Studying Control

[From Bruce Gregory (2010.03.18.0800 EDT)]

[From Rick Marken (2010.03.16.2200)]

Tracking tasks are simply a nice way to see all the connections in
closed loop behavior. I see the behavior in all experiments as being
closed loop, the same as that in the tracking task. But unlike in the
tracking task it is harder to see the components of the closed loop in
a conventional experiment, particularly the controlled variable.

I assume then that thinking is a closed-loop behavior. Suppose I say to you, “Think of a number between one and ten.” What is the source of the number you arrive at? How do you know (a) that it is a number; and (b) that it lies between one and ten? What is the perception being controlled? Does the reference level for this perception originate at a higher level? How do I avoid an infinite regress?

No doubt these questions seem like sniping. Is this a controlled perception? To what end are you controlling this perception?

Bruce

[From Rick Marken (2010.03.18.1100)]

Bruce Gregory (2010.03.18.0800 EDT)--

Rick Marken (2010.03.16.2200)--

Tracking tasks are simply a nice way to see all the connections in
closed loop behavior. I see the behavior in all experiments as being
closed loop, the same as that in the tracking task. But unlike in the
tracking task it is harder to see the components of the closed loop in
a conventional experiment, particularly the controlled variable.

I assume then that thinking is a closed-loop behavior.

I don't see how this is relevant to my post. But, yes, PCT views
thinking as closed loop too. See the discussion (and diagram) of the
proposed "imagination connection" in B:CP. This is the PCT model of
what we are doing when we're just sitting in our armchair
ratiocinating.

Suppose I say to you, "Think of a number between one and ten." What is the
source of the number you arrive at?

I think the first question should be "why do I do it, if I do it"? I
do it if I am controlling for cooperating with you. Then what I do
(according to PCT), is follow your instructions (as I understand them)
by controlling for a program called "picking a number between one and
ten", which is completed when a reference for a number is selected and
sent to a lower level system which acts to produce a perception via
the imagination connection that matches that reference (or, if I say
it out loud, forgetting that you said, "think", then I would produce
the number as an actual perception by saying it).

How do you know (a) that it is a number; and (b) that it lies
between one and ten?

I don't believe you have to consciously "know" these things. You just
have to have control systems set up that can do this. Just like
catching a ball; you don't have to "know" that you are controlling
vertical optical velocity or that you have to keep it increasing at
the rate of .01rad/sec. It's just something you learned (through
reorganization, presumably) when you played catch with your Dad.

What is the perception being controlled?

As I said, there are several. But the main one's are 1) cooperation
with the person requesting that you think this 2) the selection
program and 3) the number perception itself.

Does the reference level for this perception originate at a higher level?

The reference specification for each of these perceptions arises from
a higher level; your reference for cooperation (a relationship)
probably originates at the system concept level, which is controlling
for a perception of the kind of social world you want to live in. The
reference for the selection program may come from a principle level
system that controls for the principle of selecting randomly.

How do I avoid an infinite regress?

Just have a top level to the model, as in PCT.

No doubt these questions seem like sniping.

Yes, they do, not least because they are irrelevant to the point of my
post, which was that the behavior in all experiments is like that in a
tracking task. I wrote my "Studying Control" post to answer your
assertion that PCT researchers have a narrow view of experimentation
because they are only interested in tracking experiments. I was trying
to explain that the tracking experiment is just a particularly nice
"test board" for studying closed loop behavior because all the
variables and relationships between them are quite transparent.

It seems to me that the more constructive, non-sniping approach to
replying to my post (and thanks for replying, even if it was a non
sequitur) would be to say what you thought about what I said in the
post. Do you agree, for example, with my arguments for thinking that
the behavior in all experiments is closed loop? If not, why not. But
it's not for me (or anyone) to tell you how to use that closed-loop
noodle of yours;-)

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bill Powers (2010.03.18.1246 MDT)]

Rick Marken (2010.03.18.1100) --

RM: I was trying to explain that the tracking experiment is just a particularly nice "test board" for studying closed loop behavior because all the
variables and relationships between them are quite transparent.

BP: Pardon an off-topic remark. Window glass is transparent, while walls are opaque. When you look at a highly-transparent window, what you see is the scene outside (or inside if you're outside). What you don't see is the window glass; if you do, it's dirty or scratched or otherwise partly opaque.

If you say "variables and relationships between them are quite transparent," you're saying we can't see them.

I know everyone is using this new buzzword, but don't people ever ask themselves what the words are supposed to mean? The word everyone wants is "visible." We don't want government to be more transparent; we want the walls that mask government from view to be more transparent, so government becomes more visible.

Best,

Bill P.

[From Fred Nickols (2010.03.18.1211 MST)]

For crying out loud, Bill! Now you expect people to be literate, too!
Don't you think that's a bit much!

