Eric -
Interesting paper, but I'm not sure if I follow. The basic argument
seems to be that we often explain things by imagining (with the help
of statistics) hypothetical constructs that cannot be directly
measured. As those constructs can't be measured directly, they don't
help us predict things. Thus, predictive models are limited to using
things that actually exist, while explanatory models are not so limited.
I think I have been considering it from the opposite perspective... that
predictive models aren't *necessarily* explanatory... I think that
*explanatory* models are at least minimally *predictive* (how else can
they be validated)? They may not be sufficient to predict (m)any of the
things we are interested in, but I think (though I'm not sure exactly
why) the are in fact predictive (by definition?).
That seems like a really good argument for coming up with better
explanations, not an argument for distinguishing and reifying two
distinct modeling tasks.
I think Science is always seeking models that have more explanatory
power. The business of Science is *understanding* (my contention) with
*prediction* being merely a useful mechanism for hypothesis testing and
even generation. Most of us get excited when Science *predicts*
something, but I am not sure it is an end in itself (except for
engineering, technology development, business purposes).
I think Bruce's description of Mendeleev's Periodic Chart is my easiest
example... the "model" in this case was an almost numerological
geometric arrangement of the known elements based on correlations among
their properties without any *explanation" of the underlying reasons or
causes for those properties. It was very effective for predicting
elements yet to be discovered (recognized, identified?) as well as
properties of known elements not yet explored, and thereby helped to
structure the *search* for new elements.
This is a topic I am quite interested in. I would presume that an
ideal explanatory model would be identical to an ideal predictive
model, though I grant that non-ideal cases might differ. What am I
missing?
I think you are correct that an "ideal" model has strong predictive and
explanatory properties.
For Engineers perhaps, predictive models are sufficient, they may not be
(very?) interested in explaining *why* a particular material has the
properties it does, merely *what* those properties are and how reliable
the properties might be under a variety of conditions.
I'm sure there are others here with good perspective on this question...
Bruce?
- Steve
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
