The Meaning of Cause and Prevent: The Role of Causal Mechanism

by Clare R. Walsh University of Plymouth, U.K. and Steven A. Sloman Brown University, U.S.A.

September 24th, 2008

Address for correspondence: Dr. Clare Walsh, Centre for Thinking and Language School of Psychology University of Plymouth Plymouth, PL4 8AA U.K. Telephone: + 44 1752 233175 E-mail: [email protected]

Causal Attribution 2 Abstract How do people understand questions about cause and prevent? Some theories propose that people affirm that A causes B if A’s occurrence makes a difference to B’s occurrence in one way or another. Other theories propose that A causes B if some quantity or symbol gets passed in some way from A to B. The aim of our studies is to compare these theories’ ability to explain judgments of causation and prevention. We describe six experiments that compare judgments for causal paths that involve a mechanism, i.e., a continuous process of transmission or exchange from cause to effect, against paths that involve no mechanism yet a change in the cause nevertheless brings about a change in the effect. Our results show that people prefer to attribute cause when a mechanism links cause to effect. In contrast, prevention is sensitive both to the presence of an interruption to a causal mechanism and to a change in the outcome in the absence of a mechanism. In this sense, ‘prevent’ means something different than ‘cause not’. We discuss the implications of our results for existing theories of causation.

Keywords: Causation, Prevention, Mechanism, Counterfactual

2

Causal Attribution 3 Causal attributions are central to our ability to make sense of the world: to explain events, to make predictions and to plan for the future. Without this ability, we could not hope to influence others around us or our environment nor could we assign legal or moral responsibility (Hart & Honore***). But, how do people decide whether an event caused or prevented a particular outcome? Despite the pervasiveness of causal questions, the answer to this question is still disputed. Our aim is to examine how people understand the terms “cause” and “prevent” when making attributions of who or what is the cause of an event. Hence, like all work on causal attribution, our project straddles the frontier between psycholinguistics and the study of thinking. We want to know how people understand and use particular words in the service of a conceptual task that determines belief and action. Our results are revealing about both the meaning of terms and the cognitive structures that make such meaning possible. We aim to distinguish two broad theoretical views of how people make judgments of causation and prevention. One view is based on the idea that a cause is something that makes a difference to the effect. This view underlies covariation (e.g., Cheng & Novick, 1990) and counterfactual (e.g., Lewis, 1973) theories of causation. A second is based on the idea that a cause involves an interaction between two events (e.g., Dowe, 2000; Wolff, 2007, see also Barbey & Wolff, 2007) and may involve an exchange of some conserved quantity. We present a set of experiments that distinguish these two views and provide tests of specific theories of each type. Our focus is on “actual” cause rather than “generic” cause. Events are inevitably preceded by many influences and, in the generic sense, each is a “cause.” However, like most work on causal attribution, we focus on actual causes, the subset of events that

3

Causal Attribution 4 people select when asked for the cause of an event. For example, if I slip on the pavement, I may attribute cause to the fact that there was ice-cream on the pavement or to the person that dropped it there. I am unlikely to attribute cause to other events in the causal sequence, to the ice cream seller or to my decision to follow that particular route. But all of these events may covary to some extent with the outcome and undoing any of them may lead to a change in the outcome. When we think about the cause of a forest fire, we are more likely to attribute it to the match that was dropped than to the oxygen in the air or the fact that it hasn’t rained for some time. The latter are sometimes referred to as enabling conditions (psychological theories distinguishing cause from enable include Cheng, 1997; Goldvarg & Johnson-Laird, 2001; Sloman, Barbey, & Hotaling, 2008; and Wolff, 2007). A number of cues are relevant to selecting an actual cause: Causes should be proximal in time and space to the outcome (Hume, 1748/1988). A cause of an abnormal effect must itself be abnormal (Gavanski & Wells, 1989) and stand out against other background conditions (Hilton & Slugoski, 1986). In cases of intentional action, people tend to attribute cause to voluntary human action (Hart & Honore, 1959/1985). How and when these cues are used depends on what people are trying to achieve when they attribute cause. Our concern is this more general question: Do people attribute cause to events merely because they know they make a difference, or must people believe that an event is linked to an effect via a mechanism before they are willing to call it a cause? Of course, the answer may not be univocal. For instance, people may prefer to make attributions based on mechanism but they may rely on difference-making in the absence of mechanistic knowledge.

4

Causal Attribution 5 “Make a Difference” Theories of Causation: Covariation and Counterfactuals A number of theories of causation incorporate the notion that a cause is something that makes a difference to the effect, without regard to how the effect is brought about. Following Lewis (1986), we refer to this cluster of theories as “make a difference” theories. According to Hume (1748/1988), causation cannot be directly perceived, hence people infer causation from the regular co-occurrence of distinct events even though this practice cannot ultimately be justified. According to this view, causal inference makes no reference to the mechanism or process through which one event causes another. Instead, we may infer that C caused E if all objects similar to C are followed by objects similar to E. Imagine, for example, that a ball is thrown at a window and the window breaks. You may attribute causation to the ball if on previous occasions when you saw balls being thrown at windows, the windows broke. The great problem of induction according to Hume is that past experience cannot be used to justify prediction without incoherence or circularity. Nevertheless, past experience is all we have and is therefore what we use. In the philosophical literature, this idea has been incorporated into regularity theories like that of Hume as well as probabilistic (Reichenbach, 1938; Suppes, 1970), counterfactual (e.g., Lewis, 1986), and manipulability theories of causation (e.g., Pearl, 2000; Woodward, 2003). Our experiments examine both covariation and counterfactual theories of causation. Covariation theories propose that people attribute causation to events that in the past have tended to occur when the outcome occurred. For example, according to the delta P rule (Allan, 1980), people judge causation by calculating the probability that the effect occurs given that the cause occurred minus the probability that the effect occurs

5

Causal Attribution 6 given that the cause didn’t occur (e.g., Cheng, 1997; Cheng & Novick, 1990). The greater the value is, the stronger the causal link between the two events is judged to be. This idea can be applied directly to the problem of causal attribution: Assume that people choose the cause with the highest delta P. Other covariation theories have dominated social psychological models of causal attribution (especially Kelley, 1973). For example, in judging why John laughed at the comedian, we are more likely to make an attribution to the comedian if other people laughed at him too and if John laughs at him more than at other comedians. Covariation theories can and have been implemented in connectionist models (van Overwalle, 1998; Read & Montoya, 1999). Covariation on its own is not enough to determine the causal structure relating a set of variables. Most fundamentally, it doesn’t tell us anything about causal direction. Therefore additional cues may be required to infer causal relations. Temporal order, for instance, is informative because causes tend to precede their effects (Goldvarg & Johnson-Laird, 2001; see Lagnado, Waldmann, Hagmayer, & Sloman, 2007, for a review). Covariation also fails to tell us whether one event caused another or whether both are the result of a common cause. To overcome this, we need to control for other possible factors (e.g., Spirtes, Glymour & Schienes, 1993) and on occasion we have enough knowledge to do this by mentally running a counterfactual simulation. Counterfactual theories of causation are a second class of make-a-difference theories. These theories also assume that causation can be judged without appealing to the particular mechanism relating cause to effect. Counterfactual theories propose that causation can be defined in terms of a counterfactual conditional. In their original form, they state that “event c caused event e” provided that “if c hadn’t occurred then e

6

Causal Attribution 7 wouldn’t have occurred” or in other words that e would not occur in the closest possible world to our own in which c does not occur (Lewis, 1973; see Bennett, 2003, for a review). According to this view, throwing the ball caused the window to break if when you imagine that the ball wasn’t thrown at the window, the most likely outcome is that the window does not break. Covariation theories compare the set of situations in which the antecedent occurred to the set of situations in which the antecedent did not occur. Counterfactual theories compare the facts with a single imagined alternative in which the antecedent does not occur, for example, the closest possible world. There have been a number of reformulations of Lewis’s original counterfactual theory that aim to solve problems such as preemption (see Experiment 5 below). Lewis’s (2000) theory proposes that the degree of causal influence reflects the extent to which whether, when and how one event happens depends on whether, when and how another event happens. The modified theory states that c causally influences e if an alteration to c would have led to an alteration of e. Another proposal is that people may evaluate cause using counterfactuals that hold certain events constant (Halpern & Pearl, 2001; Hitchcock, 2001; Woodward, 2003). Similarly, manipulability theories propose that “event c causes event e” provided that if I manipulate or intervene in a particular way that involves changing c, then this should change e also. Rather than relying on the notion of a closest possible world, Pearl (2000) proposes that an intervention to change “c” leads to a counterfactual in which the consequences of “c” are changed but its causes are not. For example, if I had thrown a ball at a window, it would most likely have broken but the reverse is not true. In our studies, manipulability theories make the same predictions as counterfactual theories.

