[From Bill Powers (930326.0830)]
Greg Williams (930325.0850) --
Back on the net in fine fettle, I see.
It is arrogance to suppose that one's own goals are the only,
or even the most, important ones.
Oh, heck. Well, is it arrogance to suppose that one's own
definition of arrogance is the only, or even the most important
one? What is the most important goal, then? And where can I look
out there to find it?
Just try selling an insurance exec on the method of
modeling -- he/she does fine with statistical descriptions of
populations and doesn't care about predicting individual
behaviors.
Seems to me that I've commented a number of times on the fact
that people who deal with populations do fine (for themselves)
with statistics, because statistics concerns the properties of
whole populations, not individuals. When that's your concern, who
cares if a few of the ants get stepped on?
Insurance executives do use a model to predict behavior. It's
called the actuarial tables. It predicts the behavior of
populations.
In fact, the method of modeling individuals would be unwieldy
here, to say the least, not to mention the fact that it is
still in its infancy and cannot be applied to "high-level"
behavior now (or, I predict, for decades to come).
If enough people don't adopt the method of modeling and try to
apply it to understanding human nature, you can change "decades"
to "centuries."
In the 19 March _Science_, p.1773, there's an interesting review
of _Testing testing_ by F. Allan Hanson. "Hanson goes on to argue
that 'institutional analysis' reveals that history can be 'read'
as the gradual perfection of the manipulation and subordination
of human beings to serve institutional ends." One of the primary
tools for achieving this end is statistical analysis of
population tests.
As Phil Runkel pointed out in his book, there is a lot which
can be accomplished by the "Grand Method" (his term) of
descriptive statistics of populations.
I agree. I don't agree with much of what it is used to
accomplish. And Phil agrees that casting nets is no way to
understand specimens.
I would suggest that there are some human goals which can be
met more efficiently by using descriptive statistics than by
using the (again, "Grand Method") of individual models.
Whose goals are you talking about? If you own an insurance
company, it's most efficient to use statistics to set the rates,
to raise the rates on people who make claims, and to reject
applicants who belong to high-risk populations. That's the most
efficient way to satisfy your personal human goal of making money
out of selling insurance to a population.
But I believe that statistics of populations IS a kind of
knowledge..
Yes, it is. It is knowledge about populations.
It is not as detailed a knowledge as models of the population
individuals ...
It is not detailed at all. It is a mass measure, applying only to
the entire mass.
Surely you don't expect epidemiologists sometime in the
(far?!?!) future to replace statistical description with
individual models.
It might help them if they could say why person A catches the
disease while persons B ... Z do not. Identifying individual
carriers has occasionally been important.
Some physicists still work with thermodynamics and statistical
mechanics ...
Yes, and who cares if a molecule or two drops through the cracks?
No problem here.
When will your modeling and associated methodolgy be
sufficiently sophisticated to predict when I will COOPERATE and
when I won't?
Probably never. Predicting behavior is not what PCT is about. PCT
is about understanding behavior, and what it is being used to
control. If I ask you to do a tracking task and you do something
else, it will be clear that your goal is not cooperation, but
something else. I might be able to find out what the something
else is by interacting with you long enough. I'll never find out
by studying "people like you" who show "behavior like yours" in
"circumstances like these."
And when PCT has leapt that hurdle, the next one is to make
models for THOUSANDS of individuals and combine them some way
to predict population measures. Lots of luck.
Why would I want to do that? I'm not trying to make a living by
selling insurance or proving that I am -- on the average-- a
successful doctor or educator or politician. My interest is in
understanding the next individual I meet, by some means that
doesn't involve formalized prejudice.
No, it is better to follow the example of physics and stick
with descriptive statistics for generating SOME kinds of
knowledge.
Better for everyone? Is this one of those "more important human
goals?"
But I am not arrogant enough to think that modeling individuals
is the only path to knowledge. It IS the only path to SOME
kinds of knowledge. But some people don't need that kind of
knowledge.
Not for their professions, I agree. What about for getting along
with their families and friends and the salesperson and the
waitperson and themselves? It seems to me that by relying on
statistical generalization in such person-to-person
circumstances, people create more problems for each other than
they solve.
I think what you're saying basically makes sense: use statistics
for appropriate purposes, and models for other appropriate
purposes. Partly this is just a practical matter of what we
currently know how to do. Weather modeling doesn't work very
well, and perhaps can't, so weathermen also use statistical data.
Even when a model would in principle be the best tool, if you
don't have a model developed well enough to use you fall back to
relying on generalizing from experience just as people have
always done.
It's really a question of how well you need to understand things.
In principle the statistical approach, given huge populations and
easily-measured mass effects (as in thermodynamics) can yield
very precise predictions for the population, with small error
spreads. The laws relating gas pressure, volume, and temperature
do extremely well in predicting the behavior of large enough
packets of gas. They say nothing about the behavior of molecules,
of course. But the precision of the mass measures is quite good
enough to serve almost any practical purpose.
This is not true of statistical measures of human populations. I
don't know what the smallest packet of people is that would
provide an accurate characterization of a population, but it must
be at least as large as largest studies so far done. You probably
need hundreds of thousands, perhaps millions, of people as
subjects, and even then, human circumstances vary so much that in
a population of millions you may still have subpopulations that
are too small to provide accurate data.
What people really want is a way to predict the behavior of
others within the very small subpopulations of people with whom
they are likely to interact. With how many people will you have
an important face-to-face interaction during your lifetime? A few
hundred? A few thousand? That, for you, is the relevant
subpopulation. It is far too small to characterize reliably by
any statistical means. Whatever generalizations might be made
about the behavior of that number of people, you are likely to
find more negative than positive instances of it. To do better
than that, you need a model that says true things about every
single person you could possibly meet, or pretty close to it.
The PCT model is that kind of model. At present, it can say such
universally true things only in simple circumstances like the
rubber-band experiment. Even then, the kind of knowledge it gives
you isn't the sort that psychologists look for: what specific
actions people will perform under certain circumstances. The
knowledge that PCT gives is about relationships and processes,
not events. It's more like relaizing that the other person is
controlling the knot connecting the rubber bands. Knowing that,
you then know that the other person will do anything necessary
with that end of the rubber bands, in order to keep the knot
where that person wants it. When you have figured out where the
person wants the knot, or how the person wants the knot to
behave, you can then do some very accurate predicting over a
limited time span, and you can predict the result of doing things
to your end of the rubber band that have never occurred before.
This has nothing to do with what any other person might do:
you're talking strictly about this person.
You might say that where we can't use the PCT model, why not use
the statistical approach, because it's all we have? I've heard
this a lot: what else can we do? My answer is always the same:
nothing. You just do what you do, and get the results you get, as
has always been the case. This doesn't make the results any
better than they have ever been. Where you can't figure out how
to use the PCT model, you're stuck with life as it was before
PCT. That's doesn't mean you have to be happy with it.
ยทยทยท
-------------------------------------------------------------
Best,
Bill