Running Head: Classifying with Essentialized Categories Classifying with Essentialized Categories

A central tenet of modern categorization research is that real-world categories have no defining features, and that features are better thought of as probabilistically related to their category. Nevertheless, the belief in properties that are invariably associated with categories has persisted. One view— psychological essentialism—further states that essential properties and structures constrain or generate the observable features of objects. We argue that categorization is often an instance of causal inference, as classifiers reason from observable features to essential ones (and then to category membership). We demonstrate that observable features viewed as causally related to underlying structures carry more weight on classification than unrelated features, and that this phenomenon is responsible for changes in some key macroscopic properties of categories, such as their variability. But because these inferences are made with less than complete certainty, we suggest that essentialism is compatible with a probabilistic view of classification, suitably amended. An account of these findings in the form of a generative model of classification is offered. Generative Essentialism 3 Classifying with Essentialized Categories A central dogma of modern categorization research is that real categories have no defining features. Once upon a time, investigators assumed that people determine category membership on the basis of a small number of defining features that covary more or less perfectly with category labels, and their research focused on how people form and test the rules that govern classification (Bourne, 1970). But the Roschian revolution in the 1970s demonstrating people's apparent ignorance of defining features, and the existence of unclear cases and better or worse examples of a category, led most of the field to the alternative view that features are better thought of as probabilistically related to their category (Rosch & Mervis, 1975). This view has come to dominate all major theories of category representation, regardless of whether they assume a single summary representation (e.g., a prototype) or the memory traces of all the exemplars of a category that a classifier has observed in the past. Nevertheless, the belief in properties that are invariably associated with categories has persisted in various forms. For example, Miller and Johnson-Laird (1976) distinguished between categories’ core properties (which could be used during, e.g., reasoning, determining the meaning of conceptual combinations, and so on) versus their identification procedures which inferred category membership on the basis of perceptual information (also see Armstrong, Gleitman, & Gleitman, 1983; Osherson & Smith, 1981; Smith, Shoben, & Rips, 1974). More recently, the view known as psychological essentialism also states that people view categories as having properties, or some kind of underlying structure, that is essential, that makes an object the kind of thing it is (Medin & Ortony, 1989; Gelman, 2003). This view differs from core properties in two important ways however. First, it does not assume that people necessarily know what those properties or structures are; thus, it is consistent with the apparent absence of real defining features. Second, it assumes that essential properties and structures constrain, or generate, the features of objects that can be observed. Evidence in support of psychological essentialism comes from a variety of sources. For example, Gelman and Markman (1986) demonstrated that children's inductive inferences reflect an assumption that members of the same kind are likely to share many hidden properties. When preschool children were told, Generative Essentialism 4 for instance, that a flamingo feeds its baby mashed-up food and that a bat feeds its baby milk, the children were more likely to judge that a blackbird would feed its baby mashed-up food (despite the fact that the blackbird was more perceptually similar to the bat than the flamingo), apparently because the children believed that objects are likely to share properties when they are the same kind (e.g., flamingos and blackbirds are both birds). A second line of evidence concerns children's beliefs about the innate potential of kinds. For example, preschool children judged that a baby kangaroo raised by goats would be more likely to be good at hopping than climbing (Gelman & Wellman, 1991). Finally, the assumption of underlying structure has also been shown to affect children's judgments of category membership. Keil (1989) performed a transformation experiment in which he told second-graders about, for instance, doctors who dyed a raccoon's fur black, bleached a white stripe down its back, and put a sac of super smell odor in its body. The children judged that the transformed animal was still a raccoon, despite that it now looked like a skunk (see Rips, 1989, for a similar finding with adults). In fact, evidence that the internal structure of animals is more important than the "outsides" has even been found with 3 year old children (Gelman & Wellman, 1991). The fact that all these findings involve young children suggests that their judgments were based on their default assumptions about the internal structure of the categories rather than specific knowledge (e.g., that raccoons have raccoon DNA) that might be acquired, for example, through formal education. But although these studies provide strong evidence that even children view categories as possessing internal structure, the second part of the essentialist claim—that internal structure generates or causes observable properties—is less well established. For example, the transformation experiments demonstrate that "insides" matter, but not that the insides are necessarily responsible for the outsides. It is also uncertain whether the effect of internal structure on classification is limited to unusual hypothetical situations (e.g., transformations) or whether it also extends to acts of classification that people perform everyday. Clearly, the case for the importance of essentialism would be strengthened if it were found to influence the classification of objects that have not undergone transformations. The purpose of the present research, then, is to assess how classification is affected when an object's normal (untransformed) features are known to be causally related to an unobserved property or Generative Essentialism 5 structure that is defining of, or essential to, membership in a category. Of course, one might start off wondering how such causes could have any influence at all, exactly because they are unobserved. Nevertheless, underlying causes might make their influence felt indirectly by changing how observable features are used in classification. In this article we assess the hypothesis that classifiers can engage in a kind of two-step inferential process