What Scientific Explanation Isn't

When a biologist (or any scientist) says she's explained some phenomenon of interest, what's that mean? That is, what does she intend by "explained?" What a scientific explanation is, is a traditional problem in the philosophy of science. Nowadays, I think it's close to a purely philosophical problem. And by that I mean, among other things, that no scientific problem hinges upon understanding what a scientific explanation is. Rather, there is a philosophical project of understanding scientific knowledge and its production, and understanding explanation is part and parcel of fleshing out that project.

Now, this is not to say that what scientists themselves mean by "explanation" is not important. It is, and there are implications on the directions the philosophical project might take in reflecting scientists' views. But philosophers have their own lore over the last several decades. And while a lot of that lore relies on examples from biology, psychology, chemistry, physics, and all that, the question seems otherwise insulated from scientific practice. Ultimately, what I want to explore here is a piece of that philosophical project. In particular, I want to stake a claim about what scientific explanation isn't. That is this: The Deductive Nomological Model, or DN Model, is a failure as a model of scientific explanation.

The vast majority of philosophers of science, and surely every philosopher of science I know personally, will think my claim about what scientific explanation isn't is not news. And they're right. But philosophers of science aren't my audience. Rather, (some) philosophers of mind are my audience. As it happens, there are some philosophers of mind who seem quite sure that the DN Model of explanation, suitably refined, is a quite fine model. And they proceed to use it in the course of fleshing out various theses on specific puzzles about physicalism, the doctrine that anything can be explained in purely physical terms, which usually means reductively explained by physics. I have one such philosopher in mind, namely, Jaegwon Kim, who in his relatively new book Physicalism or Something Near Enough, relies uncritically on the DN model to criticize prominent views on the nature of reductive explanation and to develop his own. Indeed, the DN model seems at the heart of much of Kim's discussion of reductive explanation in the book. And, so, if Kim is wrong about explanation, then we need to re-examine the prominent views of reductive explanation he criticizes and set his own aside.

Before I get started, I need to make some brief side-remarks. Friends of mine who know me and my work are probably wondering whether I've gone around the bend. Why worry about what philosophers of mind think about explanation and physicalism and such things as mental causation? I have a terrible answer to that question: It's Tom Polger's fault. Tom is a colleague of mine and one of those philosophers of mind. However, Tom is a quite sensible philosopher of mind and we've been talking a lot about what explanation is and isn't and how we might be able to say something interesting about reductive explanation if we straightened out what we take to be serious troubles at the foundation of some of the work in philosophy of mind. So, what I'm doing here is really a piece of a larger, and apparently growing project with Tom on how to solve the explanatory gap problem, or the problem of providing reductive explanations of the mental in terms of the physical. For the nonce, I'm just worried about bits of our negative project. Once those bits get organized, we can move on to the positive project. Let's get on with razing the DN Model, or at least Kim's refinement of it.

The DN Model is one part of the broader Covering Law Model. Now, I happen to think, as do many, that the general Covering Law Model of scientific explanation is a failure. But here I'll focus only on the DN Model given the context of the issue with which I'm concerned. The DN Model was introduced by Carl Hempel and Paul Oppenheim in their 1948 paper, "Studies in the Logic of Explanation"; the Covering Law Model was fully articulated by the time Hempel put together his 1965 collection, Aspects of Scientific Explanation. According to the DN Model, a scientific explanation has two parts, namely, an explanandum and an explanans (Hempel 1965, 247). On the DN Model, the explanandum is a sentence describing some phenomenon to be explained and the explanans is the set of sentences adduced to explain the explanandum. The basic structure of a DN explanation consists of a deduction of the explanandum from the explanans, which must contain at least one "universal law of nature" and which all must be true. There are other constraints, but the basic structure is all that is necessary for my purposes.

Ultimately, a DN explanation looks like this:

C1, C2, ..., Cn      (antecedent conditions)
L1, L2, ..., Lp      (laws of nature)
----------------
E                             (explanandum)

Suppose one wants to explain the length of a shadow cast by a flagpole. That is, one wants to know why the shadow is the length it is. On the DN model, one cites the relevant antecedent conditions, such as the height of the flagpole, the position of the sun in the sky, then cites, e.g., the law of the rectilinear propagation of light. From these statements, the current length of the shadow may be deduced. And there you have it: The length of the shadow is explained. The length of the shadow is expected given the antecedent conditions and general law cited.