Regards,

Fred Nickols
fred@nickols.us

···

-----Original Message-----
From: Control Systems Group Network (CSGnet)
[mailto:CSGNET@LISTSERV.ILLINOIS.EDU] On Behalf Of Bill Powers
Sent: Thursday, March 18, 2010 11:56 AM
To: CSGNET@LISTSERV.ILLINOIS.EDU
Subject: Re: Studying Control

[From Bill Powers (2010.03.18.1246 MDT)]

BP: Pardon an off-topic remark. Window glass is transparent, while
walls are opaque. When you look at a highly-transparent window, what
you see is the scene outside (or inside if you're outside). What you
don't see is the window glass; if you do, it's dirty or scratched or
otherwise partly opaque.

If you say "variables and relationships between them are quite
transparent," you're saying we can't see them.

I know everyone is using this new buzzword, but don't people ever ask
themselves what the words are supposed to mean? The word everyone
wants is "visible." We don't want government to be more transparent;
we want the walls that mask government from view to be more
transparent, so government becomes more visible.

Best,

Bill P.

[From Rick Marken (2010.03.18.1245)]

Bill Powers (2010.03.18.1246 MDT)--

Rick Marken (2010.03.18.1100) --

RM: I was trying to explain that the tracking experiment is just a
particularly nice "test board" for studying closed loop behavior because all
the variables and relationships between them are quite transparent.

BP: Pardon an off-topic remark. Window glass is transparent, while walls are
opaque. When you look at a highly-transparent window, what you see is the
scene outside (or inside if you're outside). What you don't see is the
window glass; if you do, it's dirty or scratched or otherwise partly opaque.

If you say "variables and relationships between them are quite transparent,"
you're saying we can't see them.

I know everyone is using this new buzzword, but don't people ever ask
themselves what the words are supposed to mean? The word everyone wants is
"visible." We don't want government to be more transparent; we want the
walls that mask government from view to be more transparent, so government
becomes more visible.

You caught me. I was actually remember consciously searching for the
correct word to use in that sentence and, because of all this buzz
going on about government being "transparent" I came up with that word
rather than "visible" which was (really!!) the word I was searching
(controlling in imagination) for. But "transparent" turned up first --
probably due to all the news talk-- and it satisfied my reference
because it has come to mean the same as "visible" to me. So how about:
the tracking experiment is a particularly nice "test board" for
studying closed loop behavior because all the variables and
relationships between them are VISIBLE.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.18.1616 EDT)]

[From Rick Marken (2010.03.18.1100)]

Bruce Gregory (2010.03.18.0800 EDT)–

Rick Marken (2010.03.16.2200)–

Tracking tasks are simply a nice way to see all the connections in
closed loop behavior. I see the behavior in all experiments as being
closed loop, the same as that in the tracking task. But unlike in the
tracking task it is harder to see the components of the closed loop in
a conventional experiment, particularly the controlled variable.

I assume then that thinking is a closed-loop behavior.

I don’t see how this is relevant to my post.

BG: This is very helpful. The fact that an experiment involving thinking is not relevant to your claim that behavior in all experiments is closed loop suggests that you do not see thinking as behavior.

But, yes, PCT views
thinking as closed loop too. See the discussion (and diagram) of the
proposed “imagination connection” in B:CP. This is the PCT model of
what we are doing when we’re just sitting in our armchair
ratiocinating.

BG: The model of memory in B:CP is based on an analogy with a computer. Memories are stored in locations with addresses and are accessed by calling up the contents of this address. (As Bill writes, “*Memory, in order to be useful, must be under control of some orderly addressing mechanism.” *(Emphasis in original.)) Unfortunately neuroscience provides no support for the existence of such a mechanism. It may be there, but we sure ain’t found it. But I don’t want to be picky. Let’s imagine that there is such a mechanism and that somehow it “knows” what the results of a “search” are.

Suppose I say to you, “Think of a number between one and ten.” What is the
source of the number you arrive at?

I think the first question should be “why do I do it, if I do it”? I
do it if I am controlling for cooperating with you.

BG: agreed.

Then what I do
(according to PCT), is follow your instructions (as I understand them)
by controlling for a program called “picking a number between one and
ten”, which is completed when a reference for a number is selected and
sent to a lower level system which acts to produce a perception via
the imagination connection that matches that reference (or, if I say
it out loud, forgetting that you said, “think”, then I would produce
the number as an actual perception by saying it).

BG: So this program was just sitting there waiting to be activated. I won’t ask how it got there. Activated by what? What decides which program to activate and how does it make this decision? No wait. I know. It’s the mechanism isn’t it? The mechanism knows which program to run and the mechanism knows the result of the calculation. I think I’m following you.

How do you know (a) that it is a number; and (b) that it lies
between one and ten?

I don’t believe you have to consciously “know” these things. You just
have to have control systems set up that can do this. Just like
catching a ball; you don’t have to “know” that you are controlling
vertical optical velocity or that you have to keep it increasing at
the rate of .01rad/sec. It’s just something you learned (through
reorganization, presumably) when you played catch with your Dad.

BG: So the assumption is that the control system already exists and it was somehow produced by reorganization. The mechanism knows the answer, but it doesn’t have to let you in on it. Unless it decides to. O.K.

What is the perception being controlled?

As I said, there are several. But the main one’s are 1) cooperation
with the person requesting that you think this 2) the selection
program and 3) the number perception itself.

Does the reference level for this perception originate at a higher level?

