Over the past three decades, Patricia Cheng has made highly original contributions to the field of cognitive science through her work on the psychology of thinking and reasoning. In the 1980s, Cheng and Holyoak (1985) demonstrated that human reasoning, although apparently deficient for apparently simple problems relative to the norms of formal logic, is in fact highly robust when the content (even if unfamiliar and abstract) evokes certain types of socially meaningful contexts. In particular, people (preschool children as well as adults; Chao & Cheng, 2000) appear to be well-equipped to reason about conditional permissions and obligations. This empirical work was the basis for the theory of “pragmatic reasoning schemas”, which provides a middle ground between theories of reasoning based on domain-independent mechanisms and theories based on highly specific world knowledge. This line of research also demonstrated that teaching of logic can be improved if instruction is tied to concepts at this intermediate level of generality (Cheng et al., 1986; Nisbett et al., 1987). This work identified a link between the knowledge representations used in reasoning and those used in everyday social understanding.

Since the early 1990s, Cheng has conducted a sustained programmatic combination of empirical and theoretical research directed at understanding causal learning and reasoning. As Hume emphasized, causal relations are not available to direct perception (e.g., our senses cannot tell us that one billiard ball launches another billiard ball after colliding with it—our senses can only tell us that when a billiard ball moves into contact with another, the latter moves away). Causal relations must necessarily be inferred. But although particular cause-effect links are neither innately known nor directly perceived, the system that acquires them is indeed likely (as Kant emphasized) to operate according to domain-independent innate constraints. These constraints are likely to be specialized to capture properties of causal relations in the world useful to an intelligent organism, because knowledge of cause-effect relations in the world is of tremendous functional significance to any intelligent organism.

In her early work on causal learning, Cheng and her collaborator Laura Novick developed the probabilistic contrast model (Cheng & Novick, 1992). This model was closely related to the classic “delta-p” model of contingency learning, as well as to Kelley’s “ANOVA model” of causal attribution in social reasoning. The basic empirical phenomenon was that causal strength estimates are closely related to the degree to which the presence of a potential cause “makes a difference” to the probability of the effect (where delta-p, or probabilistic contrast, is defined as the difference between the probability of the effect conditional on the presence versus absence of the cause).

However, rather than focusing solely on the successful predictions of the putatively normative delta-p model, Cheng’s research revealed a set of boundary conditions where its predictions fail (Cheng, 1993, 1997). For example, causal strength estimates systematically deviate from delta-p when the base rate of the effect varies, and become unrelated to delta-p at extreme values of the base rate (e.g., ceiling effects arise when the effect is invariably present in the absence of the cause). Moreover, Cheng and colleagues demonstrated a number of striking asymmetries in causal learning depending on whether the cause is generative (makes the effect more probable) or preventive (makes the effect less likely).

Based on extensive empirical findings and insights in the psychological and philosophical literatures, Cheng proposed the “power theory of the probabilistic contrast model”, or power PC theory (Cheng, 1997). The basic assumption of her theory is that rather than interpreting probabilistic contrast as a direct measure of causal strength, people (and possibly some non-human animals) use the observed contingencies as evidence concerning the influence of unobservable causal links. For the important special case of cause and effect variables that take on binary values (present or absent), it follows (under the default assumption that each individual cause operates on the effect independently) that the net influence of multiple generative causes will follow a noisy-OR function, whereas the net influence of a generative and preventive cause will follow a noisy-AND-NOT function. Later work extended the theory to situations involving interactive causes (Novick & Cheng, 2004) and to other types of causal queries (e.g., causal attribution, diagnostic inference), relating the power PC theory to (and contrasting it with) Pearl’s (2000) analysis of causal sufficiency and necessity.  

Following the contingency-learning tradition in psychology, one of the causes in these noisy logical functions is an abstract composite cause that includes unknown or unobserved causes in a variable context. The partitioning of all causes of an outcome into the candidate causes and this contextual cause may be a distinct property of intelligent biological systems that enables efficient causal learning and category formation, despite the limited processing capacity of such systems (Carroll & Cheng, 2010).

