| |
I will argue that a strong distinction needs to be made between causal attribution and causal explanation (cf. Hilton, 1990). For example, attributing the 9/11 attacks to someone is not the same as explaining them to him. Whereas causal attribution is a cognitive process that involves referring an event to its source, whether it be a painting to its author, or an event to its origin, explanation is a three-place predicate describing a social interaction whereby someone explains something to someone else. As such causal explanation must obey the rules of conversation (Grice, 1975); a good explanation must be probably true, informative given an interlocutor’s state of knowledge, relevant to her interests, and expressed clearly. Whereas causality is objective, explanation is subjective (or intersubjective) in nature.
In the accounts of causal attribution and explanation that I have given elsewhere (Hilton; 1990; Hilton & Slugoski, 1986), I distinguish two phases in the explanation process: a first phase of counterfactual reasoning, and a second phase of contrastive explanation that follows conversational rules. I wish to argue below that the first phase of counterfactual reasoning is objective because the quality of a counterfactual depends on its being an exact description of the objective world. Later I will expand the notion of contrastive explanation by showing how the contrasts of interests are constrained by the implicit purposes behind the causal question posed with respect to a given event.
Causality in the objects: Counterfactuals and causal models
Recent work on causal reasoning in artificial intelligence, philosophy and psychology has substantiated the importance of counterfactual reasoning in causal attribution (see the contribution by Sloman to this seminar, and his 2005 book for a review). This work has done much to clarify and extend the counterfactual analyses of causation given by philosophers such as Hart & Honoré (1985) and Mackie (1980). Nevertheless we should remember that, to echo Austin (1962), causal models are essentially constative in nature: they are more-or-less accurate descriptions of reality, and their quality resides in the predictions they support. So if I believe that the cock’s crowing is not the cause of the sun’s rising, I can creep out and tape the cock’s beak to stop him crowing at dawn – and then observe what happens. Reality will be the judge of whether my causal hypothesis, which can be expressed as the indicative conditional If I stop the cock crowing, the sun will still rise, is correct or not.
More generally, we judge the correctness of constative conditionals by their accuracy as descriptions of the nature of the world according to the Ramsey test (Evans, Handley & Over, 2003); that is the probability of the consequent given the antecedent. So if, in a given world of discourse (e.g. France), the indicative conditional If a restaurant has two stars (p) then the food will be excellent(q) will be judged good if the consequent (q) is probable given the antecedent (p). Counterfactual conditionals are also constatives in that they are assertions about the nature of reality, even if they describe events that have not actually happened. They are still judged good if they describe what would have happened if a factor that actually was the case had been “undone”. For example, if we agree that it is unlikely that the USA would not have entered the Second World War if Japan had not attacked Pearl Harbour, then we would agree with the counterfactual conditional if Japan had not attacked Pearl Harbour, then the USA would not have entered the war.
Causal models can of course influence what constative conditionals (whether expressed as indicatives or counterfactuals) are judged probable or not. For example, many people might be surprised by my assertion If Joan of Arc had lost the Battle of Orleans, then the world would be speaking French. How silly! Or how come? After all, it is normally the conquerors in battle that get to impose their language. Well, to support the counterfactual conditional, my argument would be to provide you with facts that motivate a new causal model: England at this time was ruled by French-speaking kings of Norman origin. Had they won the Battle of Orleans, England and France (with its larger population) would have been amalgamated into one kingdom, whose language of administration and education would have been French. This kingdom would have dominated Europe and then colonised the world. To the extent that you buy this causal model, then you may accept the Joan of Arc counterfactual conditional as being rational and true rather than silly and idiotic.
Causal models thus support indicative and counterfactual conditionals, and all function as descriptions of reality that can be true or untrue. They are thus objective in nature in that their quality depends on their truth-value; that is, their degree of approximation to that reality. While causal models and counterfactual conditionals certainly support explanations, explanations also have an essentially subjective component, having to do with the knowledge and interests of causal inquirers. Causal models and constative conditionals are thus necessary but not sufficient for understanding causal explanation processes.
