Strong and weak causation

[From Bruce Abbott (971206.0310)]

Can't sleep. (Nothing to do with present discussions.)

Bill Powers (971205.1806 MST) --

The success of predictions and control naturally depend on how reliable a
cause the IV is.

True.

Psychologists are used to predictions with large errors,
and control that works only as a suggestion of influence.

True.

Some seem to
think that this absolves them of criticisms that they want to be dictators
-- after all, they say, we can't really _make_ people do anything.

Is that what you think, that psychologists want to be dictators? The
methods of dictatorial control are well known and astonishingly effective
(e.g., see Hitler, 1936); if psychologists want to be dictators, why don't
they just use these methods?

I have a different view: psychologists want to understand human perception,
cognition, emotion, memory, learning, psychological development, brain
function, to the point where effective intervention becomes possible when
things go wrong and a person needs help. All intervention assumes that
something can be done to change the way individuals function; when Dag
offers his services to management, for example, he is selling effective
intervention that will allow managers to deal more effectively with those
they must manage. If Dag is successful, he changes manager's behaviors in
the ways he wants them to change. The only way he can do this is by
manipulating the manager's perceptual inputs (e.g., by talking to them,
making suggestions, asking questions, demonstrating, etc.) Isn't bringing
about an intended change in skills, attitude, and so on an example of
control on Dag's part? If so, does this mean that, in your mind, Dag wants
to be a dictator?

But
isn't that only because they haven't found the _real_ IVs, which _always_
have reliable effects? If they were better able to predict and control,
wouldn't they do so?

That is one possibility. It is also possible that the _real_ IVs cannot be
observed, measured, or directly altered by the psychologist. It is also
possible that the behavior in question is based in part on nonlinear
functions that yield extreme sensitivity to initial conditions, placing an
inherent limit on prediction and control.

So I don't think there's really any difference between a physicist's
concept of causation and that of a psychologist. The careful definitions
you offer simply reflect the different degrees of success at prediction and
control that we find between physics and psychology.

You've skipped entirely over my distinction (which I think is important)
between testing a specific model and attempting to discover empirically the
nature of the system of variables under investigation. The distinction I
have made has less to do with "degrees of success at prediction" than with
what has been learned about how each variable relates to the others.
(Perhaps you just didn't follow what I was talking about?)

The difference is
between "weak causation" and "strong causation."

Whether one must make do with "weak causation" or can achieve "strong
causation" has to do with the information available to the one doing the
predicting. One case where strong causation can be obtained arises when the
behavior to be controlled or predicted can be shown to be organized around
the control of a particular variable (or vector). So long as the person
continues to exert control, he or she has no choice but to act so as to
offset any perceived deviations of the variable from the reference value
established for it. Even here there may be alternative ways of doing so
that are open to the person, but examination of the situation may show what
options are available. Thus knowing what a person is attempting to control
provides a powerful means for understanding, predicting, and even
controlling the person's behavior.

But what will the person attempt to control, and why? Finding out what the
person is controlling at the moment allows one to skip over these questions.
This amounts to saying that I can predict what the person will do, because I
know what the person is doing. It succeeds so long as the person continues
to control that variable. Will that attempt to control succeed or fail?
That may depend on the individual's previous experience and the lessons and
skills that were learned as a result. Likewise, how the person exerts that
control, although this depends also on the options open to the person at the
time. In the absence of information about what the person knows, the skills
attained, and how the person is interpreting the situation in which she
finds herself, one may be in a poor position for predicting what that person
is likely to attempt to exert control over or the means the individual will
choose in order to do so. The more information one has about the person,
the better one is able to predict what the person is likely to do in a given
circumstance. When you know that the person is controlling a particular
variable (as when you ask the person to do so as part of an experiment, and
observe the person attempting to comply (or succeeding), and have provided
only one basic means for doing so (e.g., moving a mouse) that can be
mastered quickly, it is not terribly surprising (to me, at least) that one
can predict his behavior with high accuracy. This ability reflects the
nature of the task and the situation more than it does any presumed
superiority of the science involved.

My remarks on this subject have amounted to saying that I'm interested only
in strong causation, like the effect of the positions of the two ends of
the rubber bands on the position of the knot. This is the kind of causation
I think we need in order to build a science of behavior. Weak causation
simply allows for too many alternative explanations, and gives subjective
interpretations too much influence. Strong causation, while probably harder
to find and prove, is much less subject to uncertainty, wishful thinking,
and the effects of belief.

