion of participants reporting a prevention causal
structure was still only about a third (64/205), and only 33
participants (16%) reported a prevention causal structure
in the absence of the prevention hint. The results from
Experiment 2 suggest that there are strong pre-existing
individual differences that determine how participants
interpret the underlying causal structure of feature nega-
tive contingencies, but they are also somewhat sensitive to
explicit hints about the effect of the feature.
Differences between causal structures
Thus far, we have argued that the differences in causal
structure captured via self-report are qualitative. However,
our modulation and prevention subgroups could instead be
seen as differing quantitatively in the extent to which they
are willing to generalise the properties they have learned
about the negative feature B. To explore these potential
quantitative differences between causal structure sub-
groups, we calculated a generalisation score for each par-
ticipant (the difference between ratings for CB and AB on
the outcome prediction test), and examined the relation-
ship between this generalisation score and participants’
causal ratings for the negative feature B. Figure 8 shows a
contour plot for Experiment 2 with the generalisation score
on the x axis and the participants’ causal ratings for B on
the y axis. The contour plot shows regions where the data
cluster. Each contour connects regions that have similar
density, and the areas in the middle of the contours have
the highest density.
From Figure 8, it does appear that the main difference
between modulation and prevention subgroups is the
degree to which participants vary in their level of generali-
sation. On the x-axis, the prevention subgroup are clus-
tered around 0, implying that the majority of participants
gave very similar (low) ratings to both AB and CB,
whereas the modulation subgroup’s scores vary across the
full range from 0 to 100. The prevention subgroup distri-
bution overlaps substantially with one end of the modula-
tion subgroup, raising the possibility that even the
participants who chose the prevention option could be con-
sidered to be modulators with a high degree of generalisa-
tion. The wide variation in generalisation scores within the
modulatory group makes sense, as it would be ambiguous
for these participants whether B will have the same effect
on a novel target C as the original target A. This ambiguity
then allows the expression of individual differences in
willingness to generalise. Both subgroups are clustered at
the negative end on the y-axis implying similar (preven-
tive) learning about B’s effects, although as noted previ-
ously, the causal rating scale is unable to differentiate
between modulatory and preventive causal structures.
In contrast to the other two subgroups, the configural
subgroup are much more dispersed on both dimensions
than the modulation and prevention subgroups. This pattern
might be expected if participants have only learned about
the AB configuration predicting no outcome, and are there-
fore agnostic about the effects of the negative feature B
alone. This may lead participants to respond either conserv-
atively (e.g., responding at 50, the midpoint of the scale) or
responding randomly, which would result in a large amount
of variability. It is also possible that in our efforts to capture
the agnosticism about the causal effects of B in the configu-
ral participants, we inadvertently captured participants who
were simply less confident about their predictions (and who
responded somewhat randomly and therefore higher in
their causal ratings for B). Thus quantitatively, the differ-
ences between our subgroups can be described as differ-
ences in variability in responding across two dimensions
(causal learning and generalisation).

162 Quarterly Journal of Experimental Psychology 74(1)
Implications for occasion-setting and
conditioned inhibition
These experiments demonstrate that some participants
endorse a causal structure consistent with occasion setting
even with simultaneous presentation of the AB compound.
Participants who chose the modulation option showed less
transfer, on average, than the prevention subgroup. If par-
ticipants who reported a modulatory causal structure have
learned something akin to occasion setting, our experi-
ments demonstrate that occasion setters can pass the typi-
cal summation test, but do so to a lesser extent than
inhibitors. This means that in human studies, passing a
summation test does not necessarily imply that partici-
pants have learned that the feature is inhibitory or preven-
tive in the general sense described by Rescorla (1969).
