Causal knowledge and the development of inductive reasoning

Weexplored thedevelopmentof sensitivityto causal relations in chil- dren’s inductive reasoning. Children (5-, 8-, and 12-year-olds) and adults were given trials in which they decided whether a property knowntobepossessedbymembersofonecategorywasalsopossessed by members of (a) a taxonomically related category or (b) a causally related category. The direction of the causal link was either predictive (prey ? predator) or diagnostic (predator ? prey), and the property thatparticipantsreasonedaboutestablishedeitherataxonomicorcau-salcontext.Therewasacausalasymmetryeffectacrossallagegroups, with more causal choices when the causal link was predictive than when itwasdiagnostic.Furthermore,context-sensitivecausalreason- ingshowedacurvilineardevelopment,withcausalchoicesbeingmost frequent for 8-year-olds regardless of context. Causal inductions decreasedthereafterbecause12-year-oldsandadultsmademoretax-onomic choiceswhen reasoning in the taxonomic context. Theseﬁnd- ings suggest that simple causal relations may often be the default knowledgestructureinyoungchildren’sinductivereasoning,thatsen- sitivity to causal direction is present early on, and that children over-generalize their causal knowledge when reasoning. (cid:2) 2014 The Authors. Published by Elsevier Inc. This is an open access and 8-year-olds) and adults about features of two artiﬁcial base creatures, followed by a target that was more similar to one base but shared a causal antecedent with the other base. Results indicated that when the causal link was explicit, all age groups preferred to make causal rather than similarity-based inductions. That is, they preferred to project a property to the target from the causally related base creature than from the more similar base creature. When


Introduction
Children make category-based inductions when they infer properties and features in novel categories based on what they know to be true about familiar related categories (for reviews, see chapters in http://dx.doi.org/10. 1016/j.jecp.2013.11.015 0022-0965/Ó 2014 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/3.0/). The effects of causal knowledge on reasoning are not very well captured by older models of category-based induction that emphasize featural similarity (Sloman, 1993) and/or class membership (Osherson, Smith, Wilkie, Lopez, & Shafir, 1990). Such similarity-based models are powerful at accounting for patterns of inductive reasoning about taxonomic properties (i.e., properties such as genes whose distribution in the population may depend on taxonomic relations) and about blank properties (i.e., properties that participants possess no knowledge about). However, they fail to capture induction across a broader variety of properties and in expert populations (see Medin, Coley, Storms, & Hayes, 2003;Rehder & Hastie, 2001;Shafto & Coley, 2003), especially when there is a causal explanation for the occurrence of shared properties (Rehder, 2006).
Causal knowledge plays a vital role in cognition from infancy onward (Sobel & Kirkham, 2007). The ability to understand causal structures provides children with tools that help them to successfully predict future events and understand the outcome of active intervention, allowing them to gain increasing control over their environment (Gopnik et al., 2004). By 4 years of age, children are capable of understanding simple causal mechanisms across the domains of biology (Wellman, Hickling, & Schult, 1997) and psychology (Flavell, Green, & Flavell, 1995) as well as causal explanations in social and physical domains (Hickling & Wellman, 2001). Similarly, children use causal knowledge to classify objects (Ahn, Gelman, Amsterlaw, Hohenstein, & Kalish, 2000) and natural kinds (Meunier & Cordier, 2009).
The fact that children make use of causal information across diverse domains and tasks underscores its potential importance in children's category-based reasoning. Indeed, evidence suggests that children can use causal knowledge when making inductive inferences. For example, Hayes and Thompson (2007) taught children (5-and 8-year-olds) and adults about features of two artificial base creatures, followed by a target that was more similar to one base but shared a causal antecedent with the other base. Results indicated that when the causal link was explicit, all age groups preferred to make causal rather than similarity-based inductions. That is, they preferred to project a property to the target from the causally related base creature than from the more similar base creature. When the causal relation was implicit, 5-year-olds did not yet show a preference for choosing the causally related items, unlike the older children and adults who made predominantly causal choices. Similarly, Opfer and Bulloch (2007) demonstrated that 5-year-olds were capable of ignoring perceptual similarity in favor of relational similarity when the latter had a causal antecedent but not when it was non-causal. Both of these studies suggest that children's category-based inductive reasoning may be affected by knowledge about causal relations, although they do not allow us to conclude that causal knowledge structures are the default. Moreover, such studies do not address the extent to which children's inductive reasoning is sensitive to the underlying structure of causal knowledge (e.g. Pearl, 2000;Sloman, 2005). Consequently, we cannot know whether children's use of causal knowledge in induction is mediated by the same underlying processes as in adults.

