About this paper symposium

---

Panel information
Panel 4. Cognitive Processes

Paper #1
Children’s Exploration of Deterministic Versus Probabilistic Causal Systems Which Work for Others but Not Themselves
Author information	Role
Caiqin Zhou, Brown University, United States	Presenting author
Sophie Bridgers, Google DeepMind, United States	Non-presenting author
Yvonne Wang, University of Toronto, Canada	Non-presenting author
Daphna Buchsbaum, Brown University, United States	Non-presenting author
Abstract
Causal learning is challenging, in part because causal relationships are often graded in strength (e.g., both wet and icy sidewalks can make pedestrians slip, but the latter causes slipping more often), which affects the certainty of learners’ causal judgments. Children successfully infer varying degrees of causal strength from different patterns of contingency (Kushnir & Gopnik, 2005), and the strength of these inductive inferences affects their exploration. For example, knowing that members of a category have an unobserved property, children attempt to elicit that property from other category members, with stronger membership cues leading to more persistent attempts (Baldwin et al., 1993). The present study investigates whether children infer different causal strengths when observing a deterministic versus probabilistic causal system, and further, whether this inference affects children’s exploration and interpretation of this system when it stops working. Four- and five-year-olds (N = 77) watched an experimenter try to activate a machine by placing two blocks on top one at a time. Either one block always activated the machine and the other block never did (Deterministic condition), or one block activated the machine more often than the other (Probabilistic condition). Children then freely explored these two blocks and the machine, during which time the machine stayed inert regardless of children’s actions. We predicted that, relative to children in the Probabilistic condition, children in the Deterministic condition would more strongly expect the previously efficacious block to work. Furthermore, when their interventions failed, children in this condition might readily infer that they were doing something wrong, and persist in trying the previously efficacious block or experimenting with the machine in various ways. Indeed, compared to the Probabilistic condition, children in the Deterministic condition intervened more often using the previously more efficacious of the two blocks (F(1, 65) = 6.17, p = .016, Figure 1). These children also tried a larger variety of alternative methods of intervention (e.g., flipping the block over to try a different surface in contact with the machine; F(1, 65) = 4.74, p = .033, Figure 2), possibly reflecting attempts to figure out the ‘correct’ way of using the causal system. Intriguingly, children in the Probabilistic condition had marginally longer overall exploration time (F(1, 65) = 3.07, p = .084). Since children in the Deterministic condition more strongly believed the previously efficacious block would work, they might have quickly attributed their failed attempts to a change in the system (e.g., running out of battery) or their own inability to properly use it. Data from a pre-registered replication study are currently being analyzed and will be included in the conference presentation. Our results show that children rely on their sensitivity to the gradations in causal strength when exploring causal systems that behave unexpectedly. When a previously efficacious cause fails to produce an effect, children explore longer if this failure seems to be due to the system’s inherent stochasticity rather than a change or error, suggesting that children’s persistence during exploration may be adaptive (Lucca et al., 2020).