Who would disagree that better knowledge of the system (permitting strong
causation) is to be preferred over poorer knowledge (permitting only weak
causation)? The problem is that for some problems it is currently the best
one can do. I'm presently having trouble recalling a particular word. It
came to me yesterday but I forgot it again before I could write it down when
I was momentarily distracted. I can tell you some of the conditions that
are likely to improve my ability to recall the word, but cannot guarantee
that if I set those conditions up for myself, the word will come to me.
Please show me how I can use PCT to guarantee that I will recall the word
(i.e. applying strong causation). If you cannot, then tell me what specific
"strong-causation" research I should undertake to discover how to predict my
memory successes and failings accurately in the future ("100% correlations"
please), so that I won't have to make do with mushy weak causation, which
permits only probabilistic statements to be made.

Eager to use strong causation,

Bruce

[From Bill Powers (971206.0459 MST)]

Bruce Abbott (971206.0310)--

Some seem to
think that this absolves them of criticisms that they want to be dictators
-- after all, they say, we can't really _make_ people do anything.

Is that what you think, that psychologists want to be dictators? The
methods of dictatorial control are well known and astonishingly effective
(e.g., see Hitler, 1936); if psychologists want to be dictators, why don't
they just use these methods?

Because they're illegal in this country, except when applied to children or
prisoners. In the old Soviet Union, psychiatric "treatment" was used to
nullify dissidents: deviation from the party line was prima facie evidence
of mental illness.

I have a different view: psychologists want to understand human perception,
cognition, emotion, memory, learning, psychological development, brain
function, to the point where effective intervention becomes possible when
things go wrong and a person needs help.

The tricky question is who decides that a person needs help.

All intervention assumes that
something can be done to change the way individuals function;

It also assumes that the way an individual functions _should_ be changed,
and that it should be changed in a particular way (often, back toward the
average behavior -- see David Goldstein's proposed project). In recent
years, this decision has been left more to the person whose behavior is to
be changed, but that has come about mainly through a long series of court
cases rather than enlightenment on the part of psychologists or psychiatrists.

Isn't bringing
about an intended change in skills, attitude, and so on an example of
control on Dag's part? If so, does this mean that, in your mind, Dag wants
to be a dictator?

No, although Dag may disagree. A dictator wants to control others to suit
his own concepts of how he wants people to behave. A helper tries to find
out what other people are trying to control, and helps them do it better if
he can (and if he agrees he wants them to be able to do it better -- I
don't think I'd want to help a burglar get better at his trade).

But
isn't that only because they haven't found the _real_ IVs, which _always_
have reliable effects? If they were better able to predict and control,
wouldn't they do so?

That is one possibility. It is also possible that the _real_ IVs cannot be
observed, measured, or directly altered by the psychologist. It is also
possible that the behavior in question is based in part on nonlinear
functions that yield extreme sensitivity to initial conditions, placing an
inherent limit on prediction and control.

That's only a practical problem from the standpoint of one who wants to
control others. My point was that if it were possible to control other
people perfectly, many psychologists would (or think they would) jump at
the chance. I spent a year in a behavioristic graduate school of
psychology, and the other students spent what seemed a large part of their
time thinking up fun ways to control other people. And their professors,
after hours, joined in the fun with their own gleeful ancedotes.

I think you misunderstood my reference to weak and strong causation. I was
talking about the elements of a theory. If one looks only for strong
causation, then one assumes it is there to be found, with sufficient
effort. I think there are definite, reliable, robust laws of behavior to be
found that are build on concepts of strong causation. They remain
undiscovered primarily because so many people have been satisfied with
finding weak causes and don't believe that there can be any rigorous laws
of behavior.

You've skipped entirely over my distinction (which I think is important)
between testing a specific model and attempting to discover empirically the
nature of the system of variables under investigation.

I suppose I failed to comment on it because I thought it obvious that I
agree with this distinction. The method of testing specific models, as we
see in the hard sciences, yields several orders of magnitude better
understanding of any process than does purely empirical research (if there
is such a thing).

The distinction I
have made has less to do with "degrees of success at prediction" than with
what has been learned about how each variable relates to the others.
(Perhaps you just didn't follow what I was talking about?)

Perhaps not. What I mean by strong causation pertains to variables that are
very reliable functions of other variables. This kind of causation is
usually found only in immediate, direct effects where essentially all
contributing variables can be identified.