Our results suggest that only a small proportion of par-
ticipants (the preventors) learn “true” inhibition in the way
described by Rescorla (1969) and the Rescorla-Wagner
model (Rescorla & Wagner, 1972) given the standard sin-
gle set of feature negative contingencies. However, further
studies with different cover stories, cues, and outcomes are
needed to verify the generality of this finding. At this
stage, whether the prevention subgroup can be considered
analogous to inhibition, or whether they are simply modu-
lators who are willing to generalise (as the analysis in
Figure 8 suggests), is unclear. Thus, if participants learn
elementally, modulation may be the default outcome for
feature negative training, contradicting traditional associa-
tive models such as the Rescorla-Wagner model, which
predict that the negative feature acquires a direct inhibitory
link with the outcome. While lending some support to the
validity of associative models, our results also demonstrate
that prevention learning is more complex and varied in
humans than in animals, potentially explaining why condi-
tioned inhibition effects as assessed in a summation test
are less robust in humans (Karazinov & Boakes, 2007; Lee
& Livesey, 2012; Williams, 1995; Wilkinson et al., 1989).
An interesting avenue for future research would be to test
whether classic manipulations known to increase the like-
lihood of occasion-setting in animals (e.g., serial vs. simul-
taneous presentation) result in a larger proportion of
participants reporting a modulatory causal structure.
Implications for configural learning
An area of continued debate in associative learning is
whether stimuli enter into associations with outcomes as
configurations (Pearce, 1994) or elements (e.g., Harris &
Livesey, 2010; McLaren & Mackintosh, 2000; Rescorla &
Wagner, 1972), or whether both types of representation are
needed (Rescorla, 1973). Our study suggests that a separate
configural strategy can be differentiated from occasion-set-
ting whereby participants learn about configurations of
stimuli and are reluctant to make any strong assumptions
about novel configurations they have not observed. Our
results are thus consistent with studies demonstrating a fail-
ure to generalise between configurations of stimuli (Shanks,
Charles, et al., 1998; Shanks, Darby, et al., 1998; Williams
et al., 1994). Melchers et al. (2008) have proposed that
whether participants learn elementally or configurally
depends on properties of the experimental task. In other
words, the degree to which participants learn configurally is
somewhat flexible. This idea is also instantiated in Wagner’s
(2003) Replaced Elements Theory. The r parameter deter-
mines the proportion of stimulus elements that is replaced
when the stimulus is presented with another cue (or context)
and therefore controls whether a stimulus is processed more
Figure 8. Density plot with individual data points overlaid from each causal structure subgroup in Experiment 2. The x-axis
represents the generalisation score (difference in outcome prediction ratings between CB and AB) and the y-axis is the causal rating
given to B in the causal ratings test (positive values indicate causation and negative values indicate prevention).

Lee and Lovibond 163
(high r) or less (low r) configurally. Rather than configural
learning being a feature of the task or the accompanying
stimulus, our results suggest that certain participants have a
predisposition towards learning configurally. This idea is
consistent with studies demonstrating individual differences
in attention to global or local stimulus features, which may
account for why some participants appear to learn elemen-
tally and others configurally (Byrom & Murphy, 2014,
2016)
Concluding remarks
The presence of reliable individual differences in our task
suggest that greater efforts should be made in human asso-
ciative learning tasks to identify learning strategies that dif-
fer in important qualitative ways (e.g., Glautier & Brudan,
2019; Shanks & Darby, 1998). We have already shown that
group-level trends can be misleading when there are sub-
groups of participants who learn and generalise in different
ways (Lee et al., 2018). Similarly here, concluding in favour
of an effect (conditioned inhibition) due to functional char-
acteristics (passing the summation test) would contradict the
explicit report and the transfer behaviour of a large propor-
tion of participants. Our results complement recent calls for
further investigation into the generality of classic associa-
tive learning effects (Maes et al., 2016; Urcelay, 2017) and
greater recognition of individual differences in theory and
analysis (Lee et al., 2018; Lee, Mills, Hayes, & Livesey,
2020). Acknowledging that humans can solve associative
learning tasks by learning qualitatively different content
will facilitate better connections with the animal condition-
ing literature, and provide a better understanding of the
diverse ways in which humans learn about causal and pre-
ventive relationships in the world.
