Child Development, xxxx 2017, Volume 00, Number 0, Pages 1–17

Causal Learning Across Culture and Socioeconomic Status

Adrienne O. Wente and Katherine Kimura Caren M. Walker
University of California, Berkeley University of California, San Diego

Nirajana Banerjee Mar
ıa Fern
andez Flecha

University of California, Berkeley Pontificia Universidad Catolica del Per
u

Bridget MacDonald Christopher Lucas

University of California, San Diego The University of Edinburgh

Alison Gopnik
University of California, Berkeley

Extensive research has explored the ability of young children to learn about the causal structure of the world
from patterns of evidence. These studies, however, have been conducted with middle-class samples from
North America and Europe. In the present study, low-income Peruvian 4- and 5-year-olds and adults, low-
income U.S. 4- and 5-year-olds in Head Start programs, and middle-class children from the United States par-
ticipated in a causal learning task (N = 435). Consistent with previous studies, children learned both specific
causal relations and more abstract causal principles across culture and socioeconomic status (SES). The Peru-
vian children and adults generally performed like middle-class U.S. children and adults, but the low-SES U.S.
children showed some differences.

Coming to understand the causal structure of the with toddlers and preschoolers, coming from a
world is a central part of cognitive and conceptual number of different laboratories (see Gopnik &
development. Causal learning plays an especially Wellman, 2012; Xu & Kushnir, 2012 for recent
important role in the development of intuitive theo- reviews). These studies show remarkably high
ries of the world, such as folk biology and “theory levels of competence in young children. In fact,
of mind.” In the past 15 years, there has been a some recent causal learning studies have yielded a
large body of research showing how computational counterintuitive pattern of findings: The ability to
systems can accurately infer causal relations from infer certain types of abstract causal relations from
statistical patterns of data (e.g., Pearl, 2009; Spirtes, evidence actually appears to decline with age, so
Glymour, & Scheines, 1993/2000; Tenenbaum, that younger children do better than older children
Kemp, Griffiths, & Goodman, 2011). There have and adults (Gopnik, Griffiths, & Lucas, 2015; Gop-
also been hundreds of causal learning experiments nik et al., in press; Lucas, Bridgers, Griffiths, &
Gopnik, 2014; Seiver, Gopnik, & Goodman, 2013).
Adrienne O. Wente and Katherine Kimura contributed equally However, to our knowledge, these studies have
to this work. all examined children in similar middle- to upper-
This research was supported by funding from the Bezos Fam-
ily Foundation and Carlos Rodriguez-Pastor. The authors would middle-class samples in North America and Eur-
like to thank Alexa Kossuth Dasso and Maria Fe Hashimoto for ope—that is, children from Western, educated,
data collection in Peru. They also wish to thank Teresa Garcia, industrial, rich, and democratic (WEIRD) cultures
Renata Botello, and Aaya Abdelshafy for assistance with data
collection and coding in the United States, as well as the parents, (Henrich, Heine, & Norenzayan, 2010). There has
students, teachers, and administrators at the Innova Schools in been important cross-cultural work on the devel-
Peru, Berkeley and Oakland Head Start Schools, the Early Child- opment of intuitive theories (e.g., Avis & Harris,
hood Education Program at UC Berkeley, the Berkeley School
Early Childhood Center, Montverde School, and Step One
School.
Correspondence concerning this article should be addressed to © 2017 The Authors
Adrienne O. Wente, Department of Psychology, University of Child Development © 2017 Society for Research in Child Development, Inc.
California, Berkeley, 3210 Tolman Hall, Berkeley, CA 94720. Elec- All rights reserved. 0009-3920/2017/xxxx-xxxx
tronic mail may be sent to [email protected]. DOI: 10.1111/cdev.12943
2 Wente et al.

1991; Callaghan et al., 2005; Coley, 2012; Gelman “framework theories” (Gopnik & Wellman, 2012;
& Legare, 2011; Medin & Atran, 2004; Wellman, Laudan, 1978) or “overhypotheses” (Goodman,
Fang, Liu, Zhu, & Liu, 2006), but not on the cau- 1955) in science. These principles do not specify
sal learning abilities that might help underpin this particular causal hypotheses, but they do constrain
development. the possible hypotheses a learner will consider
In this article, we extend research on causal (Goodman, 1955; Griffiths & Tenenbaum, 2007). For
learning to two very different types of populations: example, in intuitive psychology an “overhypothe-
a cross-cultural sample of relatively low-income sis” would specify that the causal explanation of an
Peruvian children and adults and a cross-socioeco- action is likely to involve beliefs and desires—even
nomic sample of children from low-income North if it does not specify exactly which beliefs and
American families. There is increasing recognition desires explain a particular action. Several studies
of the importance of including a wider range of from different laboratories show that preschoolers
participants in developmental studies (e.g., Henrich can also make these more abstract inferences
et al., 2010; Legare & Harris, 2016). A recent (Schulz, Goodman, Tenenbaum, & Jenkins, 2008;
review article, for example, found that < 1% of Seiver et al., 2013; Sim & Xu, 2017).
published developmental studies were conducted In particular, in three separate studies, Lucas
in South or Central American countries, with et al. (2014) contrasted two abstract overhypothe-
important implications for the generality of the ses: one that is initially more likely and one that is
conclusions drawn from those studies (Nielsen, less likely (at least for adults). Participants received
Haun, K€ artner, & Legare, 2017). Studying causal evidence of the covariation between causes and
learning, in particular, across these populations effects that supported the more likely overhypothe-
may be especially important from a practical per- sis or else received covariation evidence that sup-
spective. Causal learning is itself a central part of ported the less likely overhypothesis. Experimenters
cognitive development and is important for the then presented the participants with a new set of
development of intuitive theories. The ability to evidence and recorded whether participants of dif-
infer causal structure from evidence is also a foun- ferent ages interpreted that evidence in accord with
dational part of scientific reasoning. Differences in the overhypothesis. In more detail, preschoolers
these abilities, then, might have important conse- and adults saw evidence in a training trial that jus-
quences for education in general and science edu- tified the conclusion that a novel machine worked
cation in particular. according to one of two abstract logical principles:
a disjunctive (i.e., OR) principle or a conjunctive
(i.e., AND) principle. Almost all the earlier studies
Previous Studies
of children tested disjunctive causal relations, which
Most of the studies of causal learning in children are also often the default assumption in studies of
have focused on children’s ability to learn particular adults (see Cheng, 1997). However, one study
causal hypotheses from statistical patterns of data, (Schulz et al., 2007) found that 4-year-olds would
particularly covariation between causes and effects. infer a conjunctive causal structure given the right
Early studies found that children as young as evidence.
24 months old could make these inferences in ways Participants were first told that “blicketness”
that went beyond simple associative learning (Gop- activated the machine. If participants were assigned
nik, Sobel, Schulz, & Glymour, 2001; Gopnik et al., to the disjunctive condition, they saw that a single
2004; Sobel & Kirkham, 2007). Researchers also object would or would not activate the machine.
found that children could use statistical patterns of Those assigned to the conjunctive condition, how-
evidence to infer more complex structures, such as ever, observed that only a combination of two
common causes versus causal chains (Schulz, Gop- objects would activate the machine. After viewing
nik, & Glymour, 2007) and to infer unobserved this evidence, participants saw an ambiguous
causes (Gopnik et al., 2004; Schulz & Sommerville, activation pattern with new objects that was consis-
2006) and that children as young as 24 months old tent with both the disjunctive and conjunctive
could infer probabilistic as well as deterministic principles, and had to infer the causal structure
causal relations (Kushnir & Gopnik, 2005; underlying that pattern (see Figure 1).
Waismeyer, Meltzoff, & Gopnik, 2014). Participants demonstrated their causal inferences
More recently, researchers have investigated in two ways. First, they judged whether each of
whether children can also learn more abstract and three objects, D, E, or F, was or was not a blicket.
general causal principles, the equivalent of Second, they intervened to make the machine go.
 Causal Learning Across Culture and SES 3

Training Trials
Conjunctive Condition Disjunctive Condition

A B C A&B A&C B&C A B C A&B A&C B&C

Test Trial

D D D E D & F D, E & F D & F

Figure 1. Activation patterns in the conjunctive and disjunctive conditions and test trial.

