Chronology Current Month Current Thread Current Date [Year List] [Month List (current year)] [Date Index] [Thread Index] [Thread Prev] [Thread Next] [Date Prev] [Date Next]

# Re: [Phys-L] modeling is easier than predicting

• From: John Denker <jsd@av8n.com>
• Date: Tue, 7 Apr 2020 02:03:07 -0700

Previously I wrote "modeling is hard".

I have not changed my views on that, not at all, but I am gradually
improving the way I express myself.

Models, in this context, are hypothetical. That is, they tell you
what *would* happen under such-and-such hypothetical conditions.
This may take the form of an if-then statement. This is conceptually
the same as a conditional probability, i.e. the probability of
outcome (a) *given* conditions (b). This is written P(a | b).
It does *not* tell you the marginal probability that condition (b)
will occur; you have to figure that out separately.

Sometimes the entire point of a model is to make sure the condition
does not occur. Consider for example a model that tells the driver:
"If you don't stop and buy gas, you'll run out before you get home."
The whole point is to motivate the driver to buy gas, so the premise
of the model does not occur. Even so, the model was *not* wrong and
the modeling exercise was *not* a waste of effort.

The mirror image phenomenon also occurs: If the hypothetical outcome
(a) is sufficiently desirable, a prediction of the form "b implies a"
may motivate people to arrange things to make the premise (b) more
likely.

However, this is not guaranteed. Just because something is desirable
and "could" happen does not mean it /will/ happen. If you build a
better mousetrap, the world will not beat a path to your door. (The
proverb on this subject is quite wrong.)

With all due respect, MLK was wrong when he said the arc of the
moral universe bends toward justice. The arc has neither momentum
nor agency; it never does anything of its own accord. If we want
it to bend, we have to grab it and *make* it bend.

As for pandemic models in particular, there are a number of models
that tell us what conditions produce favorable results. The South
Korea example. The Iceland example. The IHME models. And so on.
I do not have any objection to these models. I must however object
when the highly-conditional model is treated as an unconditional
prediction. The model gives us the conditional probability
P(outcome | conditions)
but does not tell us the marginal probability
P(conditions)
so we have to figure that out separately before we can make a
prediction:
P(outcome) = P(outcome | conditions) ⋅ P(conditions)

Politics is partly about goals, but it is also very much about
pathways. You have to build a path to the goal, starting from
where we are now. Announcing a desirable goal does not mean
people will automatically do what is needed to obtain that goal.

It is a tremendous fallacy to imagine that people will act
rationally. If people were even halfway rational, we would
not be where we are now.

It is also a mistake to model each state's long-term pandemic
response on a state-by-state basis, as if it were independent
of the other states. Short term maybe; long term no. For
example, suppose California stops the spread within its own
borders. This is doable, indeed likely, in the short term.
*) Longer term, things get ugly:
-- California could try to maintain the lockdown long-term,
but this would be economically unsustainable.
-- It could lift the lockdown, and suffer re-infection coming
in from other states, which vitiates the short-term success.
https://medium.com/@tomaspueyo/coronavirus-the-hammer-and-the-dance-be9337092b56
-- It could hope that other states get their act together.
Alas, as the saying goes, hope is not a strategy, much
less a pathway.
-- It could implement measures that control the disease
but are not as crude, clumsy, and costly as a lockdown.
Such measures require a multi-step process. So far
*none* of the steps have been implemented properly in
the US.
*) Also, California cannot recover economically while the
other states are in meltdown. So once again it needs to
hope that the other states get their act together.

========

To repeat: There absolutely are models P(a | b) where the
California outcome (a) is relatively reasonable. There is
nothing wrong with these models. You do not need to explain
models to me.

My point is that all such models remain highly conditional.
A model is not a prediction.

My question is whether the conditions (b) that lead to the
preferred outcome can be obtained and sustained beyond the
short term.

I don't have an answer. I suspect any answer would be more
politics than physics, and therefore outside the charter of
this list, so perhaps that should be discussed in some other
forum.