428
The Nobel Prizes
precise probabilities. Indeed Knight (1921) posed a direct challenge to time se-
ries econometrics:
We live in a world full of contradiction and paradox, a fact of which
perhaps the most fundamental illustration is this: that the existence
of a problem of knowledge depends on the future being different than
the past, while the possibility of the solution of the problem depends
on the future being like the past.
While Knight’s comment goes to the heart of the problem, I believe the most
productive response is not to abandon models but to exercise caution in how we
use them. How might we make this more formal? I think we should use model
misspecification as a source of uncertainty. One approach that has been used
in econometric model-building is to let approximation errors be a source for
random disturbances to econometric relations. It is typically not apparent, how-
ever, where the explicit structure comes from when specifying such errors; nor
is it evident that substantively interesting misspecifications are captured by this
approach. Moreover, this approach is typically adopted for an outside modeler
but not for economic actors inside the model. I suspect that investors or entre-
preneurs inside the models we build also struggle to forecast the future.
My co-authors and I, along with many others, are reconsidering the concept
of uncertainty and exploring operational ways to broaden its meaning. Let me
begin by laying out some constructs that I find to be helpful in such a discussion.
When confronted with multiple models, I find it revealing to pose the result-
ing uncertainty as a two-stage lottery. For the purposes of my discussion, there
is no reason to distinguish unknown models from unknown parameters of a
given model. I will view each parameter configuration as a distinct model. Thus
a model, inclusive of its parameter values, assigns probabilities to all events or
outcomes within the model’s domain. The probabilities are often expressed by
shocks with known distributions and outcomes are functions of these shocks.
This assignment of probabilities is what I will call risk. By contrast there may be
many such potential models. Consider a two-stage lottery where in stage one we
select a model and in stage two we draw an outcome using the model probabili-
ties. Call stage one model ambiguity and stage two risk that is internal to a model.
To confront model ambiguity, we may assign subjective probabilities across
models (including the unknown parameters). This gives us a way of averaging
model implications. This approach takes a two-stage lottery and reduces it to a
single lottery through subjective averaging. The probabilities assigned by each of
a family of models are averaged using the subjective probabilities. In a dynamic
setting in which information arrives over time, we update these probabilities
6490_Book.indb 428
11/4/14 2:30 PM
Uncertainty Outside and Inside Economic Models 429
using Bayes’ Rule. de Finetti (1937) and Savage (1954) advocate this use of sub-
jective probability. It leads to an elegant and often tractable way to proceed.
While both de Finetti (1937) and Savage (1954) gave elegant defenses for the use
of subjective probability, in fact they both expressed some skepticism or caution
in applications. For example, de Finetti (as quoted by Dempster (1975) based on
personal correspondence) wrote:
37
Subjectivists should feel obligated to recognize that any opinion
(so much more the initial one) is only vaguely acceptable . . . So
it is important not only to know the exact answer for an exactly
specified initial problem, but what happens changing in a reasonable
neighborhood the assumed initial opinion.
Segal (1990) suggested an alternative approach to decision theory that avoids
reducing a two-stage lottery into a single lottery. Preserving the two-stage struc-
ture opens the door to decision making in which the behavioral responses for
risk (stage two) are distinct from those for what I will call ambiguity (stage one).
The interplay between uncertainty and dynamics adds an additional degree of
complexity into this discussion, but let me abstract from that complexity tem-
porarily. Typically there is a recursive counterpart to this construction that in-
corporates dynamics and respects the abstraction that I have just described. It
is the first stage of this lottery that will be the focus of much of the following
discussion.
6.1 Robust Prior analysis and ambiguity aversion
One possible source of ambiguity, in contrast to risk, is in how to assign sub-
jective probabilities across the array of models. Modern decision theory gives
alternative ways to confront this ambiguity from the first stage in ways that are
tractable. Given my desire to use formal mathematical models, it is important to
have conceptually appealing and tractable ways to represent preferences in envi-
ronments with uncertainty. Such tools are provided by decision theory. Some of
the literature features axiomatic development that explores the question of what
is a “rational” response to uncertainty.
The de Finetti quote suggests the need for a prior sensitivity analysis. When
there is a reference to a decision problem, an analysis with multiple priors can
37
Similarly, Savage (1954) wrote: “No matter how neat modern operational definitions of
personal probability may look, it is usually possible to determine the personal probabili-
ties of events only very crudely.” See Berger (1984) for further discussion.
6490_Book.indb 429
11/4/14 2:30 PM