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect
to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support
for the research, authorship, and/or publication of this article:
This study was funded by an Australian Research Council
Discovery Grant (DP190103738) awarded to Peter Lovibond.
ORCID iD
Jessica C Lee https://orcid.org/0000-0003-4253-2008
Data accessibility statement
The data and materials are available at https://osf.io/esfm3/
Supplementary material
The Supplementary Material is available at: qjep.sagepub.com
References
Baeyens, F., Vervliet, B., Vansteenwegen, D., Beckers, T.,
Hermans, D., & Eelen, P. (2004). Simultaneous and sequen-
tial feature negative discriminations: Elemental learning and
occasion setting in human Pavlovian conditioning. Learning
and Motivation, 35, 136–166.
Bonardi, C., Robinson, J., & Jennings, D. (2017). Can exist-
ing associative principles explain occasion setting? Some
old ideas and some new data. Behavioural Processes, 137,
5–18.
Byrom, N. C., & Murphy, R. A. (2014). Sampling capacity
underlies individual differences in human associative learn-
ing. Journal of Experimental Psychology: Animal Learning
and Cognition, 40, 133–143.
Byrom, N. C., & Murphy, R. A. (2016). Individual difference
in configural associative learning. Journal of Experimental
Psychology: Animal Learning and Cognition, 4, 2325–2335.
Cheng, P. W. (1997). From covariation to causation: A causal
power theory. Psychological Review, 104, 367–405.
Cheng, P. W., & Novick, L. R. (1992). Covariation in natural
causal induction. Psychological Review, 99, 365–382.
de Leeuw, J. R. (2015). jsPsych: A JavaScript library for creat-
ing behavioural experiments in a web browser. Behaviour
Research Methods, 47, 1–12. https://doi.org/10.3758/
s13428-014-0458-y
Dickinson, A., Shanks, D., & Evenden, J. (1984). Judgement of
act-outcome contingency: The role of selective attribution.
Quarterly Journal of Experimental Psychology, Section A,
36, 29–50.
Fraser, K. M., & Holland, P. C. (2019). Occasion setting.
Behavioral Neuroscience, 133, 145–175.
Glautier, S., & Brudan, O. (2019). Stable individual differences
in occasion setting. Experimental Psychology, 66, 281–295.
Harris, J. A., & Livesey, E. J. (2010). An attention-modulated
associative network. Learning & Behavior, 38, 1–26.
Holland, P. C. (1984). Differential effects of reinforcement of
an inhibitory feature after serial and simultaneous feature
negative discrimination training. Journal of Experimental
Psychology: Animal Behavior Processes, 10, 461–475.
Holland, P. C. (1989). Transfer of negative occasion setting and
conditioned inhibition across conditioned and uncondi-
tioned stimuli. Journal of Experimental Psychology: Animal
Behavior Processes, 15, 311–328.
Holland, P. C., & Lamarre, J. (1984). Transfer of inhibition after
serial and simultaneous feature negative discrimination
training. Learning and Motivation, 15, 219–243.
Holland, P. C., & Morell, J. M. (1996). The effects of feature-
target and intertrial intervals (ITI) on learning and perfor-
mance of serial feature negative discriminations. Learning
and Motivation, 27, 21–42.
Jean-Richard-dit-Bressel, P., Ma, C., Bradfield, L. A., Killcross,
S., & McNally, G. P. (2019). Punishment insensitivity
emerges from impaired contingency detection, not aversion
insensitivity or reward dominance. eLife, 8, Article e52765.
Karazinov, D., & Boakes, R. A. (2004). Learning about cues that
prevent an outcome: Conditioned inhibition and differential
inhibition in human predictive learning. Quarterly Journal
of Experimental Psychology, Section B, 57, 153–178.
Karazinov, D., & Boakes, R. A. (2007). Second-order condition-
ing in human predictive judgements when there is little time

164 Quarterly Journal of Experimental Psychology 74(1)
to think. Quarterly Journal of Experimental Psychology, 60,
448–460.