There's a bit more to say about the nature of DN explanation. But I'll get to that. Now, I want to substantiate why I think Kim is hooked on the DN Model in his Physicalism or Something Near Enough. In chapters 4 and 5 of Kim's little book, he takes up the topic of reductive explanation in its role as a solution to the problem of the explanatory gap, that is, the problem of explaining the mental, e.g., pain sensations, in terms of the physical, C-fiber stimulation (the usual philosopher's scientific account of pain). Kim uses the DN Model, among other tools, to criticize some recent work on reductive explanation by three teams of philosophers: David Chalmers with Frank Jackson (2001), Ned Block with Robert Stalnaker (1999), and Christopher Hill and Brian McLaughlin (1999). These three teams will be important in a subsequent post on this topic. Kim also bases his own view of reductive explanation on the DN Model. Kim says very little in defense of the DN Model. But his reliance is clear. Kim (2005, 106-107) says,

Explanation is plausibly regarded as involving logical derivation; at least, logical derivability is the only concrete and objective criterion available in discussions of explainability. To explain a phenomenon we must derive it, or derive a statement representing it, from a set of statements representing its explanans. As is well known, this is a crucial part of the Hempelian deductive-nomological model of explanation, and it is not uncontroversial. However, we will assume in this discussion that explanation is via logical derivation; this is not a point of contention in this debate.

My view is that Kim's first two sentences are false. The third is an understatement. And the last, if true, is part of my motivation (and Tom's) for taking up the topic.

During the 1960s and 1970s, a vast quantity of criticisms of the DN Model appeared in the philosophical literature. There are two severe problems that a vast quantity of counterexamples aimed to show. These problems are that the DN Model provides neither necessary nor sufficient conditions for explanation. More technically, the DN Model fails to account for explanatory asymmetries and explanatory irrelevancies. The first problem may be illustrated using the example of the flagpole from above. Actually, the flagpole example is a quite famous counterexample to the DN Model (due to Bromberger 1966).

While it's true that the length of the flagpole's shadow can be deduced from the height of the flagpole, the position of the sun in the sky, and the law of the rectilinear propagation of light, it's also true that the height of the flagpole can be deduced from the length of the shadow, etc. But, of course, flagpole heights aren't explained by the lengths of the shadows they cast. There is an asymmetry of explanation that the DN Model doesn't capture. In this case, the asymmetry is causal.

The second problem, the problem of explanatory irrelevancies, is often illustrated by the famous "Jones and the birth control pills" counterexample (due to Wesley Salmon 1971). The idea here is that we want to explain why Jones, a male, fails to get pregnant. The explanandum may be explained by citing that Jones has been taking birth control pills regularly and citing the "law" that all males who take birth control pills regularly fail to get pregnant. But it's obvious that the generalization is irrelevant to Jones' not getting pregnant.

It seems clear enough that the DN Model fails as an account of explanation, where that means a model of causal explanation that assumes a regularity account of causation, because it fails to identify necessary and sufficient conditions for an explanation.

Now, Kim is aware of these criticisms. They appear, after all, in every text in the philosophy of science available today. And Kim doesn't just blindly adhere to the view. In the passage I quoted above, Kim cites the paper in which he attempts to defend the DN Model. That paper is his 1999, "Hempel, Explanation, and Metaphysics." Kim's defense to the DN Model is interesting if for no other reason than because its key move is to infuse metaphysics into the model, namely, explicit citations of cause-effect relationships and causal laws, something that Hempel explicitly rejected for apparently Humean reasons (at least on the concept of cause). However, Kim thinks that the DN Model can only be saved by making this move.

Kim (1999, 10) claims that the range of counterexamples regarding explanatory asymmetries may be swept way if

the events and conditions invoked in the explanans be, jointly, a cause of the event specified by the explanandum, and/or that laws invoked in the explanans be causal laws.

Kim is right! By citing causes, we can avoid the asymmetry problem (at least in cases that the asymmetries are causal). It's easy to see how this makes the flagpole example disappear, since you can't explain the height of the flagpole as being caused by the length of its shadow. One might say that Kim's move is decidedly away from the DN Model at this point. But of course, Kim does want to keep laws, even if they are explicitly causal laws. More importantly, Kim insists that explanations be deductions.

We've seen that Kim thinks that "logical derivability is the only concrete and objective criterion available in discussions of explainability" (2001, 106). In 1999 (17), Kim says,

[t]o my mind, the crucial virtue of the D-N model is the fact that it forces an explanation to bring into the open the explanatorily salient properties of the phenomena involved -- properties that do the real explanatory work.

What does the forcing is the deductive structure. Presumably, by articulating the antecedent conditions and the (now causal) laws, we rightly bring into the picture only the salient explanatory properties involved. Notice what problem Kim is worried about here --the problem of explanatory irrelevancies.

According to Kim, by forcing a deductive structure on explanation, alongside the citation of explicit cause and effect relationships and causal laws, the irrelevancy problem falls away. Kim (1999, 17) says,

[t]he virtue of Hempelian D-N explanations consists in the fact that they wear their this crucial explanatory information on their sleeve. It makes evident what it is about the event, or events, invoked in the explanans that causally explains the explanandum event. The reason it does this is that laws, or causal laws, are actually used, not simply named or assumed to exist, in an explanation. When we know what the explanatory laws are, we know what the explanatorily, or causally, efficacious properties are. Without this knowledge we have no explanation; we do not understand why the event we want to have explained occurred.