Whether one must make do with "weak causation" or can achieve "strong
causation" has to do with the information available to the one doing the
predicting.

And that in turn depends on the theory under which one formulates questions
to be put to experiment. If you ask the right questions of nature, you will
get clear answers. If you ask the wrong questions, you will get weak
probabilities.

One case where strong causation can be obtained arises when the
behavior to be controlled or predicted can be shown to be organized around
the control of a particular variable (or vector). So long as the person
continues to exert control, he or she has no choice but to act so as to
offset any perceived deviations of the variable from the reference value
established for it. Even here there may be alternative ways of doing so
that are open to the person, but examination of the situation may show what
options are available. Thus knowing what a person is attempting to control
provides a powerful means for understanding, predicting, and even
controlling the person's behavior.

But what if some alien visitor were to tell us that ALL behavior is
actually control behavior, no other kind ever occurring? What if it told us
that behavior is really highly organized and regular? We could draw only
two possible conclusions: that the alien is mistaken, or that we have been
studying behavior from the wrong point of view and have failed to see the
regularities that are actually there. The latter is basically my thesis
about conventional psychology.

The more information one has about the person,
the better one is able to predict what the person is likely to do in a given
circumstance. When you know that the person is controlling a particular
variable (as when you ask the person to do so as part of an experiment, and
observe the person attempting to comply (or succeeding), and have provided
only one basic means for doing so (e.g., moving a mouse) that can be
mastered quickly, it is not terribly surprising (to me, at least) that one
can predict his behavior with high accuracy. This ability reflects the
nature of the task and the situation more than it does any presumed
superiority of the science involved.

So that's the real issue, is it? The superiority of the particular science.
My father once made a similar remark when I told him about some of the
tracking experiments. He said, "So they can follow instructions. So what?"

When I was in graduate school, we students were told about an experiment in
which a psychologist decided to eliminate all confounding factors and
_prove_ that stimuli really cause responses. He boiled the procedure down
to presenting a light stimulus and measuring some motor response, I forget
which one. To keep the animal from moving around and altering its relation
to the stimulus, he ended up building a narrow box into which the animal
was inserted, on its back, with a clamps to hold its body, head, and three
limbs still. To keep the animal from closing its eyes and thus failing to
receive the stimulus, he surgically removed its eyelids. And still, the
animal failed to respond every time with the same action given the same
stimulus. Conclusion? Behavior is variable.

Now that was certainly a simple situation, a put-up job; if any experiment
was set up to favor getting a regular result, this one certainly was. But
it failed.

I admit that tracking situations are very simple. It is also very simple to
use a motor output to affect the shape of a display, a rate of rotation, a
similarity of line-figures in size or proportions, a color, the pitch of a
sound, or a symmetry of movement (just some of the controlled variables
that have been tested). It is also true that in every case tested, the
model fit the data with very high accuracy and accurately predicted
behavior under changed circumstances. This was true even when more than one
person contributed to the control of a single variable, or when several
interacting variables were being controlled by two to four people at the
same time (Tom Bourbon's experiments). In Rick Marken's "Mind Readings"
book there are many simple experiments in which the model fits the data
with the same high accuracy, each one dealing with a different aspect of
control behavior in one or two dimensions.

You dismiss these results by saying that they reflect the nature of the
task and the situation. But are there any results, from any psychological
experiment that you know of, that can produce similar accuracies of
prediction over such a range of conditions, and under circumstances where
random disturbances, unpredictable by experimenter or subject, were
deliberately and routinely introduced?

Why not say instead that the reason for success in all these experiments
and demonstrations was that we did, in fact, have a good idea of what the
people were controlling and their means of control, and a model that
explained how they did this? Couldn't one guess that if we _studied_ what
people are controlling in natural situations, and worked out their means of
control, we would have equal success with even more complex and less
constrained behavior? Is it so unreasonable to assume that if we simply
continued and expanded this approach, we would eventually be able to
explain far more complex behaviors, and do so better than is now done under
other ways of conceiving and analyzing behavior?

Who would disagree that better knowledge of the system (permitting strong
causation) is to be preferred over poorer knowledge (permitting only weak
causation)? The problem is that for some problems it is currently the best
one can do.

But that's because PCT has not been tried. It's the best one can do while
not using PCT. What you mean, Kimo Sabe, the best "one" can do?