Structure of causal knowledge: Causal asymmetry effects
An important feature of causal knowledge is that it is directional, with causes always preceding or at least co-occurring with their effects (Waldmann, 2000). There is evidence that adults take such structural features into account because their causal reasoning tends to display distinctive asymmetry effects (Fenker, Waldmann, & Holyoak, 2005;Fernbach, Darlow, & Sloman, 2011;Sloman, 2005). Reasoning in line with how we experience cause-effect relations (predictive causal reasoning) appears to be less effortful than reasoning backward (diagnostic causal reasoning), suggesting that computational complexity is a key determinant of the causal asymmetry effect (Kahneman & Tversky, 1973). For example, when people are asked to verify whether two words are causally related, they respond faster when the words are presented in a predictive order compared with a diagnostic order (Fenker et al., 2005).
Causal direction also affects inductive inferences. Thus, people find category-based inductive arguments more convincing when reasoning in a predictive direction from cause to effect than when reasoning diagnostically from effect to cause (Medin et al., 2003;Shafto et al., 2008). To illustrate, people are more confident in the conclusion that bees have an unknown property given that it is present in the premise category flowers than when the roles of the two categories are reversed. People's judgments appear to accord well with the causal asymmetry effects that are predicted by formal models of causal-based property generalization (Rehder, 2009;Shafto et al., 2008), although there are arguments that, relative to a normative standard, people have too much confidence in arguments with a predictive structure (Fernbach et al., 2011). Despite the convincing evidence that the causal asymmetry effect is a robust phenomenon in adults' category-based reasoning, and evidence that children's predictive reasoning about novel mechanical systems is better than their diagnostic reasoning (see Bindra, Clarke, & Shultz, 1980;Hong, Chijun, Xuemei, Shan, & Chongde, 2005), it is unknown whether children display similar asymmetry effects when they use causal knowledge to support their inductive inferences. Our study was designed to answer this question.

Context-sensitive reasoning
Although it is clear that causal knowledge is important in induction, adults are adept at tailoring their inferences to the particular reasoning context. Consider the example with which we opened this article. If we ask a participant to decide whether lions have a certain gene on the basis that tigers have that gene, then the taxonomic relation that holds between lions and tigers is relevant. On the other hand, if we ask whether lions have a disease on the basis that antelopes have that disease, then the causal mechanism for disease transmission from antelope to lions via predation is relevant. There is much evidence that adults, and sometimes children, can show inductive selectivity by selecting the appropriate knowledge with which to evaluate category-based inductive arguments. For example, Heit and Rubinstein (1994) showed that people make stronger inferences about shared properties when the nature of the property is in accord with the nature of the relationship between two categories. According to Shafto, Coley, and Baldwin (2007); Shafto, Coley, and Vitkin (2007), changes in context affect the acute availability of different knowledge structures. Thus, when reasoning about properties such as diseases, people are more inclined to use knowledge about causal or ecological relations, whereas taxonomic knowledge tends to be invoked when reasoning about anatomical properties.
There is some evidence that children are also sensitive to the nature of the property that they are reasoning about (for a full review, see Hayes, 2006). Nguyen and Murphy (2003), for example, showed that by 7 years of age, children reason taxonomically about biochemical features but use thematic relations to guide inferences about situational properties. However, in their study younger children's use of different relations was heavily influenced by task format, making it difficult to evaluate the degree to which these children were using different knowledge structures in a systematic manner. Coley (2012) demonstrated that there are environmental influences on the age by which children show inductive selectivity; by 10 years of age, children raised in urban environments selectively project disease properties between categories that are ecologically related and so suggest a causal mechanism for disease transmission, whereas rural children show the same pattern as early as 6 years of age. In an open-ended category generation task, Vitkin, Coley, and Hu (2005) also showed that children's inferences are constrained by the nature of the property that they are asked to reason about. Thus, when told that the two categories share ''stuff inside,'' children generated categories based on similarity, whereas they gave interaction-based responses when reasoning about shared diseases.