The pragmatics of explanation
In terms of linguistic theories (Levinson, 1983), causal models and constative (indicative and counterfactual) conditionals are “semantic” (they have truth-values), whereas explanations are “pragmatic” (they have utility values). Explanations are sensitive to context and serve practical interests in conversation. First, experimental work shows that in interpersonal explanations, we follow Grice’s maxims of conversation to change our explanations to complement what the other doesn’t know. For example, in line with the maxims of quantity, people tend to focus on personal factors when explaining an act of juvenile delinquency to someone who knows of relevant situational factors but not personal ones, but vice-versa if the interlocutor is perceived to know of relevant personal factors but not situational ones (Slugoski et al., 1993). In line with Grice’s maxim of relation, people will give an explanation that may seem most relevant to their interlocutor. For example, Norenzayan & Schwarz (1999) found that when experimental participants believed that they were giving explanations of real-life mass murders to a psychologist, they referred more to personal factors than when they believed they were giving explanations to a sociologist, to whom they were more likely to give situational explanations. It seems unlikely that these changes in explanations reflect changes in the participants’ underlying beliefs about the “facts of the case” or causal models of the relevant behaviours (e.g. delinquency, mass murder).
Intrapersonal explanation is also pragmatic in that the inquirer’s satisfaction with an explanation will change with her knowledge-state. Counterfactual reasoning often reveals a plethora of conditions that are necessary for an outcome to occur, yet we tend to identify one or at most two factors in causal explanations. Given the plethora of alternatives, the problem then becomes to select the most relevant and informative explanation, and this will often be the condition that is perceived as abnormal in the circumstances (Hart & Honoré, 1985; Hilton & Slugoski, 1986). For example, many “naïve” observers will have attributed the Concorde’s crash in 2000 to the presence of débris on the runway during takeoff, as this certainly seems abnormal, and without its presence the accident would not have occurred.
However, “expert” observers in aeronautics consider the presence of débris on runways to be “normal” in that from time to time débris inevitably gets onto runways, and therefore aircraft have to be designed to withstand this. The relevant contrast case for aeronautics experts is the “ideal design” which would have enabled Concorde to withstand the shock of the débris, which they compare to the actual design of Concorde to find the weak point (in this case, the unprotected fuel tanks) which they identify as “the” cause (cf. Hesslow, 1988). The difference in the explanations favoured by naïve and expert judges in this case have nothing to do with the objective “facts of the case” (the perceptions of the crash itself and the sequence of events during take-off), but rather in their subjective interpretation (e.g. differences in what constitute “normal” comparison cases).
The privileged status of intentions in human causal explanation
A challenge for theories of causal explanation is to understand why free deliberate human actions make such attractive and powerful explanations. For example, few people appear to explain the Concorde’s crash by tracing causality through the débris through to the antecedent abnormal condition that was responsible for it (faulty maintenance on the tail of a Continental Airlines jet that took off just before Concorde). Yet experimental data obtained on similar scenarios (Hilton, McClure & Slugoski, 2005) indicates that it seems highly likely that people would have traced causality through to an antecedent condition if had been a deliberate human act of sabotage (e.g. someone placed the débris on the runway expressly to cause the accident).