As a novel normative theory of causal learning, which has also been proposed as a descriptive theory, the power PC theory has been highly influential. Cheng’s theory has led to spirited empirical and theoretical debates with proponents of associative accounts of causal learning. In recent years the basic integration rules derived from the theory (noisy-OR and noisy-AND-NOT) have been incorporated into Bayesian models of causal learning (Griffiths & Tenenbaum, 2005; Lu et al., 2008), including sequential learning models (Danks et al., 2003). Her work on the role played by causal learning in guiding category formation (Lien & Cheng, 2000) has also influenced recent computational models (Kemp, Goodman & Tenenbaum, 2010). 

Over her career, Cheng’s research has linked classical philosophical analyses of causality with elegant psychological experiments in the service of formal theory development, with implications for allied work in cognitive development, education, scientific causal inference, and artificial intelligence. She has been the recipient of a John Simon Guggenheim Fellowship, and is a Fellow of the Association for Psychological Science.

Referenced Publications (chronological order)

Cheng, P. W., & Holyoak, K. J. (1985). Pragmatic reasoning schemas.Cognitive Psychology, 17, 391-416.

Cheng, P. W., Holyoak, K. J., Nisbett, R. E., & Oliver, L. M. (1986). Pragmatic versus syntactic approaches to training deductive reasoning.Cognitive Psychology, 18, 293-328.

Nisbett, R.E., Fong, G.T., Lehman, D., & Cheng, P.W.  (1987). Teaching reasoning.  Science, 238, 625-631.

Cheng, P. W., & Novick, L. R. (1992). Covariation in natural causal induction.Psychological Review, 99, 365-382.

Cheng, P. W. (1993). Separating causal laws from casual facts: Pressing the limits of statistical relevance. In D. L. Medin (Ed.), The psychology of learning and motivation, Vol. 30 (pp. 215-264). New York: Academic Press.

Cheng, P. W. (1997). From covariation to causation: A causal power theory.Psychological Review, 104, 367-405.

Lien, Y., & Cheng, P. W. (2000). Distinguishing genuine from spurious causes: A coherence hypothesis. Cognitive Psychology, 40, 87-137.

Chao, J., & Cheng, P. W. (2000). The emergence of inferential rules: The use of pragmatic reasoning schemas by preshoolers. Cognitive Development, 15, 39-62.

Novick, L. R., & Cheng, P. W. (2004). Assessing interactive causal influence.Psychological Review, 111, 455-485.

Liljeholm, M., & Cheng, P. W. (2007). When is a cause the “same”? Coherent generalization across contexts. Psychological Science, 18, 1014-1021.

Lu, H., Yuille, A. L., Liljeholm, M., Cheng, P. W., & Holyoak, K. J. (2008). Bayesian generic priors for causal learning. Psychological Review, 115, 955-982.

Carroll, C.D., & Cheng, P.W. (2010). The induction of hidden causes: Causal mediation and violations of independent causal influence. In S. Ohlsson & R. Catrambone (Eds.), Proceedings of the 32nd Annual Conference of the Cognitive Science Society (pp. 913-918). Austin, TX: Cognitive Science Society. 

Other References:

Danks, D., Griffiths, T. L., & Tenenbaum, J. B. (2003). Dynamical causal learning. In S. Becker, S. Thrun, & K. Obermayer (Eds.), Advances in neural information processing systems (Vol. 15, pp. 67-74). Cambridge, MA: MIT Press.

Griffiths, T. L., & Tenenbaum, J. B. (2005). Structure and strength in causal induction. Cognitive Psychology, 51, 334-384. 

Kemp, C., Goodman, N. D., & Tenenbaum, J. B. (2010). Learning to learn causal models. Cognitive Science, 34, 1185-1243.

Pearl, J. (2000). Causality. New York: Cambridge University Press.