Given the ambiguous evidence, if they had inferred more blocks are blickets and placing more objects
the disjunctive principle, they should say that only on the machine in the conjunctive than disjunctive
F was a blicket and should use just that object to condition. If children produce these responses, it
activate the machine. If they had inferred the con- suggests that they are not using a simple imitative
junctive principle, they should say that more than or perceptual strategy.
one object was a blicket—in particular, both D and Lucas et al. (2014) found that middle-class U.S.
F would be more likely to be blickets than E—and preschool children readily learned both types of
they should choose multiple objects to activate the abstract causal rules depending on the activation
machine. pattern they observed. This means that, given the
Importantly, the disjunctive and conjunctive con- general causal principle that the machine was dis-
ditions were completely identical except for the junctive or conjunctive, children inferred the correct
activations in the training trials. The same objects specific causal hypothesis in the test phase, con-
were presented in the same order with the same cluding exactly which blocks were most likely to
script in the training trials, and the participants activate the machine. But significantly, it also
made judgments about exactly the same ambiguous means that they were inferring the general principle
test trial. The conditions could therefore act as con- itself from the training data, and then applying that
trols for each other. If participants responded sys- general principle to the ambiguous test data. So the
tematically differently in the two conditions, this task both tests children’s abilities to infer specific
suggests that they were influenced by the training causal hypotheses and more general overhypothe-
pattern. Similarly, comparing the judgments of the ses from data. In the current article, we extend this
three objects across the two conditions ensures that method across cultures and socioeconomic status
participants do not simply have a tendency to say (SES).
that objects are “blickets” (or not), or to behave at Notably also, adults, unlike the preschoolers,
chance, but are paying attention to the specific pat- tended to infer that the machine was activated by
tern of evidence in each condition. In addition, in one object, thus following the more likely disjunc-
the disjunctive training trials, participants actually tive principle regardless of the training evidence. A
see more examples of blocks activating the recent study (Gopnik et al., 2017) further tracked
machine, and more multiple combinations of blocks this ability across development. Six- to 11-year-olds
activating the machine, than they do in the conjunc- were less likely to infer the unusual conjunctive
tive trials, where the machine only activates once structure than 4-year-olds, and 12- to 14-year-olds
(see Figure 1). Nevertheless, the correct causal infer- as well as adults were less likely to infer the con-
ence involves the opposite response, saying that junctive structure than school-aged children.
4 Wente et al.

backgrounds may learn to attribute causation to a

Research Questions and Predictions
primary, focal object, whereas those who reason
Cross-cultural and cross-SES studies are intrinsi- more holistically may more readily assume a cause
cally and practically important, but they can also that requires a combination of objects. There is
help to differentiate among different, more specific, research suggesting that people from some South
hypotheses about development. First, it is possible American cultures reason more holistically than
that the precocious causal learning seen in earlier people from the United States (Henrich et al., 2010).
studies depends on growing up in an enriched In fact, Henrich et al. (2010) argue that mainstream
environment with many artifacts and extensive U.S. culture falls on the extreme end of the distribu-
informal science-like pedagogy. Since the ability to tion and that people from the United States reason
infer causal structure from evidence is a founda- more analytically than people from most other cul-
tional part of scientific methodology, it might be tures. Given this, another possibility is that children
encouraged in cultures or households where science and adults from Peru, particularly from groups that
is valued. Indeed, many studies have recruited chil- have recently migrated from rural indigenous com-
dren from science museums and from university munities, might actually be more likely to infer the
preschools, and it is plausible that these parents do multiple object conjunctive structure than children
provide implicit science pedagogy. Moreover, the and adults from the United States.
information-processing demands of tasks such as Finally, the pattern of results and levels of per-
that in Lucas et al. (2014) are quite high—involving formance might be similar across SES and culture,
tracking nine objects and 25 activation events. Chil- suggesting that these abilities are more generally
dren in other cultures or classes might prove to characteristic of young children.
have more information-processing difficulty than
the WEIRD children in previous studies. In either
case, we might expect to see general difficulties and Experiment 1
lower performance on all these tasks, involving
both the usual and unusual “overhypotheses,” in Method
children from other backgrounds.
Participants
For similar reasons, children in other cultural
groups might demonstrate some causal learning Ninety Peruvian 4- and 5-year-old children (45
abilities but not others, suggesting a developmental females and 45 males), and 145 Peruvian under-
ordering. In particular, children might be able to graduate college students (88 females and 57 males)
correctly infer the more common disjunctive struc- participated in this study. Three children were
ture, but have difficulty with the less common con- tested and not included in the analysis. One child
junctive one. They also might simply default to one was excluded because of experimenter error and
hypothesis or the other in the ambiguous test trials, two children were excluded because they picked up
but fail to differentiate between them, suggesting and placed one of the objects on or near the
that they had not inferred the overhypothesis from machine at an inappropriate time. No adults were
the training trials. excluded.
Alternatively, children (and adults) might be able Thirty children participated in the conjunctive
to solve these tasks in general, but they might show condition (M = 5.21, range = 4.21–5.97), 30 partici-
different patterns of response across cultures. For pated in the disjunctive condition (M = 5.18,
example, different cultural groups might have dif- range = 4.22–6.01), and 30 participated in the base-
ferent preferred “overhypotheses” about causal line condition (M = 5.2, range = 4.09–6). Forty-nine
relations. In particular, cross-cultural psychologists Peruvian adults participated in the conjunctive con-
frequently distinguish between analytic and holistic dition (M = 22.72, range = 19.97–39.76), 44 partici-
styles of causal reasoning (Henrich et al., 2010; Nis- pated in the disjunctive condition (M = 25.92,
bett, Peng, Choi, & Norenzayan, 2001; Peng & range = 16.93–45.15), and 52 participated in the
Knowles, 2003). Individuals from Western cultures, baseline condition (M = 19.45, range = 17.11–24.63).
for example, more readily engage in an analytic The published data set from Lucas et al. (2014) was
style of reason, which involves identifying a single used to compare results across cultures.
candidate cause. In contrast, the holistic style of rea- Children were recruited and tested at Innova
soning places a greater emphasis on the causal rela- schools in Lima, Peru. Innova schools are a Peru-
tions between multiple objects, and between objects vian chain of private schools designed to provide
and their environment. Individuals from analytic affordable education to Peru’s emerging middle
 Causal Learning Across Culture and SES 5