It's not at all clear to me that imposing a deductive structure does the heavy lifting Kim claims it does. What seems to me to be doing the heavy lifting here is getting clear on what the right causal relations and causal laws are in constructing an explanation. Notice that getting clear about what the causal relations are and what the causal laws are makes it clear what is wrong with the birth control counterexample. Adding a deductive structure doesn't magically impose a set of relevance relations. Kim has skipped a step. Indeed, the step of fleshing out how a model of explanation includes what's relevant and excludes what's irrelevant is not an easy one. But I think what's part and parcel of taking the step is establishing what it means to say that you have the right causal relation (or dependency relation). That's what does the heavy lifting in solving both the problems of asymmetry and relevance. There seems no special need for deduction.

Where am I? What does the heavy lifting in Kim's view of explanation seems to me to be the citation of causal relations and causal laws. Causal relations and causal laws solve the asymmetry problem. And, if I'm right in what I've said about Kim's insistence that explanations be deductions, causal relations and causal laws allow you to get relevance for free. That is, once you've determined the right causal relations and causal laws, you've got relevance. The puzzle is to understand how such relations and laws are established. And that is an ongoing problem for adherents of the causal-mechanical view of scientific explanation, an alternative to the DN Model.

Explanations aren't deductions of statements about phenomena from antecedent conditions and universal laws. Rather, they are articulations of dependency relations, often (perhaps almost always in science) causal relations. The problem is to build a model of explanation that reflects that. Again, that is an ongoing problem for adherents of the causal-mechanical view of scientific explanation. And, notably, Jim Woodward (2003) has offered a view that seems to be proving itself successful. The next question, for next (?) time, is what to say about reductive explanation and physicalist accounts of the mental assuming that explanations aren't DN explanations.

References

Block, N. and R. Stalnaker (1999), "Conceptual Analysis, Dualism, and the Explanatory Gap", Philosophical Review 108: 1-46.

Bromberger, S. (1966), "Why Questions", in R. Colodny (ed.), Mind and Cosmos: Essays in Contemporary Science and Philosophy. Pittsburgh: University of Pittsburgh Press.

Chalmers, D. and F. Jackson (2001), "Conceptual Analysis and Reductive Explanation", Philosophical Review 110: 315-361.

Hempel, C. and P. Oppenheim (1948), "Studies in the Logic of Explanation", Philosophy of Science 15: 135-175.

Hempel, C. (1965), Aspects of Scientific Explanation and Other Essays in the Philosophy of Science. New York: Free Press.

Hill, C. and B. McLaughlin (1999), "There are Fewer Things in Reality Than Are Dreamt of in Chalmers' Philosophy", Philosophy and Phenomonological Research 59: 445-454.

Kim, J. (1999), "Hempel, Explanation, Metaphysics", Philosophical Studies 94: 1-20.

Kim, J. (2005), Physicalism or Something Near Enough. Princeton: Princeton University Press.

Kitcher, P. (1989), "Explanatory Unification and the Causal Structure of the World", in P. Kitcher and W. Salmon (ed.), Scientific Explanation. Minneapolis: University of Minnesota Press: 410-505.

Salmon, W. (1971), Statistical Explanation and Statistical Relevance. Pittsburgh: University of Pittsburgh Press.

Woodward, James (2003), Making Things Happen: A Theory of Causal Explanation. New York: Oxford University Press.

Comments

[John Wilkins]

Nice intro, Rob. I'm going to post a qualified defence of the DN model on my blog tomorrow, based on your discussion here.

[Matt Dunn]

Very cool stuff. I saw Ned Block give a talk last fall and I've been fascinated ever since.

I imagine your colleague Prof. Bickle might have a thing or two to say about this. I'm unfamiliar with his work aside from the fact that he is "ruthlessly reductive". Does his reductive account of mental phenomena rely on DN?

I look forward to the "next (?)" post.

[Robert Skipper]

I imagine your colleague Prof. Bickle might have a thing or two to say about this. I'm unfamiliar with his work aside from the fact that he is "ruthlessly reductive". Does his reductive account of mental phenomena rely on DN?

John will have nothing to do with mental phenomena. For John, explanations of behavior are couched entirely in cellular/molecular terms. So the explanations are mechanistic and are what you find in neurobiology journal articles and texts. (John won't say what "mechanisms" are or what "causal relations" are, and so on.)

If there is any preservation of folk psychological terms, such as "belief," "desire," etc., they are given a purely heuristic role. If you want to know what's going on, you've got to do the neurobiology. That's the ruthless part of "ruthless reductionism."