7

Causal Attribution 8 Each of the theories discussed have in common the view that c is deemed a cause if and only if it makes a difference to whether or not e occurs. An alternative possibility, to which we now turn, is that a cause is an event that involves a causal mechanism. Generative Theories of Causation In contrast to the theories described above, generative theories define causation in terms of the process bringing about the effect. In this sense, the cause is believed to generate the effect. This idea can be traced back to Kant (1781) and found in theories of causation in philosophy (e.g., Dowe, 2000; Salmon, 1984) and psychology (e.g., Shultz, 1982). Causal process theories propose that causation involves a transmission along a causal pathway (Bullock, 1985; Salmon, 1984; Shultz, 1982) and may involve the exchange of some conserved physical quantity, such as energy or momentum (Dowe, 2000; Salmon, 1997). For example, the reason that we believe that throwing a ball at the window caused it break may be that energy or momentum is conserved along the trajectory of the ball, transmitting a force that breaks the window. According to this account a single observation may be sufficient to make an attribution. Wolff (2007) proposes a theory of this type based on force dynamics (Talmy 1988). In his model, A causes C when something has a tendency towards an end state B and A interacts with it and forces it to a different end state C. In our example, the bottle had a tendency not to break but Billy’s rock interacts with it such that it breaks. Mandel (2003) proposed an alternative theory, that judgments of causation are based on both covariation and generative processes. A cause must be a sufficient condition for an outcome as well as being connected to it by a mechanism. Theories of Prevention 8

Causal Attribution 9 The literature on causation is rich; less has been said about prevention. One possibility is that to prevent something means the same as to cause it not to occur (Collins, 2000). Make-a-difference theories tend to assume this. Both are consistent with the same possibilities (Goldvarg & Johnson-Laird, 2001). Covariation theories propose that an inhibitory cause reduces the likelihood of the effect (e.g., Cheng & Novick, 1991) and therefore on this account, prevention complements causation. Lewis’s (2000) counterfactual theory treats double prevention as a case of causation, in other words, if A prevents B and B prevents C then we should say that A causes C. On this account we may extend counterfactual theories to cover prevention. For example, we might argue that Billy prevented the window from breaking provided that if he hadn’t caught the ball, the window would have broken. Hence, make-a-difference theories assume that ‘prevent’ means the same as ‘cause not’. Generative theories are less consistent in their views of prevention perhaps because prevention tends to be problematic for generative theories. If A prevents B, then usually B does not occur and hence there can be no continuous process connecting the preventor and the effect. Wolff’s (2007) dynamics model gets around this by proposing that A prevents C if something moving towards an end state C and then as a result of an interaction with A moves towards a different end state. Wolff’s theory differs from makea-difference theories of prevention because the theory assumes that prevention, like causation, should involve an interaction between two events. But his view has in common with make-a-difference theories the assumption that to prevent something has the same meaning as to cause it not to occur (Barbey & Wolff, 2007). Mandel (2003) developed a theory of prevention based on counterfactuals, however this theory applies

9

Causal Attribution 10 only to judgments about how an outcome ‘could have been prevented’ (Mandel, personal communication). In contrast, he proposes that judgments about how an outcome ‘was prevented’ depend, like causation, on sufficiency and generative processes (Mandel, personal communication). Hence, on this view, judgments about how an event ‘was prevented’ will also be the same as for ‘caused not’. But the meanings of cause and prevent may not be symmetrical. For example, people tend to pick out a single event as ‘the cause’ but may judge that several events prevented an outcome. Dowe (2000) developed a causal process theory of prevention that defines prevention as qualitatively different from causation. He proposed that A prevented B if A occurred and B did not and if there was a causal interaction between A and another process x and if A hadn’t occurred then x would have caused B. For example, Jack prevented the child from being hit by a car if and only if (i) he grabbed the child and (ii) if he hadn’t grabbed her, she would have been hit by the car. The critical elements are the actual interaction between Jack and the child and the counterfactual interaction between the child and the car. Hence according to this view causation and prevention have quite different meanings and the definition of prevention involves a counterfactual. We describe tests of these theories in Experiments 4, 5 and 6. Judgments of Causation and Prevention: The Psychological Evidence A consensus has not been achieved about whether regularity theories or generative theories best describe how people make causal attributions (Baillargeon, Kotovsky & Needham, 1995; Cheng, 1997; Lewis, 2000). Most of the evidence comes down on the side of generativity. When people aim to identify the cause of an event, they seek out evidence about possible mechanisms rather than evidence that potential causes co-vary 10

Causal Attribution 11 with the effect (Ahn, Kalish, Medin & Gelman, 1995). There is also evidence to suggest that people assume mechanisms in the form of forces in their use of causal verbs. People tend to use ‘cause’ when an object is forced towards an outcome against its initial tendency whereas they tend to use ‘prevent’ when an object is forced away from an initial tendency (Wolff, 2007). When an outcome is overdetermined, individuals select the cause that generated the outcome (Mandel, 2003). Even young children require a generative transmission process to attribute cause. Children are more willing to make causal attributions based on inferences about a causal mechanism than on temporal or spatial contiguity (Shultz, 1982). Similarly, when a ball rolled down a runway hit a jack-in-the-box and the jack popped up, both adults and children indicated that the ball caused the jack to pop up. However, if the ball stopped before it reached the jack-in-the-box and the jack popped up, neither adults nor preschool children made a causal attribution to the ball (Bullock. 1985). This result suggests that that people are more likely to attribute causation when a continuous mechanism links an action (e.g., rolling a ball) to an outcome (e.g., jack to pop out of the box). However, this result could be explained by a purely counterfactual theory. For example, when the ball doesn’t reach the jack-in-the-box, people may infer that if the ball hadn’t been rolled the jack still would have popped. Such a counterfactual would have to reflect knowledge about causal mechanisms. Nevertheless, it is logically possible that the counterfactual is necessary to make the causal attribution. Other psychological studies have shown that attributions can depend on the availability of different counterfactual possibilities. When a set of facts is held constant, changing the availability of counterfactual alternatives can influence causal attribution.

11

Causal Attribution 12 An action is judged to be more causal if an alternative action would have led to a different outcome (Wells & Gavanski, 1989). In fact, judgments of causation and prevention may differ in this regard. Judgments of causation may focus on an event that is sufficient for the outcome and one that elucidates the mechanism (Mandel, 2003), whereas judgments of prevention may depend more on the generation of a counterfactual conditional and may focus on events that are controllable (Mandel & Lehman, 1996).

EXPERIMENTAL STUDIES Our aim is to address the question of how people make judgments of both causation and prevention. In most situations, make-a-difference and generative theories make the same predictions as usually a change to an event would entail a change to an outcome by virtue of a continuous causal process linking the event and the outcome. To distinguish the theories, we compare cases that involve a complete causal mechanism linking an action to an outcome (i.e., a clear process of generative transmission) to ones where the action interrupts a mechanism and hence there is no causal mechanism linking the action to the outcome. The counterfactual alternatives are identical in the two cases – a change in the action brings about a change in the outcome. Hence, generative and make-a-difference theories make different predictions. If causal attribution depends on understanding how the outcome comes about, i.e., on the specific process generating the outcome, people should make different judgments for the two scenarios. In contrast, if causal attribution depends on an analysis of covariation or the simulation of counterfactual alternatives, people should respond in the same way to the two scenarios.

12

Causal Attribution 13 Consider this example of an interruption scenario: John is playing ball with his friends on the street. The ball is flying through the air and John is blocking the path between it and the neighbor’s window. John jumps out of the way and the ball hits the neighbor’s window. Make-a-difference theories predict that John is the cause of the window breaking because his action made a difference to the outcome. In contrast, generative theories predict that John is not the cause because he didn’t interact with the ball (Dowe, 2000; Mandel, 2003). Wolff’s (2007) force dynamics theory implies more specifically that John enables the window to break because the force of John’s action, the agent, is in the same direction as the window’s breaking, the patient.1 Table 1 sets out the predictions of each theory for scenarios with this structure and our aim is to test these predictions. ******************** Insert Table 1 about here ******************** Experiment 1: A Test of Theories of Causation The aim of this experiment is to test make-a-difference and generative theories of causation. We constructed four scenarios involving a window being broken, a boulder falling off a cliff, a drink being knocked off a table, and a girl falling into a swimming pool. Each scenario had two versions: one involving an action linked to the outcome by a continuous causal mechanism and one involving an action that unblocks the causal pathway. A change to either action was sufficient to undo the outcome. The predictions are set out in Table 1. If people attribute causation when there is a continuous mechanism

1

Wolff (personal communication) is developing a theory which will allow for virtual as well as actual interactions. On this account, John would be judged to have caused the outcome.

13

Causal Attribution 14 linking an action to an outcome, as generative theories predict, then they should attribute causation to the action which generates the outcome and not to the one which unblocks the causal pathway. In contrast, if they attribute causation to an action that makes a difference to the outcome, then they should attribute causation to both types of action. Method Materials and Design We constructed four new scenarios, each containing an action which is linked by a continuous causal mechanism to the outcome and one which unblocks the causal pathway but is not linked by a mechanism to the outcome. One of the scenarios was as follows: Frank and Sam are playing ball on the street with friends. Frank kicks the ball hard and unintentionally sends it in the direction of a neighbour’s house. Fortunately, Sam is blocking the path of the ball. However, at that moment Sam is distracted from the game and steps away. The ball flies past Sam and it hits the window of the neighbour’s house. The window smashes. The scenario was followed by the questions: Did Frank cause the window to smash? Did Sam cause the window to smash? The full set of scenarios is presented in the Appendix. Participants were presented with each of the four scenarios in a counterbalanced order. Participants and Procedure The participants were 43 undergraduates at the University of Plymouth who took part in return for course credit. After reading each scenario, they responded to the two questions by answering “yes” or “no”. They also rated their confidence on a scale of 1 to 7.