class (e.g., families who are transitioning from low-

Procedure
income to middle-income). In some ways, this sam-
ple was similar to the children in Lucas et al. The method was identical to Lucas et al. (2014)
(2014). Children from both backgrounds were from Experiment 2. Children were tested one by one in a
urban, industrialized environments, and parents private office or classroom in their school. Adults
were heavily invested in their education. However, were tested in university classrooms in groups of
the samples also differed along key dimensions. approximately 20 subjects per testing session, simi-
Not only were children tested in different countries, lar to the procedure in Lucas et al. (2014). An
but Peruvian children also had a substantially experimenter demonstrated the procedure, and
lower SES than the Lucas et al. (2014) children. adults were instructed to write their answers in
Innova schools were frequently located in low-SES, booklets so as to avoid any contamination of
high crime neighborhoods within Lima or in less responses.
developed neighborhoods on the outskirts of Lima, The word “flipo” was used as a plausible non-
accessible only by dirt roads. sense word in Spanish rather than “blicket.” The
In general, the average income of families experimenter first introduced participants to a bucket
across all the Innova schools is around $1,200 per of nine unsorted objects and the machine. They
month. However, we chose to test in schools at explained that flipos contained a property called
the lower end of the income spectrum, so the “fliponess” which activated the machine. They also
children were from families who almost certainly explained that flipos were quite rare, and that the
made less than $1,200 per month on average. goal of the game was to figure out which of the
Only a small percentage of these parents have objects were flipos. Next, the experimenter presented
had access to higher education and most are two new buckets, said that someone else had already
small business owners. Spanish was the native tested these objects on the machine, and showed that
language of the children, but most students were one bucket contained a single flipo, while the other
second- or third-generation internal immigrants bucket contained four nonflipos.
from the Andean highlands with a strong indige- Following this, the experimenter selected three
nous heritage. objects, in what appeared to be a random order,
Peruvian undergraduate students were recruited from the first bucket, which contained the nine
and tested at Pontificia Universidad Cat
olica del unsorted objects, and asked the child to name the
Per
u, a local private university often chosen by the objects. The order in which objects were selected
emerging middle class. Although the tuition varies was counterbalanced. Children and adults then
according to the family’s income—it can vary received two training trials, with objects A, B, and
between $333 and $1,363 a month—almost 70% of C, and A1, B1, and C1. Each training event varied
students are among those who pay lower tuition. according to condition (see Figure 1), and was con-
Many of the students are also first-generation col- sistent with one of the two overhypotheses, con-
lege students. junctive or disjunctive. In particular, in the
conjunctive training, a combination of objects A
and C was necessary to activate the machine. In the
Materials
disjunctive training, either A or C activated the
A machine derived from Gopnik and Sobel machine. Each trial also involved different objects.
(2000) was constructed from a wooden box that After observing each trial, the experimenter asked
was approximately 10 9 6 9 5 in. The top of the participants whether each object was a flipo. For
box was covered with orange construction paper example, the experiment might say, “Do you think
and a wireless doorbell was placed inside. This the Star is a flipo or not a flipo?” No feedback was
doorbell was surreptitiously activated by the experi- provided.
menter. Fifteen 3-dimensional clay blocks of unique After the two training events, the experimenter
geometric shapes but of similar sizes were painted brought out another object (G) which she said she
gray for color uniformity. These objects resembled had forgotten. This object was never placed on the
those used by Lucas et al. (2014). While the exact machine, and children and adults were asked to
dimensions varied, objects were about 3 in. and guess if they thought the object was a flipo.
were made into a variety of distinctive shapes, such Finally, the experimenter selected three more
as a square, a circle, and a star. Three plastic semio- objects: D, E, and F. The experimenter then demon-
paque buckets held the objects. Buckets were strated a new and ambiguous event (see Figure 1).
approximately 7 9 5 9 3 in. During this event, it was not clear if one object
6 Wente et al.

alone activated the machine or if a combination of and they tended to select more objects than Ameri-
objects activated the machine. After demonstrating can children overall. Because of this concern, the
the event, the experimenter once again asked the intervention prompt was retranslated for adults to,
participant to say whether each object was a flipo “¿Qu
e deber
ıamos usar de aqu
ı para hacer que la
or not a flipo. Afterward, children and adults were m
aquina se encienda?” This translates in English to,
given an intervention prompt, and were asked “What should we use from here to make the
about the object(s) they would use to activate the machine turn on?” This question still excluded the
machine. singular–plural marking yet was somewhat more
The procedure for the baseline condition was comparable to the original English prompt in that it
similar to the disjunctive and conjunctive experi- explicitly guided participants to select from the
mental conditions except that it omitted the two three objects directly displayed in front of them.
training events. In the baseline, participants also However, it still did not make specific reference to
provided feedback for two test events rather than the blocks and so might have led to more variable
one. We analyzed and reported their responses to responses. This second prompt was used when test-
the first test event. ing adults.
Inspection of adult data indicates that Peruvian
adults, like the Peruvian children, tended to choose
Translation
multiple objects for interventions more than U.S.
The study protocol from Lucas et al. (2014) was adults: 44% of the Peruvian adults chose multiple
translated to Spanish by an experienced Peruvian objects, whereas only 23% of the U.S. adults did so,
linguist as well as back-translated to English in although their flipo judgments were comparable.
order to assess the translation. The English version One reason for this might be that the English word
of this question, “Do you think {insert shape name} “which” encouraged English-speaking participants
is a Blicket or not a Blicket?” was directly translated to be more discriminatory in their selection,
to, “¿Crees que {insert shape name} es un flipo o whereas the Spanish phrasing using “que” or
no es un flipo?” “what” is more suggestive of multiple items.
The second question was the intervention Because the questions were not comparable across
prompt. In English it read, “Which of these should the languages and across children and adults, we
we use to make my machine turn on?” Translating did not include further analysis of the intervention
this directly to Spanish was challenging since the data in the Results section of Experiment 1.
English word, “which,” can be translated into either
“¿Cu
al?” or “¿Cu
ales?” These, however, provide a
singular–plural marking, which could indicate to Results
participants that they should select either one object
Training Trials: A Versus B Versus C
or multiple objects for the intervention. Selection of
single versus multiple objects is a dependent vari- We first assessed participants’ responses during
able. To circumvent this, the intervention prompt the training trials. This served to test whether chil-
for the children was translated to, “¿Qu
e har
ıas dren could infer specific causal hypotheses from
para que la m
aquina se encienda?” which directly unambiguous evidence, and whether they attended
translates in English to, “What would you do to to and understood those trials. In the training trials,
make the machine turn on?” This prompt avoids participants in the disjunctive condition should con-
the use of the singular or plural marking; however, clude that both A and C activate the machine,
it also makes it much less clear that the intervention although they do so independent of each other, but
refers specifically to choosing among the three that B does not. Those in the conjunctive condition
objects on the table. In fact, the most effective tech- should, likewise, infer that objects A and C are fli-
nique for activating the machine in all conditions pos, because both are necessary together to activate
would be to place all the objects on the machine the machine, and that B is not. (Note that the actual
since that would guarantee that the machine would A, B, and C objects shapes differed across partici-
light up. pants and training trials, and were counterbalanced;
An initial inspection of children’s responses sug- see Figure 1.)
gested that this difference in language did indeed If participants are confused by the task or have
lead to different intervention patterns. There were difficulty processing the activation patterns, they
no significant differences in the intervention should fail to differentiate between objects A, B,
responses across conditions for Peruvian children, and C, and instead respond “yes” (or “no”) to all
 Causal Learning Across Culture and SES 7

three objects or respond at chance when asked reporting that objects A (children: M = 1.57,
whether the blocks were flipos. Figure 2 illustrates SD = 0.73; adults: M = 0.57, SD = 0.79) and C (chil-
the proportion of children and adults who judged dren: M = 1.43, SD = 0.73; adults: M = 0.39,
A, B, and C as flipos. SD = 0.70) were flipos more often than B (children:
Overall, we found that participants in both con- M = 0.37, SD = 0.62; adults: M = 0.06, SD = 0.24),
ditions successfully differentiated object B from A versus B—children: t(58) = 6.897, p < .001; adults:
objects A and C and thus were not simply respond- t(56.93) = 4.319, p < .001; B versus C—children: t
ing “yes,” to each question. Participants received a (58) = 6.131, p < .001; adults: t(59.28) = 3.079,
score of 0, 1, or 2 depending on how often they p = .003, respectively.
said each block was a blicket across the two trials.
As predicted, children and adults in the disjunctive
Test Trials
condition reliably judged object A (children:
M = 1.57, SD = 0.82; adults: M = 1.86, SD = 0.41) to D, E, and F flipo judgment comparisons. Partici-
be a flipo more often than they judged object B to pants were clearly differentiating the objects in the
be a flipo (children: M = 0.30, SD = 0.65; adults: training trials; we next explored whether they were
M = 0.18, SD = 0.50), t(58) = 6.639, p < .001; t also differentiating objects in the test trial. If partici-
(86) = 17.375, p < .001, respectively. They were also pants reason conjunctively, then they should infer
more likely to call object C a flipo (children: that objects D and F are flipos—since together they
M = 1.43, SD = 0.77; adults: M = 1.86, SD = 0.46) activate the machine three times—but they should
than object B, t(58) = 6.137, p < .001; t(86) = 16.47, be uncertain about object E, which is associated
p < .001, respectively. with one conjunctive activation and one nonactiva-
Similar patterns were observed in the conjunctive tion, even though E and D are equally strongly
condition, with children and adults once again associated with the activation of the machine (see
Lucas et al., 2014 for discussion). If they reason dis-
junctively, then they should infer that object F is a
A 2.0 flipo, whereas objects D and E are not. If, however,
1.8 participants are confused by the task, they should
1.6 respond “yes” or “no” to all three objects or
Proportion Flipos

