Facets of Causation
George Kampis
Dept. of History and Philosophy of Science
Eötvös University Budapest
Hungary
This is something like a transcript of the live talk. No footnotes, no
references, with
the exception of one or two, just a talk, and lakonic remarks.
Introduction and Summary
In this talk I am doing two things: I will tell my own version of a global
story about science, language, and causality etc., and I will present some
new details concerning some causal problems of a more local interest. As
a former scientist and engineer, I will be dealing with philosophical problems
of causality in the style of science; more about this later.
While navigating through these issues I will make a case for what can
be summarized in the following statements:
(1) science is the domain of causality, not propositions (e.g. "theories",
"representations")
(2) causality is a primitive concept prior to an different from explanation
and description
(3) causal models (hence scientific models) are indications, rather than
descriptions.
Note that I am not using words like "scientific theory".
Origins
This work grew out of my 1991 book "Self-Modifying Systems in Biology
and Cognitive Science", Pergamon, Oxford - New York, pp. 546 + xix.
A better title would have been "Anti-Essentialism for the Scientist".
The book mostly dealt with theoretical biology from the point of view of
dynamics and complex systems, and offered an analysis (a "rational reconstruction",
as it were) of modeling methodology. It defended an anti-essentialism about
entities. I tried to show how the "properties" (to keep the scientific
term) which are interesting for the biologist and the psychologist turn
out to be relational -- the ones like being a gene, or being an enzyme,
or having molecular property x or mental state y, and so on.
Derived from this view, the main argument was that causality is different
from determinism in science. For instance, by using relational properties
one can produce causal systems which are not deterministic yet not random
- causality started to escape from the usual dichotomies, raising the suspicion
that it is altogether different. Now I am extening these remarks using
what I recently call the theory of implicit variables.
Whose Causality?
Contrast "define" with "study". Definitions of causality are metaphysical
and make analysis possible. There is a temptation from philosophy to handle
the problems of causality in this way, but the concept is mainly a scientific
one, which allows an independent study.
Let me make two remarks on this. First, I am assuming the primacy of
science over philosophy, where none of the two is autonomous by itself.
Quine, Dennett and Popper argued famously for this view, supported by the
habit of even the anti-naturalists to use words freely taken from science
and technology. The same sources (Q D and P) warned us, for a variety of
reasons, not to bother with definitions. As our theme develops, we will see
that this is good advice.
Second, unlike philosophy, science has never been Humean or Russellian.
A recent convert, Judea Pearl (2000) says this: "Fortunately, very few physicists
paid attention to Russell's enigma. They continued to write equations in
the office and talk cause-effect in the cafeteria; with astonishing success
they smashed the atom, invented the transistor and the laser." (p.338) Pearl,
J. (2000). Causality: Models, reasoning, and inference. Cambridge, UK: Cambridge
University Press.
So, despite much talk to the effect of the contrary, causality is here
and alive, unreflected and mostly (maybe with the exception of quantum
physics) unproblematic in science.
How is this possible? I think this is our main question. The task, then,
is to understand and to characterize causality as a phenomenon in relationship
to science, mind, language, decription, explanation, etc.
The Descriptive Fallacy
Here I am making my first main point. The usual way causality is dealt
with in philosophy is bound to an error, which we can aptly mark by borrowing
Austin's words. It will be easy to point out that under causality we understand
something different than just some kind of description. (Hence, "causal
description" will be an oxymoron.)
I am taking just one example, or one and a half, not to obfuscate the
point by covering it with much debris. Take a Cellular Automaton (CA). A
CA is a regular structure, like a two-dimensional grid, with identical finite
automata sitting at each grid point, each changing its state by way of a
transition function, on the basis of its own state and the state of the neighbours.
By these local transitions one often gets meaningful global behaviors. Ed
Fredkin, a BU physicist devised a particular CA transition rule that can
only produce symmetrical patterns. The actual pattern that arises is always
a function of the initial conditions and changes with them, but the way
the pattern looks always stays the same (see illustrations).
