1
Running head: CAUSAL AND COUNTERFACTUAL EXPLANATION
Mental Simulation and the Nexus of Causal and Counterfactual Explanation
David R. Mandel
Defence R&D Canada – Toronto
For correspondence:
Dr. David R. Mandel
Leader, Thinking, Risk, and Intelligence Group
Adversarial Intent Section
Defence R&D Canada – Toronto
1133 Sheppard Avenue West
P.O. Box 2000
Toronto, ON M3M 3B9
Canada
Phone: (416) 635-2000 ext. 3146
Fax: (416) 635-2184
Email:
Acknowledgement. I wish to thank Jim Woodward and the editors for their insightful comments on an earlier draft of this paper.
1. Introduction
Attempts to make sense of specific episodes in the past, especially when they entail consequential, surprising, or unwanted outcomes, tend to involve an inter-related set of causal and counterfactual questions that people may pose to themselves or to others: Why did it happen? How could it have happened? How might it have been prevented? And, so on. Given the transactional nature of such questions, the answers provided may be regarded as explanations (Keil, 2006). Such explanations have long been explained themselves in terms of the functional benefit of prediction and learning that they confer when they are accurate (Heider, 1958). However, such explanations, especially in cases involving harm, also underlie people’s moral cognitions and ‘prosecutorial mindsets’ (Tetlock et al., 2007), serving as bases for addressing other related ‘attributional’ questions such as: Who is responsible? Who is to blame? What response—for instance, in terms of punishment or compensation—would be fair? And, so on.
For a few decades now, experimental psychologists have sought to understand the cognitive, motivational, and functional bases for such post-event querying. An important part of that endeavor has focused on elucidating the nature of the relationship between the various forms of causal and counterfactual thinking, which appear to give rise to the answers people provide to such queries. In this article, I examine the role of mental simulation (Kahneman and Tversky, 1982a)—the cognitive process whereby possibilities are brought to mind through mental construction—in causal and counterfactual explanations. I begin in Part 2 by discussing reasons for my emphasis on explanation as opposed to thinking or reasoning.
In Part 3, I trace the development of the mental simulation construct from Kahneman and Tversky’s (1982a) seminal chapter on the simulation heuristic, noting how other psychologists have drawn on their notions of simulation and counterfactual thinking. My aim is Part 3 is largely two-fold. Although Kahneman and Tversky’s brief chapter on mental simulation was highly generative of subsequent research on counterfactual thinking, many of the ideas sketched, or simply alluded to, in the chapter have not been adequately discussed. Hence, one aim here is to reflect, and possibly expand, on some of those notions. For example, I explore some process-related issues pertaining to mental simulation that have not previously been discussed in the literature. My second objective is to critically examine how theorists, largely in social psychology, have drawn on the simulation heuristic notion to make claims about the nature of causal explanation. In doing so, I review psychological research on adults (for overviews of research on children, see in this volume: Beck and Rigs; McCormack, Hoerl, and Butterfill; Perner and Rafetseder; and Sobel) that has tested these notions.
In Part 4, I summarize an alternative ‘judgment dissociation theory’ of counterfactual and causal explanations that has emerged in later work, largely in response to the earlier notions discussed in Part 3. In this account (e.g., Mandel, 2003, 2005), although mental simulations play a role in both causal and counterfactual explanations, the focus of each type of explanation is different. Specifically, causal explanations tend to focus on antecedents that were sufficient under the circumstances to yield the actual event, whereas counterfactual explanations tend to focus on (the mutation of) antecedents that would have been sufficient to prevent the actual outcome and others like it from occurring. These different foci lead to predictable dissociations in explanatory content, which have been confirmed in recent experiments (e.g., Mandel, 2003; Mandel and Lehman, 1996). The chapter concludes with a discussion of the compatibility of these ideas with the kind of interventionist account that Woodward (this volume) seeks to advance.
To set the stage for the foregoing discussion, it is important to point out, as the opening paragraph suggests, that I am mainly concerned here with explanation of tokens (i.e., particular cases) rather than of types (i.e., categories of cases). The studies I review, which were largely the result of the generative effect of Kahneman and Tversky’s work on the simulation heuristic, tend to focus on people’s explanations of negative past outcomes, such as why a particular protagonist died or how he could have been saved rather than what the most probable causes of death are or how life expectancy might generally be improved. Whereas causal and counterfactual reasoning about types focuses on ascertaining ‘causal laws’ (Cheng, 1993), causal reasoning about tokens may draw on knowledge about causal laws to answer attributional queries in ways that need not generalize to other cases, but that nevertheless constitute ‘causal facts.’ Woodward (this volume) makes a similar distinction, and applies his interventionist analysis to type rather than token causation. Towards the end of the chapter, I shall return to this issue in order to reflect on the compatibility of interventionism and judgment dissociation theory.
