122
S T A R M E R
treated in the way in which probabilities are handled in the standard theory and
enter “raw” with
u
(
x
i
)
5
x
i
for all
i
.
Edwards called this
subjective expected
value
, and in the version presented by Jagdish Handa (1977) the decision weight
attached to each outcome is determined by a
probability weighting function
“(
p
i
),
which transforms the individual probabilities of each consequence directly into
weights. As in most theories that incorporate probability weights, “(
?
) is assumed
to be increasing with “(1)
5
1 and “(0)
5
0, and I will retain these assumptions
from now on. The subjective expected value form has not been widely used, but
theories that allow nonlinear transformations of both
probabilities and conse-
quences have received much more attention. In the simplest variant of this latter
type of model, individuals are assumed to maximize the function
V
(
q
)
5
• (
p
i
)
?
u
(
x
i
).
(6)
I will call this form
simple decision weighted utility
.
14
Both this and subjective ex-
pected value, because they transform the probabilities of individual consequences
directly
into weights (i.e.,
p
i
5
• (
p
i
)), have the property that
V
(
q
) will not gener-
ally satisfy monotonicity. To see this, suppose for the sake of example that • (
?
) is
convex, then • (
p
)
1
• (1
2
p
)
,
1 and there will be some •
.
0 such that gam-
bles of the form (
x
,
p
;
x
1
• , 1
2
p
) will be rejected in favor of (
x
, 1), even though
they stochastically dominate the sure option. A similar argument applies for
any
departure from linearity, and the only way to ensure general monotonicity in this
type of theory is to set decision weights equal to objective probabilities (i.e.,
p
i
5
• (
p
i
)
5
p
i
for all
i
), in which case the theory reduces to EU. This property
was first noted by Fishburn (1978) and since then has been widely viewed as a fa-
tal objection to models that attach decision weights to the raw probabilities of in-
dividual consequences. For example, Machina (1983, p. 97) argues that any such
theory will be, “in the author’s view at least, unacceptable as a descriptive or ana-
lytical model of behavior.” The point seems to have been generally accepted, and,
while many theorists have wished to retain the idea that probabilities may be sub-
jectively weighted, the thrust of work in this stream of the literature over the past
two decades has been toward variants of the decision-weighting form that satisfy
monotonicity.
There are two distinct strands to this contemporary literature: one conven-
tional, the other distinctly nonconventional. The nonconventional
route is that
taken by Kahneman and Tversky (1979) in
prospect theory
, but that model takes
us outside the bounds of conventional theory, and so I postpone further discussion
of it until the next section. Theorists following the conventional route have pro-
posed decision-weighting models with more sophisticated
probability transfor-
mations designed to ensure monotonicity of
V
(
?
). One of the best-known models
of this type is
rank-dependent expected-utility theory
, which was first proposed
by John Quiggin (1982). Machina (1994) describes the rank-dependent model as
14
This form has sometimes been called
subjective expected utility
, but this label is now more com-
monly used to refer to L. Savage’s (1954) formulation of EU.