in which they first reason backwards from observable features to the essential properties or structures that are their causes, and then from those essential properties to category membership. On this view, one "diagnoses" category membership in the same way that one diagnoses the presence of a disease from the presence of the symptoms it causes. We will refer to this proposal as the classification as diagnostic reasoning view. It may be that diagnostic reasoning of this kind is sometimes decisive in judgments of category membership. For example, an object is definitely not a category member if it is missing a feature that is always generated by the category, and an object definitely is a category member if it has a feature that is generated by no other category. However, we suggest that such causal inferences are frequently made with less than complete certainty. The conclusion against category membership for an object missing a feature will be less than certain if the categorizer recognizes that the causal process that generates the feature is less than 100% reliable (or if knowledge about its reliability is unavailable). Similarly, the conclusion in favor of category membership for an object displaying a feature will be less than certain if the categorizer suspects that the feature can be generated by the causal processes associated with some other category. In other words, although essentialism stipulates defining features, this does not entail that categorization decisions concerning essentialized objects are therefore all-or-none, because the observable features' diagnostic value, or evidentiary weight, will vary on a continuum depending on the details of the causal beliefs involved. Moreover, the importance of causal inference to categorization will also depend on how fully essentialized the category is, that is, the extent to which the internal properties and structures are indeed considered defining of category membership. In fact, research suggests that not all categories are essentialized to the same degree, or at all. The strongest evidence for essentialism is found with natural kinds, especially biological kinds (Gelman, 2003). There is also considerable evidence that social kinds Generative Essentialism 6 such as races and ethnicities are essentialized (Eberhardt & Randall, 1997; Hirschfeld, 1996; Rothbart & Taylor, 1992; Yulill, 1992). Evidence for essentialism is more ambiguous for artifacts, because whereas more complex artifacts like cars and computers possess considerable internal mechanisms, simple artifacts like pencils and wastepaper baskets do not. (Although it has been argued that an artifact's essential feature is the intention of its human designer, Bloom, 1998; see Malt & Johnson 1992; 1998 for a dissenting view). Clearly, one can expect the two-step process in which classifiers reason from observable features to internal structure and then to category membership to be operative only to the extent that internal structure is present and thought to be defining of category membership. When a category is essentialized only partly, or not at all, then a feature's evidentiary weight will be dominated by the factors already identified by current theories of categorization, such as the frequency with the features appear in category members and nonmembers (Rosch & Mervis, 1985) and their perceptual salience (Sloman, Love, & Ahn, 1998). In summary then, our analysis suggests that although essentialism reintroduces the concept of defining features back into the psychology of concepts, it does not predict that classification will revert to the all-or-none view of category membership associated with the classical view of concepts (Smith & Medin, 1984). Instead, we suggest that the classification of essentialized objects fits comfortably within a probabilistic view of classification in which the evidentiary weight that features provide for judgments of category membership varies on a continuum. But our proposal augments current theories by claiming that a feature's evidentiary weight is determined in part by how diagnostic it is of underlying structures that are viewed as defining of category membership. The goal of this article is to establish that a feature will serve as stronger evidence for category membership when it is viewed as causally generated by that category's essential properties. In each experiment adults were taught a pair of novel categories and then asked to choose which category a particular object is more likely to belong to. Novel categories were used to experimentally control which features are considered essential but unobservable, and which observable features are causally generated by the essential ones. For example, some participants learned about two species of ants, Kehoe Ants and Argentine Ants. For Kehoe Ants, the feature “blood high in iron sulfate” was stipulated as essential by Generative Essentialism 7 stating that it was common to all Kehoe Ants and to no other species of ants. For Argentine Ants, the feature “blood high in metallic sodium” was stipulated as essential by stating that it was common to all Argentine Ants and to no other species of ants. In addition, each species of ants had two or more observable feature. For Kehoe Ants, two observable features might be “hyperactive immune system” and “thick blood," whereas for Argentine Ants they might be “fast digestion” and “short life span." The features were "observable" because they were displayed, in various combinations, by the objects that participants subsequently classified, whereas no explicit information regarding the presence of the essential features in those objects was provided. Importantly, one or more of the observable features were described as being causally generated by the essential one. For Kehoe Ants, participants might be told that “Blood high in iron sulfate causes a hyperactive immune system," and about the causal mechanism linking those two features: "The iron sulfate molecules are detected as foreign by the immune system, and the immune system is highly active as a result.” Examples of causal relations relating essential and observable features are presented in Table 1 for Kehoe Ants and Argentine Ants. In addition to ants, five different pairs of categories were tested. Table 1 Example of experimental materials. Category Essential Feature Observable Features Causal Relations Kehoe Ants Blood high in iron sulfate Hyperactive immune system Thick blood Blood high in iron sulfate causes a hyperactive immune system. The iron sulfate molecules are detected as foreign by the immune system, and the immune system is highly active as a result. Blood high in iron sulfate causes thick blood. Iron sulfate provides the extra iron that the ant uses to produce extra red blood cells. The extra red blood cells thicken the blood. Argentine Ants Blood high in metallic sodium Fast digestion Short life span Blood high in metallic sodium causes fast digestion. Because metallic