Paper #2
Caregiver impact on 3- to 6-year-old children’s mechanistic causal reasoning during contexts of failure
Author information	Role
Gauri Harindranath, Tufts University, United States	Presenting author
Paul Muentener, Tufts University, United States	Non-presenting author
Abstract
Research suggests that caregivers can support children’s causal learning and that contexts of failure support children’s mechanistic reasoning – reasoning about the visible and hidden factors that connect actions to outcomes (e.g., batteries, electricity) (e.g., Legare, 2012; Medina & Sobel, 2020). Yet, people make different inferences about the sources of failure: people are more likely to attribute their own failure to the situation and another person’s failure to the person (Jones & Nisbett, 1972). The current studies investigate whether the source of failure (i.e., experiencing failure vs. observing children fail) influences caregiver’s influence on children’s mechanistic reasoning. We reasoned that caregivers who observed their child fail may attribute it to the child’s actions. However, caregivers who experienced their own failure may instead attribute it to a causal system, better supporting their child’s mechanistic reasoning. In Study 1, 75 caregiver-child dyads (children 3- to 6-years) were assigned to a Caregiver-Observed (n = 38) or Caregiver-Experienced (n = 37) Failure condition. Dyads were shown how to use a novel toy (Figure 1a). Unbeknownst to them, there was an obscured mechanism – a switch – that also needed to be ON for the toy to work. Dyads were then given the toy with the switch OFF so they would fail to make the toy work. We manipulated the source of failure between conditions: in the Caregiver-Observed condition, caregivers watched their child fail; in the Caregiver-Experienced condition, caregivers failed while their child was distracted. Subsequently, caregivers were able to provide only verbal instruction to guide their child’s exploration. We then asked children to explain how the toy worked (to assess mechanistic reasoning) and how a novel toy with a similar mechanism worked (to assess generalization; Figure 1b). Consistent with our prediction, caregivers used more mechanistic language (χ2(1) = 4.00, p < 0.05) and children were more likely to discover the mechanism (χ2(1) = 5.16, p < 0.05; Figure 2a) in the Caregiver-Experienced condition. Yet, this facilitated discovery did not carry forward to children’s explanations (p > .05). Finally, contrary to our prediction, children were more likely to generalize their mechanistic reasoning in the Caregiver-Observed condition, χ2(1) = 4.53, p < 0.05 (Figure 2b). In Study 2 (in progress, N = 36/70), we aim to extend this finding to a joint exploration task. The procedure is similar to Study 1, except instead of only providing verbal instructions, dyads engage in joint exploration after caregivers experience or observe failure. We then assess caregiver-child language and actions during joint exploration, and children’s mechanistic causal explanations, predictions and generalizations. These current findings suggest a tension in caregivers' role in supporting children’s mechanistic reasoning. Although experiencing failure may increase caregivers’ teaching and lead to an immediate increase in children’s mechanistic exploration, this mechanistic learning may be more limited than if children learned from their own failure. Investigating how failure influences caregiver’s behavior and children’s learning can help us better understand the role caregiver-child interactions play in children's causal learning and help identify factors that contribute to children’s learning from failure.

Paper #3
Children’s use of feature-based and alternation-based rules to explain causal relationships
Author information	Role
Rebekah Gelpí, University of Toronto, Canada	Presenting author
Daphna Buchsbaum, Brown University, United States	Non-presenting author
Christopher Lucas, University of Edinburgh, Scotland	Non-presenting author
Abstract
When learning about causal relationships, children typically need to consider multiple potential causal hypotheses that could all account for the same observed outcomes, and to revise their beliefs about the true cause if new data shows their prior beliefs to be improbable. Indeed, under some circumstances, children may even be more effective at learning an unlikely hypothesis than adults (Lucas et al., 2014), partly because they may be less susceptible to “learning traps” where early beliefs preclude later learning (Liquin & Gopnik, 2022). We investigate whether 5- and 6-year-old children (current N = 60; target N = 90) avoid causal learning traps due to having different prior beliefs than adults about likely causal relationships, or whether their belief revision process takes a different form from adults. In our causal learning task, two potential rules with differing salience could each explain a portion of the observed data. Children observe 12 trays of blocks placed on a machine, which sometimes activated. Trays are initially consistent with two rules (feature: a red square must be on the tray to activate the machine; alternation: every other tray activates the machine); however, the last four trays are consistent only with the feature-based (Feature condition) or alternation-based rule (Alternation condition). Children then predict activations for eight novel trays. We predict that children will be able to disambiguate between these rules of differing salience after the last four trays are presented, making more predictions consistent with the correct rule in each condition. To characterize children’s responses on our task, we develop a family of generative models that yield choice probabilities for each tray, which assume that children use a feature-based rule (choosing trays with red squares), an alternation-based rule (choosing every other tray), an absolute rule (choosing every tray or no trays), or other uncharacterized rules to make predictions. We find that a plurality of children’s choices are consistent with the correct rule: 34% of Feature condition best fit by feature rule, LL = –280 (16% best fit by alternation rule, LL = –286); 48% of Alternation condition best fit by alternation rule, LL = –109 (26% best fit by feature rule, LL = –188) (Figure 1). However, children’s choices were more consistent with the alternation rule overall, while a moderate proportion of children (26% in the Alternation condition, 47% in the Feature condition) were best fit by a different rule. When asked to describe their choices, children who were best fit by neither the alternation nor the feature rule often articulated different patterns of alternation (i.e., “yes-yes-no-no” rather than “yes-no-yes-no”). Our findings suggest that children are able to recognize when an ambiguous causal relationship is disambiguated by later evidence, and make choices consistent with the correct rule. In addition, children may also have an especially salient prior belief in alternation as an explanation for causal relationships. More broadly, this suggests children may find certain unusual relationships more a priori plausible than adults, which may help them to learn in situations that pose challenges for adult learners.