I'm presently having trouble recalling a particular word. It
came to me yesterday but I forgot it again before I could write it down when
I was momentarily distracted. I can tell you some of the conditions that
are likely to improve my ability to recall the word, but cannot guarantee
that if I set those conditions up for myself, the word will come to me.
Please show me how I can use PCT to guarantee that I will recall the word
(i.e. applying strong causation). If you cannot, then tell me what specific
"strong-causation" research I should undertake to discover how to predict my
memory successes and failings accurately in the future ("100% correlations"
please), so that I won't have to make do with mushy weak causation, which
permits only probabilistic statements to be made.

Oh, Bruce, what a silly challenge. I'm talking about a decades-long
research effort involving thousands of people who understand PCT and are
working to develop it, starting with the simple experiments where it works
so well and trying to expand it to work better with more complex
situations, and always maintaining the highest standards for what will be
accepted as a fact. And you're challenging me to come up with parlor-game
stuff that is as far beyond us now as going to the moon was beyond Newton.
You might as well ask James Clerk Maxwell to build a radio transmitter, or
Madame Curie to design a nuclear reactor. If you need to be able to say
right now what circumstances would favor remembering a word, I suggest you
study the question empirically. But if you really want to _know_ how memory
works, I counsel patience.

You appear to believe that the future of our understanding of human nature
lies in extending and expanding current concepts of behavior and
experimental methods. While I would agree that that would lead to A future,
I have a different future in mind. I think that if we focus on finding out
what people can control and how they do it, we can build up a new
understanding of behavior that will quickly surpass anything that current
methods and concepts are likely to develop in the near, or even distant,
future. The only way to find out if my vision of the future is feasible is
to start trying to bring it about.

Best,

Bill P.

[Martin Taylor 971206 22.40]

I don't think I agree with any of the comments that have been made recently
about "cause." But I think Bill Powers has made a few significant
statements that can lead to the position I do take--that "cause" is
a nonsense word if taken literally, and a useful approximation if one
doesn't delve into what it means. Rather like "obscenity," or "beauty."
We know it when we see it, but we shouldn't expect anyone else to agree
that we have seen it.

Quotes from Bill Powers:

1.>When I was in graduate school, we students were told about an experiment in

which a psychologist decided to eliminate all confounding factors and
_prove_ that stimuli really cause responses. (971206.0459 MST)

2.>Organization has the same kind of existence that matter and energy have,

but it is a separate existence. (971206.0654 MST)

3.>The careful definitions

... simply reflect the different degrees of success at prediction and
control that we find between physics and psychology. The difference is
between "weak causation" and "strong causation."(971205.1806 MST)

How are these related?

I start from the position that no event in the history of the Universe has
ever been repeated. Every moment is different from every other moment. This
seems to be a self-evident proposition, because if the state of the
Universe ever repeated, it would repeat again and again ad infinitum
(barring the possible existence of a real, true, randomness underlying
what we see as quantum probability amplitudes).

If no event is ever repeated, why do we talk as if all sorts of events
occurred over and over again? Why do we talk about the sun rising morning
after morning, as if it were the same object, rather than an object that
has spewed off millions of tons of mass in the form of light and an ionic
wind since we last saw it? Why do we talk as if the conditions of an
experiment could possibly be replicated to try "the same thing" again?
Surely that's wrong? The conditions are _never_ replicated.

What we mean when we say that something happened more than once is that
the circumstances surrounding the "repetitions" of the event were alike
in ways that we think matter, and differed only in ways we think don't
matter. That is the nub of the first quote from Bill: "eliminating
all the confounding factors" means to make the surrounding
circumstances alike in all respects that matter. The apocryphal
psychologist tried to make everything alike except for the one change
he called the "stimulus." When that event happened, the "event" called the
"response" would occur each time. Then the "stimulus" would "cause" the
"response."

In physics, the observations of an "event" have relatively few degrees
of freedom, and most of the surrounding circumstances really don't
matter (so far as we can observe). It doesn't matter whether the electron
slept well the previous night, or whether the day is cloudy, or that
this proton and that neutron had a bar fight last week. So, many
"events" can reliably be said to belong to a labellable class, such as
a proton impacting a nucleus of lead. But even there, for such a
"replicable event" there are circumstances that matter, which cannot
be pre-observed, and that affect the result. Results in physics are
usually reported in terms of probabilities--there is, say, a 10^-8
chance that the proton impinging on the lead nucleus will react in such
a way as to produce a pair of pi mesons (I'm making up the detail) and
a pair of opposed gammas.