[CK Loo]

I don't really know much about phil of mind but I wonder if the tendancy to favor DN type views regarding models isn't a hold over from the sort of logicism that was prevalent in analytic philosophy. This might be especially true for Kim, he has tended to be a guy who appeals to logical relations to explain mental phemonena.

Also, if I remember correctly, weren't most of the people who argued against the DN view people who were interested in a more semantic approach to scientific models. If that is the case, maybe there some people outside of the phil of sci who are not familiar with the criticisms of the DN view.

Though that is no excuse I guess.

[Tom Polger]

One of the things that we/I are/am interested in exploring is what sustains the continued appeal of DN or CL explanations in some quarters. There is something right about saying (as CK does) that it is a commitment to "logicism." But that doesn't get us very far. What's so special about logicism?

The next question, then, is: What is it that motivates the "analytic" or "logicist" conceptions of explanation? Of course I mean philosophical motivation, not (or not merely) psychological motivation.

Here's an idea: Maybe DN satisfies some desiderata that other accounts do not, thus accounting for its staying power and the lack of a consensus replacement. If so, then maybe those desiderata could be satisfied by a non-DN model.

In fact I think something like this could be right. DN models get right that the explanatory relation is non-accidental; it is in some sense necessary. But they mistakenly achieve that by making the explanatory relation a deductive relation. The alternatives often fail (so say I) because they don't get that aspect of the explanatory relation.

One way to put the idea is to say that explanations characterize dependencies. Obviously, then, much remains to be said about dependence relations. I/we hope to say something about that... someday... maybe soon.

[CK Loo]

Tom,

I personally would suggest that there might not be good philosophical reasons to be committed to logicism.

However, there might be some argument than can be made for deduction providing a, for lack of a better word, stronger basis for inference. I guess if one has concerns regarding the epistemic status of other forms of inference one might be tempted to try for an account of modeling that relies upon deductive relations.

Perhaps the commitment of some people to DN approaches might be analogous to the commitment of some proponents of semantic accounts of scientific modeling to an isomorphic relation between entities and the terms that refer to them in models. Deduction, like isomorphism, are concrete relationships rather more fuzzy notions such as induction or similarity.

[Robert Skipper]

CK,

I'm a little confused about your responses to Tom. You say,

However, there might be some argument than can be made for deduction providing a, for lack of a better word, stronger basis for inference. I guess if one has concerns regarding the epistemic status of other forms of inference one might be tempted to try for an account of modeling that relies upon deductive relations.

While it's not clear to me that putting some line of inference in the form of a deductive argument gives it a stronger inferential basis, I don't think that's the issue. The issue is whether a deductive argument imparts "explanatoriness" to some purported explanation. Obviously, not all deductive arguments are explanations. But for those who think that some are, what makes them so?

Perhaps the commitment of some people to DN approaches might be analogous to the commitment of some proponents of semantic accounts of scientific modeling to an isomorphic relation between entities and the terms that refer to them in models. Deduction, like isomorphism, are concrete relationships rather more fuzzy notions such as induction or similarity.

Isomorphism is a mathematical/logical, one-to-one relation between to items, here a model and system. While there may be an isomorphism between a model and the ideal system the model can be said to literally describe, i.e., the "quasi-realistic system," there is almost never anything close to an isomorphism between a model and a real world system. At best, the relationship can be some sort of (informal) homomorphic, or similarity relationship.

From the process of relating models to real world systems, we get explanations. Indeed, we may deduce, induce, or abduce to some explanation. But what the explanation is need not be some formal deductive or other kind of argument.

[CK Loo]

Rob,

I should have thought out my response to Tom a little more. I was just speculating off the cuff regarding what desiderata that a DN approach might fulfill that others might not. The only thing that I came up with was that perhaps one might think that deduction is a better grounds for inference than others and that a good explanation is one that one can draw inferences from. I should have given it more thought before I posted a comment.

As for what I said about isomorphism, it's been a long time since I read Suppe and all I remembered was that he held the notion that there was a formal relationship that mapped entities in theories to entities in the world. I thought the relationship he suggested was an isomorphic one but it might be a homomorphic one. Though, that said, I must admit that I'm not entirely sure what the difference between a homomorphic and an isomorphic relationship is.

[CK Loo]

Correction, I meant Patrick Suppes and not Frederick Suppe. I get them mixed up.

[Tom Polger]

I agree that "there might not be good philosophical reasons to be committed to logicism." The question is whether there is a way to satisfy those reasons without the total apparatus of DN, and particularly without the deductions. (Probably also without the laws, at least taken strictly.) My suspicion is that the logicism, as you are calling it, is overkill. And, like swatting flies with a hammer, it tends to damage the furniture. So if a better tool can be found, that would be good.