Content uploaded by Cristine H Legare
Author content
All content in this area was uploaded by Cristine H Legare on Nov 10, 2017
Content may be subject to copyright.
Advanced Review
Causal learning in children
David M. Sobel1∗and Cristine H. Legare2
How do children learn the causal structure of the environment? We first sum-
marize a set of theories from the adult literature on causal learning, including
associative models, parameter estimation theories, and causal structure learning
accounts, as applicable to developmental science. We focus on causal graphical mod-
els as a description of children’s causal knowledge, and the implications of this
computational description for children’s causal learning. We then examine the con-
tributions of explanation and exploration to causal learning from a computational
standpoint. Finally, we examine how children might learn causal knowledge from
others and how computational and constructivist accounts of causal learning can be
integrated. © 2014 John Wiley & Sons, Ltd.
How to cite this article:
WIREs Cogn Sci 2014. doi: 10.1002/wcs.1291
INTRODUCTION
Children are remarkable causal learners. Despite
the fact that traditional cognitive development
research has suggested that young children are
‘precausal’,1contemporary accounts of cognitive
development have demonstrated that young children
have sophisticated domain-specic causal reasoning
abilities. Infants register particular aspects of physical
causality.2–5Toddlers recognize various causal rela-
tions in the psychological domain, especially about
others’ desires and intentions.6Preschoolers under-
stand that biological and psychological events can
rely on nonobvious, hidden causal relations.7,8
More generally, preschoolers also display a
variety of domain-general causal reasoning abil-
ities. Young children recognize the importance
of Hume’s principles—temporal priority, spatial
priority, and contingency—in making judgments
about causal relations.9,10 Preschoolers also possess
predictive,11 explanatory,12,13 and counterfactual
reasoning abilities.14,15 But what explains the process
of causal learning? How do children take the specic
and often sparse data they observe and construct
abstract representations of causal knowledge?
∗Correspondence to: Dave_Sobel@brown.edu
1Department of Cognitive, Linguistic, and Psychological Sciences,
Brown University, Providence, RI, USA
2Department of Psychology, The University of Texas at Austin,
Austin, TX, USA
Conict of interest: The authors have declared no conicts of
interest for this article.
We will attempt to answer this question in three
ways. The rst will be to summarize a set of general
theories of causal learning, which have been developed
in adult cognitive psychology. This section will con-
clude with a discussion of the application of these the-
ories to developmental science. The second will be to
consider a particular set of issues concerning the rela-
tion between the explanations children hear and gen-
erate and their exploration of the environment. This
section will conclude with a set of discussion ques-
tions attempting to integrate how children explore and
explain the environment with descriptions of the way
they may engage in causal learning. Third, we will
examine how children learn causal knowledge from
others, particularly focused on the idea that such a
learning process might be best explained by general
theories of causal learning.
THEORIES OF CAUSAL LEARNING
AND THEIR APPLICATION TO
COGNITIVE DEVELOPMENT
To describe how children engage in causal inference,
we must consider both the existing domain-specic
knowledge they possess about how causal rela-
tions work (called ‘substantive principles’ for causal
learning, see Ref 11) and the more domain-general
mechanisms by which children acquire new causal
knowledge from information in the environment
(‘formal principles’ for causal learning, see Ref 11).
Children’s substantive principles take the form of
© 2014 John Wiley & Sons, Ltd.
Advanced Review wires.wiley.com/cogsci
content information that applies to the causal infor-
mation they are learning. Such knowledge could also
be quite general regarding the inferences it licenses
(e.g., a broad piece of substantive knowledge is that
children understand is temporal priority—that causes
precede their effects). Other pieces of substantive
knowledge might specify only certain kind of causal
inferences (e.g., children might know that plants grow
without knowing a more general conceptual struc-
ture about plants, such as that they are alive). This
content knowledge is usually acquired through the
accretion of facts and information, and could differ
across domains and situations (i.e., temporal priority
might be more general than inferences about plant
behavior). Critically, substantive knowledge often
allows children to make better causal inferences when
mechanisms are familiar to them (see e.g., Ref 16).
Describing how children learn causal knowl-
edge (including the substantive principles they pos-
sess for any domain) usually involves describing the
formal principles, or domain-general mechanisms for
causal learning. In the next section, we outline sev-
eral descriptions of such formal mechanisms for causal
learning. Critically, no formal learning mechanism is
independent of the existing knowledge (i.e., the sub-
stantive principles) that the child possesses. Our goal
is to review how such formal mechanisms might inter-
act with children’s existing knowledge to formulate a
description of causal learning.
Theories Based on Associative Strength
and Parameter Estimation
One possible way children engage in causal inference is
to simply associate causes and effects, in the same way
that animals associate conditioned and unconditioned
stimuli in classical conditioning.17,18 On this view,
there is nothing to understanding causality beyond
recognizing associations. These descriptions of causal
learning assume that candidate causes and effects
have been identied (typically based on relatively
low-level substantive information such as temporal
priority), and output the strength of each cause–effect
association.
As such models only output associative informa-
tion, it is hard to determine how they make predictions
about the ways in which people use causal knowledge
or act on the world. Thus, contemporary advocates of
this approach have suggested that causal learning and
inference takes place by translating associative infor-
mation into a measure of causal structure. That mea-
sure of causal strength might then be combined with
substantive information to make causal inferences or
generate new interventions.19,20 For instance, build-
ing on the model described in Ref,18 there are various
accounts of causal learning that calculate the causal
strength of known causal relations from associative
information.21–23 More recently, similar models of
semantic cognition (including causality) have been
proposed based on neural network architectures.24,25
A related approach to causal learning relies on
estimating causal parameters based on the frequency
with which events co-occur. Two prominent proposals
in this category are the ΔPmodel26–28 and the Power
PC model29 and extensions for interactive causality
(i.e., when two causes must combine to produce an
effect, as opposed to being just additive, see Ref 30).
These models calculate an estimate of the maximal
likelihood value of the strength of a presumed causal
relation given a set of data.31,32aHow these presumed
causal relations are determined is typically a function
of prior substantive knowledge (e.g., ‘focal sets’,
see Ref 34).
Are such mechanisms plausible as developmen-
tal accounts? Associative learning mechanisms are
available to children at very early ages in the form
of their statistical learning capacities.35,36 Such sta-
tistical learning is implicated in infants’ processing
of causal data,37 social knowledge,38 and linguistic
information.39,40
Statistical learning capacities, in turn, are related
to infants’ ability to generalize. Eight-month-olds reg-
ister the appropriate generalizations from population
to sample and from samples to populations.41 Such
statistical learning mechanisms can lead to broader
inferences. For instance, nonrandom sampling leads
18–24-month-old infants to infer that the individual
has a subjective preference for those objects.42,43 Sim-
ilarly, by age 2, children can use the regularity in other
people’s choices to generalize their preferences to cat-
egories of objects.44 Although it is possible that causal
knowledge can be learned from registering associa-
tive or frequency information among events, it is
not clear how such statistical knowledge is integrated
with children’s existing substantive knowledge. For
instance, by their rst birthday, infants are clearly inte-
grating their existing physical knowledge45 or social
knowledge46,47 into their statistical inferences. Sim-
ilarly, children’s causal inferences are inuenced by
their mechanistic understanding of the domain48,49
and their familiarity with that domain.16
Theories Based on Learning Causal
Structure
Children’s ability tointegrate their existing substantive
knowledge with formal mechanisms for causal learn-
ing have led many researchers to propose that children
are learning an abstract ‘causal model’31 or ‘causal
© 2014 John Wiley & Sons, Ltd.
WIREs Cognitive Science Children’s causal learning
map’.50bCausal graphical models (CGMs)31,44,50,51
can dene causal relations over a variety of domains,
such as physics, biology, and psychology, but they
should not be taken as a domain-general represen-
tation of knowledge. An individual causal model is
domain-specic (i.e., an individual model could rep-
resent a particular piece of substantive knowledge).
The framework itself is more general (i.e., it can rep-
resent knowledge across domains). These models have
been posited as a computational description of chil-
dren’s naiuml;ve theories.50,52cBecause of the poten-
tial importance of these models, below we briey
introduce and summarize research using this frame-
work as a description of causal inference and learning.
To begin, a graphical model is a representation
of a joint probability distribution—a list of all possi-
ble combinations of events under consideration and
the probability that each combination occurs. Condi-
tional probability information can be extracted from
this list. In this formalism, events or objects are rep-
resented as nodes, and vertices represent particular
types of dependencies between such objects or events.