In this study we manipulated the nature of the property that participants were asked to reason about. Although we expected adults to show property effects, our aim in this study was not to assess the age by which children demonstrate inductive selectivity because any answer to this question will depend on the paradigm employed. Indeed, inductive selectivity effects have been demonstrated in young children (Heyman & Gelman, 2000a;Heyman & Gelman, 2000b) and even infants (see Mandler & McDonough, 1998;Rakison & Hahn, 2004). Instead, we wanted to examine, in the domain of folk biology, how children reason before selectivity emerges and whether a particular knowledge structure is more important.
On the basis of the literature, it is very difficult to answer questions about default knowledge structures. There appears to be reason for assuming that taxonomic knowledge structures are the default; for example, using a speeded response paradigm, Shafto and colleagues (2007) showed that access to knowledge about causal mechanisms for disease transmission due to ecological relations between categories is restricted when adults are asked to respond quickly, whereas taxonomic knowledge seems unaffected by response time manipulations. This finding, that taxonomic knowledge is primary, is consistent with suggestions that taxonomic knowledge structures are the default (Kemp & Tenenbaum, 2009). However, there is also reason to believe that taxonomic knowledge structures are not the default; Rehder (2006), Rehder (2009) emphasized the primacy of causal knowledge over taxonomic knowledge, and there is evidence that young children prefer to use thematic knowledge rather than taxonomic knowledge when reasoning about categories (see Greenfield & Scott, 1986;Nguyen & Murphy, 2003;Smiley & Brown, 1979).
Because the developmental literature is divided on the question of how young children reason inductively, there are also important theoretical questions about the development of inductive reasoning to which questions of default knowledge are relevant. Sloutsky (2010) theorized that inductive selectivity depends on a maturationally late learning system. Sloutsky suggested that, prior to the emergence of inductive selectivity, children rely on knowledge acquired through a simple learning system that exploits co-occurrences and similarity. He cited evidence that before 7 years of age children rely largely on such similarity between base and target category to guide their inferences (Fisher & Sloutsky, 2005;Rakison & Lupyan, 2008;Sloutsky & Fisher, 2004). This would suggest that prior to the emergence of inductive selectivity, children ought to rely heavily on superficial similarity or category membership rather than on causal knowledge. However, there is evidence that even children as young as 5 years can ignore featural similarity when explicit causal knowledge is available (Hayes & Thompson, 2007;Opfer & Bulloch, 2007).
Here we examined whether children as young as 5 years prefer to reason on the basis of taxonomic or causal relations. This group was younger than the youngest participants that Coley (2012) found to display inductive selectivity. The contradictory findings in the literature made it hard to predict whether these children will prefer arguments with taxonomic or causal relations and whether they will be sensitive to causal structure. If young children's inductive inferences are dominated by taxonomic relations, then we would except to see few effects of causal knowledge. If, on the other hand, causal knowledge dominates early in reasoning development, then the question arises as to whether young children will exhibit causal asymmetry effects.

Participants
In total, 44 Year 1 primary school children (M age = 5.4 years), 40 Year 4 primary school children (M age = 8.4 years), 54 Year 8 secondary school children (M age = 12.6 years), and a control group of 26 adults from Durham University (M age = 24.6 years) in the United Kingdom took part in the experiment.
Year 1 and Year 4 students were recruited from a suburban state primary school in the North East of England. Year 8 students were recruited from a suburban state secondary school in the same geographical area. There were approximately equal numbers of male and female participants in the various age groups.

Materials
Participants reasoned about 12 trials, each consisting of a triad of categories: a base and two targets. On each trial, participants were presented with a base category, were told that it possessed a feature, and were asked which of the target categories they thought was most likely to share that feature. One of the targets was causally related to the base in the form of a food chain relation. This causally related target category always belonged to a different superordinate category from the base. The other target category was taxonomically related to the base category and was from the same superordinate category as the base. In the predictive condition, the causal direction was ecologically consistent. For example, the base category might be a banana and the causally related target might be a monkey. In the diagnostic version of this example, the directionality of the causal link was reversed so that the monkey served as the base and the banana served as the target. To ensure that children understood the causal mechanisms by which two categories might be connected, the nature of the causal relationship was limited to transparent food chains. An example with pictures, all of which were selected from Rossion and Pourtois (2004) database, is shown in Fig. 1.
