Noisy Newtons: Unifying process and dependency accounts of causal attribution

Noisy Newtons: Unifying process and dependency accounts of causal attribution Tobias Gerstenberg 1 (t.gerstenberg@ucl.ac.uk), Noah Goodman 2 (ngoodman@stanford.edu), David A. Lagnado 1 (d.lagnado@ucl.ac.uk) & Joshua B. Tenenbaum 3 (jbt@mit.edu) 1 Cognitive, Perceptual and Brain Sciences, University College London, London WC1H 0AP of Psychology, Stanford University, Stanford, CA 94305 3 Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139 2 Department Abstract There is a long tradition in both philosophy and psychology to separate process accounts from dependency accounts of causa- tion. In this paper, we motivate a unifying account that explains people’s causal attributions in terms of counterfactuals defined over probabilistic generative models. In our experiments, par- ticipants see two billiard balls colliding and indicate to what extent ball A caused/prevented ball B to go through a gate. Our model predicts that people arrive at their causal judgments by comparing what actually happened with what they think would have happened, had the collision between A and B not taken place. Participants’ judgments about what would have hap- pened are highly correlated with a noisy model of Newtonian physics. Using those counterfactual judgments, we can predict participants’ cause and prevention judgments very accurately (r = .99). Our framework also allows us to capture intrinsically counterfactual judgments such as almost caused/prevented. Keywords: causality; counterfactuals; attribution; physics. Introduction There has been a longstanding divide in philosophy be- tween two fundamentally different ways of conceptualizing causality. According to dependency accounts of causation, what it means for A to be a cause of B is that B is in some way dependent on A. Dependence has been conceptualized in terms of regularity of succession (A is regularly succeeded by B; Hume, 2000 [1748]), probabilities (the presence of A in- creases the probability of B; Suppes, 1970) or counterfactuals (if A had not been present B would not have occurred; Lewis, 1970). For process accounts, in contrast, what it means for A to be a cause of B is that a physical quantity is transmitted along a pathway from A to B (Dowe, 2000). The psychological literature on causal learning and attribu- tion neatly maps onto the two major accounts in philosophy. On the one hand, people have been shown to use contingency information when drawing inferences about whether and how strongly two events are causally linked (Cheng, 1997). On the other hand, people display a preference to choose causes that influence an effect via a continuous causal mechanism over causes that are connected with the effect through mere dependence (Walsh & Sloman, 2011). A point that has been raised in favor of process accounts is that they are capable of capturing the semantics of differ- ent causal terms. Whereas dependency accounts have mostly focussed on causation and prevention, Wolff (2007) has pro- vided a process account that not only predicts when people use the terms cause and prevent but also enable and despite. Fol- lowing a linguistic analysis of causation by Talmy (1988) in terms of force dynamics, Wolff (2007) argues that the afore- mentioned causal terms can be reduced to configurations of force vectors. For example, what it means for a patient (P) to have been caused by an affector (A) to reach an endstate (E) is that P did not have a tendency towards E, A impacted on P in a way that their force vectors were not pointing in the same direction and P reached E. If, in contrast, the force vectors of both P and A point towards E and P reaches E, the model pre- dicts that people will say “A enabled (rather than caused) P”. Importantly, according to Wolff’s account, the core dimen- sions which underlie the different causal terms, such as P’s tendency towards E, are defined in strictly non-counterfactual terms. Hence, “tendency” is defined as the direction of P’s force rather than whether P would reach E in the absence of any other forces. While the force dynamics model has strong intuitive appeal for interactions between physical entities, it is questionable how it can be extended to capture causal attributions in situa- tions involving more abstract entities. For example, one might legitimately assert that the fall of Lehman Brothers caused the financial crisis or that Tom’s belief that he forgot his keys caused him to turn around and go back home. While it is unclear how these causal relationships could be expressed in terms of force vectors, they do not pose a problem for the more flexible dependency accounts. For example, according to a counterfactual account, Tom’s belief qualifies as cause of his behaviour if it is true that his behavior would have been different had the content of his belief been different. Hence, there appears to be a trade-off between the semantic richness of process accounts on the one hand and the generality and flexibility of dependency accounts on the other hand. Rather than fostering the divide between process accounts and dependency accounts, we propose a theory of causal attri- bution that combines the best of both worlds. In the spirit of Pearl (2000), we model causal attributions in terms of counter- factuals defined over probabilistic generative models. How- ever, we agree with Wolff (2007) that people’s causal knowl- edge is often richer than what can be expressed with a causal Bayes net. We aim to unify process and dependency accounts by showing that people have intuitive theories in the form of detailed generative models, and that causal judgements are made by considering counterfactuals over these intuitive the- ories. Here we demonstrate the superiority of our approach over existing models of causal attribution in a physical do- main. We show that people use their intuitive understand- ing of physics to simulate possible future outcomes and that their causal attributions are a function of what actually hap- pened and their belief about what would have happened had the cause not been present.