1.4
respond at chance. The crucial contrasts, then are
1.2
1.0
between D and F, which should be different in the
0.8
Peruvian Children disjunctive condition and similar in the conjunctive
0.6 Peruvian Adults one, and between D and E which should be similar
0.4 in the disjunctive condition and different in the con-
0.2 junctive one (see Figure 3 for participants’ D, E,
0.0
and F judgments).
A B C
Objects
In the disjunctive condition, Peruvian children
called object F a flipo (M = 0.77, SD = 0.43) reliably
more often than object D (M = 0.27, SD = 0.45),
B 2.0
p < .001, McNemar’s exact test, or object E
1.8
(M = 0.20, SD = 0.41), p < .001, McNemar’s exact
1.6 test. There were no significant differences between
objects D and E, p = .683, McNemar’s exact test.
Proportion Flipos

1.4
1.2 This is the pattern we would expect if learners were
1.0
Peruvian Children
making the correct inference from the evidence.
0.8
Peruvian Adults
In the conjunctive condition, children reliably
0.6
0.4
reported that D was a flipo (M = 0.87, SD = 0.35)
0.2 more often than E (M = 0.63, SD = 0.49), p = .046,
0.0 McNemar’s exact test. There was no significant dif-
A B C ference between performance on D and F, p = .683,
Objects McNemar’s exact test. Again, this is the pattern of
responses we would expect if participants learned
Figure 2. Average proportion of flipo judgments for objects A, B,
and C among Peruvian children and adults in the disjunctive the correct causal structure. Therefore, despite this
condition (A) and conjunctive condition (B). Error bars indicate challenging and unusual task, Peruvian children
1 SEM in each direction. were successfully tracking the activation patterns to
8 Wente et al.

A 1 McNemar’s exact tests were also used to analyze

0.9 adult’s responses. In the disjunctive condition, Peru-
0.8 vian adults called object F a flipo (M = 0.89,
Proportion Flipos

0.7 SD = 0.32) more often than object D (M = 0.11,

0.6
SD = 0.32), p < .001. There was no significant differ-
0.5
0.4
Peruvian Children ence between adults’ judgments of objects D and E
0.3 Peruvian Adults (M = 0.11, SD = 0.32), p = .480.
0.2 However, Peruvian adults in the conjunctive con-
0.1 dition, as in the disjunctive condition, also called
0 object F a flipo (M = 0.65, SD = 0.48) more often
D E F
than object D (M = 0.10, SD = 0.31), p < .001. There
Objects
was no difference in their judgments of objects D
and E (M = 0.12, SD = 0.33), p = 1.0.
Similar patterns were observed in the baseline
B 1
0.9
condition. Peruvian adults were more likely to say
0.8 that object F was a flipo (M = 0.77, SD = 0.43) than
object D (M = 0.12, SD = 0.32), p < .001. We found
Proportion Flipos

0.7
0.6 no reliable difference, however, in adults’ judg-
0.5
Peruvian Children
ments of objects D and E (M = 0.15, SD = 0.36),
0.4 p = .752.
0.3 Peruvian Adults
Flipo judgments of D and F across condition. The
0.2
0.1
crucial responses to determine if participants
0 inferred the overhypotheses were the judgments of
D E F D across the conditions. If participants concluded
Objects that the machine operated on a disjunctive princi-
ple, they should say that F was a flipo and D was
C not. If they concluded that the machine operated on
1
a conjunctive principle, they should say both F and
0.9
0.8
D were flipos.
As predicted, both age groups in all three condi-
Proportion Flipos

0.7
0.6 tions tended to say that F was a blicket (responses
0.5
Peruvian Children
ranging from 65% to 89%, see Table 1). To test
0.4 whether Peruvian children and adults differed in
0.3 Peruvian Adults
their judgments of D as a function of condition, we
0.2
0.1
performed a binary logistic regression with age
0 group, condition, and the interactions between each
D E F of these as predictors of judgments of D. The result-
Objects ing model was statistically significant,
v2(5) = 83.361, p < .001, Nagelkerke R2 = .427. The
Figure 3. Average proportion of flipo judgments for objects D, E,
and F among Peruvian children and adults in the disjunctive
results also yielded a main effect of age group,
condition (A), conjunctive condition (B), and baseline condition v2 = 32.032, df = 1, p < .001, and condition,
(C). Error bars indicate 1 SEM in each direction. v2 = 18.603, df = 2, p < .001, and an interaction
between the two, v2 = 9.961, df = 2, p = .007. To
more closely examine these effects, we next looked
make accurate causal inferences and did not appear at performance across conditions in each age group
to be confused. using two-tailed Fisher’s exact tests.
In the baseline condition, for which children saw
neither evidence supporting the conjunctive nor evi-
Peruvian Children
dence supporting the disjunctive principle, children
were just as likely to label object F (M = 0.77, Crucially, Peruvian children in the conjunctive
SD = 0.43) a flipo as they were to label objects D condition were significantly more likely than those
(M = 0.60, SD = 0.50), p = .131, McNemar’s exact in the disjunctive condition to call object D a flipo,
test, and E (M = 0.73, SD = 0.45), p = 1.0, McNe- p < .001, Fisher’s exact tests. Children in the con-
mar’s exact test. junctive condition also labeled object D a flipo more
 Causal Learning Across Culture and SES 9

Table 1 children and adults were just as likely to label D a

Mean Proportion of Flipo or Blicket Judgments for Peruvian Children flipo, p = .122, Fisher’s exact test. In the conjunctive
and Adults in Experiment 1 and North American Children and Adults condition, in contrast, Peruvian children were more
From Lucas et al. (2014)
likely to call the ambiguous object D a flipo than
Flipo or blicket judgments Peruvian adults, p < .001, Fisher’s exact test. Simi-
larly, in the baseline condition, children were more
Participant group D E F likely than adults to label object D as a flipo,
p < .001, Fisher’s exact test. These differences
Peruvian children
Disjunctive (N = 30) .27 .20 .77
appear to be responsible for the interaction in the
Conjunctive (N = 30) .87 .63 .87 model.
Baseline (N = 30) .60 .73 .77
Peruvian adults Object G
Disjunctive (N = 44) .11 .11 .89
Conjunctive (N = 49) .10 .12 .65 Despite initial demonstrations that flipos are
Baseline (N = 52) .12 .15 .77 rare, both children (M = 0.37, SD = 0.49) and adults
North American children (Lucas et al., 2014; Experiment 2) (M = 0.40, SD = 0.50) judged the item G to be a
Disjunctive (N = 25) .32 .28 .80 flipo at chance (i.e., 0.50) in the baseline condition,
Conjunctive (N = 25) .92 .68 .88 p = .200 and .212, respectively, by binomial test.
Baseline (N = 24) .42 .33 .75 This is similar to the pattern in Lucas et al. and
North American adults (Lucas et al., 2014; Experiment 2) suggests that the training trials raised the baseline
Disjunctive (N = 28) .11 .11 .82
probability that blocks were flipos.
Conjunctive (N = 28) .25 .11 .71
Baseline (N = 26) .08 .08 .81
Discussion
often than those in the baseline condition, p = .039, Peruvian children’s and adults’ judgments were
Fisher’s exact test. consistent with the developmental pattern observed
In general, the pattern of judgments was very in the North American sample. Peruvian children
similar to the pattern in Lucas et al. (2014). Chil- appeared to infer the correct specific causal
dren inferred the conjunctive hypothesis in the con- hypotheses from the data pattern in both the con-
junctive case and vice versa, and their performance junctive and disjunctive case, as evidenced by their
in the baseline condition appeared to fall in specific object choices of A, B, and C, and D, E, and
between. In fact, a direct comparison with the F. They also inferred the more abstract disjunctive
Lucas et al. results on every measure using two- and conjunctive overhypotheses, as evidenced by
tailed Fisher’s exact tests found only one difference: the difference between conditions. They correctly
The Peruvian children were more likely than the differentiated among the three objects across condi-
U.S. children to select E as a “flipo” in the baseline tions in the training trials and the test trials, which
condition, p = .006 (see Table 1 for a cross-cultural suggests that they were not confused by the data
comparison). and did not simply have a yes or no bias or answer
at chance.
In the baseline condition, they showed an inter-
Peruvian Adults
mediate pattern and did not differentiate among
In contrast, there were no significant differences the objects. This may suggest either that they were
among adults across conditions in their judgments unsure which overhypothesis applied or that they
of D (Fisher’s exact tests on D judgments: conjunc- were confused without the training trials.
tive vs. disjunctive, p = 1.00; baseline vs. conjunc- Peruvian adults, in contrast, like U.S. adults,
tive, p = 1.00; baseline vs. disjunctive, p = 1.00). tended to report that only object F activated the
Fisher’s exact tests also revealed no significant dif- machine, irrespective of condition. The two age
ferences on any measure between the Peruvian groups performed comparably in the disjunctive
adults and the adults in Lucas et al. (see Table 1). condition. But in the conjunctive condition, Peru-
vian children were more likely than Peruvian
adults to endorse the conjunctive principle. Over-
Peruvian Children Versus Adults
all, Peruvian children, like the North American
As in Lucas et al., we found that there was no sample, provided responses that appeared to be
age difference in the disjunctive condition, as more sensitive to the training data than Peruvian
10 Wente et al.