The reference specification for each of these perceptions arises from
a higher level; your reference for cooperation (a relationship)
probably originates at the system concept level, which is controlling
for a perception of the kind of social world you want to live in. The
reference for the selection program may come from a principle level
system that controls for the principle of selecting randomly.

BG: Yes indeed. Another pre-existing control system. It’s almost like a magician’s top hat. You reach inside and produce a control system. Voila!

How do I avoid an infinite regress?

Just have a top level to the model, as in PCT.

BG: Please forgive me, but the top level in the model sounds suspiciously like a deus ex machina. But that’s just me controlling my skeptical perception.

No doubt these questions seem like sniping.

Yes, they do, not least because they are irrelevant to the point of my
post, which was that the behavior in all experiments is like that in a
tracking task. I wrote my “Studying Control” post to answer your
assertion that PCT researchers have a narrow view of experimentation
because they are only interested in tracking experiments. I was trying
to explain that the tracking experiment is just a particularly nice
“test board” for studying closed loop behavior because all the
variables and relationships between them are quite transparent.

BG: Yes, I know. Unless an experiment bears a close relationship to a tracking experiment, it hardly even counts as an experiment, does it? So far you have confirmed my claim. The fact that you cannot see this is an example of what cognitive scientists (those charlatans) call “confirmation bias.” To put the situation in PCT terms, you are controlling the perception that all legitimate experiments resemble tracking experiments. The gain in this particular controlled perception is fairly high.

It seems to me that the more constructive, non-sniping approach to
replying to my post (and thanks for replying, even if it was a non
sequitur) would be to say what you thought about what I said in the
post. Do you agree, for example, with my arguments for thinking that
the behavior in all experiments is closed loop? If not, why not. But
it’s not for me (or anyone) to tell you how to use that closed-loop
noodle of yours;-)

I agree that, as you define a legitimate experiment, all behaviors in a legitimate experiment are closed-loop. I think Bill might call this petitio principii, however. The experiments on verbal learning that I carried out as a graduate student, for example, were not legitimate experiments. (In this case I happen to agree with you, but not because they did not resemble tracking experiments.) The example I used above was not a legitimate experiment. I see why you and Bill dismiss all cognitive science and neuroscience as smoke and mirrors. If I saw the world the way you do, I’m sure I would be equally dismissive. You can, of course, understand why those who do not share this worldview might find it slightly parochial.

Bruce

[From Rick Marken (2010.03.18.1345)]

Bruce Gregory (2010.03.18.1616 EDT)--

RM: I was trying
to explain that the tracking experiment is just a particularly nice
"test board" for studying closed loop behavior because all the
variables and relationships between them are quite transparent[visible:RM.

BG: Yes, I know. Unless an experiment bears a close relationship to a
tracking experiment, it hardly even counts as an experiment, does it?

No, I claim that every psychological experiment is like a tracking
task. In every experiment there must be a controlled variable which
can be disturbed by the independent variable and also influenced by
the dependent variable.

RM: Do you agree, for example, with my arguments for thinking that
the behavior in all experiments is closed loop? If not, why not. �But
it's not for me (or anyone) to tell you how to use that closed-loop
noodle of yours;-)

I agree that, as you define a legitimate experiment, all behaviors in a
legitimate experiment are closed-loop.

I didn't define a "legitimate" experiment. I said (if you read the
sentence above more carefully) that the behavior in _all_ experiments
is closed loop.

I think Bill might call this petitio principii, however.

He would, it it were;-)

The experiments on verbal learning that I carried out as a graduate student,
for example, were not legitimate experiments.

Why not? My guess is that they were perfectly legitimate (assuming you
got IRB approval;-) and that the behavior in these experiments was
closed loop. If you would be willing to describe one of these
experiments in some detail I could try to explain the closed loop
nature of the behavior in it.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bruce Gregory (2010.03.18.1740 EDT)]

[From Rick Marken (2010.03.18.1345)]

Bruce Gregory (2010.03.18.1616 EDT)–

RM: I was trying
to explain that the tracking experiment is just a particularly nice
“test board” for studying closed loop behavior because all the
variables and relationships between them are quite transparent[visible:RM.

BG: Yes, I know. Unless an experiment bears a close relationship to a
tracking experiment, it hardly even counts as an experiment, does it?

No, I claim that every psychological experiment is like a tracking
task. In every experiment there must be a controlled variable which
can be disturbed by the independent variable and also influenced by
the dependent variable.

That’s what I meant by a “legitimate” experiment. But I now see any experiment can be legitimized by identifying a controlled variable. Heck you don’t even need an experiment, every action must be closed loop. If it looks open loop, you haven’t tried hard enough to think of a variable that might be being controlled.

RM: Do you agree, for example, with my arguments for thinking that
the behavior in all experiments is closed loop? If not, why not. But
it’s not for me (or anyone) to tell you how to use that closed-loop
noodle of yours;-)

I agree that, as you define a legitimate experiment, all behaviors in a
legitimate experiment are closed-loop.

I didn’t define a “legitimate” experiment. I said (if you read the
sentence above more carefully) that the behavior in all experiments
is closed loop.

Got it.

The experiments on verbal learning that I carried out as a graduate student,
for example, were not legitimate experiments.