Konorski, J. (1948). Conditioned reflexes and neuron organiza-
tion. Cambridge University Press.
Konorski, J. (1967). Integrative activity of the brain. University
of Chicago Press.
Lamarre, J., & Holland, P. C. (1985). Acquisition and trans-
fer of feature negative discriminations. Bulletin of the
Psychonomic Society, 23, 71–74.
Lee, J. C., Mills, L., Hayes, B. K., & Livesey, E. J. (2020).
Modelling generalisation gradients as augmented Gaussian
functions. Quarterly Journal of Experimental Psychology.
Advance online publication.
Lee, J. C., Hayes, B. K., & Lovibond, P. F. (2018). Peak shift
and rules in human generalization. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 44, 1955–
1970.
Lee, J. C., & Livesey, E. J. (2012). Second-order conditioning and
conditioned inhibition: Influences of speed versus accuracy
on human causal learning. PLOS ONE, 7(11), Article e49899.
Lee, J. C., Lovibond, P. F., & Hayes, B. K. (2019). Evidential
diversity increases generalization in predictive learning.
Quarterly Journal of Experimental Psychology, 72, 2647–
2657.
Lee, J. C., Lovibond, P. F., Hayes, B. K., & Navarro, D. (2019).
Negative evidence and inductive reasoning in gener-
alization of associative learning. Journal of Experimental
Psychology: General, 148, 289–303.
Lee, J. C., Mills, L., Hayes, B. K., & Livesey, E. J. (2020).
Modelling generalisation gradients as augmented
Gaussian functions. Quarterly Journal of Experimental
Psychology. Advance online publication. https://doi.
org/10.1177/1747021820949470
Lenth, R. (2019). emmeans: Estimated marginal means, aka
least-squares means (R Package Version 1.3.5). https://
CRAN.R-project.org/package=emmeans
Lovibond, P. F., Lee, J. C., & Hayes, B. K. (2020). Stimulus
discriminability and induction as independent components
of generalization. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 46, 1106–1120.
Maes, E., Boddez, Y., Alfei, J. M., ’Krypotos, A. M., D’Hooge,
R., De Houwer, J., & Beckers, T. (2016). The elusive nature
of the blocking effect: 15 failures to replicate. Journal of
Experimental Psychology: General, 145(9), e49–e71.
McLaren, I. P. L., & Mackintosh, N. J. (2000). An elemen-
tal model of associative learning: I. Latent inhibition and
perceptual learning. Animal Learning and Behavior, 28,
211–246.
McLaren, I. P. L., McAndrew, A., Angerer, K., McLaren, R.,
Forrest, C., Bowditch, W., Monsell, S., & Verbruggen,
F. (2018). Mackintosh lecture -: Association and cogni-
tion: Two processes, one system. Quarterly Journal of
Experimental Psychology, 72, 98–117.
Melchers, K. G., Shanks, D. R., & Lachnit, H. (2008). Stimulus
coding in human associative learning: Flexible represen-
tations of parts and wholes. Behavioural Processes, 77,
413–427.
Mitchell, C. J., De Houwer, J., & Lovibond, P. F. (2009).
The propositional nature of human associative learning.
Behavioral and Brain Sciences, 32, 183–246.
Pavlov, I. P. (1927). Conditioned reflexes. Oxford University
Press.
Pearce, J. M. (1987). A model for stimulus generalization in
Pavlovian conditioning. Psychological Review, 94, 61–73.
Pearce, J. M. (1994). Similarity and discrimination: A selective
review and a connectionist model. Psychological Review,
101, 587–607.
R Core Team. (2020). R: A language and environment for sta-
tistical computing. R Foundation for Statistical Computing.
https://www.R-project.org/
Rescorla, R. A. (1969). Pavlovian conditioned inhibition.
Psychological Bulletin, 72, 77–94.
Rescorla, R. A. (1973). Evidence for a “unique stimulus” account
of configural conditioning. Journal of Comparative and
Physiological Psychology, 85, 331–338.