14

Causal Attribution 15

Results As Table 2 shows, participants attributed causation to the action linked to the outcome by a causal mechanism (87%) more often than to the action which unlocked the causal pathway (24%; Wilcoxon Test = 5.47, p < .001). The difference occurred for all four scenarios (all ps < .02). The effect was weaker for the scenario about the boulder falling off a cliff than for the other three scenarios perhaps because at the point at which the mechanism is initiated (i.e., the boulder is pushed), the causal pathway which links it to the outcome is blocked (i.e., the gate in its path is closed). ******************** Insert Table 2 about here ******************** The result demonstrates that individuals tend to attribute causation to events if and only if they are linked to the outcome by a continuous causal mechanism regardless of whether undoing the event is sufficient to undo the outcome. The result provides support for generative theories which propose that causation involves an interaction between events (Dowe, 2007; Mandel, 2003; Wolff, 2007) but cannot be explained by make-a-difference theories whether they are based on covariation (Cheng & Novick, 1991) or counterfactuals (Halpern & Pearl, 2001; Lewis, 2000). Experiment 2: Controlling for Outcome, Normality and Omissions The aim of Experiment 2 was to examine whether our findings generalize when the context changes in three specific ways. These contexts were chosen because they present

15

Causal Attribution 16 a challenge to generative theories or because they are likely to encourage counterfactual thinking. Hence, this study provides a stronger test of generative theories. First, we developed scenarios where the actions led to negative consequences for others. One involved scoring a goal in a soccer game and the second involved a ball breaking a champagne bottle. Negative outcomes are one of the primary triggers of counterfactual thinking (Roese, 1997) and so they may be more likely to elicit judgments based on counterfactuals. Second, counterfactual antecedents and causes may focus on abnormal events (Hilton & Slugoski, 1986; Kahneman & Miller, 1986). To test for this, in one of our scenarios, the action that did not involve a mechanism conflicted with norms or expectations. It involved a goal-keeper who failed to save a goal. If individuals attribute causation to events that are contrary to expectations, then individuals should attribute causation in the absence of a causal mechanism. For comparison, our second scenario involved a similar structure but involved events that were neutral with regard to normality. Our third aim was to generalize our findings to situations involving causation by omission. In our previous study of causation, all of the scenarios involved actions. In this study, we included a scenario involving a failure to act and as a result the actor does not interact with a mechanism but if he had, he would have brought about a change to the outcome. Compared to actions, failures to act may be more likely to be judged using a counterfactual. Dowe (2000) classifies these examples as ‘quasi-causation’ and like prevention he defines them in terms of interactions that would have occurred. For example, “if the goalkeeper had caught the ball, he would have prevented the ball from

16

Causal Attribution 17 going into the net”. He treats these as having a lower status than genuine causation which involves a complete mechanism. If Dowe’s account is correct, people should attribute causation less often to omissions than to actions that are linked by a mechanism to the outcome. In contrast, make-a-difference theories predict that causation should be attributed equally often in both cases. Hence, as Table 1 shows, the predictions for this study are the same as for Experiment 1. Method Materials and Design We constructed two scenarios, a soccer scenario and a champagne bottle scenario. As in previous experiments, each scenario had two versions. In the ‘mechanism complete’ version, the actor initiated the action and then failed to block it. Hence, there was a continuous mechanism linking the action to the outcome. For the soccer scenario, this read as follows: Jack is a goalkeeper for his indoor soccer team. In the final minute of a tied soccer match, Steve (a member of the opposing team) kicks the ball towards the net in an attempt to score a winning goal. The ball rolls toward the net and Jack kicks the ball away. The ball flies through the air, bounces off a wall, and heads back toward the net. Jack would have been able to block the ball, but he gets confused and stumbles, misses it, and the ball rolls into the net. The opposing team wins. In the ‘mechanism not blocked’ version, someone else initiated the mechanism and the actor in question failed to block it. Hence, there is no mechanism linking the agent’s action to the outcome. The soccer version was as follows:

17

Causal Attribution 18 Jack is a goalkeeper for his soccer team. In the final minute of a tied soccer match, Steve (a member of the opposing team) attempts to score a winning goal. He kicks the ball toward the net. Jack would have been able to block the ball, but he gets confused and stumbles, misses it, and the ball rolls into the net. The opposing team wins. The scenarios were followed by the question: Did Jack cause the opposing team to win? For completeness, participants were also asked two further questions, i.e., whether Steve caused his team to win and whether Jack was to blame for the opposing team’s win but we will not report responses to these questions. Participants were given either both ‘mechanism complete’ scenarios or both ‘mechanism not blocked’ scenarios and the soccer scenario always preceded the champagne bottle scenario (the two versions of the champagne bottle scenario are provided in the Appendix). Participants and Procedure Participants were ninety-four Brown University undergraduates and four Community College of Rhode Island undergraduates. They were entered into a lottery for $19.50 in return for taking part. One participant was eliminated for failing to follow instructions. After reading each of the scenarios, they responded to the questions by answering ‘yes’ or ‘no’. Results and Discussion We recorded the percentage of people who responded that the target actor, ‘Jack’ or ‘Suzy’, caused the outcome (i.e., the opposing team to win or the bottle to break). As shown in Table 3, participants attributed causation more often in the mechanism

18

Causal Attribution 19 complete versions than in the mechanism not blocked versions and this pattern occurred both for the soccer scenario (70% vs 49%, chi2 = 4.47, p < .04) and the champagne bottle scenario (76% vs 51%, chi2 = 6.53, p < .02). These results generalize the findings of our previous experiment. They provide further support for the view that people tend to think about whether a mechanism links the actor and the outcome when they make an attribution of causation. This tendency occurs even when the outcome has negative consequences. It also occurs when the antecedents were abnormal. Participants were no more likely to attribute causation when the mechanism-not-blocked scenario contained an abnormal event (soccer scenario, 49%) than a neutral event (51%). The results also generalize our findings to cases of causation by omission. They show that people are more likely to attribute causation when there is a complete causal mechanism than when, if an event had occurred, it would have resulted in an interaction with a mechanism. Nonetheless, many participants did attribute causation even when there was no mechanism linking the action and the outcome. The results suggest that there are cases when people are willing to attribute causation using covariation or counterfactuals. ******************** Insert Table 3 about here ******************** Experiment 3: A Test of Covariation Theories of Causation This experiment evaluates an alternative explanation for the results of the previous studies. In our scenarios, the two events linked by a mechanism may have represented classes of events that covary more than the two events not linked by a mechanism. For example, in general, kicking balls may covary with breaking windows more than jumping

19

Causal Attribution 20 out of the way. To test this possibility, we constructed a scenario with two identical actions (two actors throwing tennis balls), one linked to the outcome (a plastic bottle falls off the wall) by a mechanism but did not covary with the outcome on previous trials and one not linked to the outcome by a mechanism that did covary with the outcome. The scenario read as follows: There is a plastic bottle standing on a wall. Sam and Max have some tennis balls and they’re taking turns throwing them at the bottle. Sam has fantastic aim. He takes 20 shots and hits the bottle every time. Max, on the other hand, has terrible aim. He also takes 20 shots but misses each time. On their final shot, Sam and Max happen to aim at the same time. Max throws a moment before Sam. This time Max’s shot is on target and it hits the bottle and knocks it off the wall. Sam’s shot is also on target and his ball would have hit the bottle if Max had missed. Hence, if participants use covariation information to attribute causation, Sam will be judged to have caused the outcome whereas if participants rely on an analysis of the mechanisms, then Max will be judged to have caused the outcome. Method Materials and Design We used the scenario above. One action was linked to the outcome by a mechanism but did not covary with the outcome on previous trials. The second outcome was not linked by an action to the outcome but did covary with it on previous trials. The scenario was followed by two questions which were presented in a counterbalanced order: Did Sam cause the bottle to fall off the wall? Did Max cause the bottle to fall off the wall?

