[From Rick Marken (2008.12.25.1200)]
Martin Taylor (2008.12.24.16.57)
Rick Marken (2008.12.24.0940)--
However, if you accept that I do have a
pretty good understanding of PCT, you might ask yourself why I think some
studies done in the "conventional" way could have something to tell us about
the way people might work
I think you have not yet grasped (or you just don't buy) the
implications of PCT for conventional experimental psychology, as
articulated (and demonstrated) so clearly in Bill's 1978 "Spadework"
paper in _Psychological Review_.
I thought you took r2 as the square of the correlation (which is, of course
1.0). Since you don't mean this, I'm not sure what you do mean.
The correlation in this study is not 1.0. Your way of computing the
correlation in this study would lead to the conclusion that the
correlation (and r2) for every 2 level experiment is 1.0. You are
saying that the correlation is 1.0 because there are two X values (the
orderly and disorderly conditions) associated with two Y values (the
proportion littering in each condition). The correlation between these
variables is indeed 1.0, as it would be in any two level experiment
where the X variable is the two conditions (X1, X2) and the Y variable
is the average response in each condition (Y1, Y2). The correct way
to calculate the correlation (and r2) for this experiment is to look
at the relationship between the X and Y values for each subject. In
this case, each subject's X value is the condition they are in (0 =
orderly and 1 = disorderly) and their Y value is their behavior (0 =
no litter, 1 = litter). Since the variable values are binary we would
compute a point biserial correlation, which gives a result that is
very close to r. Since I didn't know the number of subjects, only the
proportion who littered in each condition, I assumed that an equal
number of subjects was in each condition and calculated the
correlation assuming a total of 20 or 100 subjects with the
appropriate proportions of litterers and non-litterers in each
condition. It turns out that the correlation value was nearly the same
despite the change in total number of subjects -- ~.33. So the r2 was
.11.
I would say that the assumption
is not that behaviour is probabilistic (though some practitioners may make
that assumption). It is that the behaviour observed is influenced by factors
the experimenter cannot determine, and that therefore the observations (not
the behaviour) must be treated as probabilistic.
I know that. But scientific psychologists have been saying this for
nearly 100 years and still they are only able to account for, on
average, about 30% of the variance in behavior in any particular
experiment. The factors that are causing the other 70% of the variance
should have been discovered by now
Here, I think, is where we diverge, not conceptually, but numerically. You
talk about THE controlled variable
I'm very well aware of the fact that people are assumed to control
many more than one CV at a time, according to PCT. My point was that
in conventional psychology researchers have never been able to explain
much more than 30% of the variance in the behavior they study,
indicating that the entire enterprise, which is based on the general
linear model of behavior, is a failure. And that is because research
based on the GLM completely misses the existence of controlled
variable. The fact that you believe that much of the variance in
behavior can be explained by identifying controlled variables doesn't
make the conventional approach suddenly informative.
There are other reasons for probabilistic output, which should be apparent
to you: "many means to the same end" might trigger one line of thought.
I would rather say it this way: according to PCT there are several
reasons why we would expect the _appearance_ of behavior in
conventional experiments to be probabilistic. I take this to mean that
the whole conventional research enterprise in psychology is based on
an illusion. You seem to take it to mean that we can now learn things
about individual behavior from conventional research that we could not
before PCT. I think we disagree on this point (the merits of
conventional research) even more deeply than we disagree about the
merits of information theory.
Another thought, again based on not expecting to be able to do The Test in
such a way as to be able to tease out all the controlled variables:
You don't have to be able to test for all controlled variables a
person is controlling in order for the Test to be successful. The goal
of the Test is to figure out what variable(s) a person is controlling
when carrying out a particular behavior. I presented my research on
catching baseballs to show how this would work in a realistic
(non-tracking) behavioral situation. People control a lot more than 2
variables when they catch balls. But once you know that they are
controlling vertical optical velocity and lateral optical displacement
you can understand about 98% of the variance in their behavior (their
movements on the field). Conventional research could never get you
there, as I explained in my 2005 JEP:HPP comment (Marken, R. S. (2005)
Optical Trajectories and the Informational Basis of Fly Ball Catching,
Journal of Experimental Psychology: Human Perception & Performance, 31
(3), 630 - 634).
No causation is suggested or inferred, though some
possible causative possibilities are dismissed as implausible.
Then why do it as an experiment?
My guess is that they wanted to see whether a long-time speculation would
have the predicted influence on the behaviour of people. Various cities have
followed New York's example of going after minor issues such as graffiti and
littering, in the hope that some major problems might be reduced if people
behaved differently when the environment was tidy, but (the authors say)
nobody had done an experiment to see whether such an influence might exist.