Rescorla, R. A., & Wagner, A. R. (1972). A theory of Pavlovian
conditioning: Variations in the effectiveness of reinforce-
ment and nonreinforcement. Classical Conditioning II:
Current Research and Theory, 2, 64–99.
Ross, R. T., & Holland, P. C. (1981). Conditioning of simulta-
neous and serial feature-positive discriminations. Animal
Learning & Behavior, 9, 293–303.
Savastano, H., Cole, R. P., Barnet, R. C., & Miller, R. R. (1999).
Reconsidering conditioned inhibition. Learning and
Motivation, 30, 101–127.
Shanks, D. R., & Darby, R. J. (1998). Feature- and rule-based
generalization in human associative learning. Journal of
Experimental Psychology: Animal Behavior Processes, 24,
405–415.
Shanks, D. R., Charles, D., Darby, R. J., & Azmi, A. (1998).
Configural processes in human associative learning.
Journal of Experimental Psychology: Learning, Memory,
and Cognition, 24, 1353–1378.
Shanks, D. R., Darby, R. J., & Charles, D. (1998). Resistance
to interference in human associative learning: Evidence of
configural processing. Journal of Experimental Psychology:
Animal Behavior Processes, 24, 136–150.
Singmann, H., Bolker, B., Westfall, J., & Aust, F. (2018). afex:
Analysis of factorial experiments (R package version 0.21-
2). https://CRAN.R-project.org/package=afex
Sosa, R., & Ramírez, M. N. (2019). Conditioned inhibition:
Historical critiques and controversies in the light of recent
advances. Journal of Experimental Psychology: Animal
Learning and Cognition, 45, 17–42.
Stegmann, Y., Schiele, M. A., Schumann, D., Lonsdorf, T.
B., Zwanger, P., Romanos, M., Rief, A., Domschke, K.,
Deckert, J., Gamer, M., & Paulo, P. (2019). Individual dif-
ferences in human fear generalization—Pattern identifica-
tion and implications for anxiety disorders. Translational
Psychiatry, 9, 307.
Swartzentruber, D. (1995). Modulatory mechanisms in Pavlovian
conditioning. Animal Learning & Behavior, 23, 123–143.
Urcelay, G. P. (2017). Competition and facilitation in compound
conditioning. Journal of Experimental Psychology: Animal
Learning and Cognition, 43, 303–314.
Wagner, A. R. (2003). Context-sensitive elemental theory.
Quarterly Journal of Experimental Psychology, Section B, 56,
7–29.
Waldmann, M. R., Hagmayer, Y., & Blaisdell, A. P. (2006).
Beyond the information given: Causal models in learning

Lee and Lovibond 165
and reasoning. Current Directions in Psychological Science,
15, 307–311.
Waldmann, M. R., & Holyoak, K. J. (1992). Predictive and diag-
nostic learning within causal models: Asymmetries in cue
competition. Journal of Experimental Psychology: General,
121, 222–236.
Ward, W., & Jenkins, H. (1965). The display of information
and the judgement of contingency. Canadian Journal of
Psychology, 19, 231–241.
Wilkinson, G., Lovibond, P. F., Siddle, D. A. T., & Bond,
N. (1989). Effects of fear-relevance on electrodermal
safety signal learning. Biological Psychology, 28,
89–104.
Williams, D. A. (1995). Forms of inhibition in animal and human
learning. Journal of Experimental Psychology: Animal
Behavior Processes, 21, 129–142.
Williams, D. A., Sagness, K. E., & McPhee, J. E. (1994).
Configural and elemental strategies in predictive learning.
Journal of Experimental Psychology: Learning, Memory,
and Cognition, 20, 694–709.
Wilson, P. N., & Pearce, J. M. (1990). Selective transfer of
responding in conditional discriminations. Quarterly
Journal of Experimental Psychology, Section B, 42, 41–58.
Wong, A. H. K., & Lovibond, P. F. (2017). Rule-based gener-
alization in single-cue and differential fear conditioning in
humans. Biological Psychology, 129, 111–120.