20

Causal Attribution 21 Participants and Procedure Forty-one participants were recruited using a webpage that advertises psychology studies and they completed the questionnaire on-line. After reading the scenario, they responded to the questions by answering ‘yes’ or ‘no’. Results and Discussion More participants attributed causation to the action linked to the outcome by a mechanism (95%) than to the action that covaried with the outcome (12%, McNemar Tests, p < .01). The results strongly support generative theories of causation and are not consistent with covariation theories of causation. They also challenge Mandel’s (2003) view that effects must occur in the presence of a cause for causal attribution. Covariation may be useful for learning about causal relations (but see Lagnado et al., 2007), for making predictions about the likely outcomes of events, and for judging what events could potentially have caused an outcome when information about the mechanisms is missing. But when mechanisms are known, people use them to judge causation. Experiment 4: A Test of Theories of Prevention Experiment 4 aimed to provide a further test of make-a-difference and generative theories by examining how people attribute prevention. We constructed three versions of a scenario about a ball rolling down a hill towards a bottle. One scenario involved an interruption to a causal mechanism (i.e., an actor catching the ball). There is a bottle at the bottom of a hill. Frank is standing close by at the top. While he is there, Billy aims to roll a ball towards the bottle. The aim is perfectly on target. Billy lets go of the ball and it rolls down towards the bottle. Frank then

21

Causal Attribution 22 runs down the hill after the ball. He manages to catch up with the ball and picks it up before it reaches the bottle. The bottle does not break. Did Frank prevent the bottle from breaking? There is an interaction between the actor (Frank) and the outcome and a change to the action will bring about a change to the effect. Hence, as shown in Table 4, all theories predict that this is an example of prevention. Our second case involved an omission (the agent thinks about initiating an action that would change the outcome but changes his mind): There is a bottle at the bottom of a hill. Billy is standing close by at the top. While he is there he thinks about rolling a ball towards the bottle. He always has a perfect aim and he will definitely hit the bottle. At the last minute Billy changes his mind. He decides not to roll the ball. The bottle does not break. Did Billy prevent the bottle from breaking? Once again a change to the decision will bring about a change to the outcome. Hence, covariation and pure counterfactual theories (Lewis, 2000) must consider this an example of prevention. However, the scenario does not involve an interaction and hence Dowe’s (2000) causal process theory and Wolff’s (2007) dynamics model do not define this as an example of prevention2. We also introduced a third version of the scenario to test another counterfactual account (the ‘agent absent’ counterfactual account) that people will attribute prevention only when it is the case that if the agent had been absent the outcome would be different (Kagan, 1989). This account makes the same prediction as generative theories for the two

2

Wolff (personal communication) is developing his dynamic model to include virtual interactions that are not actually realized. On this account, omissions may be classed as preventors.

22

Causal Attribution 23 scenarios presented above because in the omission scenario if the actor was not there, the outcome would be the same and hence the act is not an example of prevention. In the third scenario, the same agent both initiates and interrupts the roll of a ball: There is a bottle at the bottom of a hill. John is standing close by, at the top. While he is there, John aims to roll a ball towards the bottle. The aim is perfectly on target. John lets go of the ball and it rolls down towards the bottle. Within a split second he then chooses to run down the hill after the ball. He manages to catch up with the ball and picks it up before it reaches the bottle. The bottle does not break. Did John prevent the bottle from breaking? Generative accounts predict that people should attribute prevention in this case as the mechanism is interrupted (Dowe, 2000) or because John’s action forces the ball away from its initial tendency towards the bottle (Wolff, 2007). Covariation and most counterfactual theories (e.g., Lewis, 2000) also make this prediction because John’s catch made a difference to the outcome. However, the ‘agent absent’ account predicts that people should not attribute prevention because the outcome would have been the same if the agent had been absent. Finally, we introduced a control scenario in which all theories predict that people should not attribute prevention: There is a bottle at the bottom of a hill. Paul is standing close by, at the top. While he is there, Paul aims to roll a ball towards the bottle. The aim is perfectly on target. Paul lets go of the ball and it rolls down towards the bottle. At the last

23

Causal Attribution 24 minute the ball hits off a stone and it bounces in a different direction. The bottle does not break. This scenario was added to ensure that participants were reading the scenarios carefully and were not just responding ‘yes’ in all cases. The predictions for each of the theories are set out in Table 4. Mandel’s (2003) theory of prevention refers only to judgments of how an outcome ‘could have been prevented’ and therefore does not make predictions for questions about whether the actors actually prevented the outcomes (Mandel, personal communication). Method Materials and Design We constructed four scenarios, as described above. In one the action interrupted a causal mechanism, in one the agent fails to act, in the third the agent initiates and interrupts a mechanism, and in the control scenario an object accidently interrupts the mechanism. Participants were given all four scenarios. The control scenario was always presented last and the remaining scenarios were presented in a counterbalanced order. After each scenario, they were asked one question: Did ‘the agent’ prevent the bottle from breaking? In each case, ‘the agent’ was replaced by the name of the protagonist in the scenario. For instance, in the mechanism-interrupted scenario, the agent referred to Frank. Participants and Procedure Participants were 28 undergraduates at the University of Plymouth in the UK who took part in return for course credit. After reading each scenario, they responded to the question by answering “yes” or “no”.

24

Causal Attribution 25

Results and Discussion One participant was eliminated from the analysis because he responded ‘yes’ to the control scenario. As shown in Table 4, more participants attributed prevention when the agent interrupted the mechanism (84%) and when the agent both initiated and interrupted the mechanism (77%) than when the actor failed to act (46%, McNemar Tests, p < .01; p < .03 respectively). Participants did not attribute prevention significantly more often when the agent only interrupted the scenario than when he initiated and interrupted the scenario (McNemar Tests, p > .6). ******************** Insert Table 4 about here ******************** Our results demonstrate, first, a difference in attributions between the scenarios that involved an interruption and the scenario that involved an omission. This difference is consistent with generative theories (Dowe, 2000; Wolff, 2007). Make-a-difference theories cannot explain this difference. They define all of these events as prevention because they increase the likelihood of the outcome (Cheng & Novick, 1990) and a change to the action would bring about a change to the effect (Halpern & Pearl, 2001; Lewis, 2000). The ‘agent absent’ counterfactual theory (Kagan, 1989) cannot explain why people attribute prevention when the actor both initiates and interrupts a mechanism because the actor’s absence makes no difference to the outcome. The second key finding is that many participants attributed prevention even when a mechanism was not interrupted. One possible explanation for this is that, despite our

25

Causal Attribution 26 efforts, a mechanism was interrupted in this condition. Namely, a thought process was interrupted in that Billy changed his mind. A second possibility is that the meaning of prevention is less clear to participants than the meaning of causation. Participants may prefer to attribute prevention when there is an interaction with a mechanism but they may sometimes attribute prevention based on a counterfactual. Questions about prevention by definition involve counterfactual outcomes and hence, may be more likely to cue consideration of counterfactual possibilities. And for Dowe (2000), the attribution of prevention involves the consideration of both generative processes and counterfactuals. Experiment 5: Comparing Causation and Prevention The four studies reported so far are consistent with the different generative theories of causation. Our fifth study examined a case where Dowe’s causal process theory and Wolff’s dynamics model make different predictions, as Table 5 shows. The experiment had two key aims. Our first goal was to provide a new test of prevention by examining whether people understand it to mean the same thing as ‘cause not’. To do so, we constructed a new scenario with two alternative outcomes. The scenario described a spinning coin which as a result of some actions will land on heads instead of tails. We asked both whether the actions caused the coin to land on heads and whether they prevented the coin from landing on tails. Make-a-difference theories tend to define causation and prevention symmetrically (Cheng & Novick, 1991; Lewis, 2000) as does Wolff’s (2007; Barbey & Wolff, 2007) dynamics model and so these theories imply that individuals should give the same response to both questions. In contrast, Dowe (2000) defines causation and prevention as being qualitatively different. He defines causation in terms of a continuous generative process and prevention in terms of an interruption to a 26

Causal Attribution 27 process that would otherwise have produced its effect. Hence, his theory does not require that people respond the same way to both questions. Our goal is to examine whether people’s underlying conceptual representations of causation and prevention are symmetric. Our second aim was to reexamine make-a-difference and generative theories of causation using scenarios involving pre-emption. In these cases, more than one cause occurs and undoing either one is not sufficient to undo the outcome. For example, imagine both Billy and Suzy throw rocks at a bottle and both their throws are right on target. Billy throws first, and his rock hits the bottle and it breaks. In this case, the counterfactual “if Billy hadn’t thrown the rock, then the bottle wouldn’t have broken” is false even though we might say that Billy caused the bottle to break. Hence, pre-emption has traditionally been a problem for counterfactual theory. However, as we noted earlier, reformulations of counterfactual theories have been developed to resolve this problem (Halpern & Pearl, 2001; Hitchcock, 2001; Lewis, 2000). Covariation theorists have not yet explicitly addressed this issue (e.g., Cheng & Novick, 1990; Cheng, 1997). Generative theories predict that causation should be attributed to the event that generates the outcome and hence they can explain why Billy and not Suzy caused the bottle to break. We generated two scenarios with similar structures based on this classic example of late pre-emption (cf. Hall & Paul, 2003). The ‘mechanism-complete’ scenario included an actual mechanism from an action (a marble is rolled) to an effect (a coin falls over and lands on heads): There is a coin standing on its edge at the end of the table. It is unstable and it is about to fall over and land on tails. Billy and Suzy are standing close by with

27

Causal Attribution 28 marbles. Each one rolls their marble down the table towards the coin. Their rolls are perfectly on target and each one will hit the coin at exactly the same spot, knock it the other way, and the coin will land on heads. Billy happens to roll first and his marble reaches the coin before Suzy’s. The coin falls over and lands on heads. After reading the scenario, participants answered questions about both causation and prevention: Did Billy cause the coin to land on heads? Did Suzy cause the coin to land on heads? Did Billy prevent the coin from landing on tails? Did Suzy prevent the coin from landing on tails? The second scenario was similar to the first but it involved an interruption to a causal mechanism, that is an action (a marble is caught) interrupts a mechanism that would otherwise have produced an effect (a coin to fall over and land on tails). An interruption scenario was used once again to distinguish the make-a-difference and generative theories of causation. There is a coin standing on its edge at the end of the table. It is unstable and it is about to fall over and land on heads. Frank and Jane are standing close by with marbles. While they are there someone else rolls a marble toward the coin. The roll is perfectly on target and it will hit the coin, knock it over and the coin will land on tails. Frank and Jane both reach out and put their hands in front of the coin. Frank happens to put his hand in front of Jane’s and he catches the marble. The coin falls over and lands on heads.