Paper #4
Internal disposition or external disruption: young learners’ explanations of inconsistent behavior in people versus machines
Author information	Role
Dr. Elizabeth Lapidow, Ph.D., University of Waterloo, United States	Presenting author
Stephanie Denison, University of Waterloo, Canada	Non-presenting author
Abstract
The world around us is both changeable and unreliable; and while causal learning aims to arrive at explanations or hypotheses that will generalize from past to present situations, exceptions are inevitable (e.g., dropping plates usually, but not always, causes them to break and waving at strangers might, but usually doesn’t, cause waving back). Even young learners are adept at taking these variations in stride when making causal inferences. For example, preschoolers can learn preferences from inconsistent choices (e.g., choosing A over B three times and B over A once, Hu et al., 2015) and causes from probabilistic evidence (e.g., a block lights up a detector 2/3 versus one that lights 1/3 of the time, Kushnir & Gopnik, 2005). However, past work does not reveal how children understand these variations in causal relationships or the systems that give rise to them. In particular, what do young learners think about why variations occur? Work in causal cognition suggests that an understanding of potential sources of variation is inherently captured by our representations of causal relationships (Woodward, 2010; Vasilyeva et al., 2018). Importantly, sources may be either external (variations due to changing contexts, e.g., plates dropped into water will not break) or internal (variations due to strength within contexts, e.g., dropping on the floor increases likelihood of, but does not guarantee, plates breaking). Behavioral responses in infancy suggest that learners as young as 16 months are equally capable of entertaining internal or external reasons for causal variability depending on which is supported by the evidence (Gweon & Schulz, 2011). But what if the evidence is ambiguous? Do learners have prior expectations about the underlying variability of causal relationships that can guide their thinking? For example, does it seem more likely I got a rare return wave because I encountered a particularly friendly stranger or because I waved more invitingly than usual? The current study examines whether young children appeal to internal or external explanations of inconsistent events. Four- to 6-year-olds (N = 90) watched either an agent or a machine repeatedly behave in one of two ways (see Table 1). At test, they saw the opposite behavior in a potentially different context and were asked whether it was due to internal properties/disposition or a change to the external context. We found a preference for internal explanations of inconsistencies for agents (68%, p = 0.02, two-tailed binomial), but not for machines (39%, p = 0.2). Intriguingly, while both older and younger children appealed to internal sources to explain agents’ inconsistencies, younger children had no preference when explaining mechanical inconsistencies (50% external) but older children did (66% external, p = 0.09). These results suggest young learners consider both inherent and contextual possibilities for variation in reasoning about causal relationships and gradually develop complex expectations about the nature of these relationships and likely sources of variation in different domains. Follow up experiments are currently underway to investigate how children trade-off between these prior expectations and unambiguous evidence, and will be included in the conference presentation.

⇦ Back to session