sodium is a digestive enzyme which facilitates nutrition extraction, high levels of metallic sodium results in fast digestion. Blood high in metallic sodium causes a short life span. Because metallic sodium gradually corrodes the valves in the ant's heart, its life span is shorter. Although explicitly defining essential features and their causal mechanisms in this manner controls the knowledge brought to bear during classification, note that these experimentally-defined Generative Essentialism 8 materials may differ in various ways from (people's beliefs about) some real category essences and their associated causal mechanisms. For example, although adults beliefs about essences are sometimes as concrete as "blood high in iron sulfate" (e.g., DNA functions as the essence for biological kinds for many Western educated adults), preschool children's knowledge about animals' essential properties is less specific, involving only a commitment to biological mechanisms that operate on their "insides" (Gelman & Wellman, 1991; Gelman, 2003; Johnson & Solomon, 1997). And we have mentioned how one's representation of an essence may involve a placeholder only (or even less, Strevens, 2000). But while the concreteness of one's essentialist beliefs is undoubtedly important under some circumstances, it is not critical to the classification as diagnostic reasoning view that only assumes that classifiers can reason backwards from observables through some sort of causal link to some sort of underlying property that is diagnostic of category membership. The assumption is that this inference can take place regardless of how concrete or abstract those essentialist beliefs are. A second difference is that an essential property is more than one that just happens to be present in all category members (and absent in all nonmembers)—it is more than a "contingent" fact about today's distribution of properties among categories and their objects. Rather, people believe that an essential property is one that is present in all category members that could exist. But again, this important distinction is orthogonal to the classification as diagnostic reasoning view. We argue that regardless of whether underlying properties are truly versus only contingently essential (or even only present in category members with high probability), classifications will be affected when those properties can be causally inferred from observable features. Thus, although our research has direct implications for essentialism, because it treats (true) essences as a special case, it is potentially applicable to a wider variety of situations. This article is organized into three parts. In the first part we test our central claim that observable features will be more diagnostic of their category when they are perceived as being caused by the category's essential properties. Experiments 1-6 will demonstrate that classification can involve a twostep process of reasoning from observable to essential features and then to a category label, and will rule out a number of alternative interpretations. In the second part we consider implications of the greater Generative Essentialism 9 diagnosticity of causally related features and how those features result in changes to two important macroscopic properties of categories. Experiment 7 will demonstrate that such features result in boundary intensification, the phenomenon in which the boundary between categories becomes sharper or more pronounced. Experiment 8 will show that they result in a change to participants' perception of how variable or homogenous a category is. Finally, with the results from Experiments 1-8 in hand, the last section evaluates the implications they have for theories of how causal knowledge affects categorization. Two models will be considered. The first model, the dependency model proposed by Sloman et al. (1998), states that features become more important to the extent they have many dependents (i.e., effects). The second model, the generative model proposed by Rehder and Kim (2006; Rehder, 2003ab), states that classifiers estimate the likelihood that a potential category member was generated by a category's causal model. After this evaluation we derive additional predictions that are tested in the final Experiments 9 and 10. PART 1: CLASSIFICATION AS DIAGNOSTIC REASONING Experiment 1 To conduct an initial test of the proposal that features are weighed more heavily when they are causally related to an essential feature, participants in Experiment 1 were taught two novel categories with features as shown in Figure 1. Category A had three features, one of which was described as essential (EA) and two of which were observable (A1 and A2). The first observable feature (A1) was described as being caused by EA but the second (A2) was not. Likewise, category B had one essential feature (EB) which was said to cause the second observable feature (B2) but not the first (B1). For example, of the participants who learned the two species of ants, half were told that Kehoe Ants' essential feature (“blood high in iron sulfate”) caused its first observable feature (“hyperactive immune system”) but not its second (“thick blood”), and that Argentine Ants' essential feature (“blood high in metallic sodium”) caused its second observable feature (“short life span”) but not its first (“fast digestion”). As mentioned, the categories were essentialized because EA and EB were described as occurring in all members of their respective category and no nonmembers. Observable features were associated with their Generative Essentialism 10 category by stating that they occurred in 75% of category members. To eliminate the possibility that any effects are due to the particular features and causal relationships involved, the assignment of Kehoe Ants and Argentine Ants to the roles of category A and B in Figure 1 was reversed for the other half of the subjects. Five other category pairs besides ants were tested. After learning about the two categories, participants performed classification tests in which they were shown two features, one from each category, and asked which category the object belonged to. For example, an item might have features A1 and B1, which we predict will be classified as an A, because from A1 one can reason to EA via the causal link that connects them, but one cannot so reason from B1 to EB. For a similar reason, an item with features A2 and B2 should be classified a B. We refer to those test items with features whose presence is affirmed as Positive items. We also presented Negative items with features that were stipulated as absent. For example, an ant might have a normal rather than a hyperactive immune system (which we denote as ~A1) and a normal rather than fast digestion (~B1). In this case, we predict the item will be classified as a B, because from ~A1 one can reason that EA is likely to be absent. The test items and predicted classifications are listed in Table 2. Category A