Interpreting these models as representations of causal
knowledge involves making three assumptions about
the underlying structure of the connections between
nodes (events or properties/features of objects) and
vertices (dependency relations that indicate causal
structure): Mechanisms,theMarkov Assumption,and
Faithfulness.
Assumption 1: Mechanisms
The rst assumption is that any vertex represents a
causal relation between the two nodes, specically in
the form of a mechanism that can be either observed
or unobserved. That is, given a particular relation
X→Y, the arrow indicates that there is some mech-
anism whereby changing the probability of Xdirectly
affects the probability that Ywill occur. A causal graph
is consistent with an innite set of probabilistic mod-
els that specify how the variables are related. A sin-
gle representation of that causal structure is made by
parameterizing the graph: dening the probability dis-
tribution for each variable conditioned on its parents.
Critically, a graph’s parameterization reects assump-
tions about the nature of the mechanism by which
causes produce effects. For instance, consider the
hypothetical model of the weather shown in Figure 1.
When we draw an arrow from ‘rains yesterday’ to
‘rains today’, we are positing that there is some (hid-
den) mechanism that relates the probability of raining
yesterday to the probability that it rains today. This
mechanism might be remarkably simple or complex,
but critically, we reason as if such a mechanism is
Raining
Yesterday
Raining
Today
Raining
Tomorrow
FIGURE 1|A toy model of the weather.
present (and regardless of the complexity of the mech-
anism, such inferences might be available even to very
young children, see Ref 53). Positing the presence of
such causal mechanisms allows for a ‘calculus of inter-
vention’ (see Ref 44, p. 85, see also Ref 54)—a way of
interpreting how an intervention on one part of the
causal system affects the system as a whole.
Intuitions about causal mechanisms, and by
extension, the ‘calculus of intervention’ have been
examined in several psychological studies. Some
researchers have suggested that causal transmission
was inherent in particular perceptual features of a
display.55 Given those features, one could not help to
see particular sequences of events as causal. Moreover,
adults reason about causal relations by virtue of the
‘do’ operator is described in Ref 44 (see also Ref 56);
causal learning from interventions is superior to just
observing the same data.57–59
Children engage in similar inferences. There
have been various replications and extensions of the
Michotte paradigm, suggesting that from very early
ages, infants register certain congurations of percep-
tual features as causal.3,60 Shultz argued that children
understood causal relations in terms of ‘generative
transmission’10 and demonstrated that preschoolers
treated mechanism information as more important for
judging whether a causal relation was present than
correlational information (see also Ref 61). Regarding
interventions, children reason differently about inter-
vening on a causal system than when simply observing
that system.50 Four-year-olds also infer the presence of
hidden causes when shown stochastic data, suggesting
that they interpret probabilistic events as indicating
the presence of hidden mechanisms.62 Although it is
not clear that children (or adults for that matter, see
e.g., Ref 63) can articulate the mechanism underlying
any particular causal relation, it does seem clear that
young children can reason as if such mechanisms are
present.49,64
Assumption 2: The Markov Assumption
The Markov assumption translates conditional proba-
bility information into causal knowledge. It states that
the value of an event (i.e., a node in the graph) is inde-
pendent of all other events except its children (i.e., its
direct effects) conditional on its parents (i.e., its direct
causes). Let’s return to the model of the weather shown
in Figure 1. According to this model, whether it rained
© 2014 John Wiley & Sons, Ltd.
Advanced Review wires.wiley.com/cogsci
FIGURE 2|Various reactions from a child interacting with two ‘Blicket Detectors’ (originally used by Gopnik and Sobel65), from the protocol
described in Legare.66
yesterday and whether it rains today are statistically
related, as is whether it rains today and whether it
will rain tomorrow. Given those dependencies, it is
also true that raining yesterday and raining tomor-
row are dependent. The Markov Assumption states
that raining yesterday and tomorrow are independent
given the knowledge of whether it rains today. The
only inuence raining yesterday has on raining tomor-
row is through whether it rains today.
How could we examine whether children are
reasoning about the relations among events using the
Markov assumption? One difculty in answering this
question is that we need a method that allows us to
test whether children recognize the conditional inde-
pendence relations among events separately from their
prior knowledge about how these events are related.
One such experimental paradigm was developed by
Gopnik and Sobel,65 who introduced children to a
‘blicket detector’ (see Figure 2), a machine that lights
up and plays music when certain objects are placed
upon it. The detector presents a novel, nonobvious
property of each object: its activation potential. (The
machine is controlled through an ‘enabling’ switch.
When the switch is on, any object will activate the
detector. When it is off, no object will activate the
detector). Because the machine is novel, children
have few expectations about what kinds of objects
activate it.
Using this paradigm, researchers have found that
children treat objects that activate the detector by
themselves differently from objects that only activate
the detector dependent on the presence of another
object—that is they examined whether children obey
the Markov assumption. Three- and 4-year-olds were
trained to know that objects that activated the detector
were called ‘blickets’. Then, children observed a set of
trials in which objects either independently activated
the machine, or did so only dependent on the presence
of another object. Specically, on the one cause trials,
children were shown two objects. One object (A)acti-
vated the detector by itself. The other object (B)did
not. Children then saw objects Aand Bactivate the
detector together (twice). Children labeled only object
Aas a blicket. Even though object Bactivated the
detector 2 out of the 3 times it was placed on it, it only
did so dependent on the presence of object A. If chil-
dren reasoned according to the Markov assumption,
they would not use the positive association between
object Band the machine’s activation to infer the ef-
cacy of object B, but rather recognize that such efcacy
is conditionally dependent on the presence of object
A. Remove object Afrom the equation, and object B
lacks efcacy. Children reasoned in this manner and
stated that object Bwas not a blicket. In contrast,
in an analogous two cause condition, in which an
object activated the detector independently two out of
three times, children were likely to label the Bobject a
blicket. Here, object Bactivates the machine indepen-
dent from object A, and thus children should use the
associations they observe.11
Various investigations have extended these
ndings to younger children,37 other domains of
© 2014 John Wiley & Sons, Ltd.
WIREs Cognitive Science Children’s causal learning
knowledge,67 and other kinds of screening-off infer-
ences (i.e., those involving chains or common causes,
instead of the common effect structure presented
above, see Ref 68). These data all suggest that chil-
dren robustly adhere to the Markov assumption in
their causal inferences. That said, there are a variety
of ndings in adult cognition suggesting noninde-
pendence—that adults’ probabilistic inferences are
inconsistent with the Markov assumption.69–72 As
an example, suppose you know that smoking causes
thick blood vessels and smoking increases risks
for cataracts. The Markov assumption states that
the probability that a heavy smoker has a risk for
cataracts is the same regardless of whether he/she
has thick or normal blood vessels. However, if adult
participants are given examples like this, the nonin-
dependence effect is their likelihood to judge that the
probability a heavy smoker has risk for cataracts is
higher if that smoker also has thick blood vessels than
normal blood vessels. Because of these results, some
of the researchers cited above have suggested that this
framework does not describe adults’ causal inference
well (or at the least provides an incomplete account
of causal inference, see also Refs 73 and 74).
One potential explanation for this inconsistency
between the adult and developmental ndings is that
the methods used to test adults’ causal reasoning
often contextualize the problem in such a way that
adults’ prior knowledge (i.e., their substantive princi-
ples) might inuence the causal model they construct.
Rehder and Burnet71 were sensitive to this issue, and
argued that adults do reason according to the Markov
assumption, but the causal model that they build
when they are asked to make such inferences incor-
porates prior knowledge in the form of mechanism
information. That is, in the smoking example above,
adult participants assume a mechanism through which
smoking causes both increased risk for cataracts and
thick blood vessels. In this way, nonindependence is
nothing more than adults reasoning according to the
Markov assumption, just not representing the simplest
possible causal model.
Some have expanded on this hypothesis, sug-
gesting that in order to describe children’s and
adults’ causal reasoning, one underlying principle
of the CGM framework should be modied, that is
minimality. Under minimality, whenever observed
data posit a causal relation, the standard instantiation
of that relation is a single vertex linking cause to
effect.44 However, one could also specify a distribu-
tion of intermediary causal structures between cause
and effect, with each structure’s prior probability
dependent on its complexity. This modication (called
Edge Replacement) nicely explains the phenomenon
of nonindependence, as it posits various kinds of
mechanism knowledge that potentially inuences
causal inferences.53 In addition, it explains several
other novel aspects of children’s causal inference.64,75
Assumption 3: Faithfulness
Faithfulness species that the data a learner observes is
actually indicative of the causal structure in the world.