Why not? My guess is that they were perfectly legitimate (assuming you
got IRB approval;-) and that the behavior in these experiments was
closed loop. If you would be willing to describe one of these
experiments in some detail I could try to explain the closed loop
nature of the behavior in it.

I think I’ve got the hang of it. The trick is imagining a controlled variable. In one experiment subjects learned a list of nonsense syllables. When presented with the first syllable, the “mechanism” accessed the second syllable. The controlled variable was saying the second syllable. It’s not hard to do once you know the rule (there must be a controlled variable).

I can even show that the tremors in Parkinson’s are closed-loop! The mechanism in the person with Parkinson’s addresses memories of prior tremors and shakes in the same way. (Control is not very good so the same pattern rarely repeats exactly.) The great thing is that you don’t even have to be conscious. (I think that was true of more than one of my subjects.) Now what I don’t understand is why any experiment can’t be analyzed in closed loop terms. Lets consider the paired-associates learning task. How could I make it into a “real” PCT experiment?

When behavior looks to be open loop it is just because I have not been clever enough to think of a plausible controlled variable. The simplest way to identify the controlled variable is to look at the instructions. (If there are no instructions you have to be more creative.)

I think I’m ready for the test.

Bruce

[From Bruce gregory 92010.03.18.1750 EDT)]

[From Rick Marken (2010.03.18.1345)]

No, I claim that every psychological experiment is like a tracking
task. In every experiment there must be a controlled variable which
can be disturbed by the independent variable and also influenced by
the dependent variable.

I just realized that if even there were open-loop behavior, it would have to be executed closed-loop! Talk about a can’t lose model.

Bruce

[From Bruce gregory (2010.03.18.1757)]

[From Rick Marken (2010.03.18.1345)]

Bruce Gregory (2010.03.18.1616 EDT)–

No, I claim that every psychological experiment is like a tracking
task. In every experiment there must be a controlled variable which
can be disturbed by the independent variable and also influenced by
the dependent variable.

Just answered my own question! (See, I’m not as dumb as I look.) Since the controlled variable involves saying the second syllable, I could have put a rubber strap around the subject’s head. The purpose of the strap would be to keep the subject’s jaw closed! If the subject succeeded in speaking the second syllable by overcoming this disturbance, I would have showed that the subject was controlling the variable I had identified. All the data I gathered was irrelevant. (I suspected that at the time.)

Bruce

From Bill Powers (2010.03.18.1545 MST)]

Rick Marken (2010.03.18.1345) –

BG: Yes, I know. Unless an
experiment bears a close relationship to a

tracking experiment, it hardly even counts as an experiment, does
it?

RM: No, I claim that every psychological experiment is like a
tracking

task. In every experiment there must be a controlled variable which

can be disturbed by the independent variable and also influenced by

the dependent variable.

I think Bruce G’s comment has to be answered in a somewhat different way.
In a tracking experiment, there’s a moving target and the subject tries
to keep a cursor matching its position and velocity. The PCT paradigm
applies to that situation because, as you say, there’s a controlled
variable (distance between target and cursor in this case) disturbed by
an independent variable (which moves the target) and also influenced by
the actions of the control system (the cursor movements).
When we say “tracking experiment,” we mean the particular setup
with a moving (or stationary) target and a cursor. but that’s not what
remains the same from one PCT experiment to another.
In the first demo/experiment in LCSIII, one of the three possible
controlled variables is a variable shape, and another is a variable
angular orientation. These are not tracking experiments, but they are
handled by the same model that is used for analyzing tracking. You once
devised an experiment in which the controlled variable was described as
the “size” of a rectangle or square, with instructions to keep
the “size” matching the “size” of another figure (was
it a circle?). In this one you tried several possible ways in which size
might be perceived: perimeter and area, I think. I think you found that
area gave the best fit between model and real performance. In a program I
wrote for David Goldstein, the controlled variable was a color composed
of red, green, and blue. The subject, who suffered from synesthesia, used
a mouse to adjust the color of one patch on the screen so it matched the
color in which the subject perceived black-on-white numerals shown on the
same screen. In this way the subject could show an experimenter the color
that numbers (or words or letters) seemed to have, without having to name
the colors. You may remember that back in Haimowoods, in the 1980s, I
showed a demo in which a subject had to use a mouse to keep a displayed
name of a president at “Rutherford B. Hayes”, despite
disturbances that kept changing it to other presidents who served before
or after Hayes. I have shown several demos of control of velocity, of
rate of repetition, of quantities represented by printed numerals rather
that analog quantities. I have shown how control systems can balance an
inverted pendulum (and Richard Kennaway promptly showed how the same
control organization could be expanded to balance an inverted and
double-jointed
pendulum). I have shown the image of a little man who
points to a randomly moving target by using retinal images and depth
perception. I have shown crowds of individual control systems who avoid
collisions with each other, follow other people, and seek a destination
position on a playing field. I have showed how people can control the
pitch of a sound even if the pitch is being disturbed. I have
demonstrated an artificial cerebellum which stabilizes a control system
by using the convolution theorem instead of E. coli reorganization, and
of course I’ve demonstrated numerous variations on the E. coli type of
reorganizing system which achieves control through an output function
that works completely at random. I have even written a program that uses
the video image from a camera on a telescope finder to make the telescope
track satellites (of course that one is a tracking task). And
using images supplied by Warren Mansell, I’ve written a program that
allows a person with a mouse to control the expression on a human face,
against disturbances, through the range from angry to joyful.

