Cause

[From Bruce Gregory (971205.1125 EST)]

Bruce Abbott (971205.1040 EST)

For me the most important thing is not the idea that causes precede effects
(if you want to say that the changes occur simultaneously, fine with me), it
is that the effect is the result of the cause -- an influence in a
particular direction. The crucial aspect of this is that effects do not
precede their causes. If X influences Y, we expect to see the change in Y
occur before or (apparently) simultaneously with the change X, and not
before. If the causal relationship is one way, then we would not expect to
find X changing systematically as Y is manipulated. Varying the disturbance
causes variations in control-system output, but variations in control-system
output do not cause variations in the disturbance. For this reason the
airline pilot is justified in his belief that pulling the control yoke back
raises (causes to rise) the elevators, even though his input change in the
elevator servo's reference acts through a loop of circular causality to
produce a predictable change in the angle of the elevator on the tail of the
aircraft.

It might be well not to lose sight of the fact that causes are
features of models, not features of the world, as Hume pointed
out some time ago.

Bruce

[From Bruce Abbott (971205.1040 EST)]

Mary Powers 9712040 --

In support of this comment Bruce lists several phenomena: moving a switch
turns on a light, pulling back on the yoke moves the elevators of a plane,
smoking causes cancer, and dinosaurs went extinct because of a chain of
events set in motion by an asteroid collision.

These all, Bruce says, have "the essential feature" of cause: "a change in
Variable A (the cause) _leads to_ a change in variable B (the effect). The
term 'cause' asserts a directional property in the systematic relationship
that 'systematically related' does not".

What people are trying to get across to you, Bruce, is that the direction
A->B implies a temporal relationship: B follows A. That is what "leads to"
means. But the temporal relationships of this nature in the physical world
do not apply to control systems.

Let's talk about that. A control system is in a state of equilibrium when
the magnitude of disturbance to the CV begins to change so as to push the CV
farther away from its reference value, the CV, which at this moment is being
affected both by the changing disturbance and the constant system output,
begins to move further away from the system's current reference value for
the CV. This change is picked up by the perceptual input device, whose
output now also begins to change as a function of the changing input. The
change in perceptual signal propagates to the comparator, where it is
combined with the current reference signal to produce an output change in
error signal. This change in signal propagates to the output device, which
produces a change in output according to the output function, and this
change in output now begins to affect the CV.

Meanwhile, the disturbance continues to change, and while the initial change
has been propagating around the loop, further changes to the CV have
occurred and these also are propagating around the loop. At some point the
effect of the original change completes the circle and begins to influence
the value of the CV, and at this point the CV is being influenced
simultaneously by the present change in disturbance and by the recirculating
and smeared out effects of the earlier changes in the disturbance.
Everything all around the loop is changing simultaneously, and while B is
being affected by the changing value of A, A itself is being affected by the
changing value of B, and all changes must be treated simultaneously to
derive the proper equilibrium value that all will eventually reach under
constant disturbance.

Although all these changes are happening concurrently, this does not mean
that there is no temporal ordering -- effects propagate around the loop at a
particular rate (not instantly) and in a particular direction (clockwise in
the usual diagram). Causes precede effects, because physical interactions
take time.

For me the most important thing is not the idea that causes precede effects
(if you want to say that the changes occur simultaneously, fine with me), it
is that the effect is the result of the cause -- an influence in a
particular direction. The crucial aspect of this is that effects do not
precede their causes. If X influences Y, we expect to see the change in Y
occur before or (apparently) simultaneously with the change X, and not
before. If the causal relationship is one way, then we would not expect to
find X changing systematically as Y is manipulated. Varying the disturbance
causes variations in control-system output, but variations in control-system
output do not cause variations in the disturbance. For this reason the
airline pilot is justified in his belief that pulling the control yoke back
raises (causes to rise) the elevators, even though his input change in the
elevator servo's reference acts through a loop of circular causality to
produce a predictable change in the angle of the elevator on the tail of the
aircraft.

Regards,

Bruce

[From Rupert Young (971205.1300 UT)]

My turn.

(Bruce Abbott (971205.1040 EST)

>Mary Powers 9712040 --
>What people are trying to get across to you, Bruce, is that the direction
>A->B implies a temporal relationship: B follows A. That is what "leads to"
>means. But the temporal relationships of this nature in the physical world
>do not apply to control systems.