Put simply, the faithfulness assumption is that the data
children observe indicate the actual causal structure
of the world. To illustrate, suppose that children
observe two events Xand Y. The actual structure
of the world is that there is a generative relation
between Xand Y, such that raising the probability of
Xwould therefore raise the probability that Yoccurs.
However, it is possible that there is another event (call
it Z), which children do not observe, but that also
affects Y.EventZhas a preventative relation with
Y, such that increasing the probability of Zdecreases
the probability of Y. Suppose further that the extent
to which Xaffects Yand Zaffects Yare exactly
the same and that they always affect Yin tandem.
Thus, even though raising the probability of Xshould
raise the probability of Y,Xand Ywould be seen
as independent from one another. The faithfulness
assumption is that this sequence of coincidental events
never occurs. The causal relations among X,Y,andZ
will never work out such that Xand Zcancel each
other’s effects on Yexactly.
To our knowledge, there are no direct psycholog-
ical investigations dedicated primarily to faithfulness.
That said, because it essentially involves positing the
presence of Cartesian demons, we do feel, however,
that it is safe to assume this principle. More psycholog-
ical investigation is warranted, however, to be certain
of this assumption.
Integrating the Two Accounts
At this point in our review, we wish to speculate on a
way of integrating accounts of causal learning based
on recognizing associations among events and those
based on building CGMs. The statistical learning lit-
erature suggests that infants have associative learn-
ing capacities, potentially even from birth.35,36 This
is supported by classic work demonstrating that very
young infants can learn and remember associations
they observe among events (see Ref 76, for a review).
Thus, one possibility is that such a mechanism can
account for both the acquisition of substantive prin-
ciples of causal knowledge and the way in which chil-
dren come to recognize any kind of causal relation.
Yet for the same reason that we lecture our intro-
ductory statistics students that ‘correlation doesn’t
© 2014 John Wiley & Sons, Ltd.
Advanced Review wires.wiley.com/cogsci
equal causality’, we believe that such accounts cannot
solely describe the formal principles through which
children learn causal structure. As mentioned in the
section on the Markov Assumption above, any case
in which the dependence relation between two events
(Xand Y) switches given the presence or absence
of a third (Z) suggests a causal structure in which
a direct causal relation does not exist between them
(e.g., X→Z→Y). Children must have a mechanism for
recognizing statistical regularities among events, but
also parsing out conditional independence and depen-
dence relations, and making causal inferences based
on observed data. Thus, another possibility is that chil-
dren’s causal reasoning is explained by a formal mech-
anism that is described by the CGM framework. That
is, from birth, the way in which children learn causal
knowledge and make causal inferences is guided by
principles from the CGM framework.
One way of integrating these two approaches
comes from rening Piaget’s descriptions of the devel-
opment of causal reasoning.77,78 In the earliest stages
of the sensorimotor stage, Piaget described the infant
as only experiencing causality as a form of associa-
tion of experiences: ‘there is no causality for the child
other than his own actions; the initial universe is not
a web of causal sequences but a mere collection of
events arising in extension of the activity itself’ (see Ref
78, p. 220). As the infant learns to act on the world,
he/she might move beyond such associations to recog-
nize deeper relations among events.
Pearl’s description of CGMs supporting infer-
ences about intervention (see Ref 44, see also Ref 54)
nicely t with this description. When the child realizes
that objects themselves can have efcacy on the envi-
ronment (roughly consistent with Piaget’s substage 4
of the sensorimotor stage, or around 8 months old),
they might also begin to extract the conditional prob-
ability information inherent in associations among
events.
There are two lines of support for this hypoth-
esis. The rst comes from several investigations
generated by Sommerville and colleagues, suggesting
that infants’ emerging actions predict their causal
inferences about others’ intentional actions.79,80
Sommerville et al.,81 in particular demonstrated that
providing infants with the ability to act on the envi-
ronment changed whether they perceived others’
actions as goal-directed.
Second, there are now numerous ndings in
the infant literature suggesting that infants are capa-
ble of sophisticated causal inferences.82–85 What is
interesting about these ndings is that they all come
from infants in the second half of the rst year of
life (usually 8-month-olds or older), and rarely are
developmental differences investigated.dCohen and
colleagues60,87,88 have suggested that the perception
of simple causal relations develops between approxi-
mately ages 5–10 months. Sobel and Kirkham37 found
that children’s ability to recognize conditional inde-
pendence and dependence in statistical regularity
developed between the ages of 5 and 8 months. Similar
ndings in infants’ statistical learning capacities using
a different (but one could argue simpler) paradigm,
show analogous development between 4.5 and 6
months.89 This developmental trajectory is consistent
with the constructivist interpretation outlined above.
Two Caveats About CGMs as a Description
of Children’s Causal Knowledge
Examining how this computational framework
describes children’s causal inference engages a particu-
lar debate within the psychological and computational
literatures concerning the level of representational
breadth. Thinking about CGMs as a description
of children’s causal knowledge contrasts with asso-
ciative accounts of causal reasoning20 or similar
neural network descriptions,24 which suggest that
such a computational description of causal knowl-
edge must be more domain-general. Because causal
knowledge can cross or link domains, however, this
approach should also be treated differently from
modularity or certain starting-state nativist accounts
that propose that there are separate domain-specic
causal structures, which potentially have neural
correlates.86,90,91
Thus, one important caveat to the proceeding
discussion is that CGMs are a way of represent-
ing causal structure, not a specic commitment to
how causal knowledge is learned. Most psychologi-
cal research on learning such causal structures have
relied on algorithms based on Bayesian inference (see
e.g., Refs 49, 92 and 93). That said, other algo-
rithms do exist to describe such structure learning.50,51
Although Bayesian approaches have been instrumen-
tal in facilitating our understanding of causal learning,
they should not be taken as a model for what algo-
rithm or calculations children are computing when
faced with causal learning problems. That is, these
models provide a description of the causal learning
process and have been useful in generating psycholog-
ical theory. They should not be considered algorith-
mic or process accounts of how children are exactly
make causal inferences. For instance, it is possible that
neural network simulations could model many of the
psychological phenomena described in this paper (see
Ref 25 for example). It is also possible that algorithms
that simulate Bayesian updating will provide a good
© 2014 John Wiley & Sons, Ltd.
WIREs Cognitive Science Children’s causal learning
algorithmic level description for the same inferences
(see Refs 92 and 94 for example). We are agnostic as
to which modeling architecture will ultimately pro-
vide the best and most descriptive algorithmic-level
approach. Any computational description must suc-
cessfully account for extant data, but also make novel
predictions concerning its psychological implications,
and perhaps the most exciting work in the eld comes
from this endeavor.
Second, CGMs have the potential to describe
how children represent their causal knowledge at
multiple levels. One could imagine a CGM model
describing a particular event (e.g., Abe dials a number
on his phone, which causes Bob’s phone to ring). One
could also imagine a CGM describing this kind of
event (phones can cause each other to ring). Such a
distinction between specic theories and framework
theories6is critical to thinking about the role the
CGM framework plays in describing children’s causal
knowledge.93,95 It is likely that learning about specic
events using an algorithm from the CGM framework
is guided not only by the data that children observe,
but also the knowledge they possess regarding what
kinds of specic causal models can be built (or that
have greater aprioriprobability).
EXPLANATION AND EXPLORATION IN
CHILDREN’S CAUSAL LEARNING
Having presented a basic description of computational
models of causal learning, we now turn to several pro-
grammatic lines of constructivist research examining
the relation between explanation, exploration, and the
development of causal learning.
Explanation
The search for explanations motivates the causal
learning process.96,97 The tendency to seek and gen-
erate explanations is so pervasive and compelling that
some psychologists have posited a ‘drive to explain’.98
At very early ages, children generate appropriate
domain-specic explanations 13,15,81 and use ques-
tions effectively to elicit explanations from others as a
means of acquiring new knowledge.99–101 Given that
young children generate and seek out explanations,
how might explanations benet causal learning?
