As early as the age of 3, children form hypotheses about the causes of events happening in their
world. This concern with understanding causes persists throughout people's lives, because it is one
way people impose structure on what would otherwise be an endless stream of unconnected
elemental events. Our attributions of the causes of events have important implications for how we
behave in the world. To take a simple example, you would react quite differently toward a colleague
who failed to show up for an important meeting with you depending on whether you attributed the
cause to the person being irresponsible or swamped with work or very ill.
In collaboration with Patricia Cheng, I have investigated how adults induce both simple and complex causes.
A Covariation-based Model of Causal Inference
Early research on causal inference began with the proposal that people determine the causes of events much as scientists do, by assessing (intuitively) the degree of covariation between a potential causal factor and the effect. Subsequent research, however, identified many ways in which people's causal attributions deviated from the predictions of covariational models. Patricia Cheng and I hypothesized that, contrary to the claims in the literature, people's causal inferences are in fact rational.
We argued that to address this issue, it is necessary to distinguish between the data on which the inference process operates and the process of inference computation itself. Previous research either had not made this distinction or had not accurately identified the information people use to assess causality. We developed a new covariation-based model of causal inference, which we named the probabilistic contrast model. We also introduced the concept of a focal set, which is the set of events considered by the reasoner to be relevant to the causal question. Our probabilistic contrast model specifies that causal inference involves an assessment of covariation computed for the events in the focal set (which might differ for different people). In particular, a main-effect contrast for a candidate cause i with respect to effect e is defined as the difference between the probability, in the focal set, of e in the presence of i versus in its absence. If the value of this contrast is noticeably greater than 0, i is considered to have a generative influence on e; if the value is noticeably less than 0, i is considered to have a preventive influence on e; and if the value is not noticeably different from 0, i is considered to be noncausal. Similarly, a two-way interaction contrast evaluates a pair of candidate causes, i and j, as a conjunctive cause of e. This contrast is a difference between differences -- the main-effect contrast for i when j is present minus the main-effect contrast for i when j is absent.
Two studies generated from our theorizing provided strong support for our probabilistic contrast model and thus for our hypothesis that human causal inference is unbiased. In our first study (Cheng & Novick, 1990a, 1990b), we used unfamiliar scenarios and provided subjects with (what we assumed would be) all the relevant information, so there was little possibility of different subjects making different assumptions about unspecified information. In a later study (Novick, Fratianne, & Cheng, 1992), we provided incomplete information about more common situations and assessed both subjects' assumptions about the missing information and their causal attributions.
Our probabilistic contrast model, combined with the concept of a focal set, also can explain the distinction people make between causes of events and enabling conditions. For example, in answer to the question of what caused a particular airplane to crash, investigators are likely to reserve the term "cause" for factors such as the malfunctioning of a critical component of the plane, pilot error, or wind shear. Gravity, they might say, was merely a condition that enabled the crash to occur. In two experiments, we (Cheng & Novick, 1991) showed that neither of the two then-dominant explanations of this distinction could account for subjects' conceptions of factors as causes versus enabling conditions. However, the assumption that covariation is computed over events in (different) focal sets can explain this distinction (Cheng & Novick, 1991).
With this empirical work as background,
and I published a theoretical paper (Cheng & Novick, 1992) comparing our
covariational account of causal inference to several competing models in the social
psychology, cognitive psychology, philosophy, and animal behavior literatures. We showed that our probabilistic contrast model, applied to a focal set of events, provided a better explanation of causal inference than did any previously proposed model.
A Causal Power Theory of Conjunctive Causes
Since our 1992 paper, Patricia Cheng and I have radically revised our view of causal induction. Although our previous model, which is purely covariational, explains a wide range of findings regarding causal inference, it cannot explain why even untutored reasoners do not equate covariation with causation. An opposing approach -- the power or mechanism approach -- has attempted to address reasoners' intuitive understanding of this fundamental inequality, but it has been unable to specify the process that transforms information from the available noncausal input to a causal judgment. Cheng (1997; also see Cheng & Novick, 2005) formulated a revised version of our probabilistic contrast model -- the power PC theory -- that demonstrates that an integration of the covariation and power approaches can overcome the problems confronting each approach.