I hope we can now finally put an end to the statements that PCT is just
about tracking behavior. It’s about any behavior we know how to simulate
on a computer, and it would be about other kinds too if someone could
tell us what to simulate.

The other side of Bruce’s criticism is that since we’re in love with
control theory, we see everything as a control system. That, of course,
is a popular human foible and a constant danger in science – of which we
are as aware as anyone, perhaps even a little more aware than average.
This is why we use models to test our ideas instead of just endless
strings of verbal arguments and plausible propositions. A model or
simulation is the most serious test to which an idea can be put, and we
use modeling because if our ideas are wrong, we want to be the first to
know. Being the last to know is embarrassing. Models operate by the
properties you put into them, not out of knowledge of the behavior you
want them to produce. The first thing that needs to be done with any new
theory is to try to prove that it’s wrong. A working model, a simulation,
is the quickest way to do that, because all assumptions are out in the
open for anyone to see, the reasoning is public, and the behavior of the
model “is what it is,” with no chance to reach in and tweak it
if it starts to behave in an unwanted way. By using models, those of us
in PCT who do that are subjecting our own explanations of behavior to far
more rigorous tests than any other theories we know of are ever asked to
pass.

Do Rick and I and others ever talk off the top of our heads without
testing our ideas all the way? Of course we do; so does everyone else.
But you won’t often find Rick or me or others who follow the same
principles claiming something as a fact if we don’t have some pretty
solid way of demonstrating that it is a fact. And I don’t mean a
statistical fact.

Best,

Bill P.

[From Mike Acree (2010.03.18.2005 PDT)]

Bill Powers (2010.03.18.1246 MDT)--

We don't want government to be more transparent;
we want the walls that mask government from view to be more
transparent, so government becomes more visible.

MA: Transparency isn't a property of government, but of politicians:
At least from certain perspectives, we can see through every one of
them.

Mike

[From Bruce Gregory (2010.03.19.1043 EDT)]

From Bill Powers (2010.03.18.1545 MST)]

The first thing that needs to be done with any new
theory is to try to prove that it’s wrong. A working model, a simulation,
is the quickest way to do that, because all assumptions are out in the
open for anyone to see, the reasoning is public, and the behavior of the
model “is what it is,” with no chance to reach in and tweak it
if it starts to behave in an unwanted way. By using models, those of us
in PCT who do that are subjecting our own explanations of behavior to far
more rigorous tests than any other theories we know of are ever asked to
pass.

Do Rick and I and others ever talk off the top of our heads without
testing our ideas all the way? Of course we do; so does everyone else.
But you won’t often find Rick or me or others who follow the same
principles claiming something as a fact if we don’t have some pretty
solid way of demonstrating that it is a fact. And I don’t mean a
statistical fact.

BG: I think your last statement highlights a fundamental problem. I understand and appreciate where you are coming from, but the rest of the world is interested in statistical facts. Let me give you an example. The book The Spirit Level: Why Greater Equality Makes Societies Stronger by Richard Wilkinson and Kate Picket contains a wealth of statistical data that supports the claim that many of the problems facing this country (low educational achievement, high crime rates, physical and health problems, drug use) are to be expected when we take into account the fact that this country enjoys the greatest income inequality of any developed nation. (Number two for no obvious reason that I know is Portugal.) Obviously the evidence is statistical and since we cannot conveniently manipulate income inequality, confined to correlations. Nevertheless, there are so many correlations that the conclusion is likely to be persuasive to anyone who takes the time to read the book. What does PCT have to add to this picture? Very little as far as I can tell. All the demonstrations you describe are based on tasks with little apparent connection to the “real world.” If such connections could be made, and if those connections led to surprising results, I suspect PCT would gather a lot more attention.

Bruce

[From Richard Kennaway (2010.03.19.1517)]

[From Bruce Gregory (2010.03.19.1043 EDT)]
The book _The Spirit Level: Why Greater Equality Makes Societies Stronger_ by Richard Wilkinson and Kate Picket contains a wealth of statistical data that supports the claim that many of the problems facing this country (low educational achievement, high crime rates, physical and health problems, drug use) are to be expected when we take into account the fact that this country enjoys the greatest income inequality of any developed nation. (Number two for no obvious reason that I know is Portugal.) Obviously the evidence is statistical and since we cannot conveniently manipulate income inequality, confined to correlations. Nevertheless, there are so many correlations that the conclusion is likely to be persuasive to anyone who takes the time to read the book.

How do the authors go about deducing causality from correlations in non-interventional data?