The direction of the causal link was counterbalanced across items. For List A, the causal direction was predictive for Items 1 to 6 and was diagnostic for Items 7 to 12; for List B, the opposite was the case. Thus, although direction was a within-participants factor, participants never saw any of the causal pairs twice. The task was self-paced and run on a laptop using purpose-built software.
So that the causally related pairs could be presented in either a predictive direction (e.g., carrot ? horse) or a diagnostic direction (e.g., horse ? carrot), different categories served as the base (e.g., carrot or horse). Because the taxonomically related target category belonged to the same superordinate category as the base category, this meant that the taxonomically related target category was different on diagnostic and predictive trials. Thus, it was crucial to equate the similarity between the base instance and the taxonomic alternative targets across predictive conditions (e.g., carrot ? onion) and diagnostic conditions (e.g., horse ? sheep).
A group of 18 Durham University students rated the similarity on a scale from 1 (not at all similar) to 9 (highly similar) between pairs of categories presented verbally in a pretest. Only triads in which the similarity ratings were approximately equal across both conditions were selected (all paired-sample t tests had ps > .05), resulting in 12 triads. Mean similarity ratings by triad are presented in Appendix Table A1). Although we did not collect similarity ratings from children, there is evidence that children and adults perceive similarity relations in similar ways. For example, using materials pretested for similarity with adults by Osherson and colleagues (1990), López, Gelman, Gutheil, and Smith (1992) showed effects of similarity on reasoning in 5-year-olds that were identical to the similarity effects originally found in adults by Osherson and colleagues.
The strength of the causal relation between the base and target categories was examined in a separate pretest on 19 Durham University students who rated the extent to which properties might be transmitted between the base and both target categories in each triad on a scale from 1 (not at all) to 9 (a great deal). In all cases, ratings were stronger for causally related categories than for taxonomically related categories. These differences were statistically significant (p < .05) in 9 of 12 cases and were marginally significant in the remaining 3 cases (see Appendix Table A2 for ratings).

Procedure
Parents received written information about the study with an opt-out form in case they did not want their children to take part. Children gave their assent at the beginning of the session.
Participants were randomly allocated to either the cell or disease condition. The task was explained verbally, and written instructions were presented on-screen. Participants completed 2 practice trials before commencing the main task. They saw a base picture at the top of the screen and were told that the animal, plant, or object had an unfamiliar disease or special cells (e.g., ''This horse has a disease called talio/has talio cells inside''). A different artificial cell or disease name was used for each trial. The base picture was followed by two arrows pointing to two target pictures. Participants decided which of the two pictures was more likely to share the disease or special cells with the base category. The location of the causally and taxonomically related targets on the screen was randomized. Participants pressed 1 if they thought that the left-hand picture was more likely to share the property and pressed 9 if they chose the right-hand picture.

Results
For each individual, the proportion of causally related targets that were chosen was calculated for the six predictive and six diagnostic triad items. These were analyzed using a 2 (Direction) Â 4 (Age Group) Â 2 (Property) Â 2 (List) mixed-design analysis of variance (ANOVA) with direction as the only within-participants variable. For the item analysis, proportions of causal choices were averaged across items rather than participants.
All effects involving the counterbalancing list variable were nonsignificant (all ps > .10), so there is no further reference to this variable.
Post hoc comparisons using Bonferroni adjustments on the means involved in the significant effect of age group showed that 8-year-olds made more causal choices than both 12-year-olds (p = .05) and adults (p < .0005). There was no significant difference in the proportion of causal choices made by 12year-olds and adults (p = .17), and 5-year-olds made significantly more causal choices than adults (p = .014). Both 5-year-olds, t(43) = 2.69, p = . 001, and 8-year-olds, t(39) = 3.91, p < .0005, made more causal choices than predicted by chance alone.