"Influence" is not "causation" in my language, and they were looking to see
whether there was an influence. There is.
I agree with this if we are dealing with this research as sociology.
At the population level fewer people litter in an orderly than in a
disorderly environment. The fact of the matter, however, is that this
exact kind of research is used by psychologists as a way of
determining the causes of individual behavior. They do this in
multi-subject experiments just like this where the goal is to
determine whether the IV has an effect on the DV. When they do this
they are testing the general linear model, which is simply that DV =
a+b1*IV1+b2*IV2...+bn*IVn+e for each individual subject. They test
several subjects and look at the relationship between the IVs and the
average DV under the assumption that e (error) is random and will,
thus, be canceled out by the averaging. So the average results for
many subjects is supposed to show something close to the "true" causal
relationship between IV and DV for each individual. Causality can
presumably be inferred if all variables other than the IVs have been
controlled (held constant at all levels of the IVs).
Since the research on the effect of the environment on littering was
done by psychologists I presume the main aim was to tell us something
about the cause of littering in individual people. And you have been
trying to tell me that this research should be taken as useful by
those of us trying to understand individual behavior. So I assume that
you yourself think that this research is telling us something about
individuals. I think Bill's point (and mine) is that it doesn't. But
both of us have said that this research could certainly be useful to
policymakers (like city mayors) who deal with populations. It looks
like littering goes down in orderly environments. And this is
presumably a "causal" relationship in the sense that, if properly
done, this research rules out other possibly confounding variables
(like the type of people in the two environments) as explanations of
the result. So the mayor can be relatively confident that littering
will be reduced if there is investment in increasing the orderliness
of neighborhoods.
The whole idea of doing an experiment
in the context of the GLM is so that one can infer that any observed
>relationship between the IV and DV is causal.
I might ask on what PCT foundation you infer that this is a controlled
variable in all "conventional" psychological experimenters?
PCT has nothing to do with it. Read any text on Research Methods in
Psychology (mine is particularly good but I hear Bruce Abbott has one
that sells well;-)) and you will see that the reason for doing an
experiment (manipulating an IV under controlled conditions and
measuring the DV in each condition) is to determine a causal
relationship between IV and DV. Since all experiments in psychology
are analyzed used ANOVA or t test then they are all being done in the
context of the GLM (see equation above).
Incidentally, I see in this study no dependence on a model of any kind,
linear or nonlinear.
The model on which this experiment is based is the GLM: DV = f(IV)
The analysis, it is true, used a chi-square, but that
is necessary only to satisfy journal editors. The IOT (InterOcular
Traumatic) test suffices in this case.
What you are testing with the IOT is the match of the observed results
to what you would expect if DV = f(IV), that is, what you would expect
if the GLM were true.
The reason I brought this paper into the discussion is that it seemed to me
to be an example of a statistical result that seemed to be quite clear
It's not the "clarity" of the result that is at issue; it's what the
results say about human nature. The results do clearly show that there
is less littering in orderly as compared to disorderly environments.
What they don't show is that the orderliness of the environment has
any effect (or influence) on the behavior of individuals. Bill makes
this point by showing that, based on these results, you would be wrong
more often than you are right in predicting an individual's littering
behavior based on the orderliness of the environment in which the
individual is found. I make the point by showing that, as a model of
individual behavior, the GLM, which says that littering =
f(orderliness of environment) accounts for only about 11% of the
variance in the littering behavior of individuals, which means that
nearly 80% the variation across individuals in terms of their
littering behavior (given the GLM model) is in the e (error) term of
the GLM rather than the IV. This shows either that littering is mainly
a function of other variables besides the orderliness of the
environment (this is the preferred "other variables" explanation of
the noisy results of psychological experiments) or that the GLM is
wrong (my preferred explanation since there is simply no case I know
of where psychologists have been able to add other variables to the
GLM prediction and increase the amount of behavioral variance
accounted for much more than a few percent, like going from 30% to
33%).
No. I disagree. What I agree about (and I said this in my initial posting)
is that you can't use the results to say much about any individual
Good. Then you agree that it's useless to a control theorist
interested in understanding individual behavior. It's good information
for the mayor, though.
You can, however, use the results as a
starting point to look for what may be a quite generic effect of an
apparently irrelevant environmental variable when different control systems
in an individual are in conflict.
I agree that it can be a starting point for control research; but you
could have gotten that starting point just from the title of the
research; you could start testing to see whether a person is
controlling for keeping certain aspects of the environment at the
status quo or tryng to make it cleaner, etc. But the data from this
study is basically irrelevant to a control theorist.
Best regards
Rick
···
--
Richard S. Marken PhD
rsmarken@gmail.com