adults and were more likely to endorse a conjunc- condition (M = 4.66, range = 4.01–5.78). An addi-
tive principle. tional 13 children were tested but not included in
analysis for failure to complete the task (n = 2), for
machine malfunction (n = 4), or for language com-
prehension issues (n = 7). Given the relatively small
Experiment 2
number of available Head Start children, the base-
In Experiment 2, U.S. children from low-income line condition was not administered.
families were compared to U.S. children from mid- Children from low-SES families were recruited
dle- to upper-middle-class families. Additional and tested at Head Start programs in Berkeley
measures were also administered to determine if and Oakland, California. Children are only eligible
differences in SES might be related to other moder- for enrollment in a Head Start program if the fam-
ating factors that might differ between the groups, ily income falls below the federal poverty line,
such as general cognitive abilities, executive func- which in 2015 was below $24,250 for a family of
tioning, and language comprehension. If we did four (Poverty Guidelines, 2015). The higher SES
find SES differences in causal learning tasks, those sample was recruited from private preschools
might reflect something relatively specific to causal throughout Berkeley, California. This population
learning, such as differences in informal science differed slightly from the Lucas et al. (2014) sam-
experience, and so would not be correlated to ple, which largely consisted of preschools affiliated
other abilities. Alternatively, such differences might with the University of California and, as a result,
reflect broader SES differences in children’s cogni- included a substantial number of low-income stu-
tive abilities, rather than differences particular to dent parents with subsidized care. In order to
causal learning. Executive function and language ensure strong SES differences, and to replicate the
ability have been shown to be related to the for- Lucas et al. results with another clearly middle-
mation of intuitive theories, particularly, “theory of class sample, recruitment was expanded to include
mind” (Astington & Jenkins, 1999; Carlson & private, unsubsidized preschools. In 2015, the
Moses, 2001), and so might also be responsible for Berkeley population had a median household
differences on causal learning tasks and correlate income of $66,237 (QuickFacts, Berkeley, Califor-
with performance on those tasks. Similarly, if the nia, n.d.). The family incomes of the non-Head
low-SES children had difficulties with causal learn- Start children in the private preschools were prob-
ing tasks because of more general cognitive diffi- ably higher, although official demographic infor-
culties, then performance on other cognitive tasks mation was not collected.
such as Piagetian conservation tasks should corre-
late with performance on the causal learning tasks.
Materials and Procedure
On the other hand, children might do well on cau-
sal learning tasks in spite of difficulties in general Children were tested either in a quiet corner of
cognitive tasks, executive function tasks, or lan- their classroom or in a separate side room. Each
guage tasks, and that pattern would also be infor- session lasted approximately 15 min and included a
mative. causal learning task, a number conservation task,
and an executive function task.
All children were reported to be fluent English
Method speakers by their teachers, but many of them came
from non-English-speaking backgrounds. Upon test-
Participants
ing children, we observed varying degrees of Eng-
Two hundred 4- and 5-year-olds from the San lish language proficiency. To more systematically
Francisco Bay Area participated in Experiment 2. assess this, a vocabulary assessment measure was
Ninety-six of these children were from low-SES added midway through data collection. Forty-nine
families, of which 49 children participated in the Head Start children and 46 non-Head Start children
conjunctive condition (M = 4.66, range = 3.95–5.40) completed an expressive vocabulary task to ensure
and 47 participated in the disjunctive condition their comprehension of the task instructions, as well
(M = 4.67, range = 3.99–5.45). The remaining 104 as to measure language development. Participants
children were from middle- to upper-middle-class completed the tasks in the following order: (a) cau-
families. Fifty-six of these children were assigned to sal learning task, (b) number conservation task, (c)
the conjunctive condition (M = 4.67, range = 4.02– executive function test, and (d) expressive vocabu-
5.74) and 48 children were in the disjunctive lary test.
 Causal Learning Across Culture and SES 11

Causal learning task. Materials for this task everyday objects and were asked to label these
were identical to those described in Experiment 1; objects in English. Pictures were presented in order
however, the procedure varied slightly among two of difficulty, and testing was terminated once the
subgroups of children. For 59 Head Start children child provided four consecutive incorrect answers.
and 59 non-Head Start children, the task was iden- Scores reflect the total number of trials adminis-
tical to that described in Experiment 1 (as well as in tered.
Experiment 2 of Lucas et al., 2014) with the excep-
tion that participants were tested in English and
that the baseline condition was omitted. For 37 Results
Head Start and 45 non-Head Start children who
Training Trials: A Versus B Versus C
participated in the other version, the task was simi-
lar but, as in Lucas et al. Experiment 1, used a As in Experiment 1, we first examined chil-
more streamlined presentation that did not include dren’s judgments in the training trials to assess
the initial demonstration that blickets were rare nor whether they correctly interpreted the causal rela-
included the G item after the training trials. This tions. Participants should infer that objects A and
was motivated by the possibility that the simpler C are blickets, but B is not, in both conditions. In
version might involve fewer information-processing the disjunctive condition, non-Head Start children
demands. However, preliminary analyses compar- were more likely to call objects A (M = 1.83,
ing the performance between children tested in the SD = 0.48) and C (M = 1.79, SD = 0.50) blickets
two versions of the study showed no significant dif- than object B (M = 0.29, SD = 0.62, ps < .001 for
ferences. We therefore collapsed the two experimen- contrasts between A vs. B and B vs. C using
tal versions. independent t tests). Similarly, Head Start children
Number conservation task. A Piagetian conser- called C a blicket (M = 1.43, SD = 0.74) more
vation of numbers task was administered to assess often than B (M = 0.98, SD = 0.79), t(92) = 2.815,
general cognitive skills. Children were shown two p = .006, and called object A a blicket (M = 1.28,
rows of varying lengths, each containing five pen- SD = 0.80) marginally more often than B, t
nies. Responses were scored as correct if the child (92) = 1.813, p = .073.
said that both rows had the same number of pen- In the conjunctive condition, both Head Start and
nies after a transformation, and incorrect if he or non-Head Start participants labeled object A (Head
she said the rows contained different amounts. Start: M = 1.27, SD = 0.86; non-Head Start:
Day–night executive function test. To measure M = 1.32, SD = 0.77) and C (Head Start: M = 1.43,
children’s ability to inhibit responses, we also SD = 0.74; non-Head Start: M = 1.45, SD = 0.69) as
administered an executive function task similar to blickets more often than object B (Head Start:
the classic Stroop test. This day–night task (Ger- M = 0.84, SD = 0.75; non-Head Start: M = 0.79,
stadt, Hong, & Diamond, 1994) consisted of SD = 0.78), A versus B: t(96) = 2.634, p = .010;
twenty-two 3 9 3 in. cards. Half of the cards t(110) = 3.670, p < .001, respectively; B versus C:
depicted a yellow sun, whereas the other half t(96) = 3.954, p < .001; t(110) = 4.763; p < .001,
depicted a blue moon. respectively; A versus C: t(96) = 1.009, p = .315;
The experimenter instructed the child to say t(110) = 0.910, p = .365, respectively.
“day” when shown a card with the moon and say
“night” when shown a card with the sun. Next, the
Test Trial: Blicket Judgments
experimenter practiced the game with the child
using four cards, two of each kind, and provided Figure 4 shows the mean proportion of blicket
feedback. Then the experimenter presented 16 test judgments in the test trial separated by condition
cards in a quasi-random order (8 with the sun and and by SES.
8 with the moon). Trials were coded as correct or Test trial judgments: D versus E versus F. A ser-
incorrect, and scores were aggregated across the 16 ies of McNemar’s exact tests examined whether
trials. children correctly discriminated between objects in
Kaufman Assessment Battery for Children–II. We the test trial. If they are reasoning disjunctively,
administered the Kaufman Assessment Battery for then they should be less likely to say that D is a
Children, 2nd ed. (KABC–II) to measure expressive blicket than F, but should be equally likely to say
vocabulary and act as a proxy for language com- that D and E are blickets. If they are reasoning con-
prehension (Kaufman, Lichtenberger, Fletcher-Jan- junctively, they should be equally likely to say that
zen, & Kaufman, 2005). Children saw drawings of D and F are blickets, but more likely to say that D
12 Wente et al.