We are now interested in events when two cells (at two arbitrary points)
change their state in the same way at a given time instance (such as when
the one turns red and the other turns red). By counterfactual theory, for
instance, we see that this symmetric behavior is caused by the initial
state. This is true in the sense that if we don't turn on the right kind
of initial conditions at time zero (or any time less than t) we will never
get the specified behavior at time t. (As with practically every initial
condition there is a different pattern, one can always find events that
counterfactually depend on particular initial states.)
But the cause of the symmetry is none of these. In the Fredkin CA we
always get symmetry, no matter what initial conditions we choose. This
shows that counterfactual dependence is not what we understand (or want
to understand) by causality with respect to these selected events. Counterfactuals
prove to be insensitive to certain kind of dependencies like every initial
condition producing some symmetry.
Our example bears upon a number of other classical questions about causation
and explanation, such as the Reichenbachian common-cause industry and W.
Salmon's theory of explanations etc. Of course, one would be tempted to
say that it is the transition rule that causes the globally symmetrical
transitions. The transition rule is not an event-like cause and it is easy
to save counterfactuals and Reichenbach etc. by excluding this kind of
dependance from the realm of causes. However, one should be careful here.
It is possible to devise more complicated CA's where the rules change, so
for the transition rules we can retain all intuitive elements of a causal
dependance like manipulability, propagation of effects and so on, even
counterfactuals, without their having to be event-like (here is one score
for the slogan "don’t trust definitions").
The Descriptive Fallacy II.
This is just an extension of the earlier point in order to show how the
CA example is linked to more ovbious and well-known matters.
Take a geometrical form such as a trajectory. Under mild mathematical
conditions it can be rewritten in a generative form such as an ordinary differential
equation. There is a complete equivalence between the two representations:
all theorems about the one are theorems about the other. In the differential
equation, if one takes a given value x(t) this generatively sets all other
values; by interpreting t as time we can say x(t) sets all "future" values.
Causal properties, held by whatever descriptions of causality (such as
deducibility, counterfactual dependence etc.) usually hold for these simple
differential systems, which is no wonder because theories of causality
use such kinds of systems as the home page to start with.
However, as every value of an equation is just a point in a geometry,
we are left with this question: what can be causal about a geometry? Of course,
this (or something similar) is the basis of Russell's criticism of causality.
CA-s are no exception. For every CA, there is an equation similar to a differential
equation which defines a geometry.
By bringing up this short note I just want to motivate why I now believe
that ALL descriptive theories of causation must collapse to Russell elimination.
The CA example shows less than that, it shows that this is the case in
many situations. Either way, we are left with a picture where counterfactual
dependance is neither sufficient nor necessary for something to be a cause.
If we want to look for causes, we have to look elsewhere.
How Does Causality Enter Science? (remark)
So far we see, or start to see, what causality is not. Before turning
to the question, what it is, I stop to pick up some more ammunition about
how causality makes its entry in science. Causality enters science not via
some descriptions (or some theories) as we might think but it does so independently
via the actions of the scientist. This must be surprising in the light
of the fact that scientific theories are about the description of causal
events (and processes…), see the J. Pearl quote.
But it can't be otherwise. Many have noted that causality is conceptually
prior to any explanation (e.g. Cartwright, Rudder-Baker). But it was I.
Hacking (and, low and behold, B. Latour) who did the most in order to dissect
the myth that science equals theory, especially if the latter is understood
in the sense of the Vienna Circle as a set of propositions or statements
that can be true or false, and that's all there is. Hacking makes it clear
that nothing in science works without being embedded inseparably in a melange,
of which Nature is a silent, but fundamental, component. Hacking's focus
is on experiment and its role in questions of realism about theoretical
entities, but we can be more general here.
Experiments are just examples of actions, and actions of the scientist
include ones that bring forth observations. Much of philosophy cum philosophy
of science today is based on a peripathetic epistemology where "facts" are
found on the road and are taken as given. Contrast this with biology where
it can get you a Nobel Prize to make, literally speaking, just an observation
(such as which gene does what). Note that by current science funding standards,
more effort goes into this than into any theory.