People have been trying to do that for years, and while there are major books on causal reasoning (Pearl's "Causality" and Spirtes et al's "Causation, Prediction, and Search", the question is still disputed (e.g. papers by the late David Freedman, and the book "Causality in Crisis?" eds. McKim et al, a large proportion of which is a ding-dong between Freedman and Spirtes' group). And I have pointed out on CSGNET that when x and dx/dt are bounded they have zero long-term correlation with each other, so you can't argue from lack of correlation to lack of causal connection, although all methods of causal discovery I've seen assume as an axiom that you can. And there are the "paradoxical" correlations to be expected when control systems are present, whereby disturbance and output have high correlation with each other, despite the only causal path between them going via the perception, which typically has very low or zero correlation with either.

This all suggests to me that causal analysis of non-interventional data obtained from a dynamical system is a non-starter. If correlations on that type of data are bad evidence of causation, they remain bad evidence however much of it is piled up.

But I have not seen the book. (My university library has two copies, but they're both out on loan until June.) What do the authors say about this fundamental methodological issue?

···

--
Richard Kennaway, jrk@cmp.uea.ac.uk, Richard Kennaway
School of Computing Sciences,
University of East Anglia, Norwich NR4 7TJ, U.K.

[From Rick Marken (2010.03.19.0850)]

Bill Powers (2010.03.18.1545 MST)--

Wow, no one comments on this for days and then, suddenly, fire hose;-)

Rick Marken (2010.03.18.1345) --

RM: No, I claim that every psychological experiment is like a tracking
task. In every experiment there must be a controlled variable which
can be disturbed by the independent variable and also influenced by
the dependent variable.

I think Bruce G's comment has to be answered in a somewhat different way.

Yes, this is a nice answer to the claim that PCT is only about
tracking tasks. Some other examples of PCT applied to non-tracking
tasks (other than the ones you mention) are 1) modeling ball catching
(Baseball Catch) 2) discrete
behavior with random consequences
(Selection of Consequences) 3) economic
behavior (http://www.mindreadings.com/ControlDemo/Economics.html) 4) a
non-tracking version of the behavioral illusion
(Behavioral Illusion) 5)
non-tracking version of the "the test"
(Behavioral Illusion and 6) and a
model of coordinated movement
(Bimanual Coordination).

I hope we can now finally put an end to the statements that PCT is just
about tracking behavior. It's about any behavior we know how to simulate on
a computer, and it would be about other kinds too if someone could tell us
what to simulate.

Yes. But my aim in this discussion is specific to considering the
possibility that the behavior in all psychology experiments is closed
loop. Based on what you say next I think it's important for me to
always emphasize that what I am trying to show is that experiments
_might_ be closed loop; the closed loop nature of experiments is, of
course, a hypothesis that has to be tested. For example, my
description of the closed loop nature of classical conditioning, where
what I propose as being controlled is the viscosity of the food in the
mouth, is a hypothesis that must be tested. My point is only to show
that the behavior in conventional experiments, which looks so
obviously open loop (the DV appearing to be caused, at least
stochastically, by the IV) _might_ actually be closed loop. And I will
present some observations (in my talk at Manchester) that suggest,
before any serious testing is done, that the behavior in conventional
psychology experiments is closed loop.

The other side of Bruce's criticism is that since we're in love with control
theory, we see everything as a control system. That, of course, is a popular
human foible and a constant danger in science -- of which we are as aware as
anyone, perhaps even a little more aware than average. This is why we use
models to test our ideas instead of just endless strings of verbal arguments
and plausible propositions.

Yes. And I think the way I will eventually test the idea that the
behavior in an "obviously" open loop experiment is closed loop is to
compare open and closed loop models of the experiment.

Do Rick and I and others ever talk off the top of our heads without testing
our ideas all the way? Of course we do; so does everyone else. But you won't
often find Rick or me or others who follow the same principles claiming
something as a fact if we don't have some pretty solid way of demonstrating
that it is a fact. And I don't mean a statistical fact.

Well put.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[Martin Taylor 2010.03.19.11.46]

[From Richard Kennaway (2010.03.19.1517)]

[From Bruce Gregory (2010.03.19.1043 EDT)]

The book The Spirit Level: Why Greater Equality Makes Societies
Stronger
by Richard Wilkinson and Kate Picket contains a wealth of
statistical data that supports the claim that many of the problems
facing this country (low educational achievement, high crime rates,
physical and health problems, drug use) are to be expected when we take
into account the fact that this country enjoys the greatest income
inequality of any developed nation. (Number two for no obvious reason
that I know is Portugal.) Obviously the evidence is statistical and
since we cannot conveniently manipulate income inequality, confined to
correlations. Nevertheless, there are so many correlations that the
conclusion is likely to be persuasive to anyone who takes the time to
read the book.

How do the authors go about deducing causality from correlations in
non-interventional data?

People have been trying to do that for years, and while there are major
books on causal reasoning (Pearl’s “Causality” and Spirtes et al’s
“Causation, Prediction, and Search”, the question is still disputed
(e.g. papers by the late David Freedman, and the book “Causality in
Crisis?” eds. McKim et al, a large proportion of which is a ding-dong
between Freedman and Spirtes’ group). And I have pointed out on CSGNET
that when x and dx/dt are bounded they have zero long-term correlation
with each other, so you can’t argue from lack of correlation to lack of
causal connection, although all methods of causal discovery I’ve seen
assume as an axiom that you can. And there are the “paradoxical”
correlations to be expected when control systems are present, whereby
disturbance and output have high correlation with each other, despite
the only causal path between them going via the perception, which
typically has very low or zero correlation with either.