A growing body of research conrms that the
process of generating explanations, for others or
for oneself, has educational benets.102–106 This
‘self-explanation effect’ has been documented in
a variety of learning contexts, ranging from the
acquisition of scientic content knowledge107,108
to conceptual transitions in early childhood.109,110
Given the intimate relation between explanation and
conceptual representation,111–113 generalization,114
and learning,102,115 an understanding of explanation
is foundational for causal learning (see also Ref 116).
Even though there are many documented effects
of explanation,103–105,117–119 the process by which
explanations benet learning is underspecied.12,115
If explanation is a mechanism for learning, children
should benet from providing explanations for events
that afford new learning opportunities. Events that are
inconsistent with prior knowledge provide just such
an opportunity. The ability to explain such events
could aid in learning by focusing children on cur-
rent causal knowledge and provoke causal reasoning.
For instance, Legare12,66,120 examined the triggers that
motivate children to construct explanations. Legare
and her colleagues found that preschoolers generated
more explanations when faced with outcomes that
were inconsistent with their prior knowledge. More-
over, these explanations tended to refer to unobserved
causal mechanisms and internal causal properties, and
not external perceptual appearances. This provides
promising evidence that explanation provides chil-
dren the opportunity to articulate new hypotheses for
events that, at rst, disconrm their current knowl-
edge. These data are consistent with the proposal that
children’s explanations play an active role in the learn-
ing process and provide an empirical basis for inves-
tigating the mechanisms by which children’s explana-
tions function in the service of discovery.e
Despite this evidence, merely attending to incon-
sistency does not always lead to belief revision and
theory change.122–128 Explaining inconsistency may
serve as a critical mechanism for integrating and rec-
onciling discordant information with existing theories
and reduce engagement in theory-preserving strategies
like rejection and postponement. But how might the
process of explaining inconsistency generate amended
beliefs?
One possibility we endorse is that explaining
inconsistency triggers a process of hypothesis gen-
eration that encourages learners to formulate and
entertain hypotheses they would not have sponta-
neously considered otherwise. Generating hypotheses
in the service of explanation may inuence the kinds
of hypotheses formulated. Both children and adults
have strong intuitions about what makes something
a good explanation,129,130 and explanation-triggered
hypothesis generation may promote the production of
hypotheses that make for informative explanations.
In this way, there might be some deep, but as of
yet unexplored connections between explanations and
how children represent their causal knowledge (based
on the CGM framework described above). These two
© 2014 John Wiley & Sons, Ltd.
Advanced Review wires.wiley.com/cogsci
areas of causal learning are not seen as connected; as
Wellman and Liu point out, ‘causal Bayes nets seem
silent on how to characterize explanations and on
what role explanations might play in causal learning’
(Ref 119, p. 261). Although it is not clear how children
generate explanations from the way they represent
their causal knowledge, one way of interpreting the
self-explanation effect is that the child treats the act of
explaining in the same manner as observing analogous
data. That is, children who generate an explanation
might treat that explanation as data, which might
affect the existing model of causal structure they
possess (i.e., strengthening it if they believe their
explanation is good, or weakening it if they believe
their explanation is weak or uncertain, see Ref 131,
for a version of this argument).
Exploration
We have proposed that the act of explaining serves
as data for causal learning or conceptual change.
We view the child as an active participant in the
construction of causal knowledge, and not a passive
viewer of information. Casting the child in this light is
a hallmark of many constructive accounts.50,52 It sug-
gests that children also seek out data when faced with
ambiguity or uncertainty, and such exploration can
inform the ways in which children learn new causal
structure. That is, the weaker the representation of
children’s causal knowledge, the more likely they will
explore their environment (presumably to strengthen
that representation).
Building on classic research on children’s
play,132,133 there is converging evidence that
inconsistent or problematic events also trigger
exploration.134,135 Children’s exploratory play is
affected by the quality of the evidence that they
observe. When multiple candidate causes are available
for the same outcome and underlying causal struc-
ture is ambiguous, children preferentially explore
confounded (as opposed to unconfounded) causal
relations, show more variable play behavior when
presented with probabilistic (as opposed to deter-
ministic) information,136 and can spontaneously
disambiguate confounded variables.135 Moreover,
they recognize when it is necessary to explore the
environment as opposed to seek help from a more
knowledgeable source.137
Despite the evidence that anomalous or incon-
sistent information motivates both explanation and
exploration, the way in which the two processes may
jointly facilitate or drive causal knowledge acquisi-
tion has remains underdetermined. Does the process
of constructing a causal explanation for inconsistent
outcomes inform and constrain children’s exploratory
behavior? Do causal explanation and exploratory
behavior operate in tandem as hypothesis-generating
and hypothesis-testing mechanisms? Is this process
associated with tangible learning outcomes?
Bonawitz and colleagues138 examined this inter-
action between explanation and exploration. They
rst assessed children’s understanding of balance
events (i.e., do children know that center of mass indi-
cates balance). Next they presented children with a
free play environment that provided evidence either
for a geometric center-based or center of mass-based
account of balance (by, following Ref 139), using stim-
uli that appeared to support one kind of balance rela-
tion, but by virtue of their mass, actually supported
a counterintuitive relation). They then assessed how
children learned from their explanations and explo-
ration of these objects. They found that older chil-
dren (6–7-year-olds) could revise their beliefs in light
of theory-inconsistent evidence, but also that children
would discount such evidence if they explain these
events in terms of potential auxiliary hypotheses to
their existing theory. Preschoolers, in contrast, strug-
gled to revise their beliefs given this evidence.
Preschoolers, however, do trade-off exploration
and explanation in certain ways. When children lack
explanatory information, they can learn causal struc-
tures from exploration,140 and this learning is facil-
itated when their exploration uncovers new knowl-
edge as opposed to conrming information they
have already observed.66,141 Moreover, young children
explore novel toys more when given ambiguous data
about how the toy works135 and explore environments
in systematic ways to resolve that ambiguity.142
Critically, however, explanations and explo-
ration interact when learning.143 Bonawitz and
colleagues144 demonstrated that children who heard a
particular set of instructions regarding a novel object’s
function were less likely to explore the object (and
discover novel functions) than children who heard
incomplete explanations. They suggest that children’s
exploration was affected by their understanding of
the pedagogical intent of the individual who gener-
ated the explanations (i.e., someone who was more
knowledgeable about the object than they were). Such
effects of pedagogy—taking a teacher’s intentions
into account to determine why they are presenting
the information they are—are detectable in adults145
as well as toddlers146,147 and are potentially part of a
natural human communicative process.148
Given the trade-off between explanations
heard from others and exploration of the environ-
ment demonstrated in these experiments, there are
a number of compelling reasons to examine the
© 2014 John Wiley & Sons, Ltd.
WIREs Cognitive Science Children’s causal learning
interaction between explanation and exploration for
learning, particularly when faced with the problem
of inconsistency. Encouraging children to explain
inconsistency confronts children with the inconsistent
evidence most likely to foster theory revision, guides
the hypothesis-testing process, and promotes learning.
These explanatory intuitions may constrain learners
to focus on some aspects of what they are trying
to explain over others. In particular, explanation
may focus learners on causal mechanisms66,106 and
on abstraction.115,130 Generating hypotheses in the
service of explanation may inuence the kinds of
hypotheses formulated, as well as their impact on
cognition.
That said, how might the process of explaining
inconsistency with prior knowledge inform children’s
exploratory, hypothesis-testing behavior? Legare66
demonstrated that children’s explanations and subse-
quent exploratory behavior following events that are
consistent with their existing knowledge differ from
those following inconsistent events. When children
observed inconsistent events, the kind of explanation
children provided differentially predicted the kind of
exploratory behavior they engaged in. The kind of
explanations children provided also inuence rates of
spontaneous, hypothesis-testing exploratory behavior
and the tendency to modify existing explanatory
hypotheses in the face of disconrming evidence. For
example, children who provided explanations that
referred to problems with causal function engaged
in more extended and more variable exploratory
behavior than children who provided different kinds
of explanations (e.g., explanations referring to cate-
gory membership). Encouraging children to explain
inconsistency confronts children with the inconsis-
tent evidence most likely to foster theory revision,
guides the hypothesis-testing process, and promotes
exploration.143
CAUSAL LEARNING FROM OTHERS
Much of the constructivist research we have described
so far focuses on the problem of causal learning as
being directed by the child. When children encounter
novel data or events from which they learn, they do
so by integrating that information with their existing
knowledge to make novel inferences, generate novel
explanations, or engage in specic actions. Such a
description might be correct, but it assumes that all the
data children use to learn causal structure is directly
observable. This is obviously not the case, and the
fact that children learn from explanations (both their
own and others) provides evidence for the power of
learning through testimony from others.