A 1
SD = 0.45) and F as blickets (M = 0.76, SD = 0.43),
0.9 p = 1.00, respectively. There were also no significant
0.8 differences in the Head Start and non-Head Start
Proportion Blickets

0.7 judgments of all three objects (see below). However,

0.6 the difference between Head Start children’s judg-
0.5
Head Start ments of objects D and E (M = 0.63, SD = 0.49) did
0.4
0.3 Non-Head Start not reach significance, p = .332.
0.2 F judgments across conditions. As in Experiment
0.1 1, and as predicted, participants generally said that
0 F was a blicket across SES and condition (values
D E F ranged from .68 to .90). There was one significant
Objects
difference: Head Start children were less likely than
non-Head Start children to say that F was a blicket
in the disjunctive condition, p = .012, Fisher’s exact
B
test.
1
D judgments across conditions. As in Experi-
0.9
0.8
ment 1, if children had inferred the conjunctive prin-
ciple, then they should say that objects D and F are
Proportion Blickets

0.7
0.6 blickets. If they inferred the disjunctive principle,
0.5 then they should say that object F is a blicket and
Head Start
0.4 D is not.
0.3 Non-Head Start
To test whether SES and condition influenced
0.2
children’s judgments of D, we performed a binary
0.1
0
logistic regression with SES, condition, and the
D E F interaction between these two as predictors. The
Objects resulting model was statistically significant,
v2(3) = 54.433, p < .001, Nagelkerke R2 = .319.
Figure 4. Average proportion of blicket judgments for objects D, Analyses further revealed a main effect of condi-
E, and F among Head Start and non-Head Start children in the
disjunctive condition (A) and conjunctive condition (B). Error
tion, v2 = 6.911, df = 1, p = .009, of SES, v2 = 9.323,
bars indicate 1 SEM in each direction. df = 1, p = .002, and an interaction between condi-
tion and SES, v2 = 7.729, df = 1, p = .005. Below a
series of two-tailed Fisher’s exact tests further
is a blicket than E. If they are confused by the task, examines responses.
then they might fail to differentiate the objects. Condition differences. Non-Head Start children
In the disjunctive condition, non-Head Start chil- in the conjunctive condition called object D a blicket
dren called object F a blicket (M = 0.90, SD = 0.31) more often than those in the disjunctive condition,
more often than object D (M = 0.17, SD = 0.38), p < .001. Head Start children also judged object D to
p < .001, and judged D and E (M = 0.19, be a blicket more often in the conjunctive than dis-
SD = 0.39), p = 1.0, to be similar. Similarly, Head junctive condition, p = .012.
Start children judged object F to be a blicket Head Start versus non-Head Start children. In
(M = 0.68, SD = 0.47) reliably more often than the disjunctive condition, non-Head Start children, as
object D (M = 0.47, SD = 0.50), p = .044, but did noted earlier, were more likely to call object F a
not differentiate between D and E (M = 0.51, blicket than Head Start children, p = .012, although
SD = 0.51), p = .803. they were less likely to call objects D and E blickets,
In the conjunctive condition, non-Head Start chil- p = .002 and .001, respectively. This reflected that
dren were equally likely to call objects D (M = 0.80, fact that although both groups discriminated
SD = 0.40) and F (M = 0.86, SD = 0.35) blickets, between the two conditions, the non-Head Start chil-
p = .450. Non-Head Start children in the conjunctive dren did so more dramatically.
condition were also significantly more likely to In the conjunctive condition, in contrast, children
label D a blicket than E (M = 0.66, SD = 0.48), enrolled in Head Start did not reliably differ from
p = .043, indicating that they did not simply those who were enrolled in non-Head Start pro-
respond “yes” to every question. grams in their judgments of whether objects D, E,
Head Start children in the conjunctive condition and F were blickets, p = .487, .839, and .218,
were also equally like to label objects D (M = 0.73, respectively.
 Causal Learning Across Culture and SES 13

SD = 0.40), suggesting once again that children cor-

Interventions
rectly inferred the causal structure that was best
As an additional measure of children’s causal supported by the evidence.
learning, we explored their response to the inter- Next, we explored whether SES and condition
vention question, which asked them to choose the were a reliable predictor of whether children
object(s) that would activate the machine. We inferred the correct disjunctive rule (i.e., selecting
examined first whether children selected multiple only object F). To do so, we performed another
objects, the correct response if they had inferred the regression with SES, condition, and the interaction
conjunctive rule, and second whether they selected between these two as predictors. This model was
just object F, the correct response if they had significant, v2(3) = 37.111, p < .001, Nagelkerke
inferred the disjunctive rule (Figure 5 shows the R2 = .231, with a main effect of condition,
results). v2 = 7.352, df = 1, p = .007. Children in the disjunc-
To examine whether children selected multiple tive condition (M = 0.57, SD = 0.50) were more
objects as a function of SES and condition, we per- likely to select just object F, which was the correct
formed a binary logistic regression on their inter- disjunctive rule, than children in the conjunctive
vention choices. The model, which included SES condition (M = 0.19, SD = 0.40). However, there
and condition as fixed factors, was statistically sig- was also a main effect of SES, v2 = 5.485, df = 1,
nificant, v2(3) = 37.811, p < .001, Nagelkerke p = .019, as non-Head Start children chose only F
R = .235, with a main effect of condition,
2
(M = 0.42, SD = 0.50) more often than Head Start
v2 = 15.061, df = 1, p < .001, and a trending effect children (M = 0.31, SD = 0.47). We found no signifi-
of SES, v2 = 3.456, df = 1, p = .063. Overall, children cant interaction between these predictors,
in the conjunctive condition reliably selected multi- v2 = 1.954, df = 1, p = .162.
ple objects (M = 0.55, SD = 0.50) more often than As with the blicket judgments, these results sug-
children in the disjunctive condition (M = 0.20, gested that the Head Start children differentiated
the two conditions less clearly than the non-Head
Start children; they produced fewer just F
A 1 responses, and more multiple responses in the OR
0.9
condition. However, when considered separately,
Proportion of Participants

0.8
0.7
the Head Start children, like the non-Head Start
0.6 children, still made significantly more multiple
0.5 responses in the conjunctive than disjunctive condi-
Head Start
0.4 tions, p < .001, Fisher’s exact test, and significantly
Non-Head Start
0.3 more just F responses in the disjunctive than con-
0.2
junctive conditions, p = .008, Fisher’s exact test,
0.1
0
suggesting that they could succeed at the task.
Disjunctive Conjunctive
Condition
Number Conservation Task
Scores ranged from 0 to 2. An independent sam-
B 1
ples t test suggests that non-Head Start children
0.9
(M = 1.20, SD = 0.61) and Head Start children
Proportion of Participants