28

Causal Attribution 29 After reading the scenario, participants answered the same four questions as after the ‘mechanism complete’ scenario. Did Frank cause the coin to land on heads? Did Jane cause the coin to land on heads? Did Frank prevent the coin from landing on tails? Did Jane prevent the coin from landing on tails? The predictions made by each of the theories are set out in Table 5. When a preemptive cause is present, neither action in either scenario is necessary to bring about a change in the outcome; the effect is independent of both causes. Hence, covariation theories predict that neither of the events will be judged as causal (Cheng & Novick, 1990). Counterfactual theories have been designed to address the issue of pre-emption by using counterfactuals that hold certain events constant (e.g., the fact that Suzy’s marble didn’t hit the coin; Halpern & Pearl, 2001; Hitchcock, 2001) or by considering whether a change to how and when the action occurred (e.g., the exact timing or force of the throw) would influence the outcome (Lewis, 2000). Hence they predict that only the first actor in each scenario will be judged as causal. All of these theories assume that causation and prevention are symmetrical in meaning and hence they make the same predictions for both. Generative theories assume that an interaction is required and for this reason they do not attribute causation or prevention to the second actor. But they make different predictions regarding the first actor. Wolff (2007) proposes that causation occurs when there is an initial tendency towards an outcome (e.g., for the coin to land on tails) and it is forced to a different outcome (e.g., the marble hits it and forces it to land on tails). Hence

29

Causal Attribution 30 his theory predicts that causation will be attributed to the first actor in the ‘mechanismcomplete’ scenario. The prediction is less clear in the ‘mechanism interrupted’ scenario because the theory is ambiguous with regard to when the analysis should be carried out. Assuming it is at the start of the scenario (i.e., before the marble is rolled), then the initial tendency is the same as the final outcome. Hence, the action in this scenario will be judged as an enabler. Wolff (2007) defines causation and prevention symmetrically and therefore he predicts that prevention will also be attributed in the first scenario only. Dowe (2000) defines causation in terms of a continuous causal mechanism but prevention in terms of an interruption to a causal mechanism. Only the mechanismcomplete scenario contains a continuous causal mechanism and therefore he predicts that causation will only be attributed to the action in this scenario. In contrast, there is an interruption to a causal mechanism in both the mechanism-complete scenario (the marble interrupts the coin’s movement) and the mechanism-interrupted scenario (the catch interrupts the roll of the marble) and hence on this account, prevention should be attributed in both scenarios.3 ******************** Insert Table 5 about here ******************** Method Materials and Design We constructed two scenarios, as described above. In one there was a complete mechanism from the action to the outcome and in the second the action interrupted a

3

As Dowe’s (2000) theory of prevention has a counterfactual element, we assume that he relies on methods used by counterfactual theorists to cope with pre-emption.

30

Causal Attribution 31 causal mechanism. Both scenarios involved pre-emption. Participants were given both scenarios in a counterbalanced order and after each one, they were asked four questions, that is, whether each of the two actors caused and prevented an outcome. Half of them answered the cause questions first and half answered the prevention questions first. Participants and Procedure Fifty seven participants were recruited mainly through a University campus-based electronic newspaper and they completed the study on-line. Participants were offered a small chance to win $25. After reading each scenario, they responded to the four questions by answering “yes” or “no”.

Results and Discussion As Table 6 shows, the majority of participants attributed causation to the first actor in the mechanism-complete scenario where there was a transmission along the pathway from an action (Billy throws a marble) to the outcome (the coin lands on heads; 84%). However, in the mechanism-interrupted scenario where an action (Frank catches the marble) interrupts a causal mechanism and hence the outcome is unchanged (the coin lands on heads), the majority of participants attributed causation to neither (77%). The percentage of participants who made an attribution to the first actor was significantly higher in the mechanism-complete scenario than in the mechanism-interrupted scenario (McNemar Test, p < .001). Once again, the results support the predictions of generative theories that causal judgments depend on the presence of a complete causal mechanism. Covariation and counterfactual theories provide no account of the difference between the mechanism-

31

Causal Attribution 32 complete and interrupted scenarios. Other make-a-difference accounts also fail. Spellman’s (1997) crediting causality theory predicts that causal attribution depends on the change in probability before and after the event. In our scenarios, the outcomes were deterministic so a change in the action would lead to a certain change in the outcome in both cases. Hence, this account also cannot explain our findings. Despite the large difference in attributions across the scenarios a significant minority (23%) of participants attributed causation in the no-mechanism scenario. The result suggests that in a small number of cases, people do use counterfactuals to attribute causation. In contrast, the pattern of prevention attributions was similar for the mechanismcomplete and mechanism-interrupted scenarios. Approximately two-thirds of the participants attributed prevention to the first actor (67% and 65% for the mechanism and no-mechanism scenarios respectively, and these did not differ significantly, p > .9). Fewer participants attributed prevention when the mechanism was interrupted in this study than in the previous study. The main difference is that in the previous study no outcome occurs and this may provide the best example of prevention (Dowe, 2000). In the scenarios in the present studies, an alternative outcome occurs (e.g., the coin falls on heads instead of tails). Nonetheless, the modal response was for participants to attribute prevention in both scenarios. The result is predicted by counterfactual theories (e.g., Lewis, 2000) and Dowe’s (2000) causal process theory. The result cannot be explained by covariation theories given the existence of a pre-emptive cause. Nor can it be explained by Wolff’s (2007) dynamic model. Once again, participants were less consistent in their responses to questions about prevention compared to questions about causation providing further evidence that the meaning of prevention is less clear-cut.

32

Causal Attribution 33 ******************** Insert Table 6 about here ******************** Across the two scenarios, the second actor was judged to have prevented the outcome (15%) more than to have caused it (8%, Wilcoxon Test, z = 2.00, p <.05). People often talk about prevention in the sense of having the potential to block some event even if that event does not occur (e.g., the lock is preventing the bike from being stolen). The result is consistent with our earlier suggestion that although people tend to select one event as being ‘the cause’ of an outcome, people may be less likely to do so for prevention. Finally our results showed clearly that participants do not judge prevent to mean the same thing as to cause it not to occur. When a mechanism was interrupted they tended to attribute prevention but not causation. The result is contrary to most theories of prevention but is predicted by Dowe (2000). When the causal mechanism was interrupted, people judged the first actor to have prevented the coin from landing on tails (65%) more often than to have caused the coin to land on heads (23%, McNemar Test, p < .001). According to generative theories, people should attribute causation instead to an alternative mechanism. To verify this, we ran a follow-up experiment using the same two scenarios but using open questions, e.g., ‘what caused the coin to land on heads?’ Consistent with the predictions of generative theories, the modal response in the interruption scenario was to attribute causation to another mechanism, i.e., the spinning of the coin. We carried out a further follow-up study using a scenario similar to the interruption scenario but rather than interrupting the roll of the marble, the actor removes

33

Causal Attribution 34 a book which is blocking the path between the marble and the coin. Once again there was no mechanism linking the action to the outcome but this time there was a change to the default outcome: the marble hits the coin and it falls4. The pattern of results was the same as for our interruption scenario. Participants failed to attribute causation in the absence of a mechanism even when the action led to a change in the outcome.

Experiment 6: Controlling for the Description of the Outcome Generative theories predict the difference in attributions for causation and prevention that we have observed so far. Counterfactual theories cannot. However, prevention questions by their nature refer to counterfactual outcomes whereas causation questions refer to factual outcomes. One possibility is that by referring explicitly to a counterfactual outcome in the question, we are encouraging people to think about counterfactual possibilities more in our prevention than causation questions. The aim of our next experiment was to control for the actuality of the outcomes in attributions of causation and prevention. To do so, we used the same scenarios as in Experiment 5 but we asked whether the actor caused or prevented an event rather than the specific outcome. The predictions are identical to those in Experiment 5, as Table 5 shows. All theories except Dowe’s (2000) predict that people should respond the same way to questions about causation and prevention.

Method Materials and Design

4

We thank Jim Woodward for suggesting this.