Isn’t that a bit of a red herring? Bruce isn’t talking about a
situation in which there is a clear causal path but no correlation. He
is talking about a situation in which there is a structure of clear
correlations. As we discussed a while back, if A and B are correlated,
there are only a few possibilities:

  1. A influences B, or vice versa

  2. There is a variable X that influences both A and B

  3. There is a selection bias in the statistics, such as would happen
    if the values of A and B were measured while holding C constant, and C
    = A+B+e (e = other causal influences on C, unmeasured and possibly
    unknown).

causal_connections.jpg

In each diagram, an arrow means “has a direct causal influence on”. Of
course, all the variables may be involved in other causal conenctions,
including feedback loops, but I can’t think of any other possibilities
for the existence of correlation between A and B than that one of these
three types of direct influence exists. Can you?

This all suggests to me that causal analysis of
non-interventional data obtained from a dynamical system is a
non-starter. If correlations on that type of data are bad evidence of
causation, they remain bad evidence however much of it is piled up.

I disagree. They are evidence that there exists a causal relationship,
but the evidence does not discriminate among the three possibilities
described. On the other hand, as you and I have frequently shown and
pointed out, lack of correlation is no evidence of lack of causality.

To put it another way: “If A then B” implies “If not B then not A” but
it does not imply “If not A then not B”. So, “If correlation then
causality” implies “If no causality then no correlation” but does not
imply “If no correlation then no causality”.

Martin

···

On 2010/03/19 11:28 AM, Richard Kennaway wrote:

[From Bill Powers (2010.03.18.1015 MDT)]

Bruce Gregory (2010.03.19.1043 EDT) –

BP earlier: Do Rick and I and
others ever talk off the top of our heads without testing our ideas all
the way? Of course we do; so does everyone else. But you won’t often find
Rick or me or others who follow the same principles claiming something as
a fact if we don’t have some pretty solid way of demonstrating that it is
a fact. And I don’t mean a statistical
fact.

BG: I think your last statement highlights a fundamental problem. I
understand and appreciate where you are coming from, but the rest of the
world is interested in statistical facts.

BP: At one point in history there were people who pointed out, quite
correctly, that the rest of the world believes in the flatness of the
earth.

That didn’t make the rest of the world right.

BG: Let me give you an example.
The book The Spirit Level: Why Greater Equality Makes Societies
Stronger
by Richard Wilkinson and Kate Picket contains a wealth of
statistical data that supports the claim that many of the problems facing
this country (low educational achievement, high crime rates, physical and
health problems, drug use) are to be expected when we take into account
the fact that this country enjoys the greatest income inequality of any
developed nation. (Number two for no obvious reason that I know is
Portugal.) Obviously the evidence is statistical and since we cannot
conveniently manipulate income inequality, confined to correlations.
Nevertheless, there are so many correlations that the conclusion is
likely to be persuasive to anyone who takes the time to read the book.

BP: I would much rather find out what it is in the organization of
individuals in our society that leads to this inequality of income. The
same goal-oriented behaviors that lead to income inequality are probably
causing other problems like those you mention. Population data are
useful, but only with respect to predicting population variables. It
doesn’t explain why any individual is very rich, predict what any rich
individual will do, or explain why one person’s way of being very rich
causes problems for others. I would rather explore rich individuls as
specimens who do specific things for specific goal-oriented reasons. When
I have studied enough of them I will be able to create fact-based
generalizations about their properties and range of properties, based on
actual measurements that will be much more useful than any knowledge of
population means, trends, correlations, and standard deviations.

I think that misuses of population statistics have come about because of
a misconceived desire for the immediate gratification of
impressive-looking results achievable without a lot of time and expense.
Population statistics are far easier and cheaper to study than individual
characteristics. I couldn’t afford to fund either kind of research, but
if I could, I would go for the deferred gratification of high-quality
results obtained after a longer period of investigation of individual
characteristics.

What does PCT have to add to
this picture? Very little as far as I can tell. All the demonstrations
you describe are based on tasks with little apparent connection to the
“real world.” If such connections could be made, and if those
connections led to surprising results, I suspect PCT would gather a lot
more attention.

I’m not going to counter that with a list of real-world applications or
testimonials on the other side. If you can’t see anything that PCT can
add to a statistical study, then you can’t and I can’t do it for you. If
PCT is that irrelevant and useless, why have you wasted so much time on
it? That doesn’t say much for your taste in theories. I don’t think
you’re talking about what really bothers you.

Best,

Bill P.

[From Bill Powers (2010.03.18.1132 MDT)]

Martin Taylor 2010.03.19.11.46 –

MT: As we discussed a while
back, if A and B are correlated, there are only a few possibilities:

  1. A influences B, or vice versa

  2. There is a variable X that influences both A and B

  3. There is a selection bias in the statistics, such as
    would happen if the values of A and B were measured while holding C
    constant, and C = A+B+e (e = other causal influences on C, unmeasured and
    possibly unknown).

f10098.jpg

BP: This leaves out the following case, which appears at first
statistically impossible:

 <------------ very high negative r

------------->

 A ------------->e -----> [system]

--------------> B

 <-- r very low--><-----------r very low

-------->

This happens because of two undiagrammed conditions.

