Expertise and Category-Based Induction 1 Running Head: EXPERTISE AND CATEGORY-BASED INDUCTION Expertise and Category-Based Induction

This paper examines inductive reasoning among experts in a domain. Three types of tree experts (landscapers, taxonomists, and parks maintenance personnel) completed three reasoning tasks. In Experiment 1, participants inferred which of two novel diseases would affect “more other kinds of trees” and provided justifications for their choices. Experiment 2 used modified instructions, asking which disease would be more likely to affect “all trees.” Experiment 3 eliminated the conclusion category altogether, asking participants to generate a list of other affected trees. Among these populations, typicality and diversity effects were weak to non-existent. Instead, experts’ reasoning was influenced by “local” coverage (extension of the property to members of the same folk family) and causal/ecological factors. We conclude that domain knowledge leads to the use of a variety of reasoning strategies not captured by current models of category-based induction. Expertise and Category-Based Induction 3 Expertise and Category-Based Induction Cognitive psychologists are increasingly interested in conceptual functions beyond categorization (e.g., Barsalou & Hale, 1992; Markman, Yamauchi, & Makin, 1997; Pazzani, 1991; Ross, 1996, 1997; Wisniewski, 1995). Particularly, they have focused on the use of categories in reasoning and have proposed a number of formal models of category-based reasoning (e.g., Heit, 1998; McDonald, Samuels, & Rispoli, 1996; Osherson, Smith, Wilkie, López, & Shafir, 1990; Sloman, 1993; Smith, Shafir, & Osherson, 1993). These models have been developed to account for patterns of reasoning observed in experimental psychologists’ favorite research population, American undergraduates. Whether these patterns, and hence these models, generalize to a broader population is open to question. In this paper, we are particularly interested in the effect of extensive domain knowledge on reasoning. To explore this issue, we examine expert reasoning on inductive tasks and evaluate the ability of existing models to account for experts’ reasoning behavior. In general, category-based induction requires that information about one set of categories be used to make inferences about another category. A set of premises establishes that one or more categories possess a certain property. The premises are followed by an assertion (the conclusion) that a target category also possesses that property. To ensure that such inference tasks do involve induction and not simply knowledge retrieval, researchers have typically employed novel, socalled "blank" properties. This puts the emphasis on reasoning about the categories. For example, people might be told that scientists have discovered a new disease that affects horses and be asked to judge whether it would affect all mammals. Experiments frequently require participants to judge which of two complete arguments is stronger. Using this procedure, Rips (1975) observed typicality effects---inferences from typical category members were stronger than inferences from atypical category members (see also Osherson et al., 1990). For example, consider the following arguments: (i) Robins are susceptible to disease A. Expertise and Category-Based Induction 4 Therefore, all birds are susceptible to disease A. (ii) Turkeys are susceptible to disease B. Therefore, all birds are susceptible to disease B. In the US, robins are seen as more typical than turkeys. Accordingly, when reasoning about unfamiliar properties, undergraduate subjects rate (i) as higher in inductive strength than (ii). The diversity phenomenon involves inferences from pairs of premises. Consider the following pair of arguments: (iii) Robins are susceptible to disease X. Sparrows are susceptible to disease X. Therefore, all birds are susceptible to disease X. (iv) Cardinals are susceptible to disease Y. Turkeys are susceptible to disease Y. Therefore, all birds are susceptible to disease Y. Cardinals and turkeys represent a more diverse set of birds than robins and sparrows do. For problems like this undergraduates usually rate arguments like (iv) as stronger; greater diversity among the premises leads to greater confidence in the general conclusion. One very successful model of these effects is Osherson et al.’s (1990) similarity-coverage model (SCM). The SCM explains typicality effects in terms of coverage, defined as the similarity between the premise category and members of the lowest-level category that includes both the premise and conclusion categories. In evaluating arguments (i) and (ii), the average similarity (or coverage) of robin to sampled instances of the conclusion category bird is compared to turkey’s coverage of bird. Because more birds resemble robins than turkeys, (i) is rated stronger than (ii). Thus, the typicality phenomenon results from the tendency of typical items to have high average similarity to other category members. Note that by this definition, good coverage corresponds to high central tendency: an item that is highly similar to other members of a category will have a high coverage score and will also be judged to be very typical. Typicality ratings have frequently Expertise and Category-Based Induction 5 been interpreted as estimates of central tendency, but see Barsalou (1985) and Lynch, Coley & Medin (in press) for another perspective. The SCM also accounts for the diversity phenomenon in terms of coverage. In this case the principle is average maximal similarity; for each sampled instance, the most similar of the two premise categories gets entered into the calculation of average similarity. According to the SCM, argument (iv) is stronger than (iii) because cardinals and turkeys are highly similar to a wider range of birds than robins and sparrows are. In other words, in (iii) sparrow adds little coverage to robin because birds that are very similar to robins will also be very similar to sparrows. In contrast, turkey adds a fair amount of coverage to cardinal because turkeys are similar to large, edible birds (e.g., pheasants, chickens, ducks, etc.) that are quite different from cardinals. In brief, cardinal and turkey cover the category bird better than robin and sparrow. Thus, the SCM handles diversity effects with the same mechanism, coverage, that it uses to explain typicality effects. Although this paper will be focusing on predictions made by the SCM, it is worth describing a second model that is also admirable in