34

Causal Attribution 35 Participants were given the same two scenarios as used in Experiment 5, one in which a causal mechanism is complete and one in which a causal mechanism is interrupted. But in this case we framed the questions without reference to a factual or counterfactual outcome. Instead, in each case we asked: Did Billy cause an event to occur? and Did Billy prevent an event from occurring? Participants read both scenarios in a counterbalanced order. After each one they answered four questions, that is, whether each actor caused or prevented an outcome. Half of the participants answered the causation questions first and half answered the prevention questions first. Participants and Procedure Fifty participants read both scenarios and again in each case, they responded to each of the four questions by answering “yes” or “no”. They also rated their confidence on a scale of 1 to 7.

Results As Table 7 shows, attributions of causation and prevention differed even when the prevention question was framed without reference to an explicit counterfactual outcome. Once again, participants attributed causation to the first actor more often when the mechanism was complete (82%) than when the mechanism was interrupted (51%, Wilcoxon Test, z = 3.27, p = .001). Many participants attributed causation to both actors particularly in the mechanism-complete scenario probably because the question left open

35

Causal Attribution 36 the possibility that any event in the scenario could be caused and each actor might have caused something. ******************** Insert Table 7 about here ******************** In contrast to previous experiments, participants attributed prevention to the first actor more often when the mechanism was interrupted (84%) than when it was complete (62%, McNemar Test, p < .01). Once again, the difference between this result and that found in previous experiments is likely to be due the fact that any event in the scenario could be prevented. Some people may judge that Frank prevented Jane from catching the marble but not that he prevented the coin from landing on tails. The results suggest that the difference between causation and prevention found in Experiment 5 is not triggered by the explicit mention of counterfactual outcomes in the case of prevention questions and factual questions in the case of causal questions. Instead the results appear to arise because of differences in the focal events during judgments of causation and prevention.

General Discussion The results of our experiments suggest, first, that causal attribution tends to rely on people’s understanding of the process involved in bringing about an outcome. Judgments of whether A caused B tend to depend on the presence of a causal mechanism. When some conserved quantity, such as force, is transferred from A to B, then people make attributions of causation. When nothing is passed from A to B, people are more reluctant to make attributions of causation even though a change to A would bring about a change 36

Causal Attribution 37 to B. Only generative theories predict this result (e.g., Dowe, 2000; Mandel, 2003; Wolff, 2007). Second, individuals are less consistent in their judgments of prevention. Attributions are most likely to occur when there is an interruption to a causal mechanism as generative theories predict but sometimes also occurs when there is no interaction with a causal mechanism as make-a-difference theories predict. Third, cause and prevent are not always interpreted symmetrically. Individuals understand ‘prevent’ to mean something different from ‘cause not’. This pattern of results is consistent with Dowe’s (2000) causal process theory but inconsistent with theories based on counterfactuals (e.g., Halpern & Pearl, 2001; Lewis, 2000), covariation (e.g., Cheng & Novick, 1990) and force dynamics (Wolff, 2007). Table 8 summaries the match between our results and the predictions of each of the theories. Dowe’s (2000) theory clearly fairs the best explaining all but a few of the responses to questions about prevention. In our first experiment, we compared scenarios where there was a continuous causal process with one without a continuous mechanism linking cause and effect. Consistent with generative theories but not with make-a-difference theories, people attribute causation more often when a continuous causal process is present. In Experiment 2, we generalised the results of previous studies to new contexts involving negative outcomes, abnormal actions, and to cases of causation by omission. Again, people attributed causation more often when there was a continuous mechanism linking the action (e.g., kicking the ball) to the outcome (e.g., the ball landing in the goal) than when the actor failed to interrupt the mechanism. In Experiment 3, we showed that people rely on an analysis of the mechanisms to attribute causation even when the result is inconsistent with the earlier pattern of

37

Causal Attribution 38 covariation. The result raises difficulties for the view that causation is attributed to events that sufficient for or co-occur with the outcome (Cheng & Novick, 1990; Mandel, 2003). In Experiment 4, we tested generative and make-a-difference theories of prevention using three scenarios: the agent interrupts an action, the agent initiates and interrupts an action, or there is a failure to act. Participants ascribed prevention more often in the first two scenarios which involved an interruption to the mechanism than in the third which did not. The result is consistent with generative theories of prevention and cannot be accounted for by make-a-difference theories. There was no significant difference between the first two scenarios ruling out the possibility that people attribute prevention only if the agent’s absence would have made a difference to the outcome. Nonetheless, many participants made attributions of prevention even when there was a failure to act as predicted by make-a-difference theories only. The results overall therefore suggest that the meaning of prevention may be less clear-cut than the meaning of causation perhaps because even the generative definition involves judgments about a counterfactual outcome. In our fifth experiment, we compared judgments for two scenarios involving preemption, one with a causal mechanism connecting the cause (the marble is rolled) to the effect (the coin lands on heads) and one in which a causal mechanism is interrupted, and we asked causation and prevention questions about each. Once again, the results showed that causation questions were highly sensitive to the presence of a causal mechanism corroborating the predictions of generative theories but not counterfactual theories. Participants attributed a cause to the actor only when the scenario involved a causal mechanism. Many participants attributed prevention in both scenarios, that is, regardless

38

Causal Attribution 39 of whether the action interacts with the spinning of the coin (mechanism-complete scenario) or with the rolling of the marble (mechanism-interrupted scenario). The results of Experiment 5 showed that ‘prevent’ means something different from ‘cause not’, a finding that is inconsistent with covariation (e.g., Cheng & Novick, 1990), counterfactual (e.g., Lewis, 2000) and dynamics theories (Wolff, 2007). The difference is predicted by Dowe (2000). In Experiment 6, we examined whether the difference between judgments of causation and prevention arose because prevention questions explicitly mentioned a counterfactual outcome and for this reason tended to elicit counterfactual reasoning. We reframed the prevention question by asking whether an event was prevented. Despite this change, the difference in attributions of causation and prevention persisted. The result supports the view that the difference between these attributions does not arise merely because prevention questions tend to refer to an explicit counterfactual event. Overall, the results show that people are sensitive to the presence of causal mechanisms and they make different attributions depending on them. The results are consistent with the predictions made by generative theories. In contrast, make-adifference theories predict attributions of causation regardless of the nature of the mechanisms and hence have no means to explain the complete pattern of judgments people made in our studies. Generative theories can account for most of the results for causation but not all. Some participants attributed causation even when nothing was passed from A to B, particularly for cases of causation by omission. As previously discussed, even generative theorists such as Dowe (2000) allow that when the cause or effect is an absence, there must be a counterfactual element in the judgment. And

39

Causal Attribution 40 sometimes people seem to use counterfactuals even when the events do not involve absences. One explanation for this is that all of our scenarios involved actions. Actions may lead attributions of responsibility or blame to contaminate causal attributions, and responsibility and blame may be more likely to depend on counterfactuals. The results for prevention were more complex and comprised four main findings. First, individuals are less consistent in their responses to questions about prevention than causation and hence no theory can explain the responses of all individuals. Despite this some theories faired better than others. Our second finding was that attributions of prevention were more likely when there was an interruption to a causal mechanism than when there was a failure to act. This result is predicted by generative theories and cannot be explained by make-a-difference theories. Third, half of participants made attributions of prevention following a failure to act. This account can only be explained by make-adifference theories. The results suggest that individuals are more willing to attribute prevention in a wide range of situations including those that do not require an analysis of mechanisms. One likely reason for this is that all theories of prevention, including generative accounts, rely on counterfactuals in their definitions. Finally, our results showed that the meanings of causation and prevention are not symmetrical. When a mechanism is interrupted, prevention is usually attributed whereas causation is not. Another difference between the judgments is simply that prevention judgments showed more variability than causation judgments. One possibility is that the meaning of “prevent” is vaguer than the meaning of “cause.” “Prevent” is normally used in negative contexts, when outcomes do not occur. “Cause” is at least as likely in positive contexts,

40

Causal Attribution 41 when outcomes do occur. Negative contexts may be less well articulated by language and thought than positive contexts. Theories of causal attribution sometimes assume that attributions are more likely to be made to actions that are intentional (Hart & Honore, 1959/1985; McClure, Hilton & Sutton, 2005) and in legal contexts intention is often an important factor in determining whether an accusation is made. In a study not reported here, we compared attributions to the same actions described either as purposeful or accidental; for example, a marble is dropped on purpose or by accident. Attributions to the actors were slightly lower when described as accidents but the pattern of results was the same as in the experiments reported here. The results support the view that, at least in this context, people make judgments in the same way for intentional and unintentional actions. The examples used in our experiments all involved visually observable physical interactions where the mechanisms are very salient. In addition, our scenarios all involved actions and it may be the goal-directed nature of actions that provides or enhances the representation of the mechanisms involved. Indeed, there is neuroscientific evidence that different brain areas are activated during the learning of social and physical causal relations (Overwalle, in press). Hence, it may be that our examples are particularly likely to elicit causal judgments based on an analysis of the mechanisms involved. Clearly many causal relations are not of this nature. For example, causation can involve social influence or people may have reasons which lead them to carry out a particular action. Furthermore, physical mechanisms can involve a transmission process involving observable physical contact between cause and effect as in our examples of the marble hitting the coin, or the transmission process may be invisible if wind from a fan caused