First, B has an effect on e nearly equal and opposite to the effect of A,
so the variations in e are far less than the variations in A and B. This
is because of an undiagrammed feedback connection from B to e.

Second, there is an irreducible noise level in the system, such that when
e is being kept very small, its effects on B are much less than the
effects of the system noise on B. This makes the correlation between B
and e small. The varible e is also affected by the system noise via B;
the result is that the correlation of A with e is very small.
Alternatively, the feedback could involve an integration, which will
lower the correlations by shifting the phase of signals.

In this case there is a correlation between A and B, but it is spurious
because the link from e to B is not what it appears to be (since the
feedback connection is left out). This is similar to your cases in which
e is usually left out of a causal diagram.

MT: I disagree [ with Kennaway].
They (the correlations) are evidence that there exists a causal
relationship, but the evidence does not discriminate among the three
possibilities described. On the other hand, as you and I have frequently
shown and pointed out, lack of correlation is no evidence of lack of
causality.

BP: You have left out the case relevant to PCT, which is the case (easily
and very reliably demonstrated in experiments with human subjects) in
which there is correlation, but causality does not run from A to B as it
apparently does. This is because B is in part causing itself.

MT: To put it another way:
“If A then B” implies “If not B then not A” but it
does not imply “If not A then not B”. So, “If correlation
then causality” implies “If no causality then no
correlation” but does not imply “If no correlation then no
causality”.

BP: However, the PCT-relevant case shows high correlation when the
apparent causality is an illusion. It shows that there can be correlation
without causality. I think you were the one who demonstrated that this is
also true if there is an integration in the feedback loop, even without
system noise.

I wonder if the problem here isn’t simply a difference in what we mean by
causality. If you mean that causality is involved in some way, visible or
not, when correlations exist, I would have to agree with you. The
feedback path is a causal path. What I am talking about is apparent
causality in which there is an input-output path that is obvious, and a
hidden feedback path which is not observed. In that case, the appearance
is that of a simple causal chain running from input, through the system,
to output. In a lineal chain of processes, we would expect all noise
sources to add their effects in quadrature, so it would be impossible for
the correlations to be observed as in my diagram above.

So the point here is not about a violation of any principle of causality.
It is about how an incorrect appearance of causality can be created when
there is an unobserved feedback path and system noise or a phase shift –
the actual case in all known examples of human control systems.

This is not a proof that all apparent input-output processes in human
beings are really control processes. It simply shows that if a
control-system organization is a significant possibility, then that
possibility has to be ruled out before any apparent causal connection
suggested by a correlation (or in any other way) can be taken at face
value.

Best,

Bill P.

[From Bruce Gregory (2010.03.19.1434 EDT)]

[From Bill Powers (2010.03.18.1015 MDT)]

BG: What does PCT have to add to this picture? Very little as far as I can tell. All the demonstrations you describe are based on tasks with little apparent connection to the "real world." If such connections could be made, and if those connections led to surprising results, I suspect PCT would gather a lot more attention.

BP: I'm not going to counter that with a list of real-world applications or testimonials on the other side. If you can't see anything that PCT can add to a statistical study, then you can't and I can't do it for you. If PCT is that irrelevant and useless, why have you wasted so much time on it? That doesn't say much for your taste in theories. I don't think you're talking about what really bothers you.

In the words of that philosophical assembly Buffalo Springfield,

"Paranoia strikes deep
Into your life it will creep"

When, pray tell, did I say that PCT is irrelevant and useless? You yourself acknowledge that PCT has little to say about the question of inequality and its social consequences. How come you can get away with telling the truth, but I can't? I suppose CSGnet is your baby, so you get to set the rules. Fair enough. Unfortunately there are real world problems that need addressing. PCT's answer, seems to be "Hold on, eventually we will have something relevant to say about these problems." That's fine, but don't get bent out of shape when people leave to your own experiments. Call in when you have something significant to report.

I'm sorry you have such contempt for statistics and population studies. Image if physicists had said, "Enough with the statistics Einstein, we plan to study individual photons until we understand their behavior in detail, then we will be able to say something about collections of photons." Needless to say, we'd still be waiting.

Bruce

[From Bruce Gregory (2010.03.19.1448 EDT)]

[From Richard Kennaway (2010.03.19.1517)]

This all suggests to me that causal analysis of non-interventional data obtained from a dynamical system is a non-starter. If correlations on that type of data are bad evidence of causation, they remain bad evidence however much of it is piled up.

I’d like to think you are joking, but I am afraid you are not. I have no problem with you not believing statistical data. Physicists gave up on causality a long time ago. PCT enthusiasts are still enthralled by it. Can I guess? You dismiss all the data linking smoking with lung cancer as a non-interventional non-starter. Same for nicotine being addictive. In fact all of medicine is just smoke and mirrors. Computer science is clearly the right place for you. Mazel tov!

Bruce