2. Why Explanation?
I use the term explanationrather than other terms such as thinking or reasoning in this chapter for two reasons. First, I believe that much of the emphasis on counterfactual and causal thinking about tokens, at least, functions to support explanation. Explanations, as noted earlier, are transactional (Keil, 2006), and subject to conversational norms (see, e,g., Grice, 1975; Hilton, 1990; Wilson and Sperber, 2004). Thus, explanations not only depend on the explainer’s understanding of the topic, but also his or her assumptions or inferences regarding what the explainee may be seeking in a response. A good explanation for one explainee therefore may not be so for another, provided their epistemic states differ (e.g., Gärdenfors, 1988; Halpern and Pearl, 2005) or they seek different kinds of explanation (see also Woodward, this volume). For instance, harkening back to Aristotle’s four senses of (be)cause (see Killeen, 2001), an explainer might give one individual seeking a mechanistic ‘material cause’ account of an event quite a different explanation than he or she would give to another individual seeking a functional ‘final cause’ explanation of the same event.
The transactional quality of explanation also leads to my second reason for focusing on explanation, and that is to better reflect the reality of the experimental context in which participants are asked to provide responses to questions posed by researchers. In studies I subsequently review, participants are usually asked to read a vignette about a chain of events that culminate in the story’s outcome. Participants are then asked to indicate what caused the outcome and/or how the outcome might have been different ‘if only....’ Thus, the participant in a psychological experiment faces many of the same challenges that any explainer would face.
The challenges, however, are in many ways much greater in the experimental context because the tasks imposed on the participant often violate conversational rules that would normally help explainers decide how to respond appropriately. For instance, in many everyday situations the reason why an explanation is sought may be fairly transparent and well indicated by the question itself. When it is not, the explainer can usually ask for clarification before formulating their response. In contrast, the experimental context often intentionally obscures such cues and denies cooperative opportunities for clarification so that the purpose of the experiment or the hypotheses being tested may remain hidden from the participant, and also so that all participants within a given experimental condition are treated in the same way. Moreover, given that the experimenter both provides participants with the relevant case information and then requests an explanation of the case from them, it may suggest to participants that they are being ‘tested’ in some manner (which of course they are). As Woodward (this volume) correctly observes, in many of the vignettes used in psychological studies the causal chain of events leading from the story’s beginning to its ending are fairly complete. Thus, asking for an explanation may seem as odd as the answer would appear obvious. While I don’t think the peculiarities of psychological research necessarily invalidate the exercise, it is important to bear in mind that the data produced by participants are attempts at explanation that are not only constrained by ‘causal thinking’, but also by other forms of social, motivational, and cognitive factors that may have little, if anything, to do with causal reasoning per se.
Trabasso and Bartalone (2003) provide a good example of this. For years, it has been widely accepted that counterfactual explanations that ‘undo’ surprising outcomes tend to do so by mentally changing abnormal antecedents. This ‘abnormality principle’ traces back to influential papers in the psychological literature on counterfactual thinking—namely, Kahneman and Tversky’s chapter on the simulation heuristic and Kahneman and Miller’s (1986) norm theory. Trabasso and Bartalone, however, observed that abnormal events described in vignettes in experiments on counterfactual thinking tended to have more detailed explanations than normal events. This is unsurprising, since they were unusual. When the level of explanation was properly controlled, they found that counterfactual explanations no longer favored abnormal antecedents. Of course, their findings do not prove the unimportance of abnormality as a determinant of counterfactual availability, but the findings do illustrate the ease with which contextual features in experimental stimuli that influence participants’ explanations can be misattributed to fundamental aspects of human cognition. It would be useful for experimenters and theorists to bear this in mind, and I would hope that a focus on explanation, with all that it entails, may be of some use in doing that. For instance, the vignette experiments described in Hitchcock (this volume) might be profitably examined in these terms.
3. Mental Simulation: Towards a Psychology of Counterfactual and Causal Explanation
In the psychological literature, sustained interest in understanding the relationship between counterfactual and causal thinking can be traced back to a brief, but influential, chapter by Kahneman and Tversky (1982a), entitled ‘The Simulation Heuristic.’ In it, the authors attempted to differentiate their earlier notion of the availability heuristic (Tversky and Kahneman, 1973) from the simulation heuristic. Whereas the availability heuristic involves making judgments on the basis of the ease of mental recall, the simulation heuristic involved doing so on the basis of the ease of mental construction.