0.8
0.7 (M = 1.00, SD = 0.80) responded differently,
0.6 t(175.94) = 1.986, p = .049, on the conservation task.
0.5
Head Start
0.4
0.3 Non-Head Start Day–Night Executive Function Task
0.2
0.1
According to an independent samples t test,
0 there was a significant difference in performance on
Disjunctive Conjunctive the executive function day–night task for Head
Condition Start and non-Head Start children, t(194) = 4.306,
p < .001. Non-Head Start children were more accu-
Figure 5. Average proportion of Head Start and non-Head Start
children who selected two or more objects (A) or just F (B) when rate at inhibiting their responses (M = 11.25 correct
asked to activate the machine. Error bars indicate 1 SEM in each of the 16, SD = 4.10), than Head Start children
direction. (M = 8.63 of the 16, SD = 4.43).
14 Wente et al.

and class contexts. Children from all the groups,

Kaufman Assessment Battery for Children–II
290 children in all, demonstrated some ability to
There was also a significant difference between learn both specific causal hypotheses and general
Head Start and non-Head Start children on their per- causal frameworks as evidenced by their discrimi-
formance on the KABC–II. An independent samples t nation among the different objects and between the
test revealed that non-Head Start children (M = 23.07, two conditions. In contrast, adults in Peru, like the
SD = 3.78) outperformed Head Start children U.S. adults, did not significantly differentiate
(M = 16.84, SD = 3.94), t(93) = 7.855, p < .001. between the three conditions. These results appear
to be robust and replicable.
One alternative potential explanation for the dif-
Relations Between Additional Tasks and Causal
ference across conditions, in particular, should be
Learning Tasks
considered, however. It was possible that the chil-
Despite the differences between SES groups on dren did appropriately make disjunctive inferences
the conservation, executive function, and vocabu- in the disjunctive case, either because of the training
lary tests, there were no significant correlations or because that was their default assumption. This
between the executive functioning and vocabulary in itself would imply some significant cross-cultural
tasks on any of the crucial blicket judgment or competence in causal learning. However, in the
intervention measures (choosing D, choosing multi- conjunctive case they might have simply became
ples in AND, and choosing just F in OR). There confused and showed a “yes” bias, saying that all
was a marginal correlation between conservation the objects were “blickets” rather than inferring the
and choosing just F in the OR condition (p = .052). conjunctive structure correctly.
Several pieces of evidence weigh against this
Discussion alternative hypothesis. First, consider the middle-
class U.S. children. Replicating the results of Lucas
Despite differences in SES, these findings largely
et al., they correctly differentiated between A, B,
replicate the main pattern described by Lucas et al.
and C in the conjunctive training trials, and
(2014). Preschoolers from low-income backgrounds
between D and E in the conjunctive test trials. Like
appear to be able to infer both specific causal
the Lucas et al. children they also correctly inter-
hypotheses and abstract causal principles. Both
vened, producing multiple object interventions in
groups of children judged the conjunctively active
the conjunctive condition, versus just F interven-
object D to be a blicket more often in the conjunctive
tions in the disjunctive one.
condition than in the disjunctive condition. More-
The Peruvian children also correctly differenti-
over, all the U.S. children made significantly more
ated between the three choices in the conjunctive,
multiple object interventions in the conjunctive than
training, and test trials, in fact, as noted their
in the disjunctive condition, and made more “just F”
judgments were indistinguishable from those in
interventions in the disjunctive condition than in the
Lucas et al. However, due to the linguistic differ-
conjunctive one, paralleling their judgments.
ences it was not possible to assess their interven-
The low-SES and high-SES children performed in
tions, and so the evidence was less strong for
a very similar way in the conjunctive condition (see
them.
Figure 4B). However, several pieces of evidence
The Head Start children, in contrast, did show
suggested that, unexpectedly and unlike the Peru-
the correct pattern in their interventions, producing
vian children, the low-SES children behaved differ-
multiple object interventions in the conjunctive con-
ently than the higher SES children in the disjunctive
dition and just F interventions in the disjunctive
condition. The Head Start children were more likely
one. They also significantly differentiated A, B, and
to say that D and E were blickets in the disjunctive
C in the conjunctive training trials, though less
condition, and were less likely to say that F was a
clearly than the non-Head Start children. However,
blicket. They were also more likely to use multiple
they did not significantly differentiate D and E in
objects to intervene in the disjunctive condition,
the conjunctive test condition, although their blicket
and less likely to select just F.
judgments were also not significantly different from
those of the middle-class children. So again, the evi-
dence was less strong for them. Further studies
General Discussion
would be necessary to definitively rule out this pos-
Overall, findings suggest stronger commonalities sibility for the Peruvian and Head Start children,
than differences in causal learning across culture but it appears to be less likely than the alternative
 Causal Learning Across Culture and SES 15

possibility: that children are correctly inferring the to design comparable intervention instructions in
conjunctive structure. Experiment 1. In Experiment 2, differences in infor-
In terms of the predictions, there was some evi- mation-processing may have influenced the Head
dence for the Head Start children that difficulties Start children’s responses in the disjunctive condi-
with inhibition and information-processing might tion. Furthermore, there were some unexpected
have influenced their performance. However, this English language comprehension issues with the
involved SES differences on the disjunctive task lower income U.S. children. Moreover, all these
rather than the conjunctive one. Although they children were growing up in industrial urban cen-
appeared to infer the structure correctly, they did ters with enough parental support to ensure that
so less clearly than the higher SES children, and they were enrolled in high-quality preschools. It
this might reflect the inhibitory demands of the would be important to see how children with less
task. There was no evidence for such differences support would behave. Similarly, it would be
between the Peruvian and American children. There important to test children from small-scale agricul-
was also little evidence that the children defaulted tural or forager backgrounds. These results do sug-
to either the disjunctive or conjunctive structure— gest, however, that children’s causal learning
all the groups of children differentiated between abilities extend beyond the WEIRD.
the two training conditions.
Similarly, there was no direct evidence for cross-
cultural differences in analytic versus holistic styles
References
of reasoning. However, there was a cultural differ-
ence in children’s E judgments in the baseline con- Astington, J. W., & Jenkins, J. M. (1999). A longitudinal
dition, and both Peruvian children and adults made study of the relationship between language and
more multiple interventions than U.S. participants. theory-of mind development. Developmental Psychology,
This raises the possibility that Peruvians could be 35, 1311–1320. https://doi.org/10.1037/0012-1649.35.5.
more open to a conjunctive principle than U.S. par- 1311
Avis, J., & Harris, P. L. (1991). Belief-desire reasoning
ticipants. However, the linguistic differences made
among Baka children: Evidence for a universal concep-
the interventions very difficult to interpret and with tion of mind. Child Development, 62, 460–467. https://
lack of support from other measurements, it is hard doi.org/10.1111/j.1467-8624.1991.tb01544.x
to tell if this finding is actually indicative of such a Callaghan, T., Rochat, P., Lillard, A., Claux, M. L.,
difference. Additional studies should explore this Odden, H., Itakura, S., . . . Singh, S. (2005). Synchrony
further. Overall, however, the patterns were strik- in the onset of mental-state reasoning: Evidence from
ingly similar in children and adults from a wide five cultures. Psychological Science, 16, 378–384.
range of cultural and economic backgrounds, and https://doi.org/10.1111/j.0956-7976.2005.01544.x
suggest that such causal learning abilities may be a Carlson, S. M., & Moses, L. J. (2001). Individual differ-
feature of young children in general. ences in inhibitory control and children’s theory of
The first, most striking, and most practically mind. Child Development, 72, 1032–1053. https://doi.
org/10.1111/1467-8624.00333
important result of these studies is that low-income
Cheng, P. W. (1997). From covariation to causation: A
children in Peru and in Head Start programs are causal power theory. Psychological Review, 104, 367–495.
able to perform causal learning tasks. These chil- https://doi.org/10.1037/0033-295X.104.2.367
dren were certainly able to infer the correct struc- Coley, J. D. (2012). Where the wild things are: Informal
ture in the disjunctive case, and also appeared able experience and ecological reasoning. Child Development,
to infer the conjunctive structure and to differenti- 83, 992–1006. https://doi.org/10.1111/j.1467-8624.2012.
ate between the two conditions, suggesting that 01751.x
they could infer “overhypotheses” as well as speci- Gelman, S. A., & Legare, C. H. (2011). Concepts and folk
fic hypotheses. In the case of the low-SES U.S. chil- theories. Annual Review of Anthropology, 40, 379–398.
dren, this was true in spite of lower conservation, https://doi.org/10.1146/annurev-anthro-081309-145822
KABC–II and executive function scores. This, in Gerstadt, C. L., Hong, Y. J., & Diamond, A. (1994). The
relationship between cognition and action: Performance
turn, suggests that these children have basic infer-
of children 3.5–7 years old on a Stroop-like day-night
ential capacities that could be leveraged in science test. Cognition, 53, 129–153. https://doi.org/10.1016/
education. 0010-0277(94)90068-X
However, the results also point to some of the Goodman, N. (1955). Fact, fiction and forecast (Vol. 74).
difficulties and complexities of such cross-cultural Cambridge, MA: Harvard University Press.
and cross-SES comparisons. The linguistic differ- Gopnik, A., Glymour, C., Sobel, D., Schulz, L., Kushnir,
ences between Spanish and English made it difficult T., & Danks, D. (2004). A theory of causal learning in
16 Wente et al.