The trend line in Fig. 2 suggests that causal induction shows a curvilinear development in the shape of an inverted U, which was confirmed with a polynomial contrast. This involves starting with the linear contrast and then sequentially carrying out higher power contrasts until the highest-power contrast is nonsignificant. There was a significant linear trend (p < .0005) as well as a significant quadratic term (p = .014). The cubic contrast was nonsignificant (p = .18). The quadratic effect across age group was also highly significant across items, F i (1, 11) = 15.7, p < .0005. This pattern confirms that the use of causal knowledge in inductive reasoning follows a curvilinear developmental trajectory, peaking at 8 years and decreasing thereafter. The results of the ANOVA contained a significant main effect of property, F s (1, 146) = 8.6, p = .004, effect size f = .24, F i (1, 11) = 60.9, p < .0005. Causal choices were significantly more frequent when participants reasoned about diseases (M = .60, SE = .033) than when they reasoned about cells (M = .47, SE = .033). However, this was qualified by a significant interaction between property and age group, F s (2, 106) = 6.2, p = .001, effect size f = .36, F i (3, 33) = 62.7, p < .0005. As Fig. 2 shows, both 4-and 8-year-olds made a similar number of causal inductions in both the cell and disease conditions (ps > .25, effect size d < 0.2), suggesting that they are not exhibiting strong inductive selectivity effects. In contrast, both adults (p = .031, effect size d = 0.8) and 12-year-olds (p = .0005, effect size d = 1.2) showed strong inductive selectivity effects, making significantly more causal inductions in the disease condition than in the cell condition.
Furthermore, the proportion of causal choices was similar across all age groups when reasoning about diseases (all pairwise comparison ps > .20). In contrast, when reasoning about cells, 4-and 8-year-olds made significantly more causal choices than both 12-year-olds and adults (ps < .022). The latter two age groups were not significantly different from each other (p = .64). Thus, it appears that the increased selectivity is driven by a developmental decrease in causal choices when reasoning about cells rather than a change in the application of causal knowledge when reasoning about diseases.
Our results appear to show that children do not display inductive selectivity until 12 years of age and that before that age they prefer to use causal knowledge when evaluating inferences. In addition, children as young as 5 years are as sensitive to causal structure as are adults. To rule out alternative accounts of children's preference for causally related targets, we carried out a number of posttests. One possibility is that the causally related category pairs were more strongly associated or co-occur more frequently than the taxonomically related category pairs. This would then give the impression that children were reasoning causally, whereas in fact the results would be explicable in terms of co-occurrence and associations (Sloutsky & Fisher, 2008). Remember that in our pretests we ensured that the different taxonomically related targets in the predictive and diagnostic conditions were equally associated with the base category, but we did not check whether the degree of association between the base category and each of the targets was approximately equal. Associations between words tend to be reflected in how we use language; words that co-occur more frequently are more strongly associated in semantic memory (Spence & Owens, 1990). One way to guard against the possibility that our materials are confounded by associative strength is to look at the frequency of cooccurrence between two words in a text or sentence (Church & Hanks, 1990;Ide & Veronis, 1998;Rapp & Wettler, 1991). The World Wide Web provides an almost infinite source of linguistic data. Using a Google proximity search specifying that the categories must co-occur in any order within a window of six consecutive words (Canas, 1990), we calculated the conditional co-occurrence of the taxonomic and causal category pairs using the following formula suggested by Heylighen (2001): In this equation, P(w 1 & w 2 ) represents the probability that a text contains both words w 1 and w 2 , and P(w 1 ) represents the probability that it contains w 1 on its own. To calculate the conditional probability, one can simply count the number of times w 1 and w 2 co-occur and divide this by the number of times w 1 occurs by chance in the same text sample. This is repeated for w 2 . We took the mean of the two conditional probabilities for each pair and calculated z scores. Using an independent-samples t test, we compared the mean co-occurrence index for the causally related category pairs (mean z score = À0.3, SD = 0.4) with the mean co-occurrence index for the taxonomically related category pairs (mean z score = 0.15, SD = 1.2). This showed that there was no significant difference between the two co-occurrence indexes, t = 1.68, df = 31.3 (adjusted for unequal variances), p = .10, effect size d = 0.5. However, as indicated by the medium effect size, if anything the taxonomic category pairs co-occurred slightly more frequently, strengthening our claim that children and adults were drawing on structured causal knowledge rather than simple associative knowledge.
Another possible alternative explanation of our results is that the youngest participants preferred causally related targets because they did not possess the knowledge about taxonomic relations required to recognize the strength of arguments based on such relations. To rule out this possibility, we checked taxonomic and causal knowledge levels in a separate sample of the same age as our youngest age group. A group of 20 5-year-olds were shown the color pictures of the category pairs used in the main experiment and were asked about biological group membership (''Do this [name of first category] and this [name of second category] belong to the same group?'') and causal relatedness (''Does this [name of first category] eat this [name of second category]?''). Paired-samples t tests across the 20 children and by items showed that there was no difference between the causal relatedness endorsement proportion (M = .78, SD = .13) and the biological group membership endorsement proportion (M = .79, SD = .17), t s (19) = À0.68, p = .51, effect size d = 0.15, t i (11) = À0.037, p = .97. Thus, it is unlikely that the 5-and 8-year-olds in the main experiment preferred causal targets because, relative to their causal knowledge, they lacked taxonomic knowledge.