The effort argument says, in chorus with Hacking, that science, understood
properly, invites elments of realism (sweat is real) - but realism about
what? It's not entities (Hacking never quite made it to the point where
he could show that experiments prove the existence of electron, which is
theory-dependent, after all). I am suggesting that it’s not entities but
causality and actions, such as our own (like using the srew-driver or sweating,
which are not theory-dependent) that science teaches us to be realists
about. We can express this by saying „nothing in science is real unless
it happens”.
If this is correct, causality may be complementary to the more usual
"text mode" of science. But we fail to see yet why causality is any different
from what is described in the text mode.
Manipulability and Agency (remark)
It is at this point where I believe the answer begins. Experiments and
actions differ from theories in the first place by the fact that they are
performed. There is a recent development which puts the actions of an agent
rather than utterances into the focus of a causal relation. I am rushing
through this to slow down later.
The manipulability-and-agency view of Price and Menzies (xx) extends
von Wright's (1971) well known position "No causation without manipulation".
Their agent view focuses on our own actions to identify causal dependence.
This is much in line with what I have in mind for the rest of this talk.
See my criticism so far of the counterfactual theory, where counterfactual
causes were somewhow not "active" enough to reach everything. Compare my
quoting Austin and Hacking for acts being fundamentally different from descriptions.
This is the gist in which we will continue.
However, let us note a few points well: The agency view of causation
is anthropomorhic. It fails to give account of non-man-made causes. It has
a giuse of anti-naturalism. Agent causation has traditionally been the realm
of Reid, Chisholm and (sic) Brentano and Freud. Besides these problems, the
idenficiation of causes by the acts of an agent has a pragmatic element,
which makes this kind of account context-dependent (Carla Galavotti xx, W.
Salmon).
To see how actions and causes can work in Nature we need to get rid of
these problems. But that will be easy if we understand what we should expect
from causality. We can do this by looking at what it means for the scientist.
What do we Mean by Causality?
I am now making my second main point. I show that a necessary (and perhaps
also sufficient) condition for a relationship between events (etc) to be
causal is to have some kind of depth. Therefore, realism about causality
is depth realism.
Take experiments in Hacking's style. A most-recent question is whether
genes are real. This is what many opponents of the human genome frenzy
ask. (Skip the entity aspect here, individuation is not just causality's
problem. Ask something like this instead: are the phenomena reported about
the genes real?)
I think the only possible answer is this: go and see (try) yourself.
(In passing, we see a difference between the causal view of science and
the textual view of science. The latter would be concerned with intersubjectivity
and language as a vehicle at this point. The causal view suggests that all
this is private business. The only thing that matters is to achieve the
causal part. If theories and textbooks help in this, so it is better, but
they are just used as stepping stones – we may end up with instrumentalism
about theories combined with realism about causality.)
To repeat an experiment, or to make a different experiment with the same
effect, means that different interactions with the same entities or the
same processes are instantiated each time. For the detection of the presence
of a given molecular compound there are zillions of ways: electrochemical,
biochemical, X-ray christallography, etc. One can never actually repeat
an experiment, nor is this what we want: it is always a different experiment
that we are interested in.
The requirement that different experiments should be possible on the
same causal relation means that a causal phenomenon should always be capable
of having many different effects (mediated by implicit or latent variables
activated in the interaction), and that many different ways for achieving
a given effect should exist. In short, no causality without multiple
and implicit causation. What science means by causation is not a single cause-and-effect
link but an (inexplicable) modal unity of materially related cause-effect
links. As a side effect of depth, causality understood in the style of science
becomes context-independent.
We also start to see what role (human) agency plays in causality. The
human agent's role is twofold: (1) it offers a template for causality by
establishing depth via the several interactions by which an action can be
realized (2) it offers a testbed for natural causation without humans.
What Causality?
To this latter pont, let us reverse things to show the relation: no depth
- no causation.
What would be your reaction if you were not allowed to do things in an
experiment your own way? Where every instruction must be carefully followed
in the proverbial way, it’s magic, not science. "Conspiracy of scientists"
is a fiction exactly because there is no obligatory way of how to get to
a given effect. An effect being "real" is synonymous in science with its
being achievable by an endless variety of known and still unkown ways.