After enough trials, the physicist reports that the number of such
events has passed a 1% significance level (against the background) and
a new particle has been discovered. The "cause" of this particle is
the proton hitting something, equally non-observable in itself, inside
the lead nucleus. Highly probabilistic. The "event" that becomes less
indeterminate with more measurement is how often this new particle is
produced when you fire a proton at the lead nucleus. That measure is
called a "cross-section", jusgt like the measurement of the cross-section
of a stick of wood. It sounds precise. So: the "degree of prediction"
in physics is often something like (using psychologists' terminology")
"stimulus A will generate response B about one time in 100 million."
And that it does is important.

In psychology, it may well matter to the result whether the subject
slept well last night, or whether the subject and the experimenter
met unhappily in the bar. There are lots more circumstances that matter.
In most S-R experiments that are reported, R follows S a lot more
often than one time in 100 million trials! So you can't use (quote 3)
"predictability" to differentiate the "scientificality" of the two
disciplines. There are other reasons, but that's not one.

ยทยทยท

--------------------

A "cause" is whatever changes within an otherwise constant constellation
of surrounding circumstances, shortly before some describable
phenomenon occurs. But that thing could have been constant, and some
other part of the constellation might have changed before the same
describable phenomenon, and then _that_ would have been the cause.
I think it was Michotte who did a lot of studies on the perception
of causality perhaps 30 or 40 years ago, with, as I remember it, the
basic finding that people saw as "causing" an event whatever happened
just beforehand. That was if they observed the event just once.

But many things have to be just so, before a particular describable
phenomenon occurs.

A glass shatters. What was the cause? Impacting a concrete surface at
considerable speed? The person holding it lost their grip? Was the cause
that someone startled the person holding the glass? Was it that the person
was holding the glass improperly, failing to put a finger under the
base? Was it a lack of attention? Was the glass too heavy? Was it that
the glass was made of a material that supports rapid fracture
propagation under sufficient local acceleration? Was it that the
concrete floor was normally covered by a thick carpet that had been
removed?

If any one of these surrounding circumstances had been otherwise,
the glass would not have shattered. But which was "the cause" of the
glass shattering?

All of this is (quote 2) "organization." The whole "event" is an
organization of factors. Change one, and the event doesn't happen.
Which is the "cause?" Are they all? Is it the structural organization
of the surrounding circumstances? Perhaps it depends on what interests you.
If you want to sue someone for the loss of the glass, the "cause" is
different from what it is if you are interested in the physics of
crack propagation in amorphous materials.

"Organization has existence," yes, and a different organization leads to
very different things happening in the world.

But slightly different organizations may lead either to the same "thing"
happening, or to something quite different happening. And here we come
again to the third quote from Bill. The wider the range of organizations
that lead to the same "event" when one particular aspect changes, the more
strongly we can say that this particular aspect is a "cause" of the "event."
If learn that the person is often startled but seldom drops what was held
at the time, we come to perceive the startle to be a very weak "cause" of
the dropped glass.

PCT can deal in strong causes of behaviour because it is possible to define
with few restrictions a constellation of circumstances in which, say, a
person will cause a cursor to track a target very much as a PCT model
would do. Stimulus-response studies can deal only in weak causes, because
there are so many ways in which any particular control loop can exercise
its control, as seen at the sensory input and muscular output interfaces
to the world. To get a repeatable relation between stimulus and response,
the experimenter has to constrain the surrounding circumstances very
tightly, like the apocryphal psychologist of the first quote.

In summary, the notion of "cause" to me depends on the description of
what may and may not change for an event of the class of interest to
occur. The wider the range of what may change, and the narrower the range
of what may not, between "repetitions", the more reasonable it is to
identify one of the critical items as a "cause" of the event. But
never can one say literally that any one thing is "the cause" of an
event. Never.

Looking from this viewpoint, it is fruitless to argue whether a stimulus
causes a response, or a disturbance an output, or the organization of
a control system a stabilization of a perception. They do, more in some
cases and less in others, and sometimes the "cause" is pre-empted by
another cause, or is not followed by the event because some part of
the constellation of surrounding circumstances is changed in a way not
anticipated, but that matters.

All in all, "cause," like "beauty" is in the eye of the beholder.

Martin