Information gain accounts in unfolding causal chains
Counterfactual reasoning appears to be unable to explain this difference, as we would agree in either case that if a) the debris had not fallen off the tail of the preceding airliner, or b) the saboteur had not placed the débris on the runway, the accident would not have occurred. Probabilistic models of the kind proposed by Spellman (1997) would probably do better on these examples, as data suggests that in these kinds of unfolding causal chains, distal causes appear to increase the probability of outcomes more when they are free deliberate actions than when they are accidental, “natural” causes (Hilton et al., 2005). The probabilistic criterion, that causes are those conditions which most increase the probability of outcomes appears to work well in these unfolding causal chains, where the successive events follow normally and foreseeably from each other in a directed causal sequence. There is a logic of mechanism in their sequence, and they cannot be switched around and still produce the effect. They can even be thought of as sequenced “recipes” for bringing about effects, either in the sense that a saboteur may intentionally concoct a plan to bring about an aircraft crash, or simply in the sense that an unintended piece of debris on runway is a “recipe for disaster”. Once the first cause has acted, the chain of causation only has to run normally on for the outcome (e.g. accident) to happen (e.g. the debris pierces the fuel tank, leading to fuel to escape, leading to a catastrophic fire as in the case of the Concorde crash in July 2000). Since free deliberate actions lead to a greater increase in the probability of the outcome they have high information-value. And since the remaining parts of the chain now become predictable, they become redundant details that can be safely omitted from an economical explanation. Focusing on intentional factors in explanations can here be explained in terms of Gricean constraints, such as informativeness.
Social utility vs. information gain accounts in opportunity chains
However, probabilistic analyses fail in what we term opportunity chains where the action of a first cause creates an opportunity for the second cause to act. For example, a natural cause (the refraction of light through broken glass) may cause some shrub to smoulder, allowing a second cause to create a bush fire. This first cause creates an opportunity for a second cause to act, which could either be an intentional action (someone pours petrol on the flames) or a natural cause (e.g. wind springs up). What is interesting here is that people rate the second factor as more strongly causal if it is a deliberate action even if it does not increase the probability of the fire any more than does the natural event. The same applies to experimental variations in the first event: someone deliberately igniting the flames is judged to have increased the probability of a forest fire just as much as does the refraction of light, yet people consider the human action to be more causal (Hilton et al., 2005).
It may therefore be that another criterion is at work that leads judges to favour intentional actions as explanations. In particular, following Tetlock’s (2002; but see also Smith, 1789/2002) “intuitive prosecutor” analogy, we hypothesised that people may focus on human actions in explanation that allow social control of negative outcomes through the menace of sanctions. We therefore asked participants to evaluate the factors in the event chain in terms of how much social control and preventability they allowed. For example, for the distal cause in the forest fire scenario, the question read “How much do you think society can control and prevent the occurrence of events like a youth setting fire to the shrub in the future (e.g., by warnings, punishments, etc). Judgments were made on a scale from 0% to 100%, with anchor points of ‘Not at all’ at 0% and ‘Completely’ at 100%. Mediational analyses showed that social controllability judgments correlated with causality judgments, and mediated the tendency for human actions to be preferred to natural events as causes.
Pragmatic accounts of the preference for human actions as explanations
The finding that perceptions of social controllability mediate perceptions of causality suggests that social utility – and not informativeness – may explain why human actions are preferred as explanations. Future research should address the question of whether social utility would explain the preference for human actions in unfolding causal chains – for example, would the perceived social controllability and preventability of sabotage attempts predict our preference for them as explanations over natural events? Perhaps not – as faulty aircraft maintenance also seems socially controllable through the menace of sanctions, yet this nevertheless does not seem to make it a strong candidate as a cause. Whatever the answer to this question, the selection of explanations will probably depend on subjective characteristics of the hearer rather than objective characteristics of the world.
Conclusions
In this essay, I have attempted to relate recent work in cognitive science on causal modelling and conditional reasoning to earlier work in ordinary language philosophy and social psychology on the pragmatics of explanation. My conclusion is that causal explanation requires both kinds of analysis. For example, recent cognitive science approaches have addressed the question of how we identify actual causes in pre-emptive causal chains, where an unfolding causal chain, which would have produced the outcome, is interrupted by a new causal chain which takes over and actually produces the effect. An example would be of a fire, which, if left to itself, would inevitably burn down a house. However, another fire reaches the house first, thus pre-empting the first fire from running its course and burning down the house (Halpern and Pearl in press, a; b; see also Mandel 2003). Causal modelling approaches thus seem to be very useful for resolving the attribution question that I raised at the beginning of this essay, in the sense of tracing an outcome to its origins.
