Our central claim is that the causal status effect derives from the supporting role that causal information plays in explanation. Although this idea is a long-standing theme in the study of categorization , relatively little research has directly examined how explanatory goodness relates to categorization. In this paper, we study the relationship between explanatory goodness and categorization. We demonstrate that the magnitude of the causal status effect depends on the quality of the explanation in which the causal information is embedded. We first briefly review the causal status hypothesis and present arguments that the causal status effect depends on explanatory goodness. Then we present a study to support this claim. Finally, we discuss the implications of this study for categorization more broadly. Intuitively,plastic flower buckets wholesale some components of a concept do indeed appear more important for categorization judgments. As Medin and Shoben point out, it is much easier to imagine a robin that is not red than it is to imagine a robin that lacks the appropriate genetic structure of robins: imagining a robin with, say, zebra DNA, would make many other typical properties of robins implausible .
There is a long history of research on categorization demonstrating that some features associated with category members are particularly important for categorization judgments . One prominent perspective on centrality is based on the theory- or knowledge-based view of conceptual structure . According to this framework, knowledge of a category is, in important ways, like a scientific theory, comprising a “host of mental explanations” . In this view, category knowledge is seen, not in terms of a prototypical member or collection of exemplars, but rather in terms of “an explanatory principle common to category members . Accordingly, categorization is typically viewed as an inference to the best explanation. Murphy and Medin illustrate this idea with a well-known example: a man at a party who jumps into a pool fully clothed in a business suit would probably be classified as intoxicated, not because he is similar to the prototype or to instances of a “drunken behavior” category, but because being intoxicated is the best explanation for his behavior . Despite the importance of the theory-based view for orienting research on categorization and centrality, the framework leaves unspecified the specific constraints related to background knowledge that influence classification decisions.
One important response to this challenge is the idea that information that participates in causal relations is of greatest centrality . In Ahn’s causal status hypothesis, causes are weighted more heavily in categorization decisions than are the corresponding effects. For example, in the simple case of a single causal relationship between two features, such as having wings and flying , having wings would be given greater weight than flying, and would more greatly influence categorization decisions. Why should causes play a more important role than the corresponding effects in the underlying principles of a category? According to Ahn, et al.,causal properties may be seen as generating other features: e.g., DNA produces external features such as hair color. Features such as DNA may thus be regarded psychologically as most defining or diagnostic of category membership, because they form part of the essence or core of a concept. It follows, Ahn, et al. claim, that possessing the most central features would provide better evidence for category membership than more superficial features. While not disputing the empirical validity of the causal status effect, and the critical role of causal information for many if not most knowledge-dependent categorization decisions, we worry that the exclusive focus on causal knowledge risks obscuring the supporting role that causal information plays in explanatory processes.As Lombrozo notes, explanations can’t be reduced to just the supporting causal information, as explanations entail a set of factors that go beyond the causal information per se .
For one thing, there are many kinds of causal structures —e.g., causal chains versus feedback systems—that draw on different kinds of domain knowledge. An additional factor is the preference for simple or parsimonious causal explanations over those that would invoke more assumptions . In a recent study that illuminates the relationship between the causal status effect and explanatory structure, Lombrozo has shown that the strength of the causal status effect depends partly on the type of explanation— mechanistic or teleological —in which the causal knowledge is embedded. In one experiment, participants were presented with a novel category characterized by two features and were told that one feature causes the other . Participants were then asked to explain why the animal has the second feature , and to estimate the probability that an object missing one of the features was a member of the category. Participants who provided a mechanistic explanation showed a larger causal status effect than those who provided a teleological explanation . In addition to the type of explanation, the causal status effect may also depend on the quality of an explanation. We suggest that some of the puzzling results obtained in studies of the causal status hypothesis might be best explained in terms of explanatory goodness. In one such study , Ahn, Kim, Lassaline, and Dennis gave participants a standard and two alternatives. Participants were instructed to choose which of the two alternatives should be categorized with the standard. One standard read: “This object has a high-intensity light bulb and a pouch that can contain liquid because it was designed to kill bugs.” The first alternative shared the cause, but not the effects ; the other alternative shared the effects, but not the cause . Ahn, et al. predicted that people would prefer to match an alternative to the standard on the basis of the shared cause, rather than the shared effects. Although the overall results confirmed the prediction, inspection of the results reveals that only half the items conformed to the prediction. Table 1 shows the materials used in the study and, for each item, the percentage of choices for the matching cause. While there is a noticeable causal status effect for three of the items,black flower buckets two items exhibit chance responding, and one even suggests a preference for the shared effect alternative. We suggest that the item variability can best be understood in terms of the explanatory role of the causal property being asserted for each item. One hallmark of a good functional explanation is that the intended function is causally connected to the effects to be explained. Ideally, the facts can then be understood as subsumed under a general causal law . In addition, good explanations exhibit breadth—the extent to which an explanation accounts for most if not all of the available facts ; and depth—the extent to which the local explanation fits within a larger explanatory framework, and can itself be explained . These considerations can shed light on why some items in Ahn et al.’s study conformed to their prediction, while others did not. Consider the item with the strongest causal status effect, the first item . In this case, a clear, general causal relationship links the intention with the effects: it’s generally known that bugs are attracted to light, and that bug-killing agents typically take the form of liquid pesticides. Second, the explanation is broad and deep .