Furthermore, children make inferences about
unobservable biological events,149 psychological
events,119 and even supernatural events.150 Children
also appeal to and easily learn culturally constructed
explanations151 and social conventions.152,153 All
of these events are not directly observable and
could not be learned just from interacting with the
environment.fIn order to learn all of this informa-
tion, children must rely on information generated
from others. How do children learn causal knowledge
from others? Do the same processes we have described
in learning from observation and interaction with the
world apply to learning from others?
A signicant number of studies now show that
children are not simply credulous of others’ infor-
mation (for reviews, see Refs 154–158). Children
as young as 2 years are capable of judiciously using
different sources by tracking informants’ history of
past accuracy.159–162 Children’s rapid cultural learning
potentially emanates from their ability to learn selec-
tively from others (see e.g., Ref 163).
We propose that how children learn from oth-
ers is as rational as how children learn from inte-
grating their own knowledge with observed data.158
Preschoolers’ beliefs about whom to trust are inu-
enced by their existing knowledge about people (e.g.,
adults are more knowledgeable than children164),
kinds (e.g., speakers knowledgeable about objects’
labels should also be knowledgeable about those
objects’ functions, 161), and expertise (e.g., speak-
ers with a certain specic knowledge base might
not be more knowledgeable overall, just about that
base165,166). These ndings all suggest that young chil-
dren can integrate what they know about the world
with the data they receive about the world from others.
Indeed, consistent with this rational account,
some have suggested different computational
accounts of the way in which children update their
beliefs—including their causal knowledge—given
information generated from others.167–169 Many
of these accounts use CGMs as a representation of
children’s existing causal knowledge, and promote
different kinds of rational learning algorithms on this
representation to explain how children’s knowledge
changes. As with our discussion of the CGM frame-
work, such computational accounts should not be
considered process models, but rather descriptions of
how children might learn from others, which in turn
could inform new psychological investigations.
CONCLUSION
To answer the question of how children construct
abstract representations of causal knowledge from the
© 2014 John Wiley & Sons, Ltd.
Advanced Review wires.wiley.com/cogsci
data they observe, we have appealed to computational
(i.e., CGMs), constructivist, and social learning frame-
works to describe the process of causal learning. Our
objective is to illustrate the striking sophistication of
young children’s causal learning capacities, as well
as demonstrating how useful computational modeling
can be for making predictions about those capacities.
There are many outstanding open questions, such as
how to translate between a causal graphical model and
a verbal explanation provided by a child, or the role
of children’s developing cognitive capacities, such as
attention or memory, in the process of such learning.
The present review, however, suggests an important
conclusion, that very much emanates from construc-
tivist theories of cognitive development: The child is an
active seeker of information—regardless of what kind
of knowledge they are acquiring. Regarding causal
knowledge, children begin to generate ‘why’ ques-
tions around the time they themselves offer causal
explanations.101,119,170 But, similar behavior is also
seen for children in the naming spurt—the acquisition
of many labels often coincides with children generat-
ing a lexical item soliciting an object’s name.171 Such
active learning allows children to recognize and ll the
gaps in their knowledge and construct new represen-
tations of the causal structure of the world.
NOTES
aDescriptions of causal learning based on calculating
causal strength from associative information typically
estimate these asymptotic parameters. For example,
the results of the Rescorla-Wagner equation can con-
verge to ΔPgiven innitely many randomly inter-
mixed trials.33
bWe prefer the ‘causal map’ designation, as it does not
suggest that children’s causal knowledge is represented
by a graphical model explicitly. Instead, it suggests
that the features of such a computational account
describe that representation.
cInterestingly, there are several domains of knowl-
edge like number, language, or spatial relations in
which causal maps potentially have little applicabil-
ity because those domains typically do not represent
knowledge in terms of causal relations. For example,
the development of numerical knowledge (such as
knowledge of the mapping between the natural num-
bers and numerals) is not represented in terms of
causal relations. How the CGM framework poten-
tially relates to the representation of this knowledge
is beyond the scope of this article.
dTwo points should be made about these ndings.
First, they are usually interpreted as evidence for
causality being part of ‘core knowledge’ or causal rea-
soning as an innately specied capacity (see e.g., Ref
86). Second, to our knowledge, all of these ndings are
consistent with the CGM framework described above,
although they are usually not described in that way.
eIt is also interesting to consider cases in which chil-
dren learn from explaining information they already
know, particularly when teaching another person.
There are many demonstrations of preschoolers
being able to make sophisticated inferences about
teaching,121 but only few demonstrations that chil-
dren’s actual teaching affects their learning. We believe
that this is an important line of research for future
investigation.
fCritically, children also learn language, which, while
not causal in nature, does involve understanding a
set of culturally constructed, but arbitrary mappings
between phonology and meaning.154
ACKNOWLEDGMENTS
The authors were supported by NSF (DLS/REESE-1223777 to D.M.S.) and by an ESRC (Economic and Social
Research Council) Large Grant (REF RES-060-25-0085 to C.H.L.) during the writing of this article.
REFERENCES
1. Piaget J. The Child’s Conception of the World. Lon-
don: Routledge and Kegan Paul; 1929.
2. Hespos SJ, Baillargeon R. Reasoning about contain-
ment events in very young infants. Cognition 2001,
78:207–245.
3. Leslie AM, Keeble S. Do six-month-old infants per-
ceive causality? Cognition 1987, 25:265–288.
4. Needham A, Baillargeon R. Intuitions about support in
4.5-month-old infants. Cognition 1993, 47:121–148.
5. Spelke ES, Breinlinger K, Macomber J, Jacobson
K. Origins of knowledge. Psychol Rev 1992, 99:
605–632.
6. Wellman HM. The Child’s Theory of Mind.Cam-
bridge, MA: MIT Press; 1990.
© 2014 John Wiley & Sons, Ltd.
WIREs Cognitive Science Children’s causal learning
7. Gelman SA, Wellman HM. Insides and essences: early
understandings of the non-obvious. Cognition 1991,
38:213–244.
8. Gopnik A, Wellman HM. The theory theory. In:
Hirscheld L, Gelman S, eds. Mapping the Mind:
Domain Specicity in Cognition and Culture.New
York: Cambridge University Press; 1994, 257–293.
9. Bullock M, Gelman R, Baillargeon R. The develop-
ment of causal reasoning. In: Friedman WJ, ed. The
Developmental Psychology of Time. New York: Aca-
demic Press; 1982, 209–254.
10. Shultz TR. Rules of causal attribution. Monogr Soc
Res Child Dev 1982, 47:1–51. Serial No. 194.
11. Gopnik A, Sobel DM, Schulz L, Glymour C. Causal
learning mechanisms in very young children: two,
three, and four-year-olds infer causal relations from
patterns of variation and co-variation. Dev Psychol
2001, 37:620–629.
12. Legare CH, Gelman SA, Wellman HM. Inconsis-
tency with prior knowledge triggers children’s causal
explanatory reasoning. Child Dev 2010, 81:929–944.
13. Schult CA, Wellman HM. Explaining human move-
ments and actions: children’s understanding of the
limits of psychological explanation. Cognition 1997,
62:291–324.
14. Harris PL, German T, Mills P. Children’s use of
counterfactual thinking in causal reasoning. Cognition
1996, 61:233–259.
15. Sobel DM. Exploring the coherence of young chil-
dren’s explanatory abilities: evidence from generating
counterfactuals. Br J Dev Psychol 2004, 22:37–58.
16. Berzonsky MD. The role of familiarity in children’s
explanations of physical causality. Child Dev 1971,
42:705–715.
17. Mackintosh NJ. A theory of attention: variations in
the associability of stimuli with reinforcement. Psychol
Rev 1975, 82:276–298.
18. Rescorla RA, Wagner AR. A theory of Pavlovian con-
ditioning: variations in the effectiveness of reinforce-
ment and nonreinforcement. In: Black AH, Prokasy
WF, eds. Classical Conditioning II: Current Theory
and Research. New York: Appleton-Century-Crofts;
1972, 64–99.
19. Cramer RE, Weiss RF, William R, Reid S, Nieri
L, Manning-Ryan B. Human agency and associative
learning: Pavlovian principles govern social process in
causal relationship detection. Q J Exp Psychol 2002,
55:241–266.