the level of detail it brings to the study of inductive reasoning. Sloman’s Feature-Based Induction Model (FBIM, 1993) seeks to explain the same set of phenomena, but uses a connectionist mechanism. As just described, the SCM explains typicality and diversity effects by computing similarity between premise categories and sampled members of a more general category that encompasses both conclusion and premises. In contrast, the FBIM makes direct comparisons between features of the premise and conclusion categories. The argument with greater featural overlap between premise and conclusion is rated to be stronger. To decide between (i) and (ii) above, the FBIM would directly compare features associated with robin and turkey with those associated with bird. Presumably robin would share more features with bird, and so (i) would be a stronger argument than (ii). Although the calculation of argument strength for dual premise items is a bit less intuitive, it roughly amounts to the proportion of the conclusion category’s features that the premise categories possess. These models differ in terms of the precise mechanisms they use to explain induction, and in some cases make different predictions (see Sloman, 1993, 1998). However, they both rely Expertise and Category-Based Induction 6 heavily on the notion that similarity among categories--be it computed in a global way or based on featural overlap--is the principal mechanism driving inductive reasoning. The importance of similarity in induction has garnered support from research with undergraduates from American universities (e.g., Osherson et al., 1990; Sloman, 1993, 1998; Smith et al., 1993). The generality of these findings has been implicitly assumed but rarely tested (see Choi, Nisbett & Smith, 1997, for an exception). However, recent findings with different populations have raised questions about the hegemony of similarity in induction. López, Atran, Coley, Medin, and Smith (1997) studied category-based induction among the Itzaj Maya, an indigenous Amerindian population in central Guatemala. Among the Itaj, López et al. observed typicality effects in reasoning, but found no evidence of diversity effects. In fact, in some cases the researchers found reliable negative diversity, where the less diverse set of premises was preferred. This dissociation of typicality and diversity effects is surprising from the perspective of the SCM because it attributes both typicality and diversity to a common mechanism, coverage. López et al. (1997) suggest that the knowledge base of the participants and the kind of properties used may account for the absence of diversity-based reasoning. In their study, López et al. asked the Itzaj, who subsist largely on hunting and farming, to make inferences about novel diseases affecting different kinds of mammals. In their responses, the Itzaj revealed a great deal of practical knowledge about local animal species. Their answers on the induction task frequently depended on arguments about mechanisms that might spread disease, rather than on similarity among categories. These results suggest that participants with extensive knowledge in a domain may employ a variety of strategies--involving causal knowledge of the properties and kinds involved as well as considering taxonomic similarity--when solving reasoning problems. Thus, the findings of López et al (1997) raise important questions about the generality of models of induction that rely solely on similarity-based mechanisms. Of course, there are myriad cultural differences between the Itzaj and the undergraduates, which may account for the observed differences in reasoning. To evaluate the possibility that expert knowledge may lead to Expertise and Category-Based Induction 7 different patterns of inductive reasoning, the present studies focus on North American tree experts reasoning about trees. Like the Itzaj, this group has extensive category-relevant knowledge, but like undergraduates, they come from a mainstream American cultural background. As an aside, some readers may object that the unfamiliar diseases used by López et al. (1997) did not constitute truly “blank” properties for the Itzaj. Their knowledge base may have enabled them to reason by analogy, using what they know about the epidemiology of real diseases in order to reach conclusions about hypothetical ones. A similar concern would hold for U.S. tree experts. We believe that this objection misses the point of what models of induction are trying to do, which is to account for how people go beyond their present knowledge to reason about novel properties, events, and categories. If people with knowledge in a domain extrapolate from what they know, then models of induction should describe the extrapolation process (see Heit, 1998, for an example). Surely inductive reasoning is not confined to knowledge-poor domains; rather, we expect it to be most powerful when used in conjunction with a rich knowledge base. The present studies have two principal goals. First, we wanted to explore typicalityand diversity-based reasoning in another expert population in an attempt to expand upon and clarify the provocative results of López et al. (1997). It may be that their findings are due to some unidentified cultural artifact, and that U.S. experts will exhibit reasoning patterns quite similar to those of U.S. undergraduates. If so, we should expect that coverage will do a good job of predicting experts’ reasoning. Alternatively, experts may draw on a set of reasoning strategies completely unrelated to coverage. If extensive category-relevant knowledge leads to alternative content-based reasoning strategies, then tree experts may be more like the Itzaj Maya than like undergraduates. That is, coverage-based responding may be diminished or absent in tree experts, who may instead show causal/ecological reasoning. A second question of interest was whether different kinds of tree experts would exhibit different reasoning behavior. Medin, Lynch, Coley, and Atran (1997) found different patterns of spontaneous sorting among three groups of tree experts: landscapers, taxonomists, and parks maintenance personnel. Landscapers differed from the other two groups by tending to sort in Expertise and Category-Based Induction 8 terms of goal-derived categories (good street trees, shade trees, specimen trees, etc.). Parks personnel and taxonomists used morphological properties, though the two groups differed in which properties were most salient. Medin et al. found that for parks personnel and taxonomists the sorting