Kahneman and Tversky (1982a) did not say much about what specifically characterizes a simulation, though it is clear from their discussion of the topic that they regarded mental simulation as closely linked to scenario-based thinking, or what they have in other work (Kahneman and Tversky, 1982b) referred to as the ‘inside view,’ and which they distinguish from the ‘outside view’—namely, thinking that relies on the aggregation of statistical information across multiples cases, and which they argue is more difficult for people to invoke in the service of judgment and decision making. From their discussion, however, it would seem reasonable to infer that their notion of mental simulation was less restrictive than the manner in which representation is depicted in mental models theory (Johnson-Laird & Byrne, 2002), which, as I discuss elsewhere (Mandel, 2008), mandates that the basic unit of mental representation is expressed in terms of possibilities depicted in rather abstract form. Mental simulations would appear much more compatible with the representation of scenes or stories (with a beginning, middle, and end) than with the mere representation of possibilities.
A central theme running through all of Kahneman and Tversky’s program of research on heuristic and biases is that a person’s experience of the ease of ‘bringing to mind’ is often used as a proxy for more formal bases of judgment (e.g., see Kahneman, Slovic, and Tversky, 1982). For instance, in judging the probability of an event class, one might be inclined to judge the probability as relatively low if it is difficult to recall exemplars of the class (via the availability heuristic) or if it is difficult to imagine ways in which that type of event might occur (via the simulation heuristic). These heuristics ought to provide useful approximations to accurate assessments if mental ease and mathematical probability are highly correlated. However, they will increasingly lead people astray in their assessments as that correlation wanes in magnitude. Or, as Dawes (1996) put it, for a counterfactual—and even one about a particular instance or token—to be regarded as normative or defensible it must be ‘one based on a supportable statistical argument’ (p. 305).
Kahneman and Tversky (1982a; Kahneman and Varey, 1990) proposed that mental simulation played an important role in counterfactual judgments, especially those in which an event is judged to be close to having happened or having not happened. In such cases, they noted, people are prone to mentally undoing the past. Mental simulations of the past tend to restore expected outcomes by mutating unusual antecedents to more normal states and they seldom involve mutations that reduce the normality of aspects of the simulated episode. They referred to the former norm-restoring mutations as downhill changes and the latter norm-violating mutations as uphill changes to highlight the respective mental ease and effort with which these types of counterfactual simulations are generated. A number of other constraints on the content of mental simulations may be seen as examples of the abnormality principle. Some of these factors, such as closeness, are discussed by Hitchcock (this volume) and reviewed in depth elsewhere (e.g., Roese & Olson, 1995).
It is clear, even from Kahneman and Tversky’s brief discussion of mental simulation, that they do not regard all mental simulation as counterfactual thinking. The earlier example of using mental simulation to estimate the likelihood of an event by gauging the ease with which one can conjure up scenarios in which the judged event might occur offers a case in point. There is no presumption in this example of a counterfactual comparison. Nor does mental simulation even have to be an example of hypothetical thinking since the representations brought to mind might be regarded as entirely veridical. In this regard, mental simulation seems to be conceptually closer to the notion of imagining, but with the constraint that the function of such imagining is to inform judgments of one kind or another, often by using the ease of construction as a proxy for what otherwise would be a more laborious reasoning exercise.
Kahneman and Tversky (1982a) also proposed that mental simulation could play a role in assessments of causality:
To test whether event A caused event B, we may undo A in our mind, and observe whether B still occurs in the simulation. Simulation can also be used to test whether A markedly increased the propensity of B, perhaps even made B inevitable. (pp. 202-203).
Clearly, their proposal was measured. For instance, they did not propose that causal assessments required mental simulations. Nor did they propose that the contents of such simulations necessarily bound individuals to their seeming implications through some form of intuitive logic. Thus, at least implicitly, they left open the possibility that an antecedent that, if mutated, would undo the outcome could still be dismissed as a cause (and certainly as the cause) of the outcome.
Later works influenced by their ideas were generally less measured in their assertions. For instance, Wells and Gavanski (1989, p. 161) stated that ‘an event will be judged as causal of an outcome to the extent that mutations to that event would undo the outcome’ [italics added], suggesting that a successful case of undoing commits the antecedent to having a causal status. Obviously, there are many necessary conditions for certain effects that would nevertheless fail to be judged by most as causes. For instance, oxygen is necessary for fire. In all everyday circumstances where there was a fire, one could construct a counterfactual in which the fire is undone by negating the presence of oxygen. Yet, it is widely agreed that notwithstanding the ‘undoing efficacy’ of the antecedent, it would not be regarded as a cause of the fire in question, unless the presence of oxygen represented an abnormal condition in that instance (e.g., see Hart and Honoré, 1985; Hilton and Slugoski, 1986; Kahneman and Miller, 1986).
In other cases, antecedents that easily pass the undoing test would be too sensitive to other alterations of the focal episode to be regarded as causes (Woodward, 2006). For example, consider a case in which a friend gives you a concert ticket and you meet someone in the seat next to you who becomes your spouse and with whom you have a child. If the friend hadn’t given the ticket, the child wouldn’t have been born. But few would say that the act of giving the ticket caused the child to be born. Other intriguing cases of counterfactual dependence that fail as suitable causal explanations are provided in Bjornsson (2006).