Let's talk about that. A control system is in a state of equilibrium when
the magnitude of disturbance to the CV begins to change so as to push the CV
farther away from its reference value, the CV,....

Although all these changes are happening concurrently, this does not mean
that there is no temporal ordering -- effects propagate around the loop at a
particular rate (not instantly) and in a particular direction (clockwise in
the usual diagram). Causes precede effects, because physical interactions
take time.

For me the most important thing is not the idea that causes precede effects
(if you want to say that the changes occur simultaneously, fine with me), it
is that the effect is the result of the cause -- an influence in a
particular direction.

You are driving down a straight road and there is a strong wind from the left.
Your behaviour, wheel turned into wind, counteracts the effects of the wind.
You (Bruce) would say, it seems, that the behaviour is 'caused' by the wind,
there is a temporal ordering from disturbance to behaviour,
        Disturbance ----> Behaviour
ie. whenever A exists B follows.

Now drive down the same road in the same car with the same wind but with your
eyes closed. The disturbance is still present but there is no behavior (or at
least not the same behaviour). In this case the disturbance is plainly NOT
causing the behaviour. The temporal ordering is broken, B does not follow from
A. All the wind is "doing" is applying a force to the car moving it to the
right. I think what the PCTers are saying is that it is meaningless to use
the term "cause" (from environment) when applied to control systems. It is
the control system that is doing the doing not the disturbance. The perceived
link
between a disturbance and the resulting behaviour only has meaning in the
context of an active control system that is controlling a variable affected by
the disturbance. Deactivate away the control system and the disturbance isn't
"causing" anything (within the control system).

Regards,
Rupert

[From Bill Powers (971205.1111 MST)]

Bruce Abbott (971205.1040 EST)--

Let's talk about that. A control system is in a state of equilibrium when
the magnitude of disturbance to the CV begins to change so as to push the CV
farther away from its reference value, the CV, which at this moment is being
affected both by the changing disturbance and the constant system output,
begins to move further away from the system's current reference value for
the CV. This change is picked up by the perceptual input device, whose
output now also begins to change as a function of the changing input. The
change in perceptual signal propagates to the comparator, where it is
combined with the current reference signal to produce an output change in
error signal. This change in signal propagates to the output device, which
produces a change in output according to the output function, and this
change in output now begins to affect the CV.

As Bruce Gregory reminds us, causation itself is an aspect of a model. If
you have a model in mind, you naturally tend to see things in terms of that
model. The lineal causation model emphasizes temporal lags, which allow you
to think of causal events followed by their effects in the form of later
events. Even if the lag is very small, it allows enough conceptual space
between cause and effect to keep them separated in one's mind.

The PCT concept of behavior is not based on events, but on continuous
variations in variables. Let's take the same example you use, but look at
it in terms of continuous variations. We begin the same way:

A control system is in a state of equilibrium when
the magnitude of disturbance to the CV begins to change so as to push the CV
farther away from its reference value, the CV, which at this moment is being
affected both by the changing disturbance and the constant system output,
begins to move further away from the system's current reference value for
the CV.

Now let's follow in some detail. As the disturbance rises from zero, there
is a change in the error signal which produces a change in the output which
produces a change in the effect of output on the input, with some delay
representing the total propagation time from input back to input. During
the loop delay time, the error keeps increasing. After the initial delay,
the output begins to rise in a direction opposite to the disturbance. This
rising output slows the the increase of the error signal, until finally the
output is rising at nearly the same rate as the disturbance (in the
opposite direction); the error signal now is rising at a much lower rate,
heading toward an asympotote. If we plot the disturbance on the same scale
as the output, with the sign of one reversed, we will see that as the
disturbance rises there is a brief delay and then the output begins to
increase along a parallel curve with the same shape but slightly delayed.
When the disturbance reaches it maximum value, the output keeps rising for
one or two delay times, and then levels out at the equilibrium
(steady-state) value. When the disturbance begins to level out, the error
signal declines from the steady value it had during most the transition,
until it drops to the equilibrium value.

So the error signal follows a particular waveform, something like this:

ยทยทยท

*******************************
          * <------------------------------> *
      * output lagging dist by delay-time *
**** ****************
    > >
  start of change end of change

This whole plot should have a slight positive slope if the integrating
output function is leaky, but it's hard to indicate with asterisks.