20. Dickinson A, Shanks D. Instrumental action and
causal representation. In: Sperber D, Premack D,
Premack AJ, eds. Causal Cognition: A Multidisci-
plinary Debate. New York: Clarendon Press/Oxford
University Press; 1995, 5–25.
21. Kruschke JK, Blair NJ. Blocking and backward block-
ing involve learned inattention. Psychon Bull Rev
2000, 7:636–645.
22. Van Hamme LJ, Wasserman EA. Cue competition
in causality judgments: the role of nonpresentation
of compound stimulus elements. Learn Motiv 1994,
25:127–151.
23. Wasserman EA, Berglan LR. Backward blocking and
recovery from overshadowing in human causal judge-
ment: the role of within-compound associations. QJ
Exp Psychol 1998, 51:121–138.
24. McClelland JL, Thompson RM. Using domain-general
principles to explain children’s causal reasoning abili-
ties. Dev Sci 2007, 10:333–356.
25. Rogers TT, McClelland JL. Semantic Cognition: A
Parallel Distributed Processing Approach. Cambridge,
MA: MIT press; 2004.
26. Allan LG. A note on measurement of contingency
between two binary variables in judgment tasks. Bull
Psychon Soc 1980, 15:147–149.
27. Jenkins HM, Ward WC. Judgment of contingency
between responses and outcomes. Psychol Monogr
1965, 79:1.
28. Shanks DR. Is human learning rational? QJExp
Psychol 1995, 48:257–279.
29. Cheng PW. From covariation to causation: a causal
power theory. Psychol Rev 1997, 104:367–405.
30. Novick LR, Cheng PW. Assessing interactive causal
inuence. Psychol Rev 2004, 111:455–485.
31. Glymour C. The Mind’s Arrows: Bayes Nets and
Graphical Causal Models in Psychology. Cambridge,
MA: The MIT Press; 2001.
32. Tenenbaum JB, Grifths TL. Structure learning in
human causal induction.In:Proceedings of the 13th
Annual Conference on the Advances in Neural Infor-
mation Processing Systems, Vancouver, Canada, 2001.
33. Danks D. Equilibria of the Rescorla-Wagner model. J
Math Psychol 2003, 47:109–121.
34. Cheng PW, Novick LR. A probabilistic contrast
model of causal induction. J Pers Soc Psychol 1990,
58:545–567.
35. Haith MM. Future-oriented processes in infancy: the
case of visual expectations. In: Granrud C, ed. Visual
Perception and Cognition in Infancy. Hillsdale, NJ:
Erlbaum; 1993, 235–264.
36. Kirkham NZ, Slemmer JA, Johnson SP. Visual statisti-
cal learning in infancy. Cognition 2002, 83:B35–B42.
37. Sobel DM, Kirkham NZ. Bayes nets and blickets:
infants developing representations of causal knowl-
edge. Dev Sci 2007, 10:298–306.
38. Johnson S, Slaughter V, Carey S. Whose gaze will
infants follow? The elicitation of gaze-following in
12-month-olds. Dev Sci 1998, 1:233–238.
39. Goldstein MH, Schwade JA, Bornstein MH. The
value of vocalizing: ve-month-old infants associate
their own noncry vocalizations with responses from
caregivers. Child Dev 2009, 80:636–644.
© 2014 John Wiley & Sons, Ltd.
Advanced Review wires.wiley.com/cogsci
40. Saffran JR, Aslin RN, Newport EL. Statistical
learning by 8-month-old infants. Science 1996,
274:1926–1928.
41. Xu F, Garcia V. Intuitive statistics by 8-month-old
infants. Proc Natl Acad Sci USA 2008,
105:5012–5015.
42. Kushnir T, Xu F, Wellman HM. Young children use
statistical sampling to infer the preferences of others.
Psychol Sci 2010, 21:1134–1140.
43. Ma L, Xu F. Preverbal infants infer rational agents
from the perception of regularity. Dev Psychol 2013,
49:1330–1337.
44. Pearl J. Causality: Models, Reasoning, and Inference.
New York: Cambridge University Press; 2000.
45. Denison S, Xu F. Integrating physical constraints in
statistical inference by 11-month-old infants. Cogn Sci
2010, 34:885–908.
46. Wu R, Gopnik A, Richardson DC, Kirkham NZ.
Infants learn about objects from statistics and people.
Dev Psychol 2011, 47:1220–1229.
47. Tummeltshammer KS, Kirkham NZ. Learning to look:
probabilistic variation and noise guide infants’ eye
movements. Dev Sci 2013, 16:760–771.
48. Schulz LE, Bonawitz EB, Grifths T. Can being scared
make your tummyache? Naiuml;ve theories, ambigu-
ous evidence and preschoolers’ causal inferences. Dev
Psychol 2007, 43:1124–1139.
49. Sobel DM, Munro SA. Domain generality and speci-
city in children’s causal inferences about ambiguous
data. Dev Psychol 2009, 45:511–524.
50. Gopnik A, Glymour C, Sobel DM, Schulz LE, Kush-
nir T, Danks D. A theory of causal learning in chil-
dren: causal maps and Bayes nets. Psychol Rev 2004,
111:3–32.
51. Spirtes P, Glymour C, Scheines R. Causation, Predic-
tion, and Search (Springer Lecture Notes in Statistics.
2ndrevised ed. Cambridge, MA: MIT Press; 2001.
52. Gopnik A, Wellman HM. Reconstructing con-
structivism: causal models, Bayesian learning
mechanisms, and the theory theory. Psychol Bull
2012, 108:1085–1108.
53. Buchanan DW, Tenenbaum JB, Sobel DM. Edge
replacement and nonindependence in causation. In:
Proceedings of the 32nd Annual Meeting of the Cog-
nitive Science Society, Portland, OR, 2010.
54. Woodward J. Making Things Happen.NewYork:
Oxford; 2003.
55. Michotte A. Causalite, Permanence Et Realite Phe-
nomenales. Oxford: Publications Universitaires; 1962.
56. Sloman SA, Lagnado D. Do we “do”? Cognit Sci 2005,
29:5–39.
57. Lagnado DA, Sloman S. The advantage of timely
intervention. J Exp Psychol Learn Mem Cogn 2004,
30:856–876.
58. Steyvers M, Tenenbaum JB, Wagenmakers EJ, Blum
B. Inferring causal networks from observations and
interventions. Cognit Sci 2003, 27:453–489.
59. Waldmann MR, Hagmayer Y. Seeing versus doing:
two modes of accessing causal knowledge. JExp
Psychol Learn Mem Cogn 2005, 31:216–227.
60. Oakes LM, Cohen LB. Infant perception of a causal
event. Cognit Dev 1990, 5:193–207.
61. Madole KL, Cohen LB. The role of object parts in
infants’ attention to form-function correlations. Dev
Psychol 1995, 31:637–648.
62. Schulz LE, Sommerville J. God does not play dice:
causal determinism and children’s inferences about
unobserved causes. Child Dev 2006, 77:427–442.
63. Rozenblit L, Keil F. The misunderstood limits of folk
science: an illusion of explanatory depth. Cognit Sci
2002, 26:521–562.
64. Buchanan DW, Sobel DM. Mechanism-based causal
reasoning in young children. Child Dev 2011,
82:2053–2066.
65. Gopnik A, Sobel DM. Detecting blickets: how young
children use information about causal powers in
categorization and induction. Child Dev 2000,
71:1205–1222.
66. Legare CH. Exploring explanation: explaining incon-
sistent information guides hypothesis-testing behavior
in young children. Child Dev 2012, 83:173–185.
67. Schulz LE, Gopnik A. Causal learning across domains.
Dev Psychol 2004, 40:162–176.
68. Sobel DM, Sommerville JA. Rationales in children’s
causal learning from others’ actions. Cognit Dev 2009,
24:70–79.
69. Mayrhofer R, Hagmayer Y, Waldmann MR. Agents
and causes: a Bayesian error attribution model of
causal reasoning. In: Proceedings of the Thirty-Second
Annual Conference of the Cognitive Science Society,
Portland, OR, 2010.
70. Park J, Sloman S. Mechanistic beliefs determine adher-
ence to the Markov property in causal reasoning. Cog-
nit Psychol 2013, 67:186–216.