data was a good predictor of performance on reasoning tasks. Given these findings, we were interested in the degree to which type of expertise affects inductive reasoning. Based on prior work, we expect that individuals will draw on what knowledge they have to recruit a variety of reasoning strategies. We do not expect that a single type of strategy (i.e., coverage-based) will be the most natural solution for all problems. In addition, we expect that because experts vary in the knowledge they bring to bear on a problem, different kinds of experts will exhibit different reasoning patterns. Both outcomes pose potentially serious problems for existing models of category-based induction, which, although they can account for the diversity phenomenon, are unable to provide a satisfactory account for the diversity of reasoning strategies that experts bring to such tasks. Experiment 1 The SCM explains both typicality and diversity effects in reasoning via the mechanism of coverage. Central to this paper is the question of whether coverage can adequately predict reasoning behavior across a variety of populations. López et al. (1997) found that these effects are not universal in human reasoning, but did not directly assess the role of coverage in Itzaj Maya induction. The goal of Experiment 1 was to assess the degree to which U.S. experts evaluated argument strength on the basis of coverage, as predicted by the SCM. We use a forced-choice paradigm in which participants indicated which of a pair of arguments provided better support for a general conclusion. We presented participants with singleand dual-premise items (corresponding to typicality and diversity effects) to examine the impact of coverage on reasoning. In contrast to examples (i) (iv) given above, participants were given two sets of premises, and asked which provided a better basis for generalizing to a common inclusive category. Specifically, we described two hypothetical diseases, which were known to Expertise and Category-Based Induction 9 affect different species of trees, and asked, “which disease would affect more of the other kinds of trees found around here?” The premise sets varied in their coverage of the category tree. We also asked experts to explain their reasoning; analyses of those justifications serves to illuminate the choice data. Of interest is the degree to which experts’ choices correspond to the coverage-based prediction, and whether types of experts differed in their patterns of choices and justifications. To preview, we find that experts’ reasoning is not well predicted by coverage as construed by the SCM. Experts do not seem to be calculating coverage over the entire conclusion category (tree). Instead, it appears that experts view disease as spreading within smaller taxonomic groups, such as families of plants. As a result, they frequently base their reasoning on what we call “local coverage,” which roughly constitutes a form of coverage, but based on a subset of the conclusion category. To distinguish between these two senses of the term, from here on we will refer to coverage as described by the SCM (i.e., calculated over the category bridging the conclusion category and premises) as “global coverage.” Method Participants Two women and 21 men (mean age 46 years, range 30 to 74 years old) having occupations related to trees completed Experiment 1. All were participants in an on-going research project investigating tree expertise (see Medin et al., 1997; Lynch et al., in press) and were paid for their participation. The average amount of experience dealing with trees was 22 years and ranged from 5 to 57 years. Seven participants had completed high school, four had some college work, nine had completed college, and three had advanced degrees (including two with doctorates). Almost all had some tree-related training. Participants were drawn from a variety of sources, including the Morton Arboretum, several private tree maintenance companies, and two Chicago-area parks districts (Evanston and Skokie). Participants had a range of occupations, but earlier work with these tree experts (Medin et al., 1997) had identified taxonomists, landscapers, and parks maintenance personnel as showing Expertise and Category-Based Induction 10 distinct patterns of spontaneous sorting (see Medin et al., 1997, for details on differences among groups). Taxonomists are scientists (primarily working at institutions like the arboretum) who conduct research on trees, as well as doing some teaching and other educational activities. Landscape workers focus on selecting appropriate plants for installation based on the aesthetic and utilitarian characteristics of trees. Maintenance personnel are engaged in planting, removing, and generally caring for city trees. Although we use these groupings as a form of shorthand, this partitioning represents an oversimplification. In practice, the three categories may not be mutually exclusive. For the purpose of this study, experts were classified based on their performance on two identical hierarchical sorting tasks, one run as part of a previous interview session with each participant, and a second performed immediately prior to experiment 1. Both sorting tasks followed the same procedure, reported in full detail in Medin et al. (1997). Participants reviewed cards with the names of 48 trees printed on them. The set of trees was selected to be broadly representative of trees in the Chicago area and covering a broad spectrum of scientific taxa. Most were native trees, but some were common introduced species. The experimenter gave instructions to “Put together the trees that go together by nature into as many different groups as you’d like.” After sorting the cards, the participant provided labels or explanations for each of the groups created. The process (combining piles of cards and labeling the resulting category) was repeated until the participant was no longer willing to combine groups. The experimenter then recreated the piles resulting from the initial sorting and gave instructions to “Split as many of the groups as you’d like into smaller groups of trees that go together by nature.” Again, labels were offered for the resulting groups and the process was repeated until no further divisions made sense to the participant. This procedure yielded a different hierarchical taxonomy of trees for each expert. These sortings were used to classify the experts and yielded nine landscapers, eight parks maintenance personnel, and six taxonomists. Expertise and Category-Based Induction 11