I recommend looking at all the variables in a simulation containing a
delay. The overall picture is one of a continuous relationship with a
slight lag between the changes in the variables. The lag is visible only
when the disturbance is changing; when the disturbance reaches a steady
state, the effect of the lag disappears.

Clearly, in this way of seeing the behavior there are no significant
events, except those we may mark off at arbitrary points in the plot.
Instead we have continuous variations with time lags. The only hint of
normal cause-effect sequences appears when the disturbance is varying;
otherwise the behavior is just as if there were no delays.

So how could we _force_ this system to behave in a way that was more
clearly causal? The most popular way is to think of disturbances as on-off
events, with instantaneous rise times. And it helps to think of the delays
as being quite long:

             dist
          ********************************************************
                                  ###############################
                              # output
          <---- delay ----> #
                           #
***********################

Now we clearly have a stimulus and a response, with a definite onset time
for the disturbance and a somewhat less definite onset time for the
response. The error signal would look like this:

           ****************
                            *
                               *
*********** **********************************

Now we can see that the disturbance causes an error, which, after a delay,
is corrected by the rise in the output action. That's the description we
most often hear from people who want to create definite events in a
cause-effect sequence.

There's clearly a continuum here, with a sequence-of-events concept at one
end and a continuous-variation-with-lags concept at the other. If you want
to think in terms of discrete cause-effect sequences, you pick one end as a
source of examples. If you want to consider continuous variations, you pick
the other.

Best,

Bill P.

[From Bruce Abbott (971205.1630 EST)]

Rupert Young (971205.1300 UT) --

You are driving down a straight road and there is a strong wind from the left.
Your behaviour, wheel turned into wind, counteracts the effects of the wind.
You (Bruce) would say, it seems, that the behaviour is 'caused' by the wind,
there is a temporal ordering from disturbance to behaviour,
       Disturbance ----> Behaviour
ie. whenever A exists B follows.

Now drive down the same road in the same car with the same wind but with your
eyes closed. The disturbance is still present but there is no behavior (or at
least not the same behaviour). In this case the disturbance is plainly NOT
causing the behaviour. The temporal ordering is broken, B does not follow from
A. All the wind is "doing" is applying a force to the car moving it to the
right. I think what the PCTers are saying is that it is meaningless to use
the term "cause" (from environment) when applied to control systems. It is
the control system that is doing the doing not the disturbance. The perceived
link
between a disturbance and the resulting behaviour only has meaning in the
context of an active control system that is controlling a variable affected by
the disturbance. Deactivate away the control system and the disturbance isn't
"causing" anything (within the control system).

Yes, I agree. (Surprised?) I think that where we are getting into trouble
is that "cause" as usually used in psychological research refers to a
conclusion reached via a particular arrangement called an experiment, in
which some variable is manipulated (varied under control of the
experimenter), other potentially confounding variables are held constant (or
as constant as possible), and any resulting systematic changes in some other
variable under observation are noted. Because other variables have been
held constant, the observed relationship indicates that some kind of
influence, running from the manipulated variable to the observed variable,
relates the values of the two variables. It is said that varying the
independent (manipulated) variable causes changes in the observed (dependent
variable). This term "cause" is not meant to indicate that the IV has a
direct physical influence on the DV as in your example where the wind exerts
a side-force on the car. It means only that variations in the DV result
systematically from varying the IV when other potential explanations for the
observed variations in the DV are ruled out. Thus, when one is controlling
the position of the car on the road, changes in disturbance (windforce) are
systematically related to changes in steering angle. When one's eyes are
closed (so that one is not able to observe the car's deviation from its
reference position), changes in disturbance may very well show no systematic
relationship. A researcher observing these changes in relationship would
wonder what conditions must hold before the relationship between windforce
and steering angle appears.

I think that knowing that B changes with A when other potential influences
on B are held constant is an important thing to know about the relationship.
It differs from the knowledge gained from merely observing two variables as
they exist in nature and noting that the two tend to change values together
in a systematic way. There may be no way to tell from these observations
whether B would vary in response to manipulating A but not vice versa, A
would vary in response to manipulating B but not vice versa, A would vary in
response to varying B _and_ vice versa, or that neither A nor B would change
in response to manipulating the other. (In the last case, A and B vary
together as observed in nature because both are changing -- independently --
in response to changes in a third, unobserved variable. Such
naturally-observed relationships are termed "correlational."

