Inducing causal relationships from observations is a classic problem in scientific inference, statistics, and machine learning. It is also a central part of human learning, and a task that people perform remarkably well given its notorious difficulties. People can learn causal structure in various settings, from diverse forms of data: observations of the co-occurrence frequencies between causes and effects, interactions between physical objects, or patterns of spatial or temporal coincidence. These different modes of learning are typically thought of as distinct psychological processes and are rarely studied together, but at heart they present the same inductive challenge—identifying the unobservable mechanisms that generate observable relations between variables, objects, or events, given only sparse and limited data. We present a computational-level analysis of this inductive problem and a framework for its solution, which allows us to model all these forms of causal learning in a common language. In this framework, causal induction is the product of domain-general statistical inference guided by domain-specific prior knowledge, in the form of an abstract causal theory. We identify 3 key aspects of abstract prior knowledge—the ontology of entities, properties, and relations that organizes a domain; the plausibility of specific causal relationships; and the functional form of those relationships—and show how they provide the constraints that people need to induce useful causal models from sparse data.

The ability to derive predictions for the outcomes of potential actions from observational data is one of the hallmarks of true causal reasoning. We present four learning experiments with deterministic and probabilistic data showing that people indeed make different predictions from causal models, whose parameters were learned in a purely observational learning phase, depending on whether learners believe that an event within the model has been merely observed (“seeing”) or was actively manipulated (“doing”). The predictions reflect sensitivity both to the structure of the causal models and to the size of their parameters. This competency is remarkable because the predictions for potential interventions were very different from the patterns that had actually been observed. Whereas associative and probabilistic theories fail, recent developments of causal Bayes net theories provide tools for modeling this competency. Causal knowledge underlies our ability to predict future events, to explain the occurrence of present events, and to achieve goals by means of actions. Thus, causal knowledge belongs to one of our most central cognitive competencies. However, the nature of causal knowledge has been debated. A number of philosophers and

Several theories have been proposed regarding how causal relations among features of objects affect how those objects are classified. The assumptions of these theories were tested in 3 experiments that manipulated the causal knowledge associated with novel categories. There were 3 results. The 1st was a multiple cause effect in which a feature’s importance increases with its number of causes. The 2nd was a coherence effect in which good category members are those whose features jointly corroborate the category’s causal knowledge. These 2 effects can be accounted for by assuming that good category members are those likely to be generated by a category’s causal laws. The 3rd result was a primary cause effect, in which primary causes are more important to category membership. This effect can also be explained by a generative account with an additional assumption: that categories often are perceived to have hidden generative causes.

computation. In preparation. Address for correspondence:

The majority of psychological studies on causality have focused on simple cause-effect relations. Little is known about how people approach more realistic, complex causal networks. Two experiments are presented that investigate how participants integrate causal knowledge that was acquired in separate learning tasks into a coherent causal model. To accomplish this task it is necessary to bring to bear knowledge about the structural implications of causal models. For example, whereas common-cause models imply a covariation among the different effects of a common cause, no such covariation between the different causes of a joint effect is implied by a common-effect model. The experiments show that participants have virtually no explicit knowledge of these relations, and therefore tend to misrepresent the structural implications of causal models in their explicit judgments. However, an implicit task that only required predictions of singular events showed surprisingly accurate sensitivity to the structural implications of causal models. This dissociation supports the view that people's sensitivity to structural implications is mediated by running simulations on mental analogs of the causal situations.

www.elsevier.com/locate/cogpsych

Numerous studies of how people reason with statistical data suggest that human judgment often fails to approximate rational probabilistic (Bayesian) inference. We argue that a major source of error in these experiments may be misunderstanding causal structure. Most laboratory studies demonstrating probabilistic reasoning deficits fail to explain the causal relationships behind the statistics presented, or they suggest causal mechanisms that are not compatible with people’s prior theories. We propose that human reasoning under uncertainty naturally operates over causal mental models, rather than pure statistical representations, and that statistical data typically support correct Bayesian inference only when they can be incorporated into a causal model consistent with people’s theory of the relevant domain. We show that presenting people with questions that clearly explain an intuitively natural causal structure responsible for a set of statistical data significantly improves their performance. In particular, we describe two modifications to the standard medical diagnosis scenario that each eliminates the phenomenon of base-rate neglect, merely by clarifying the causal structure behind false-positive test results.

A theory of categorization is presented in which knowledge of causal relationships between category features is represented as a Bayesian network. Referred to as causal-model theory, this theory predicts that objects are classified as category members to the extent they are likely to have been produced by a category s causal model. On this view, people have models of the world that lead them to expect a certain distribution of features in category members (e.g., correlations between feature pairs that are directly connected by causal relationships), and consider exemplars good category members when they manifest those expectations. These expectations include sensitivity to higher-order feature interactions that emerge from the asymmetries inherent in causal relationships.

Competence and performance in causal learning The dominant theoretical approach to causal learning postulates the acquisition of associative weights between cues and outcomes. These associative weights reflect the amount of covariation between the learning events. In the past few years, the associationist approach to causal learning has been criticized by a number of researchers

. This paper describes an approach to acquire qualitative and quantitative knowledge from verbally stated models in complex, probabilistic domains. This work is part of the development of an intelligent environment, MEDICUS 2 , that supports modelling and diagnostic reasoning in the domains of environmental medicine and human genetics. These domains are two yet new subdomains of medicine receiving increasing research efforts, but still consisting of largely fragile and uncertain knowledge. In MEDICUS, uncertainty is handled by the Bayesian network approach. Thus the modelling task for the user consists of creating a Bayesian network for the problem at hand. But since we want mathematically untrained persons to work with MEDICUS, the user may alternatively state propositions verbally and let the system generate a Bayesian network proposal. This differs from existing reasoning systems based on Bayesian networks, i.e. in medical domains, which contain a built-in knowledge bas...