71. Rehder B, Burnett R. Feature inference and the
causal structure of categories. Cognit Psychol 2005,
50:264–314.
72. Walsh CR, Sloman SA. Revising causal beliefs. In:
Forbus K, Gentner D, Regier T, eds. Proceedings of
the 26th Annual Conference of the Cognitive Science
Society. Mahwah, NJ: Erlbaum; 2004, 1423–1427.
73. Bes B, Sloman S, Lucas CG, Raufaste E. Non-Bayesian
inference: causal structure trumps correlation. Cognit
Sci 2012, 36:1178–1203.
74. Walsh C, Sloman SA. Updating beliefs with causal
models: violations of screening off. In: Gluck MA,
Anderson JR, Kosslyn SM, eds. Memory and Mind:
A Festschrift for Gordon H. Bower. Mahwah, NJ:
Lawrence Erlbaum Associates; 2008, 345–357.
© 2014 John Wiley & Sons, Ltd.
WIREs Cognitive Science Children’s causal learning
75. Erb C, Buchanan DW, Sobel DM. Variability indicates
complexity in children’s inference about causal mech-
anisms. Cognition 2013, 129:494–500.
76. Rovee-Collier C. The development of infant memory.
Curr Dir Psychol Sci 1999, 8:80–85.
77. Piaget J. The Origins of Intelligence in Children.
Madison, WI: International Universities Press; 1952.
78. Piaget J. The Construction of Reality in the Child.New
York: Basic Books; 1955.
79. Sommerville JA, Woodward AL. Pulling out the inten-
tional structure of action: the relation between action
processing and action production in infancy. Cogni-
tion 2005, 95:1–30.
80. Sommerville JA, Woodward AL, Needham A. Action
experience alters 3-month-old infants’ perception of
others’ actions. Cognition 2005, 96:B1–B11.
81. Shultz T, Wells D, Sarda M. Development of the ability
to distinguish intended actions from mistakes, reexes,
and passive movements. Br J Soc Clin Psychol 1980,
19:301–310.
82. Kosugi D, Ishida H, Fujita K. 10-Month-old infants’
inference of invisible agent: distinction in causality
between object motion and human action. Jpn Psychol
Res 2003, 45:15–24.
83. Muentener P, Carey S. Infants’ causal representa-
tions of state change events. Cognit Psychol 2010,
61:63–86.
84. Saxe R, Tenenbaum JB, Carey S. Secret agents: infer-
ences about hidden causes by 10- and 12-month-old
infants. Psychol Sci 2005, 16:995–1001.
85. Teglas E, Vul E, Girotto V, Gonzalez M, Tenenbaum
JB, Bonatti LL. Pure reasoning in 12-month-old
infants as probabilistic inference. Science 2011,
332:1054–1059.
86. Carey S. The Origin of Concepts. Oxford: New York;
2009.
87. Cohen LB, Amsel G. Precursors to infants’ perception
of the causality of a simple event. Infant Behav Dev
1998, 21:713–732.
88. Cohen LB, Oakes LM. How infants perceive simple
causality. Dev Psychol 1993, 29:421–433.
89. Denison S, Reed C, Xu F. The emergence of probabilis-
tic reasoning in very young infants: evidence from 4.5-
and 6-month-olds. Dev Psychol 2013, 49:243–249.
90. Atran S. Cognitive Foundations of Natural History:
Towards an Anthropology of Science.NewYork:
Cambridge University Press; 1990.
91. Leslie AM, Roth D. What autism teaches us
about metarepresentation. In: Baron-Cohen S,
Tager-Flusberg H, Cohen D, eds. Understanding
Other Minds: Perspectives from Autism. Oxford:
Oxford University Press; 1993, 83–111.
92. Bonawitz EB, Denison S, Chen A, Gopnik A, Grifths
TL. A simple sequential algorithm for approximating
Bayesian inference. In: Proceedings of the 33rd Annual
Cognitive Science Society, Boston, MA, 2011.
93. Kemp C, Goodman ND, Tenenbaum JB. Learning to
learn causal models. Cognit Sci 2010, 34:1185–1243.
94. Ullman, T. D., Goodman, N. D., & Tenenbaum,
J. B. Theory acquisition as stochastic search. In:
Proceedings of the 32nd Annual Cognitive Science
Society, Boston, MA, 2011.
95. Grifths TL, Tenenbaum JB. Two proposals for causal
grammars. In: Gopnik A, Schulz L, eds. Causal
Learning: Psychology, Philosophy, and Computation.
Oxford: Oxford University Press; 2007.
96. Brewer WF, Chinn CA, Samarapungavan A. Expla-
nation in scientists and children. Minds Mach 1998,
8:119–136.
97. Keil F, Wilson R. Explanation and Cognition.Cam-
bridge, MA: MIT Press; 2000.
98. Gopnik A. Explanation as orgasm. Minds Mach 1998,
8:101–118.
99. Chouinard MM. Children’s questions: a mechanism
for cognitive development. Monogr Soc Res Child Dev
2007, 72:vii.
100. Legare CH, Mills CM, Souza AL, Plummer LE,
Yasskin R. The use of questions as problem-solving
strategies in early childhood. J Exp Child Psychol
2012, 114:63–76.
101. Hood L, Bloom L. What, when, and how about why:
a longitudinal study of early expressions of causality.
Monogr Soc Res Child Dev 1979, 44:1–47. Serial No.
181.
102. Chi MTH. Self-explaining expository texts: the dual
processes of generating inferences and repairing men-
tal models. In: Glaser R, ed. Advances in Instructional
Psychology. Hillsdale, NJ: Erlbaum; 2000, 161–238.
103. Crowley K, Siegler RS. Explanation and generalization
in young children’s strategy learning. Child Dev 1999,
70:304–316.
104. de Leeuw N, Chi MTH. The role of self-explanation
in conceptual change learning. In: Sinatra G, Pintrich
P, eds. Intentional Conceptual Change. Hillsdale, NJ:
Erlbaum; 2003, 55–78.
105. Lombrozo T. The structure and function of explana-
tions. Trends Cognit Sci 2006, 10:464–470.
106. Legare CH, Lombrozo T. Selective benets of expla-
nation on learning in early childhood. J Exp Child
Psychol. In press.
107. Chi MTH, Bassok M, Lewis M, Reimann P, Glaser
R. Self-explanations: how students study and use
examples in learning to solve problems. Cognit Sci
1989, 13:145–182.
108. Chi MT, De Leeuw N, Chiu MH, LaVancher C. Elicit-
ing self-explanations improves understanding. Cognit
Sci 1994, 18:439–477.
© 2014 John Wiley & Sons, Ltd.
Advanced Review wires.wiley.com/cogsci
109. Amsterlaw J, Wellman HM. Theories of mind in tran-
sition: a microgenetic study of the development of false
belief understanding. JCognDev2006, 7:139–172.
110. Siegler RS. How does change occur: a microgenetic
study of number conservation. Cogn Psychol 1995,
25:225–273.
111. Carey S. Conceptual Change in Childhood.Cam-
bridge, MA: MIT Press/Bradford Books; 1985.
112. Lombrozo T. Explanation and categorization: how
“why?” informs “what?”. Cognition 2009, 110:
248–253.
113. Murphy GL, Medin DL. The role of theories in
conceptual coherence. Psychol Rev 1985, 92:289–316.
114. Sloman SA. When explanations compete: the role of
explanatory coherence on judgments of likelihood.
Cognition 1994, 52:1–21.
115. Williams JJ, Lombrozo T. The role of explanation in
discovery and generalization: evidence from category
learning. Cognit Sci 2010, 34:776–806.
116. Koslowski B. Theory and evidence: the development
of scientic reasoning. Cambridge, MA: MIT Press;
1996.
117. Rittle-Johnson B, Saylor M, Swygert K. Learning from
explaining: does it matter if mom is listening? JExp
Child Psychol 2007, 100:215–224.
118. Roy M, Chi MTH. Self-explanation in a multi-media
context. In: Mayer R, ed. Cambridge Handbook of
Multimedia Learning. Cambridge, MA: Cambridge
University Press; 2005, 271–286.
119. Wellman HM, Liu D. Causal reasoning as informed by
the early development of explanations. In: Gopnik A,
Schulz LE, eds. Causal Learning: Psychology, Philoso-
phy, and Computation. New York: Oxford University
Press; 2007, 261–279.