Now, consider the worst-performing item: “This painting has four pillars and is red because the painter intended to draw a dog.” The causal relation between four pillars and the intention to draw a dog is extremely weak, and there is no clear explanatory framework that would link the intention to paint a dog with any of the facts. Of course, these are after-the-fact suggestions. To make this account plausible, what is needed is a manipulation of explanatory goodness. In the current study, we test the hypothesis that the magnitude of the causal status effect depends on explanatory quality: we predict that the better the explanation in which causal information is embedded, the stronger the causal status effect. The design was a 2 × 2 mixed factorial with explanatory quality , manipulated within-subjects and task order,manipulated as a between-subjects counterbalancing variable.The materials consisted of 24 stimulus sets, each based on a standard “fact-explanation” pair, which comprised a set of facts joined with a potential explanation for those facts. In the explanation-rating task the facts were presented as short sentences describing a particular situation, which participants were instructed to assume as true; and the explanation was presented as a separate sentence describing a potential explanatory account of those facts. For example, one set of facts read: “This object has a high-intensity light bulb and a pouch that can contain liquid”; and the associated explanation read: “This object was designed to kill bugs.” For the categorization task, the facts and explanation were combined into a single sentence, following Ahn, et al. . Of the 24 standards, 16 were test items, and 8 were fillers. The fillers were constructed to minimize the likelihood of demand characteristics in the categorization task . Of the 16 test items, 6 standards were taken without modification from the materials used in the Ahn, et al. study described above. In addition to the six standards from Ahn et al.’s study, we constructed an additional ten items, five of which were designed to express good explanations; and five to express poor explanations. The good-explanation items were designed so that the explanation specified a general and plausible causal relation that accounted for all of the stated facts . . These additional ten standards were adapted from materials used in studies of category-based inference by Sloman and by Patalano, Chin-Parker, and Ross . The experiment consisted of two paper-and-pencil tasks, counterbalanced for order across participants: evaluating explanatory quality and carrying out a two-alternative forced-choice categorization task. For the explanation rating task, participants were given all 24 standards, each arranged on a separate page of the test booklet. The facts were presented at the top of each page in a separate sentence, with the candidate explanation presented immediately below the facts. Participants were instructed to read the facts and the explanation, and then to provide a rating of how good or satisfying they found the explanation to be on a scale from 1 to 7 . They were then instructed to briefly explain their response. The categorization task was identical to that of Experiment 4 in Ahn, et al. as described earlier.Specifically, it was a forced-choice task between two alternatives, in which participants were instructed to choose the one alternative that should be categorized with the target. Each triad—a standard together with both alternatives—was presented on a separate page of the test booklet, with the standard at the top of the page, and the alternatives below it. For the test items, one alternative shared the cause with the standard, but not the effects; the other alternative shared the effects, but not the cause. Sample items are presented in Table 2.The eight filler items were constructed to vary the pattern so as to minimize the possibility of task demands. To this end, for four fillers, the alternatives both shared the cause with the standard, but differed on the effects; and for the remaining four fillers, both alternatives shared the effects, but differed on the cause. For each participant, item order was randomized in both tasks; for the forced choice task, left-right presentation of alternatives was also randomized.Participants completed the tasks at their own pace: the categorization task required roughly 15 minutes to complete; the explanation-rating task, 25 minutes. The results of this study provide support for the main claim of this paper: that the causal status effect derives from the role that causal information plays in an explanation; and specifically, that the magnitude of the causal status effect depends on the goodness of the supporting explanation. This finding adds support to the position that explanatory structure is important in categorization . According to this explanation-based perspective, categorization is treated as an inference to the best explanation. This perspective suggests that a primary research focus should be on investigating properties that characterize the quality of explanations, and the role that those properties play in cognitive tasks, such as categorization. A possible objection to our study would be that we simply constructed explanations that were inconsistent with participants’ prior knowledge. Ahn et al., found that the causal status effect can be eliminated if the stated causal relationship contradicts background knowledge.