Discussion
We had two aims in carrying out this study. First, we wanted to examine whether causal or taxonomic knowledge is the default prior to the age at which children develop inductive selectivity. Second, we wanted to see whether children are sensitive to aspects of causal structure when evaluating category-based inductive arguments. In showing that 5-and 8-year-old participants preferred arguments that were strong because of a causal mechanism for transmission via a predation relation even when the property reasoned about was a cell, our findings suggest that causal knowledge structures are the default in children's reasoning prior to the emergence of inductive selectivity. In addition, 5-year-olds are just as sensitive to causal direction as are adults. In fact, all age groups in our experiment endorsed significantly more predictive arguments than diagnostic arguments. Interestingly, causal reasoning seems to show a curvilinear development, peaking at 8 years of age and decreasing thereafter due to the emergence of inductive selectivity. Although we did not anticipate this last finding, as we show below, it has parallels in the literature on children's understanding of folk biology.
The pattern of results we observed has very interesting implications for theories about how inductive reasoning develops. Sloutsky (2010) suggested that ''adult-like'' inductive reasoning depends on a maturationally late learning system. This suggestion can explain why we did not observe inductive selectivity until the age of 12 years, driven by an increase in taxonomic inferences when reasoning about cells. Under Sloutsky's account, reasoning about cells might require more abstract conceptual knowledge about mechanisms that are not perceptually observable such as genetics and inheritance. However, this account also seems to suggest that prior to the emergence of inductive selectivity children ought to rely heavily on superficial similarity or category membership rather than on causal knowledge. Our data suggest otherwise, with 5-and 8-year-olds making the most causal inductions regardless of context. Alongside recent evidence from Hayes and Lim (2013), who showed that inductive selectivity even in relatively transparent contexts depends on conscious awareness of the relevance of contextual clues, the current findings cast doubt on the claim that early category-based induction is driven exclusively by simple learning mechanisms based on co-occurrence and perceptual similarity. At the very least, it requires a substantial downward revision of the age at which conceptual knowledge such as simple causal structures can support category-based inductive reasoning.
In addition to their tendency to prefer causal relations rather than taxonomic relations as the basis for inference, 5-and 8-year-old participants, like older participants, made more causal inferences when the causal relation between the categories was predictive than when it was diagnostic. The influence of different kinds of knowledge seems to be fundamentally related to the way in which this knowledge is organized in long-term memory (Fenker et al., 2005). Although causal knowledge itself does not have a homogeneous structure and can vary in complexity from direct cause-effect links to more complex common cause relations (Sloman, 2005), one might expect the abstract organization of knowledge about food chains involving simple causal transmission to be similar in adults and children. For example, children as young as 3 years show an understanding of the importance of causal order in which effects follow or co-occur with their causes (Bullock & Gelman, 1979;Bullock, Gelman, & Baillargeon, 1982). Reasoning about relations that are in line with this ordering of events might be cognitively simpler than needing to reason about possible causes given an outcome (Kahneman & Tversky, 1973). Shafto and colleagues (2008) suggested that causal asymmetry effects are driven by a multilevel understanding of causal knowledge-concrete knowledge about the existence of causal relations and more conceptual knowledge about how and when causal relations are most likely to warrant an inference from one agent to another.
The simple causal structure used in the current task might explain why some of our findings are different from previous results. For example, whereas the current study showed no age-related changes in the frequency of causal reasoning about diseases, work on ecological reasoning by Coley, Vitkin, Seaton, and Yopchick (2005) demonstrates a clear developmental trend in the use of non-taxonomic knowledge. They argued that this is driven by an experiential increase in ecological knowledge. It is conceivable that the structure of the causal relations underlying the ecologically related categories in Coley and colleagues' study (e.g., between tiger and parrot) was more complex, involving indirect causal pathways and common causes rather than simple and direct cause-effect relations. Thus, increased complexity may render non-taxonomic knowledge less available to reasoning processes in younger children.
