[Martin Taylor 940929 10:30]
Tom Bourbon [940927.1542]
Rick Marken (940926.1545)
Yes. No question. We implicitly assume that understanding the fact that
people control their own perceptual experience is important; we think that
the results of statictical studies are not.
Agreed, Rick. There is nothing _implicit_ about our focus on the phenomenon
of control; we make it as explicit as we can. When we identify an
occurrence of control, we are equally explicit in our attempts to model it
with PCT. All of that being so, it follows directly that data gathered in
traditional statistical studies and in conventional tests of the null
hypothesis are of little use to us, hence they are of little interest to us.
There are very good reasons for rejecting "conventional tests of the null
hypothesis." Long before I had heard of PCT, I was promoting one criterion
for judging a psychological paper: if it mentioned (p<.???) I would be
very wary. The ONLY thing such a test can tell you is whether your
experiment is sensitive enough to detect a difference that most assuredly
exists. The question one should ask (staying within the S-R mode of thought
of the authors) is not "is there an effect" but "what is the most probable
size of the effect, how sure am I of the size of the effect, and is an
effect of that size interesting?"
A more valid complaint, which separates PCT-specific studies from others,
is the search for the controlled perception. However, good, reliable
and (I think) useful data can be obtained on the components of control
loops, without studying the whole control loop in action. When an
electronic engineer wants to examine an unknown control system, one
way is to break the feedback connection and examine the S-R behaviour
of the broken loop. Likewise, I think, with living control systems, one
should usefully be able to examine the S-R action of a system and draw
some conclusions about the behaviour of a control loop that incorporates
the tested components.
When I mentioned Rick's old comment of "good data, bad theory" last week,
a comment that he then disavowed, I was thinking of some early 1960's
studies on the perception of ambiguous figures of several different kinds.
One kind, for example, occurs when you ask the subject to listen to a
short tape loop containing some phrase like "splendid gladiola" (one of
the ones we used). If you get the subject to report what is heard, it
changes erratically over time. Now, when you plot the number of changes
against the number of different things heard, you get a very precise
relationship: the number of changes is proportional to n(n-1), where
n is the number of types. Often, this relationship is so close that
you need a big graph in order to plot the points as visibly different
from the curve. And the same relationship holds for visual figures like
the Necker Cube (if people haven't been trained to believe that only two
percepts are possible), to patterns of visual movement, and I forget
what else. These are powerful and consistent effects.
Another related study used a visual pattern that most people see in only
two forms (and believe me, such images are VERY hard to find). It consisted
of a (real, not photographed) side-lit piece of plasticene whose flat
surface had been dented many times with a ping-pong ball. When you look
at it, you see either a dented or a bubbly surface, and the two possibilities
flip very precisely at uncontrollable moments (at least, some people say
that they can control the flip for a few moments, but nobody has claimed
to manage to do so for long). When you analyze the time between flips,
you get what is called a "survival curve," which depicts the probability
that a flip has not happened, given that it is n msec since any given flip.
The survival curves are often different for the bubbly and for the dented
surface, and have a shape that cannot be accounted for by any simple theory
of "fatigue" (a dormitive principle if ever there was one!). Also, the
shapes of the survival curves change rather abruptly and dramatically from
time to time. We found that the survival curves and their dramatic changes
could be fitted quite accurately (without statistics) by a random walk
model, which said that there are exactly K perceiving units, each of which
independently decides whether the image is of bumps or of dents, and a
single decision unit that provides the overt perception of bubbles or
dents based on a biased majority decision with history. In other words,
if more than B units said "bubble," that's what the perception was, and
if less than D units (D<B) said "bubble," dents would be seen, and if the
number of units was between D and B, what was seen was whatever had previously
been seen. The number K turned out to be something like 28 for one subject,
and something like 32 or 33 (it changed during the week of experimentation)
for another. The abrupt changes in the survivorship curve could always
be accounted for by a shift of exactly 1 unit in either B or D (and, I
think once both of them).
These results seemed both rather precise and hard to "model" in other ways we
could think of. We did not do any statistical tests, largely because we
couldn't think of any valid tests to do! But we felt then, and feel now,
that the data were sufficiently bizarre and the model sufficiently simple
that I am prepared to bet that the two subjects did indeed have 28 and 32
or 33 units working on that perception.
Another related "fact(?)" is that in a large variety of different perceptual
phenomena involving prolonged observation, a major index of performance
changes linearly with the square root of the time since the start of the
observation. Again, no statistics, but the trend lines are awfully close
to a slope of 0.5 on a log-log plot. Some of these involve data combined
over subjects, some don't.
The point I am making is that both in principle and in practice, not
all research contributing to the understanding of perceptual control systems
need involve studying the control system in action. That has, of course,
to be the final arbiter of what people "do," but to study the components
of control systems in isolation is also valid and useful.
Complaints against statistics should be directed against improper uses of
statistics (e.g. any "significance test," or a generalization from a trend
found in a set of everages to an implied average of a set of trends).
Martin