However, while the AI and cognitive approaches may solve the causality problem, they do not to our knowledge give fully satisfactory answers to the explanation problem, nor tell us how to attribute responsibility. For example, given that we accept that Fire B is the actual cause, what is our preferred attribution in the sense of picking out the factor that is to be held “responsible” for the damage it caused to the neighbour’s house? Would it matter if we knew that Fire B came about because someone deliberately fanned some smouldering flames, rather than because a breeze sprang up and brought them to life? Our experimental data suggest that it would, and that a pragmatic analysis in terms of the causal inquirer’s knowledge-state and interests is necessary to understand how and why. Indeed, our finding that social controllability predicts causal attribution indicates that it is not just individual interests, but group interests that may drive this process. We would expect the above questions to be posed even if we knew that another concurrent fire would have had exactly the same result, just as we would not dismiss a trial for murder simply because we knew the victim would have been later murdered by someone else anyway.
Finally, it seems to me that an attention to pragmatics helps understand what is specifically human about causal explanation. Human beings may share the same kinds of associative learning processes with lower animals such as rats (Shanks & Dickinson, 1988) and even bees. Insofar as they are uniquely capable of representing counterfactual events (that were predicted by context but did not happen) or of engaging in conditional reasoning (Gardenfors, 2003), then they may dispose of cognitive processes that distinguish them qualitatively from other animals. But we will do well to remember that people are also social cognitive animals that communicate through language, who can recognize the intentions and belief-states of others and engage in co-operative communication. They will also seek to control the behaviour of others through expressing anger, holding them responsible for their actions, and identifying the causes and consequences of their behaviour. This social nature will determine human explanation processes in ways that cannot be captured by a purely cognitive perspective that treats people as isolated individuals who use language simply to construct and test representations of their environments.
References
Austin, J.L. (1962). How to do things with words. Oxford, England: Clarendon Press.
Evans, J.St.B.T., Handley, S.H., & Over, D. (2003). Conditionals and conditional probability.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 29, 321-335.
Gärdenfors, P. (2003). How homo became sapiens: On the evolution of thinking. Oxford: Oxford University Press.
Grice, H.P. (1975) “Logic and conversation”, in P. Cole and J.L. Morgan (eds) Syntax and Semantics 3: Speech Acts (pp. 41-58), New York: Academic Press.
Halpern, J. and Pearl, J. (in press, a) “Causes and explanations. A structural model approach. Part 1: Causes”. British Journal for the Philosophy of Science.
Halpern, J. and Pearl, J. (in press, b) “Causes and explanations. A structural model approach. Part 11: Explanations”. British Journal for the Philosophy of Science.
Hart, H.L.A. and Honoré, A.M., (1959/1985) Causation in the law (2nd ed.), Oxford University Press.
Hesslow, G. (1988) “The problem of causal selection”, in D. Hilton (ed.) Contemporary science and natural explanation: Commonsense conceptions of causality (pp. 33-65), Brighton: Harvester Press.
Hilton, D.J. (1990) “Conversational processes and causal explanation”, Psychological Bulletin 107: 65-81.
Hilton, D.J. McClure, J.L. & Slugoski, B.R. (2005). The course of events: Counterfactuals,
causal sequences and explanation. In D. Mandel, D.J. Hilton & P. Catellani (Eds.).
The psychology of counterfactual thinkng. London: The Psychology Press. (pp.44-73).
Hilton, D.J. and Slugoski, B.R., (1986) “Knowledge-based causal attribution: The abnormal conditions focus model”, Psychological Review 93: 75-88.
Levinson, S.C. (1983). Pragmatics. Cambridge: Cambridge University Press.
Mackie, J.L. (1980) The cement of the universe: A study of causation, Oxford: Clarendon Press.
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.