Cheng's (1997) power PC theory applies only to single-factor causes. However, a single factor is often perceived as contributing towards producing an effect, yet insufficient to produce it on its own. For example, the presence of a flu virus is not by itself sufficient to cause one to have the flu; neither is low body resistance in the absence of a flu virus. But the two in conjunction often do cause one to come down with the flu. Similarly, hard work per se typically does not produce success; it must be combined with talent and opportunity. Most causes in the real world, like these examples, involve a conjunction of factors acting in concert. How do reasoners come to know that a conjunction of factors has the power to produce or prevent an effect?
Patricia Cheng and I worked together to extend the power PC theory to account for the induction of conjunctive causes involving two factors and an effect, all of which are representable by binary variables (Novick & Cheng, 2004). Our analysis concerned the typical situation in which the conjunctive cause itself, if it exists, is observable only indirectly through observations of the component causes (i and j) and the effect (e). In these situations, whether i and j interact to influence e can only be inferred from the deviation of the observed probability of e from that expected if i and j, possibly counterfactually, exerted only independent influences on e. The estimation of this expected probability is therefore a critical step in our analyses.
The discovery of causes that in conjunction produce an effect has previously been explained by covariational theories, among which our interaction contrast model (Cheng & Novick, 1990, 1992) arguably provides the most accurate description of untutored human inference. In our more recent paper (Novick & Cheng, 2004), we (a) discuss problems with covariational theories of conjunctive causes in general, and interaction contrasts in particular, (b) propose a causal power theory of such causes that overcomes these problems, and (c) report empirical evidence in support of the new theory. Like Cheng's (1997) causal power theory of simple causes, but unlike all purely covariational models, the new theory specifies how a reasoner infers a causal relation, rather than merely a covariational relation, from the available input. Because standard statistics is purely covariational, one implication of our theory is that statistics for categorical data are appropriate for testing hypotheses regarding interactive causal influence only under restricted circumstances. The intuitiveness of the predictions based on our measures, as well as the results of the experiments reported, suggest that untutored reasoners adopt the causal power approach and make causal judgments that are normative (also see Cheng & Novick, 2005; Cheng, Novick, Liljeholm, & Ford, 2007).
Cheng, P. W., Novick, L. R., Liljeholm, M., & Ford, C. (2007). Explaining four psychological asymmetries in causal reasoning: Implications of causal assumptions for coherence. In M. O'Rourke (Ed.), Topics in contemporary philosophy, vol. 4: Explanation and causation (Ch. 1, pp. 1-32). MIT Press.
Cheng, P. W., & Novick, L. R. (2005). Constraints and nonconstraints in causal reasoning: Reply to White (2005) and to Luhmann & Ahn (2005). Psychological Review, 112, 694-707.
Novick, L. R., & Cheng, P. W. (2004). Assessing interactive causal influence. Psychological Review, 111, 455-485.
Novick, L. R., Fratianne, A., & Cheng, P. W. (1992). Knowledge-based assumptions in causal attribution. Social Cognition, 10, 299-333.
Cheng, P. W., & Novick, L. R. (1992). Covariation in natural causal induction. Psychological Review, 99, 365-382.
This paper has been reprinted in:
Goldstein, W. M., & Hogarth, R.M. (Eds.). (1997). Research on judgment and decision making: Currents, connections, and controversies (pp. 285-321). Cambridge, UK: Cambridge University Press.
Cheng, P. W., & Novick, L. R. (1991). Causes versus enabling conditions. Cognition, 40, 83-120.
Cheng, P. W., & Novick, L. R. (1990). A probabilistic contrast model of causal induction. Journal of Personality and Social Psychology, 58, 545-567.
Cheng, P. W., & Novick, L. R. (1990). Where is the bias in causal attribution? In K. J. Gilhooly, M. T. G. Keane, R. H. Logie, & G. Erdos (Eds.), Lines of thinking: Reflections on the psychology of thought (Vol. 1, pp. 181-197). Chichester, England: Wiley.