If you have a good quantitative model of the system under investigation, the
model makes specific predictions as to what relationships will be observed
to exist between variables if the model is valid. The model specifies which
variables directly cause (in the engineering sense) changes in which, and
which only have influences (together with other variables). Observing that
all the variables in the system (that are observable) vary together as
predicted will provide support for the model. However, if you do not know
much about the system and have not yet formulated a model, it is important
to know how the variables in the system under study interact. The way to
learn about the system is (a) to observe how the variables relate as the
system performs under natural conditions, and (b) to experimentally vary
some of the variables while holding others constant, and observe others, in
order to determine the nature of the linkages between the variables (direct,
indirect, linear, nonlinear, etc.). Based on this information, a model
having the "right" web of influences (as empirically determined) can be
developed.

In this research context, "cause" and "correlation" thus refer to different
states of knowledge about the observed linkage between variable -- cause if
one variable changes as another is varied when other possible influences are
held constant, and correlation if the observed relationship disappears under
this condition, or if the status of the linkage is as yet untested. This
distinction may not be as important when well-developed models are being
tested for their ability to predict, but provides crucial information when
one is attempting to infer the nature of the system under study by examining
how its variables relate as the system operates.

Regards,

Bruce

[From Bruce Abbott (971205.1840 EST)]

Bruce Gregory (971205.1125 EST) --

It might be well not to lose sight of the fact that causes are
features of models, not features of the world, as Hume pointed
out some time ago.

Hume also pointed out that there is no logical justification for scientific
induction, although as he also noted, it seems to work much of the time,
which is justification enough. There is no reason to believe that
gravitational attraction will continue in the universe tomorrow, other than
the apparent fact that it has been doing so for a very long time in the
past. (Bill P.'s particular nightmare, I would guess, is that he will wake
up tomorrow to find that everyone has become an S-R system. I don't know
how he feels about the prospects of losing gravity.)

I don't see how losing sight of the fact that causes are features of models
and not features of the world would alter either the current discussion of
causes or the conclusions being drawn. Perhaps you could explain what you
see as its relevance.

Regards,

Bruce

[From Bruce Gregory (971205.2020 EST)]

Bruce Abbott (971205.1840 EST)]

Bruce Gregory (971205.1125 EST) --

It might be well not to lose sight of the fact that causes are
features of models, not features of the world, as Hume pointed
out some time ago.

Hume also pointed out that there is no logical justification for scientific
induction, although as he also noted, it seems to work much of the time,
which is justification enough.

If you can dismiss Hume so easily what chance would I have?

Bruce

[From Bill Powers (971205.1806 MST)]

Bruce Abbott (971205.1630 EST) --

I think that where we are getting into trouble
is that "cause" as usually used in psychological research refers to a
conclusion reached via a particular arrangement called an experiment, in
which some variable is manipulated (varied under control of the
experimenter), other potentially confounding variables are held constant (or
as constant as possible), and any resulting systematic changes in some other
variable under observation are noted.

I don't see that this very cautious way of defining the term creates any
new meaning for causation. The real test of what is meant comes when you
ask how a supposed cause of behavior is to be used in practice. If the
purpose of psychology is, as Skinner has (and many others have) put it, the
"prediction and control" of behavior, then a cause, however hesitantly
defined, is still going to be used to make predictions and achieve control.
In other words, once you know what DV is caused by the IV (however you
define that), you will either use that IV in the future as a way of
predicting behavior, or by manipulating the IV you will use it in the
attempt to control behavior. What else is there to do with it?

The success of predictions and control naturally depend on how reliable a
cause the IV is. Psychologists are used to predictions with large errors,
and control that works only as a suggestion of influence. 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. 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?

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

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.

Best,

Bill P.

[From Bruce Abbott (971206.0100 EST)]

Bruce Gregory (971205.2020 EST) --

Hume also pointed out that there is no logical justification for scientific
induction, although as he also noted, it seems to work much of the time,
which is justification enough.

If you can dismiss Hume so easily what chance would I have?

I agree with Hume's assessment. Since when is that called "dismissal"?

Regards,

Bruce