41

Causal Attribution 42 the coin to fall (Shultz, 1982). People may think about these cases by analogy to physical causation (Wolff, 2007). Although we have only studied cases involving physical interactions in our studies, there is some evidence that people draw on generative mechanisms in social contexts (Wolff, 2007). Generative theories are not without their difficulties. They are less well-specified than make-a-difference theories. They depend on an analysis of which quantities are conserved and how they survive interactions that transfer those quantities from one entity to another. For example, if a stone and a ping pong ball simultaneously hit a window, both transfer energy to it and the theory needs to distinguish which one is the cause. A further problem is that the conserved quantity that defines the mechanism at issue must be chosen so as not to beg the question. If we define causation at one level by referring to a mechanism at a lower level then this may lead to an infinite regress if the mechanism itself requires the notion of cause to specify it. Indeed, the current notion of mechanism requires further elaboration. Schaffer (2000) offers multiple examples of causes that disconnect a cause from its effect. For instance, pulling a trigger causes a gun to shoot through disconnection: It moves a part (the sear) that otherwise would inhibit a spring from uncoiling, and the action of the spring causes an explosion that propels the bullet. For a mechanism theory to be viable, it must either define people’s understanding of mechanism to include such cases, or people who fully understand the operation of such mechanisms must be less willing to make the relevant causal attributions. Of course, there are many cases where people do not know the mechanism and yet are willing to attribute causation (Keil, 2003). For example, when we swallow an aspirin for a headache, we generally do not know how it works but we will readily attribute causation

42

Causal Attribution 43 to the aspirin when our headache disappears. People need only be aware that some kind of mechanism is plausible to make a causal inference. When it is clear that there is no interaction, as in the examples in our scenarios where a mechanism was interrupted or a causal pathway was unblocked, individuals generally no longer attribute causation. Our results suggest that judgments of causation may be highly sensitive to the kinds of questions that people are asked (Hilton, 1990). People may make different judgments depending on whether they are to decide if something caused or enabled an outcome (Goldvarg & Johnson-Laird, 2001) or when they are asked to make judgments of blame or responsibility. The results may also vary depending on the framing of the question. For example, individuals often select a single event from a scenario when asked to make causal judgments even though in general many events contribute to an outcome. Hence, asking whether something is “a cause”, as compared to “the cause”, may cue them to consider more than just one event. They may also respond differently when asked to judge related terms such as ‘bring about’. The task we posed to participants was essentially linguistic, asking them whether the verb “cause” or “prevent” was an appropriate characterization of a scene. Such linguistic judgments presumably derive from a conception of the scene. If we are right that the notion of causal mechanism is necessary to explain how people attribute cause, then that suggests that people have access to a notion of mechanism that could be critical in a variety of other conceptual tasks as well, like explanation, induction, and decision making (cf. Sloman, 2005). Causal judgments are ubiquitous in our everyday thinking as well as in domains ranging from science to the law. We suggest some steps toward the development of an

43

Causal Attribution 44 understanding of this fundamental process. Future theories of causal attribution must take account of how people understand how an outcome comes about.

44

Causal Attribution 45 References Ahn, W., Kalish, C.W., Medin, D.L., & Gelman, S.A. (1995). The role of covariation versus mechanism information in causal attribution. Cognition, 54, 299-352. Allan, L. G. (1980). A note on the measurement of contingency between two binary variables in judgment tasks. Bulletin of the Psychometric Society, 15, 147-149. Baillargeon, R, Kotovsky, L. & Needham, A. (1995). The acquisition of physical knowledge in infancy. In D. Premack & D. Sperber (Eds), Causal cognition: A multidisciplinary debate. (pp.79-116). New York, NY, US: Clarendon Press/Oxford University Press. Barbey, A., & Wolff, P. (2007). Learning causal structure from reasoning. In Proceedings of the Twenty-Ninth Annual Conference of the Cognitive Science Society. Hillsdale, NJ: Erlbaum. Bennett, J. (2003). A Philosophical Guide to Conditionals. New York: Oxford University Press. Bullock, M. (1985). Causal reasoning and developmental change over the preschool years. Human Development, 28, 169-191. Cheng, P.W. (1997). From covariation to causation: a causal power theory. Psychological Review, 104, 367-405. Cheng, R W., & Novick, L. R. (1990). A probabilistic contrast model of causal induction. Journal of Personality and Social Psychology, 58, 545-567. Cheng, R W., & Novick, L. R. (1991). Causes and enabling conditions. Cognition, 40, 83-120. Collins, J. (2000). Preemptive Prevention. Journal of Philosophy, 97, 223–34 45

Causal Attribution 46 Cummins, D. D. (1995). Naïve theories and causal deduction. Memory & Cognition, 23, 646–658. Dowe, P. (2000). Physical Causation. New York: Cambridge University Press. Gavanski, I., & Wells, G. L. (1989). Counterfactual processing of normal and exceptional events. Journal of Experimental Social Psychology, 25, 314-325. Goldvarg, Y. & Johnson-Laird, P. N. (2001) Naive causality: a mental model theory of causal meaning and reasoning. Cognitive Science, 25, 565-610. Halpern, J.Y. & Pearl J. (2001). Causes and Explanations: A Structural-Model Approach – Part I: Causes. In Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence (pp. 194-202). San Francisco, CA: Morgan Kaufmann. Hall, N. and Paul, L.A. (2003). Causation and Preemption. In P. Clark & K. Hawley (Eds.), Philosophy of Science Today (pp.100–130). Oxford: Oxford University Press. Hart, H. L., & Honore, A. M. (1985). Causation in the law (2nd ed.). Oxford, England: Oxford University Press. (Original work published 1959). Hilton, D. J. (1990). Conversational processes and causal explanation. Psychological Bulletin, 107, 65–81. Hilton, D. J., & Slugoski, B. R. (1986). Knowledge-based causal attribution: The abnormal conditions focus model. Psychological Review, 93, 75-88. Hitchcock, C. (2001). The intransitivity of causation revealed in equations and graphs. Journal of Philosophy, 98, 273-299. Hume, D. (1988). An Enquiry Concerning Human Understanding. A. Flew. (Ed.) La Salle, IL: Open Court. (Originally published 1748.)

46

Causal Attribution 47 Kagan, S. (1989). The Limits of Morality. Oxford: Oxford University Press. Kahneman, D. & Miller, D. T. (1986). Norm theory: Comparing reality to its alternatives. Psychological Review, 93, 136-153. Kant, I. (1781). The critique of pure reason. 1985 Web version by Palgrave Macmillan. Kelley, H.H. (1973). The processes of causal attribution. American Psychologist, 28, 107-128. Keil, F.C. (2003). Folkscience: Coarse interpretations of a complex reality. Trends in Cognitive Science, 7, 368-373. Lagnado, D. A., Waldmann, M. R., Hagmayer Y., & Sloman, S. A. (2007). Beyond covariation: Cues to causal structure. In Gopnik, A. & Schulz, L. (Eds.), Causal learning: Psychology, philosophy, and computation (pp. 154-172). Oxford: Oxford University Press. Lewis, D. (1973). Counterfactuals. Oxford: Blackwell. Lewis, D. (1986). Philosophical Papers: Volume II. Oxford: Oxford University Press. Lewis, D. (2000). Causation as influence. Journal of Philosophy, 97, 182-197. Mandel, D.R. (2003). Judgment dissociation theory: An analysis of differences in causal, counterfactual, and covariational reasoning. Journal of Experimental Psychology: General, 132, 419–434. Mandel, D. R. & Lehman, D. R. (1996). Counterfactual thinking and ascriptions of cause and preventability. Journal of Personality and Social Psychology, 70, 450-463. McClure, J., Hilton, D.J., & Sutton, R. M. (2005). Probabilistic and social functionalist criteria for causal attribution: Judgments of voluntary and physical causes in causal chains. .Manuscript submitted for publication.

47

Causal Attribution 48 Pearl, J. (2000). Causality: Models, Reasoning and Inference. Cambridge: Cambridge University Press. Read, S.J. & Montoya, J.A. (1999). An Autoassociative Model of Causal Reasoning and Causal Learning: Reply to Van Overwalle’s (1998) Critique of Read and MarcusNewhall (1993). Journal of Personality and Social Psychology, 76, 728-742. Reichenbach, H. (1938). Experience and prediction. Chicago: University of Chicago Press. Roese, N. J. (1997). Counterfactual thinking. Psychological Bulletin, 121, 133-148. Salmon, W. (1984). Scientific Explanation and the Causal Structure of the World. Princeton: Princeton University Press. Salmon, W. (1997). Causality and Explanation: A Reply to Two Critiques. Philosophy of Science, 64, 461-477. Schaffer, J. (2000). Causation by disconnection. Philosophy of Science, 67, 285-300. Shultz, T.R. (1982). Rules of causal attribution. Monographs of the Society for Research in Child Development, 47, 1-51. Sloman, S. A. (2005). Causal models: How people think about the world and its alternatives. New York: Oxford University Press. Sloman, S. A., Barbey, A. K., & Hotaling, J. (in press). A causal model theory of the meaning of cause, enable, and prevent. Cognitive Science. Spirtes, P., Glymour, C., & Scheines, R. (1993). Causation, prediction, and search. New York: Springer-Verlag. Spellman, B. A. (1997). Crediting causality. Journal of Experimental Psychology: General, 126, 323–348.