Norenzayan, A. & Schwarz, N. (1999). Telling what they want to know: Participants tailor causal attributions to researchers’ interests. European Journal of Social Psychology, 29, 1011-1020.
Shanks, D.R. and Dickinson, A. (1988). The role of selective attribution in causality
judgment. In D.J. Hilton (Ed.). Contemporary Science and Natural Explanation: Commonsense conceptions of causality. Brighton: Harvester Press.
Sloman, S.A. (2005). Causal models: How people think about reality and its alternatives. Oxford: Oxford University Press.
Slugoski, B.R., Lalljee, M.G. Lamb, R. and Ginsburg, J. (1993). Attribution in conversational
context: Effect of mutual knowledge on explanation-giving. i>European Journal of Social Psychology, 23, 219-238.
Smith, A. (1789/2002, ed. K. Haakonssen) The theory of moral sentiments, Cambridge: Cambridge University Press.
Spellman, B. (1997) “Crediting causality”, Journal of Experimental Psychology: General 126: 323- 348.
Tetlock, P.E. (2002) “Social functionalist frameworks for judgment and choice: Intuitive
politicians, theologians and prosecutors”, Psychological Review 109: 451-471. |
 |
|
|
Setting fire to mono-causation 
Robert Stonjek
Jan 18, 2006 11:20 UT
The attribution of causation to a single initial source is an assumption that requires closer examination. Let’s consider once again the fire example.
A fire is deliberately lit by individual A. The fire subsequently burns the house down, killing everybody sleeping inside. With no other information given, can the damage to the house and death of the family be attributed to person A?
Let’s consider some additional detail. The fire is lit in a wood heater, a practice done by this family throughout the winter. On the previous 50 days, the fire was lit in much the same way, and with the consent of the whole family, and no subsequent damage occurred on those previous occasions.
But person B placed wet clothes above the fire to dry them, and it was the catching on fire of the clothes that caused the fire that caused the damage. But clothes had been dried in this way on every rainy day through winter, even though authorities warn that great caution should be exercised when placing clothes near fires.
But person C was responsible for removing the clothes from around the wood heater before going to bed, and it was after everybody went to bed that the secondary fire started – C had forgotten. But C had forgotten on other occasions with nothing more than a singed sock or two resulting.
The fire’s door seal was faulty. The manufacturer of the wood heater had recalled the heater doors for replacement of the door seal but no-one in the family was aware of this. Even so, regular maintenance would have seen the door seal renewed by about now anyway, and that was the responsibility of person D. It was the escaping hot gas through the faulty door seal that caused the additional heat on this occasion that caught the clothes on fire that burnt the house down killing persons A through to D.
Let’s look at the causation:
If A had not lit the fire or
If B had not placed clothes above the wood heater to dry or
If C had remembered to remove the clothes at bedtime or
If D had been vigilant with the maintenance or
If any of them had noticed the recall notice or If the manufacturer had not sold a faulty product,
The fire would not have gotten started in the clothes, burning the house down and killing all inside. Two such fires (house destroyed, family wiped out) occurred in our region last winter.
But what if B had worn his socks another day, or thrown them out? The fire may not have started on that night but on some other night or not at all.
All too often, in real life, the causal source is spread over a number of events which all contribute to the outcome, all are essential, but where no single or initial causative agent can possibly be identified. (lighting the wood fire was not the cause of the house fire).
Kind Regards
Robert Karl Stonjek
|
| |
|
3 replies to Setting fire to mono-causation:
|
| |
|
|
Re: The problem of causal selection
Robert Stonjek, Jan 21, 2006 23:30 UT
|
| |
|
|
Re: The problem of causal selection
Robert Stonjek, Jan 21, 2006 23:30 UT
|
| |
|
|
The problem of causal selection
Denis Hilton, Jan 18, 2006 23:01 UT
|
|
|
|
Note: yellow triangles ( ) indicate new messages that have been posted since your last visit to the site.
|
|
