The Development of Diagnostic Inference About Uncertain Causes

Young children can engage in diagnostic reasoning. However, almost all research demonstrating such capacities has investigated children’s inferences when the individual efficacy of each candidate cause is known. Here we show that there is development between ages five and seven in children’s ability to reason about the number of candidate causes whose efficacy is unknown (Study 1). We also find development between ages six and seven in these abilities when children are presented with several uncertain candidate causes in an additive causal system (Study 2). These findings demonstrate how children’s diagnostic reasoning abilities develop beyond the preschool years and illustrate possible relations between children’s developing diagnostic inference and scientific reasoning capacities. How do we infer the causes of events? Such reasoning is commonplace: A nurse infers that a patient’s elevated temperature is the result of an infection. An investigator attempts to recreate the events leading up to an accident. Both of these cases involve reasoning diagnostically—from effects to their underlying cause or causes—in situations of uncertainty. The nurse must await test results to confirm a diagnosis. The investigator must weigh contradictory accounts of the same event. Moreover, each of these cases requires the reasoner to reevaluate conclusions based on new evidence. New information about the patient’s symptoms or about the reliability of the eyewitnesses could change the conclusion. As adults, we are capable of engaging in this kind of diagnostic reasoning (e.g., Fernbach, Darlow, & Sloman, 2011; Thomas, Dougherty, Sprenger, & Harbison, 2008). However, we do not know what the origins of these abilities are in childhood, nor do we know how children’s abilities develop into full-fledged abilities to reason about causal systems, especially when the efficacy of some causes is unknown. The current paper investigates how children reason in two different situations of uncertainty and how this ability develops over the early elementary-school years. One line of research has found that children possess early-emerging diagnostic reasoning capacities. Eight-month-olds use conditional probability information to make predictions (Sobel & Kirkham, 2006) and understand what underlying distribution in a population would likely produce an observed sample of data (Xu & Garcia, 2008). Slightly older infants can integrate these learning capacities with their existing knowledge CONTACT David M. Sobel [email protected] Department of Cognitive, Linguistic, and Psychological Sciences, Brown University, Providence, RI 02912, USA. Color versions of one or more of the figures in the article can be found online at www.tandfonline.com/HJCD. JOURNAL OF COGNITION AND DEVELOPMENT https://doi.org/10.1080/15248372.2017.1387117 © 2017 Taylor & Francis D ow nl oa de d by [ U ni ve rs ity o f Pe nn sy lv an ia ] at 0 8: 14 2 6 O ct ob er 2 01 7 (e.g., Denison & Xu, 2010; Gweon, Tenenbaum, & Schulz, 2010), which is one way they come to conclusions about the nature of object properties (Wu, Gopnik, Richardson, & Kirkham, 2011) or individuals’ preferences (Kushnir, Xu, & Wellman, 2010). By the preschool years, children can explicitly use temporal priority, spatial contiguity, and contingencies among events to diagnose whether certain kinds of causal relations are present (e.g., Bullock, Gelman, & Baillargeon, 1982). Preschoolers can also infer aspects of the causal structure of a set of events from conditional probability information (Gopnik, Sobel, Schulz, & Glymour, 2001; Schulz, Gopnik, & Glymour, 2007). Although these studies show that children can engage in some forms of diagnostic reasoning before they begin formal schooling, what is critical about all of these cases is that children observe or can deduce the efficacy of all of the potential causes. A paradigmatic example is work by Gopnik et al. (2001). In this experiment, preschoolers were introduced to a machine that lit up and played music when certain objects were placed on it (a “blicket detector” based on Gopnik & Sobel, 2000). Two objects (A and B) were each placed on the machine individually. Object A activated the machine and object B did not. The two objects were then placed on the machine together, which activated. Children in this study judged object A, but not B, to be efficacious. In this case (and many similar procedures, e.g., Schulz & Gopnik, 2004), children observed the individual efficacy of both candidate causes before being asked to diagnose which one was responsible for the machine’s activation when both objects were placed on the machine together. This suggests that children’s early-developing diagnostic reasoning capacities may be robust only when they know the individual efficacy of each candidate cause. In support of this suggestion, some prior research suggests that preschoolers have difficulty engaging in diagnostic inferences when they lack certain kinds of information about the causal system. For example, and in line with the work reviewed above, Bindra, Clarke, and Shultz (1980) found that fourand five-year-olds could hold multiple possibilities in mind when making diagnostic inferences about deterministic systems. When faced with multiple