There may be a number of reasons why the younger children in our experiment relied on causal knowledge rather than taxonomic knowledge when reasoning. The properties used required knowledge about two quite distinctive biological domains: (a) food chain relations and disease contagion and (b) taxonomic relations and genetics. One likely possibility is that knowledge of the genetic basis of shared properties, inheritance, and biological processes is less elaborate in 5-and 8-year-olds compared with 12-year-olds and adults (Au & Romo, 1999;Hatano & Inagaki, 1994;Hatano & Inagaki, 1997). Indeed, educational research suggests that children have conceptual gaps and erroneous concepts in their explanations for genetics (Lewis & Kattmann, 2004;Smith & Williams, 2007) that become more elaborate and accurate only with explicit instruction (Venville & Donovan, 2007). On the other hand, the mechanism by which members of a common food chain may transmit diseases is obvious and observable and constitutes a central part of the science curriculum for 8-year-olds. In line with other work showing that children like to have an explanation for phenomena (Callanan & Oakes, 1992), the children in our study may have preferred to base their reasoning on relations for which they have a mechanistic explanation and a more coherent theory (Gutheil, Vera, & Keil, 1998). Children as young as 3 years have a basic appreciation that invisible agents can cause illness (Kalish, 1996) and understand simple mechanisms by which contamination may come about (Siegal & Share, 1990;Springer & Belk, 1994), suggesting that children do have some naive but systematic theory about biological disease transmission.
One striking aspect of our results was that younger participants over-generalized their knowledge of disease transmission mechanisms. A relevant example of knowledge over-generalization comes from Keil and colleagues (1999), who showed that increases in knowledge about the mechanisms of disease transmission simultaneously led to more accurate inductions for physical illness contagion but less accurate reasoning about the causes of mental illness. Interestingly, the 8-year-old participants in Keil and colleagues' study made the most inaccurate transmission choices. This over-generalization parallels our current finding in which simple causal transmission is also seen as a basis for sharing physiological properties such as cells. Thus, inaccurate beliefs or inappropriate causal explanations seem to be supplanted only when children have more appropriate coherent explanatory systems at their disposal.
Because they show that causal knowledge dominates taxonomic knowledge in young reasoners, our results are contrary to the claim made by proponents of the structured Bayesian approach that taxonomic knowledge structures are the default (see Kemp & Tenenbaum, 2009;Tenenbaum, Kemp, & Shafto, 2007). Nonetheless, theory-driven Bayesian approaches have the potential to explain developmental changes in patterns of inductive reasoning. The domain-specific theories captured in Bayesian knowledge structures are minimalist and open to revision, both on encountering direct evidence and through verbal instruction. These characteristics of the Bayesian account, in conjunction with the observation that children are likely to have less well-developed domain-specific theories, may allow the Bayesian approach to account for the divergence between 8-year-olds' and adults' patterns of inductive selectivity. However, it aims to explain inference at a computational level rather than in terms of process; the Bayesian approach does not address the issues of availability and processing effort (Shafto, Coley, & Baldwin, 2007;Shafto, Coley, & Vitkin, 2007) that are likely to be important when explaining developmental changes in reasoning ability.
Finally, our research made use of non-blank properties such as cells and diseases, with inductive selectivity driven by the decrease in the use of causal relations when reasoning about cells. A future interesting question will be to explore the developmental use of causal knowledge for inductive problems that use blank properties. It seems likely that changes in the application of causal knowledge would be driven by cultural factors and expertise with causal choices likely to decrease in Western and/or urban samples, whereas the use of causal knowledge is likely to persist in non-Western and more rural samples (e.g., Coley, 2012;Lopez, Atran, Coley, Medin, & Smith, 1997). To summarize, young children strongly prefer inductive arguments where there is a plausible causal mechanism for property transmission, even where older children and adults appear to judge that mechanism irrelevant to the property to be generalized. Recruitment of this salient source of knowledge may be supplanted only by more contextually appropriate knowledge when formal education offers children an alternative feasible mechanism by which two categories come to share properties. In addition, even 5year-olds are influenced by the directional nature of causal relations. All of this suggests that causal relations are at least as important to young children's reasoning as they are to adults' reasoning and that, in some respects, the manner in which causal reasoning influences reasoning is similar across development.