More than that, in science single-aspect causation counts as mere regularity.
(I am contrasting "normal", or multiple, causation with single-aspect causation
now. It the latter a cause and an effect are related by a single relation
as in the propositional form). At best, a single-aspect causal relation
is something that points towards not properly understood ceteris paribus
conditions. If an experiment only works in a certain way then the outcome
can be caused by some other factor than the one accounted for in our single-aspect
model. (Maybe it was the electric charge transferred from the hair of the
assistant, and co-varying with the "right" kind of initial conditions set
by her etc. etc.)
This explains why regularity, contrary to the popular handling of the
concept in philosophy of science, is not enough to even believe that a phenomenon
exists. Inheritance in the 19. century is an example. Regularities were
obvious by then but since they were not understod in the full causal form
they were handled with suspicion - biology was divided into "hereditarians"
and "anti-hereditarians", as Darwin himself reported. Hempel certanily had
a point in his d-n model when he demanded that what he called a relevant
explanation should be able give a ground to believe that the explanandum
exists. In other words, one needs to circumnavigate an alleged causal relationship
in Nature to see if it is real. If you can't do this, the relation will
never qualify as causal.
In the CA example we didn't find the causes because there was nothing
to causally interact with, in the sense discussed now. This changes, if one
builds a physically realized CA in which the causal factors that make the
transition occur are present. It is easy to see that causes of this system
are different from single-aspect initial states of the original CA (and
include a full depth physical realization of transition functions, among
other things).
Causality and Explanation (remark)
Remember that we have one free variable too much: agency. We should not
forget about the problem of anthropomorphism either. I am making a detour
for the sake of these, to show how I think they can be incorporated into
a broader picture.
The relation between activity, causality, and agency to explanation,
description and language is increasingly well studied in developmental
psychology and anthropology, where recent theories ground language (the
prerequisite of the textual mode in science) in activity complexes. This
repeats Hacking, extended version: what happens to language with respect
to the infant is what happens to theories with respect to Nature. Esther
Thelen (xx) and others have shown how the bodily activities of the infant
and of others play a crucial role in the early formation of concepts (much
as on the basis of Hacking we can show how causal activites of scientists
play a crucial role in the formation of scientific theories)
One particularly interesting feature is the detection of activity and
agency (without this, an agency-template based model of causation would not
only be antropomorphic but solipsistic). Let me relate the work of two Hungarian
teams on infants (Gergely et al in Nature 2002 feb 14.) and dogs, respectively
(Csányi et al, a recent cover story in New Scientist 2000 March 4,
see illustration for fun). Following Meltzoff and others' studies on the
early social competence of newborns and infants Gergely and Watson (xx) pointed
out that 5-8 month old infants recognize agency (i.e. distinctions between
movers and the moved) on the basis of contingency-detection. Agents occupy
a mid-range between regular, completely foreseeable (moved) behaviors and
irregular, completely unforeseeable behaviors (true contingencies).
For the Thelen et al picture to work, we are likely to be prewired so
as to have no doubts about own actions. Gergely et al show that soon the
doubt about other's actions also vanishes. Dogs have similar competence
in both. In a recently published work it was shown that dogs can learn to
launch tennis balls by their nose, upon viewing humans doing this by their
hand. (A big deal? Well, insofar as philosophy is the exercise of turning
the obvious into the unbelievable, this is unbelievable, if one looks long
enough.) Current work is done on the identification of the cues, but it
is now clear that (at least some enculturated) animals easily detect agency
(i.e. proactivity) and translate it to own behavior.
Out of these considerations the following statements emerge: (1) Activity
is likely to exist as a primary component of human and animal life in the
form a factor about which there can be no doubt (therefore, no -ism like
realism is necessary or possible about it) and which is fundamental to all
forms of knowledge, including language (Wittgenstein would agree). (2) In
particular, causality-as-action probably turns out to be truly fundamental
and primary to other forms of causation (3) No anthropomorphism is involved,
as we are talking about a part of our animal heritage.
Mechanisms as Causal Systems Which Are Not Descriptions
We now arrive to the third and final main point of the talk. What is
all this fuss about causality and activity? I return to the idea, mentioned
before, that everything offers itself for description, actions are no different.