48

Causal Attribution 49 Suppes, P. (1970) A probabilistic theory of causality. Acta Philosophica Fennica, 24. Amsterdam: North-Holland. Talmy, L. (1988). Force dynamics in language and cognition. Cognitive Science, 12, 49– 100. Van Overwalle, F. (in press). Social Cognition and Brain. Human Brain Mapping. Van Overwalle, F. (1998). Causal Explanation as Constraint Satisfaction: A Critique and a Feedforward Connectionist Alternative. Journal of Personality and Social Psychology, 74, 312-328. Wells, G. L., & Gavanski, I. (1989). Mental simulation of causality. Journal of Personality and Social Psychology, 56, 161-169. Wolff, P. (2007). Representing Causation. Journal of Experimental Psychology: General, 136, 82 – 111. Woodward, J. (2003). Making Things Happen: A Theory of Causal Explanation. New York: Oxford University Press.

49

Causal Attribution 50 Acknowledgments We thank Jean-François Bonnefon, Denis Hilton, Joe Halpern, Rui da Silva Neves and Jim Woodward for discussions on this topic. This research was supported by NASA grant NCC2-1217 and NSF Award 0518147 to Steven Sloman. Some of these studies were presented at the 28th Annual Conference of the Cognitive Science Society and at the 14th meeting of the European Society for Cognitive Science.

50

Causal Attribution 51 Table 1: Predictions made by theories of causation for scenarios involving a continuous causal mechanism and no continuous causal mechanism in Experiments 1 and 2

Continuous Causal

No Continuous Causal

Mechanism

Mechanism

Make a Difference Theories Covariation Theory

Cause

Cause

Cause

Cause

Cause

Not Cause

Cause

Not Cause

(e.g., Cheng & Novick, 1990) Counterfactual Theories: (e.g., Lewis, 2000)

Generative Theories Causal Process Theories: (e.g., Dowe, 2000 / Mandel, 2003) Dynamics Model (Wolff, 2007)

51

Causal Attribution 52 Table 2: The percentage of “yes” responses to the questions in Experiment 1

Mechanism Complete

Mechanism Not Blocked

Scenario:

Mean

Broken Window

91

21

Spilt Bottle

93

7

Falling Boulder

77

49

Into the Pool

88

19

87

24

52

Causal Attribution 53 Table 3: The percentage of “yes” responses to the questions about the target actor in Experiment 2

Mechanism Complete

Soccer Scenario

Mechanism Not Blocked

70

49

Champagne Bottle Scenario 76

51

53

Causal Attribution 54 Table 4: Percentage of ‘Yes’ Responses for each of the three scenarios in Experiment 4 along with the predictions made by theories of prevention

% of Yes Responses

Interrupt a

Initiate & Interrupt

Causal Mechanism

Causal Mechanism

Omission

(Frank)

(John)

(Billy)

85

77

46

Prevent

Prevent

Prevent

Prevent

Prevent

Prevent

Prevent

Not Prevent

Not Prevent

Make a Difference Theories Covariation Theory (e.g., Cheng & Novick, 1990) Pure Counterfactual Theories: - if event didn’t occur (e.g., Lewis, 2000) - if agent was absent (e.g., Kagan, 1989)

Generative Theories Causal Process &: Counterfactual (Dowe, 2000)

Prevent

Prevent

Not Prevent

Prevent

Prevent

Not Prevent

Dynamics Model (Wolff, 2007)

54

Causal Attribution 55 Table 5: Predictions made by the theories for scenarios involving a continuous causal mechanism and an interruption to a causal mechanism in Experiments 5 and 6

Continuous

Interruption to a

Causal Mechanism

Causal Mechanism

(with pre-emption)

(with pre-emption)

Not Cause & Not Prevent

Not Cause & Not Prevent

Cause & Prevent

Cause & Prevent

Cause & Prevent

Not Cause & Prevent

Cause & Prevent

Not Cause & Not Prevent

Cause & Prevent

Not Cause & Prevent

Make a Difference Theories Covariation Theory * (e.g., Cheng & Novick, 1990) Counterfactual Theories: (e.g., Lewis, 2000)

Generative Theories Causal Process Theories: (e.g., Dowe, 2000) Dynamics Model (Wolff, 2007)

Data Modal Response

* Covariation theories have yet to deal with situations involving pre-emption and hence they predict that in these cases events do not make a difference to the outcome 55

Causal Attribution 56 Table 6: The percentage of “yes” responses to the eight questions in Experiment 5

Mechanism Complete Cause

Mechanism Interrupted

Prevent Question

Cause

Question

Prevent Question

Question

First actor total

84

67

23

65

First actor only

74

55

18

52

Both actors

10

12

5

14

Second actor total

10

16

5

14

Second actor only

0

4

0

0

Both actors

10

12

5

14

16

30

77

34

(Billy / Frank)

(Suzy / Jane)

Neither

56

Causal Attribution 57 Table 7: The percentage of “yes” responses to the eight questions in Experiment 6

Mechanism Complete

Mechanism Interrupted

Cause

Prevent

Cause

Question

Question

Question

82

62

51

84

First actor only

54

48

35

70

Both actors

28

14

16

14

Second actor total

30

18

18

14

2

4

2

0

28

14

16

14

16

34

34

16

First actor total

Prevent Question

(Billy / Frank)

(Suzy / Jane) Second actor only Both actors

Neither

57

Causal Attribution 58 Table 8: The match between our results and the predictions of theories of causation. ‘ ’ indicates a match, ‘x’ indicates there was not a match and ‘p’ indicates a partial match Causation

Prevention

Prevent

depends on

depends on

does not

a Causal

an interruption

mean

process

or a change

‘Cause not’

to the outcome

Covariation Theory x

p

x

x

p

x

(e.g., Cheng & Novick, 1990)

Pure Counterfactual Theories: (Lewis, 2000/ Halpern & Pearl, 2001)

Causal Process Theory p

(Dowe, 2000)

Dynamics Model (Wolff, 2007)

p

x

-

-

Judgment Dissociation Theory Mandel, (2003)

p

58

Causal Attribution 59 Appendix: Scenarios used in Experiment 1: Frank and Sam are playing ball on the street with friends. Frank kicks the ball hard and unintentionally sends it in the direction of a neighbour’s house. Fortunately, Sam is blocking the path of the ball. However, at that moment Sam is distracted from the game and steps away. The ball flies past Sam and it hits the window of the neighbour’s house. The window smashes. 1. Did Frank cause the window to break? Yes or No? 2. Did Sam cause the window to break? Yes or No?

You are at a party and there is an open bottle of beer on the table. Mike turns around and accidentally knocks against the bottle. Jack sees that the bottle is about to fall. It is easily within his reach and he is just about to catch it when Peter accidentally knocks against him. Jack doesn’t manage to grab the bottle and it falls over and spills. 1. Did Mike cause the bottle to fall? Yes or No? 2. Did Peter cause the bottle to fall? Yes or No?

There is a boulder on a hill above a cliff. Marcus leans against it and unintentionally dislodges it and it begins to roll down the hill. There is a closed gate between the boulder and the cliff edge and the boulder will stop at the gate. John is walking by, opens the gate 59

Causal Attribution 60 to go through and accidentally fails to fasten the lock properly. The boulder rolls through the gate, it reaches the edge of the cliff and it falls off. 1. Did Marcus cause the boulder to fall off the edge of the cliff? Yes or No? 2. Did John cause the boulder to fall off the edge of the cliff? Yes or No?

The girls are on holidays and are playing around the swimming pool. Rachel accidentally pushes Helen. The ground is wet and Helen slips and is about to fall into the pool. Alison sees and is just about to catch Helen and stop her from falling. Kate also begins to slip and grabs Alison by the T-shirt. Alison doesn’t catch Helen and Helen falls into the pool. 1. Did Rachel cause Helen to fall into the pool? Yes or No? 2. Did Kate cause Helen to fall into the pool? Yes or No?

Champagne Bottle Scenario used in Experiment 2 Mechanism Complete Version Suzy and Sydney are playing catch. Sydney throws the ball to Suzy and her aim is on target. Suzy tries to catch it, but the ball flies right through her hands and hits a valuable champagne bottle behind her. The bottle breaks. Did Suzy cause the bottle to break? Yes or No?

60

Causal Attribution 61 Mechanism Not Blocked Version Suzy is bouncing a ball against a wall. She throws the ball at the wall, and the ball bounces back towards her. She tries to catch it, but the ball flies right through her hands and hits a valuable champagne bottle behind her. The bottle breaks. Did Suzy cause the bottle to break? Yes or No?

61