uncertain candidate causes (i.e., candidate causes whose individual efficacy was unknown), however, children’s inferences were at chance levels until age eight. These data suggest that children’s reasoning about uncertainty undergoes significant development during the early elementary-school years. Uncertainity, however, can be conceptualized in a number of different ways. In addition to the issue of whether individual objects have known or unknown efficacy, recognizing uncertainty in causal reasoning can also involve appreciating that causal relations may be stochastic (i.e., a cause does not always produce its effect) or appreciating that a particular cause produces different effects randomly. Children’s understanding of uncertainty about stochastic effects or the randomness of causal systems appears to improve between the ages of four and six (Beck, Robinson, Carroll, & Apperly, 2006; Hong, Chijun, Xuemei, Shan, & Chongde, 2005; Sobel, Sommerville, Travers, Blumenthal, & Stoddard, 2009). While we are not measuring these forms of uncertainty in the present study, it is interesting to note that children’s ability to reason about such phenemona potentially develops around the same time—a point we will return to in the General Discussion. Our focus in the current study is on children’s abilities to reason about potential causes when they do not know whether each individual cause is efficacious. As noted above, this ability seems to develop between ages four and eight. The general hypothesis that motivates the first study presented here is that children are more successful at diagnostic 2 D. M. SOBEL ET AL. D ow nl oa de d by [ U ni ve rs ity o f Pe nn sy lv an ia ] at 0 8: 14 2 6 O ct ob er 2 01 7 inference when they know whether each candidate cause is individually efficacious but are more likely to struggle with such inferences when one or more of the candidate causes has unknown efficacy. This hypothesis may seem inconsistent with cases in which preschoolers do show some success at reasoning about this kind of uncertain data (e.g., Buchanan & Sobel, 2011; Griffiths, Sobel, Tenenbaum, & Gopnik, 2011; Kushnir & Gopnik, 2007; Lucas & Griffiths, 2010; Schulz, Bonawitz, & Griffiths, 2007; Shultz, 1982; Sobel, Tenenbaum, & Gopnik, 2004). But what all of these studies have in common is that children have some other piece of knowledge that they can use to infer the efficacy of all the potential causes. For example, Sobel et al. (2004) showed children two objects (A and B) that activated the blicket machine together. Children were then shown that object A did not activate the machine by itself, but they did not observe object B on the machine individually. Children, however, could infer that B was a blicket and A was not in this condition based on the logical possibilities, and they likely used that information as the basis for subsequent causal inferences: When asked to make the machine activate, children chose only object B (see McCormack, Butterfill, Hoerl, & Burns, 2009, for similar results). In each case where children seem to be reasoning about uncertainty, then, other considerations involving logical inference, base rate information, or knowledge of causal mechanisms allow children to infer the efficacy of each potential cause individually. Moreover, there is some development here: Three-year-olds tend to struggle with these inferences, while older children are more successful (see Bonawitz et al., 2011; Sobel & Munro, 2009, for potential domain-specific mechanisms for this age-related change). To examine whether preschoolers struggle with diagnostic reasoning about uncertain candidate causes more directly, Fernbach, Macris, and Sobel (2012) compared how children engaged in such reasoning when the efficacy of all of the candidate causes was or was not known. They showed threeand four-year-olds three objects (A, B, and C) and a novel machine. Across two conditions, one object was shown to activate the machine (e.g., A✓) and a second object was shown not to activate the machine (e.g., B✗). What differed between the conditions was the status of a third object. In one condition, the third object was shown to activate the machine (C✓); in the other, it was never placed on the machine, so its efficacy was not known (C?). The machine and objects were then occluded from the children, and the experimenter activated the machine with one of the objects. The child was asked which object had just been used to activate the machine. Regardless of the object children chose, they were told they were incorrect and asked to make a second choice. In both conditions, children should avoid choosing the object that had previously failed to activate the machine (B✗) in response to the two test questions. When the efficacy of the third object was demonstrated (C✓), children tended to avoid choosing the inefficacious (B✗) object. When the efficacy of the third object was not demonstrated (C?), however, children often chose the inefficacious object (B✗). Three-year-olds’ responses did not differ from chance in this condition. Four-year-olds did generate a correct sequence of choices more often than chance expectations, but far from ceiling levels; they only avoided the inefficacious object about 45% of the time. In follow-up work, Erb and Sobel (2014) documented that children’s