children: Causal maps and Bayes nets. Psychological development across cultures. Psychological Review, 111,
Review, 111, 1–31. https://doi.org/10.1037/0033-295X. 960–983. https://doi.org/10.1037/0033-295X.111.4.960
111.1.3 Nielsen, M., Haun, D., K€ artner, J., & Legare, C. H. (2017).
Gopnik, A., Griffiths, T. L., & Lucas, C. G. (2015). When The persistent sampling bias in developmental psychol-
younger learners can be better (or at least more open- ogy: A call to action. Journal of Experimental Child Psy-
minded) than older ones. Current Directions in Psycho- chology, 162, 31–38. https://doi.org/10.1016/j.jecp.2017.
logical Science, 24, 87–92. https://doi.org/10.1177/ 04.017
0963721414556653 Nisbett, R. E., Peng, K., Choi, I., & Norenzayan, A.
Gopnik, A., O’Grady, S., Lucas, C., Griffiths, T., Wente, (2001). Culture and systems of thought: Holistic versus
A., Bridgers, S., . . . Dahl, R. E. (2017). Changes in cog- analytic cognition. Psychological Review, 108, 291.
nitive flexibility and hypothesis search across human https://doi.org/10.1037/0033-295X.108.2.291
life history from childhood to adolescence to adult- Pearl, J. (2009). Causality. New York, NY: Cambridge
hood. PNAS, 114, 7892–7899. https://doi.org/10.1073/ University Press.
pnas.1700811114 Peng, K., & Knowles, E. D. (2003). Culture, education,
Gopnik, A., & Sobel, D. M. (2000). Detecting blickets: and the attribution of physical causality. Personality and
How young children use information about novel Social Psychology Bulletin, 29, 1272–1284. https://doi.
causal powers in categorization and induction. Child org/10.1177/014616720325460
Development, 71, 1205–1222. https://doi.org/10.1111/ Poverty Guidelines. (2015, November 23). Office of the
1467-8624.00224 assistant secretary for planning and evaluation. Retrieved
Gopnik, A., Sobel, D., Schulz, L., & Glymour, C. (2001). from https://aspe.hhs.gov/2015-poverty-guidelines.
Causal learning mechanisms in very young children: QuickFacts, Berkeley, California. (n.d.). United States Cen-
Two, three, and four-year-olds infer causal relations sus Bureau. Retrieved from http://www.census.gov/
from patterns of variation and covariation. Developmen- quickfacts/table/IPE120215/0606000
tal Psychology, 37, 620–629. https://doi.org/10.103/ Schulz, L. E., Goodman, N. D., Tenenbaum, J. B., & Jenk-
0012-1649.37.5.620 ins, A. C. (2008). Going beyond the evidence: Abstract
Gopnik, A., & Wellman, H. M. (2012). Reconstructing laws and preschoolers’ responses to anomalous data.
constructivism: Causal models, Bayesian learning Cognition, 109, 211–223. https://doi.org/10.1016/j.cogni
mechanisms, and the theory theory. Psychological Bul- tion.2008.07.017
letin, 138, 1085–1108. https://doi.org/10.1037/ Schulz, L. E., Gopnik, A., & Glymour, C. (2007). Pre-
a0028044 school children learn about causal structure from condi-
Griffiths, T. L., & Tenenbaum, J. B. (2007). Two proposals tional interventions. Developmental Science, 10, 322–332.
for causal grammars. In A. Gopnik & L. Schulz (Eds.), https://doi.org/10.1111/j.1467-7687.2007.00587.x
Causal learning: Psychology, philosophy, and computation Schulz, L. E., & Sommerville, J. (2006). God does not play
(pp. 323–345). New York, NY: Oxford University Press. dice: Causal determinism and preschoolers’ causal
Henrich, J., Heine, S. J., & Norenzayan, A. (2010). Most inferences. Child Development, 77, 427–442. https://doi.
people are not WEIRD. Nature, 466, 29. https://doi. org/10.1111/j.1467-8624.2006.00880.x
org/10.1038/466029a Seiver, E., Gopnik, A., & Goodman, N. D. (2013). Did she
Kaufman, A. S., Lichtenberger, E. O., Fletcher-Janzen, E., jump because she was the big sister or because the tram-
& Kaufman, N. L. (2005). Essentials of KABC–II Assess- poline was safe? Causal inference and the development
ment (Vol. 94). New York, NY: Wiley. of social attribution. Child Development, 84, 443–454.
Kushnir, T., & Gopnik, A. (2005). Young children infer https://doi.org/10.1111/j.1467-8624.2012.01865.x
causal strength from probabilities and interventions. Sim, Z., & Xu, F. (2017). Learning higher-order general-
Psychological Science, 16, 678–683. https://doi.org/10. izations through free play: Evidence from two- and
1111/j.1467-9280.2005.01 595.x three-year-old children. Developmental Psychology, 53,
Laudan, L. (1978). Progress and its problems: Towards a the- 642–651. https://doi.org/10.1037/dev0000278
ory of scientific growth (Vol. 282). Berkeley and Los Sobel, D. M., & Kirkham, N. Z. (2007). Bayes nets and
Angelas California: University of California Press. babies: Infants’ developing statistical reasoning abilities
Legare, C. H., & Harris, P. (2016). The ontogeny of cul- and their representation of causal knowledge. Develop-
tural learning. Child Development, 87, 633–642. mental Science, 10, 298–306. https://doi.org/10.1111/j.
https://doi.org/10.1111/cdev.12542 1467-7687.2007.00589.x
Lucas, C. G., Bridgers, S., Griffiths, T. L., & Gopnik, A. Spirtes, P., Glymour, C., & Scheines, R. (2000). Causation,
(2014). When children are better (or at least more prediction and search, Springer lecture notes in statistics
open-minded) learners than adults: Developmental dif- (2nd ed.). Cambridge, Massachusetts: MIT Press. (Origi-
ferences in learning the forms of causal relationships. nal work published 1993)
Cognition, 131, 284–299. https://doi.org/10.1016/j.cogni Tenenbaum, J. B., Kemp, C., Griffiths, T. L., & Goodman,
tion.2013.12.010 N. D. (2011). How to grow a mind: Statistics, structure,
Medin, D. L., & Atran, S. (2004). The native mind: Biolog- and abstraction. Science, 331, 1279–1285. https://doi.
ical categorization, reasoning and decision making in org/10.1126/science.1192788
 Causal Learning Across Culture and SES 17

Waismeyer, A. S., Meltzoff, A. N., & Gopnik, A. (2014). Chinese children. Psychological Science, 17, 1075–1081.
Causal learning from probabilistic events by human https://doi.org/10.1111/j.1467-9280.2006.01830.x
infants: An action measure. Developmental Science, 18, Xu, F., & Kushnir, T. (Eds.). (2012). Rational constructivism
175–182. https://doi.org/10.1111/desc.12208 in cognitive development. Advances in child development
Wellman, H. M., Fang, F., Liu, D., Zhu, L., & Liu, G. and behavior, (Vol. 43). Waltham, MA: Academic Press.
(2006). Scaling of theory-of-mind understandings in