It is natural to ask: By extending the usual view of science towards action
causality and agency, do we end up with just more description? A fast reply
is "oh, you probably can't describe that all". The idea of depth already
implies this. But that is not the most interesting point.
So far we carefully avoided dealing with the question how causality is
handled in explanations. A closer look at causal explanations will reveal
that most often we do not even attempt to describe causality, leaving causal
transitions opaque as they are. I will discuss this idea in the context
of "mechanism" as a form of reasoning about causality. Mechanisms enjoy
a recent interest in the works of Cartwright, Elster, Bechtel and others.
In our daily folk physical dealing with causation and action we tend
to use expressions like "If I press the red button, playback begins", "Don't
remove the cover or else the water boils out", "If the wheel turns, the door
will be unlocked", and so on. These expressions take the form action event
or event event etc., and our dealings with VCRs, TV sets, home PCs and
the like are all expressed in this form. Such are the familiar mechanisms
of everyday life. A mechanism does nothing more than just gives a name to
what happens there. Why does it so happen? Who knows. Miracles, acts of God
- or causality. Instead of "if A then B" we could just say anything like "the
F2 plot" or "the B program", and we often do exactly that, replacing the
enigmatic implication structure with a mere (still more enigmatic) name.
This is not science, one would think, but there is surprise to come.
Sid and Al are two characters in a PC game in which you can build all sorts
of gadgets with which Al, the mouse, tries to kill Sid, the cat. Al unlocks
a lever that is pressed by a spring to kick Sid, who slides down a slope,
just to release a ball by his body impulse. The ball falls into water, the
water spills out, an electric drier sets on flames, the flames burn the
line that holds the trigger of a revolver that kills Sid just when he arrives
at the right place at the right time. One can build whatever fantasy permits.
Then you can ask others to tell how it all works. All things move, all things
rotate, in a sometimes very complicated fashion. To see what will happen
is a delicate mechanical problem, and then we ignored the additional difficulties
caused by chemical and electric components. Yet children can solve these
puzzles and so can we, most of us non-physicists, grown-ups, as we are. To
solve the problem by brute force physics is almost out of the question. Thinking
with mechanisms gives easy answers, however, maybe with some alternatives
left open for future testing (either the ball will hit the toaster, in
which case this will happen, or it will not, but then…).
Here we deal with an explanation structure that works opposite to the
familiar d-n picture of Hempel and Oppenheim. In the d-n approach explanations
take the form
L1,…. Ln
C1,….Cm
-------------
E1,….Ek
where L-s are statements of natural laws, C-s are contingents (initial
conditions and the like), the E-s are explananda, and crossing the line
from above means derivation. By contrast, Al and Sid make us solve physics
problems without reference to covering laws:
C1,…. Cm
--------------
E1,…..Ek
Where are all the laws gone? If a mechanism is like a causal black box,
as I think we have seen it is, then laws can be thought of as describing
the content which opens this box. Without the laws, the box remains black.
The dominant view of science is that laws are generalizations of regularities.
The d-n picture has its origin in induction over these regularities. But
the way we use mechanisms in solving complicated pre-scientific as well
as scientific problems may suggest another, alternative view of science.
Intuitively, the alternative idea of a law is this: Take the d-n form as
a set of equations, where we know the C-s and the E-s from mechanisms,
and just solve the system to get the L-s.
This is view developed in detail in my 1991 book. Under this view it
would be fundamentally mistaken to expect that a scientific theory describes
causal content: it can only indicate how events are arranged in a causal,
and hence non-transparent, scheme.
Mental Causation
Most of these remarks have arisen from the study of relatively well known
domains of science such as chemistry or biology. But finally, just to see
how such a picture could work in an untested field, let me briefly reflect
on the problem of mental causation. If mental causation exists (this issue
was not our subject here) then
- it must possess depth (as I have shown for all causal systems) and
- it will enter explanation in a form which is likely to be mechanism-like
rather than covering law-like (as I have suggested by noting that laws
are nonexistent, or less common in general, than one would expect).