ability to treat the third (C?) object as a potential cause developed between the ages of four and seven, with fourto five-year-olds only making JOURNAL OF COGNITION AND DEVELOPMENT 3 D ow nl oa de d by [ U ni ve rs ity o f Pe nn sy lv an ia ] at 0 8: 14 2 6 O ct ob er 2 01 7 a correct sequence of responses about 69% of the time, while sixand seven-year-olds made a correct sequence of responses around 90% of the time. Here we extend these findings in two ways. First, in Study 1, we examine how individual children make diagnostic inferences when all causes are known and when the number of unknown causes is varied. This way, we can more specifically map the developmental trajectory of children’s abilities to make such inferences given that previous findings have only compared reasoning about causes with known and unknown efficacy at a group level. Investigating the role that uncertainty plays in causal reasoning can allow researchers to study causal reasoning not only as an abstract topic but also as one that is embedded in everyday life. For example, many studies suggest that young children, who are quite capable of the kinds of abstract causal inferences discussed so far, struggle with general scientific reasoning problems that require them to diagnose which of several potential causes with unknown individual efficacy produced a given set of results (e.g., Klahr, 2000; Klahr, Fay, & Dunbar, 1993; Koslowski, 1996; Kuhn, Amsel, & O’Loughlin, 1988; Kuhn, Garcia, Zohar, & Andersen, 1995; Kuhn, Schauble, & Garcia-Mila, 1992; Lehrer & Schauble, 2000; Masnick & Klahr, 2003; Samarapungavan, 1992). While there are several crucial differences between the tasks used in those studies and in the body of work reviewed above (an issue we consider in more detail in the General Discussion), Study 1 investigates how the presence of uncertain candidate causes influences children’s reasoning at different ages. This begins to address the gap between the literature on abstract causal inference and generic scientific reasoning problems. Second, in Study 2, we investigate whether children can make diagnostic inferences about uncertain candidate causes in an interactive causal system. In most studies of children’s causal reasoning, including our own Study 1, the parameterization of the causal model is disjunctive: An effect is produced if one or more candidate causes are present. These are the most commonly used type of model in experiments measuring children’s (and adults’) causal reasoning (as well as, potentially, in everyday life; see Lucas, Bridgers, Griffiths, & Gopnik, 2014, for a summary of this argument). In Study 2, we considered how children reasoned about a causal model with an additive element: Two candidate causes independently produced an effect but together combined to produce a different effect. Such additive (in this case, conjunctive) causality is commonplace in everyday and scientific reasoning. One example is how a trebuchet (a slinged catapult) fires its payload. The distance a payload travels is an additive function of the mass of the counterweight, the mass of the payload, and the length of the sling. For example, the heavier the counterweight and the lighter the payload, the farther the payload travels. Such inferences are necessary not only for medieval warfare (and better appreciating certain battle scenes in Game of Thrones) but also in everyday reasoning, such as when one is trying to predict how a golf ball will travel when struck, which is based on the same additive model as a trebuchet. Study 2 examines whether children could infer an additive causal model from data that do not disambiguate exactly how the causal system worked. Other studies have examined aspects of this question, but those studies have not tested children as young as those we investigate here, and they have also tended to present causal systems with some scientific content (another difference between our work and work in the field of scientific reasoning that we probe further in the General Discussion). For example, Schauble (1990, 1996) used 4 D. M. SOBEL ET AL. D ow nl oa de d by [ U ni ve rs ity o f Pe nn sy lv an ia ] at 0 8: 14 2 6 O ct ob er 2 01 7 several examples of additive causal models in her investigations (e.g., examining predictions about how depth of a canal and mass of a boat relates to the speed with which the boat travels or how the presence of certain features on a car predicts its speed). Similarly, Kuhn (2007; see also Kuhn & Dean, 2005) presented an additive model of levels of earthquake risk, in which the presence of multiple candidate causes added together to increase the risk of earthquakes. In these examples, fourth and sixth graders often performed poorly on learning additive structures from their own investigation. The goal of Study 2 is not to replicate these prior findings in scientific reasoning, as there are other critical differences among the present studies and those that measure scientific reasoning. Rather, we aim to examine whether children have the ability to diagnose causal efficacy from observed effects when the efficacy of candidate causes is uncertain. Further, taken together, these two studies shed new light on how children develop this crucial component of everyday reasoning, which is necessary both inside and outside of the lab.