The Role of Causal Status versus Inter-Feature Links in Feature Weighting - eScholarship
Woo‐kyoung Ahn, Jessecae K. Marsh · Proceedings of the Annual Meeting of the Cognitive Science Society · 2006
The Role of Causal Status versus Inter-Feature Links in Feature Weighting Jessecae K. Marsh ([email protected]) Department of Psychology, Box 208205 New Haven, CT 06520 USA Woo-kyoung Ahn ([email protected]) Department of Psychology, Box 208205 New Haven, CT 06520 USA configuration of features in an exemplar is consistent with known causal relations of the category determines the exemplar’s membership likelihood (e.g., Rehder & Hastie, 2001). For example, Rehder’s causal model theory (2003) postulates that the likelihood that a category’s causal network produces a given exemplar determines category membership for that exemplar. Exemplars that preserve a category’s causal links would be viewed as better category members because they are more likely to be produced by the causal laws governing the category than exemplars that break such relationships. Using Rehder’s (2003) example, an animal that does not fly and yet still builds nests in trees would be less likely to be judged to be a bird than an animal that does not fly and builds nests on the ground. The two proposed mechanisms can make conflicting predictions for category membership. To illustrate this, let us consider again a category with X, Y, and Z, which forms a causal chain of X→Y→Z. Now consider two exemplars: one has Y and Z but not X (represented as 011) and the other has X and Z, but not Y (101). The causal status hypothesis predicts that 101 is a better member than 011, because X should be weighted more heavily than Y. However, 101 has two causal violations (X→Y and Y→Z did not occur), whereas 011 has only one causal violation (X→Y did not occur). Thus, the mechanism sensitive to preserving inter-feature causal links would predict that 011 is a better member than 101. Given that these two accounts for the role of causal knowledge on categorization can at times produce opposite predictions, it is important to understand which mechanism is more primary under what circumstances. Recent studies by Rehder and his colleagues (2003; Rehder & Hastie, 2001) found strong effects of inter-feature causal links in some of their experiments, to the extent that the causal status effect disappeared at times. (See the next section for details.) Such results can be taken to question the validity of the causal status hypothesis. To the contrary, we argue that individual feature weightings would be more crucial than weighting determined by inter-feature relations. Consider a category with just two prototype features, X and Y, where X causes Y. There are four possible exemplars: 11, 10, 01, and 00. The proposals make two conflicting predictions. First, although 10 and 01 have the same number of causal violations, the causal status hypothesis predicts 10 to be a better category member than 01. This prediction is based on Abstract Studies have found that the causal status of features determines what exemplars are considered good members of a category (see Ahn & Kim, 2000). However, this causal status effect was questioned in recent studies (Rehder, 2003; Rehder & Hastie, 2001), because the preservation of causal links of a category’s causal network was shown to play a significant role. We demonstrate in this study that these results are methodological artifacts arising from the use of unnatural wording of category attributes. Introduction Categories are believed to include not just a catalog of features, but also rich representations of the causal relations between features (e.g., Carey, 1985; Murphy & Medin, 1985). Recently, a number of proposals were made to specify the process of applying causal knowledge to categorization. In particular, two mechanisms have been proposed to describe how causal knowledge influences goodness-of-exemplar judgments (e.g., Ahn, 1998; Rehder, 2003; Sloman, Love, & Ahn, 1998). The goal of the current study is to re-assess the empirical support for each of these feature-weighting mechanisms. One proposal states that causal knowledge indicates a feature’s causal status, which, in turn, determines each individual feature’s weighting. Ahn and colleagues have proposed in their causal status hypothesis that, with all else equal, features which cause other features in the same category are weighted more heavily than features that are effects of other features (e.g., Ahn, 1998; Ahn, Kim, Lassaline, & Dennis, 2000). Consider a hypothetical category with three features (X, Y, Z), which form a causal chain such that X causes Y, which causes Z (X→Y→Z). X has the highest causal status, Y has the next highest, and Z has the lowest causal status. Thus, the conceptual centrality, or importance to the concept, according to the causal status hypothesis, would be in the descending order of X, Y, and Z. For instance, upon learning that Roobans’ eating of sweet fruits tends to cause Roobans to have sticky feet, which tend to allow them to climb trees, participants judged an instance missing “eating sweet fruits” to be the least likely member of Roobans, whereas an instance missing “climbing trees” to be the most likely member of the category Roobans (Ahn et al., 2000). The second proposed mechanism states that whether the