120. Legare CH, Gelman SA. Examining explanatory biases
in young children’s biological reasoning. JCognDev
2014. doi:10.1080/15248372.2012.749480.
121. Ziv M, Frye D. Children’s understanding of teaching:
the role of knowledge and belief. Cogn Dev 2004,
19:457–477.
122. Bindra D, Clarke KA, Shultz TR. Understanding pre-
dictive relations of necessity and sufciency in formally
equivalent “causal” and “logical” problems. JExp
Psychol Gen 1980, 109:422–443.
123. Chinn CA, Brewer WF. The role of anomalous data in
knowledge acquisition: a theoretical framework and
implications for science instruction. Rev Educ Res
1993, 63:1–49.
124. Klahr D, Dunbar K. Dual space search during scientic
reasoning. Cognit Sci 1988, 12:1–48.
125. Fay AL, Klahr D. Knowing about guessing and guess-
ing about knowing: preschoolers’ understanding of
indeterminacy. Child Dev 1996, 67:689–716.
126. Kuhn D. Children and adults as intuitive scientists.
Psychol Rev 1989, 96:674–689.
127. Vosniadou S, Brewer WF. Mental models of the earth:
a study of conceptual change in childhood. Cogn
Psychol 1992, 24:535–585.
128. Vosniadou S, Brewer WF. Mental models of the
day/night cycle. Cognit Sci 1994, 18:123–184.
129. Frazier BN, Gelman SA, Wellman HM. Preschoolers’
search for explanatory information within adult-child
conversation. Child Dev 2009, 80:1592–1611.
130. Lombrozo T. Simplicity and probability in causal
explanation. Cogn Psychol 2007, 55:232–257.
131. Williams JJ, Lombrozo T, Redher B. The hazards of
explanation: overgeneralization in the face of excep-
tions. J Exp Psychol Gen 2013, 142:1006–1114.
132. Hutt C, Bhavnani R. Predictions from play. Nature
1972, 237:171–172.
133. Switzky HN, Haywood HC, Isett R. Exploration,
curiosity, and play behavior in young children: the
effects of stimulus complexity. Dev Psychol 1974,
10:321–329.
134. Baldwin DA, Markman EM, Melartin RL. Infants’
ability to draw inferences about nonobvious object
properties: evidence from exploratory play. Child Dev
1993, 64:711–728.
135. Schulz LE, Bonawitz E. Serious fun: preschoolers play
more when evidence is confounded. Dev Psychol 2007,
43:1045–1050.
136. Schulz LE, Hooppell K, Jenkins A. Judicious imitation:
young children imitate deterministic actions exactly,
stochastic actions more variably. Child Dev 2008,
79:395–410.
137. Gweon H, Schulz LE. 16-Month-olds rationally infer
causes of failed actions. Science 2011, 332:1524.
138. Bonawitz EB, van Schijndel T, Friel D, Schulz LE.
Balancing theories and evidence in children’s explo-
ration, explanations, and learning. Cognit Psychol
2012, 64:215–234.
139. Karmiloff-Smith A, Inhelder B. If you want to get
ahead, get a theory. Cognition 1974, 3:195–212.
140. Schulz LE, Gopnik A, Glymour C. Preschool children
learn about causal structure from conditional interven-
tions. Dev Sci 2007, 10:322–332.
141. Sobel DM, Sommerville JA. The importance of dis-
covery in children’s causal learning from interventions.
Front Dev Psychol 2010, 1:176–183.
142. Cook C, Goodman ND, Schulz LE. Where science
starts: spontaneous experiments in preschoolers’
exploratory play. Cognition 2011, 120:341–349.
143. Legare CH. The contributions of explanation and
exploration to children’s scientic reasoning. Child
Dev Perspect. In press.
144. Bonawitz EB, Shafto P, Gweon H, Goodman N, Spelke
E, Schulz LE. The double edged sword of pedagogy:
teaching limits children’s spontaneous exploration and
discovery. Cognition 2011, 120:322–330.
© 2014 John Wiley & Sons, Ltd.
WIREs Cognitive Science Children’s causal learning
145. Shafto P, Goodman ND. Teaching games: statistical
sampling assumptions for learning in pedagogical sit-
uations. In: Proceedings of the 30th Annual Confer-
ence of the Cognitive Science Society, Washington, DC,
2008.
146. Gweon H, Tenenbaum JB, Schulz LE. Infants consider
both the sample and the sampling process in induc-
tive generalization. Proc Natl Acad Sci USA 2010,
107:9066.
147. Yang DJ, Bushnell EW, Buchanan DW, Sobel DM.
Infants’ use of contextual cues in imitation and gener-
alization of effective causal actions. J Exp Child Psy-
chol 2013, 116:510–531.
148. Csibra G, Gergely G. Natural pedagogy. Trends Cogn
Sci 2009, 13:148–153.
149. Kalish C. Causes and symptoms in preschoolers’ con-
ceptions of illness. Child Dev 1996, 67:1647–1670.
150. Woolley JD. Thinking about fantasy: are children
fundamentally different thinkers and believers from
adults? Child Dev 1997, 68:991–1011.
151. Legare CH, Evans EM, Rosengren KS, Harris PL. The
coexistence of natural and supernatural explanations
across cultures and development. Child Dev 2012,
83:779–793.
152. Herrmann PA, Legare CH, Harris PL, Whitehouse
H. Stick to the script: The effect of witnessing mul-
tiple actors on children’s imitation. Cognition 2013,
129:536–543.
153. Rakoczy H, Warneken F, Tomasello M. The sources
of normativity: young children’s awareness of the
normative structure of games. Dev Psychol 2008,
44:875–881.
154. Bloom P. How Children Learn the Meaning of Words.
Cambridge, MA: MIT Press; 2000.
155. Harris PL. Trusting What You’re Told: How Children
Learn from Others. Cambridge, MA: Harvard Univer-
sity Press; 2012.
156. Heyman GD, Legare CH. Social cognitive develop-
ment: learning from others. In: Carlston DE, ed. The
Oxford Handbook of Social Cognition.NewYork:
Oxford University Press; 2013, 749–766.
157. Mills CM. Knowing when to doubt: developing a
critical stance when learning from others. Dev Psychol
2013, 49:404–418.
158. Sobel DM, Kushnir T. Knowledge matters: how chil-
dren evaluate the reliability of testimony as a process
of rational inference. Psychol Rev 2013, 120:779–797.
159. Clement F, Koenig M, Harris P. The ontogenesis of
trust. Mind Lang 2004, 19:360–379.
160. Koenig MA, Clement F, Harris PL. Trust in testimony:
children’s use of true and false statements. Psychol Sci
2004, 15:694–698.
161. Koenig MA, Harris PL. Preschoolers mistrust igno-
rant and inaccurate speakers. Child Dev 2005,
76:1261–1277.
162. Koenig MA, Woodward AL. Sensitivity of
24-month-olds to the prior inaccuracy of the source:
possible mechanisms. Dev Psychol 2010, 46:815–826.
163. Harris PL, Koenig MA. Trust in testimony: how
children learn about science and religion. Child Dev
2006, 77:505–524.
164. Jaswal VK, Neely LA. Adults don’t always know best
preschoolers use past reliability over age when learning
new words. Psychol Sci 2006, 17:757–758.
165. Koenig MA, Jaswal VK. Characterizing children’s
expectations about expertise and incompetence: Halo
or pitchfork effects? Child Dev 2011, 82:1634–1647.
166. Sobel DM, Corriveau KH. Children monitor individ-
uals’ expertise for word learning. Child Dev 2010,
81:669–679.
167. Buttereld J, Jenkins OC, Sobel DM, Schwertfeger
J. Modeling aspects of theory of mind with Markov
random elds. Int J Soc Robot 2009, 1:41–51.
168. Shafto P, Eaves B, Navarro DJ, Perfors A. Epis-
temic trust: modeling children’s reasoning about
others’ knowledge and intent. Dev Sci 2012, 15:
436–447.
169. Shafto P, Goodman ND, Frank MC. Learning from
others: the consequences of psychological reason-
ing for human learning. Perspect Psychol Sci 2012,
7:341–351.
170. Hickling A, Wellman H. The emergence of chil-
dren’s causal explanations and theories: evidence
from everyday conversation. Dev Psychol 2001, 37:
668–683.
171. Nelson K. Structure and strategy in learning to talk.
Monogr Soc Res Child Dev 1973, 38:136.
© 2014 John Wiley & Sons, Ltd.
