Gianna Lotito, DiGSPES, University of Eastern Piedmont, Alessandria, Italy

Anna Maffioletti, ESOMAS, University of Turin, Turin, Italy

Choice under risk and uncertainty

Definition
This entry outlines what is meant by decision making under risk and uncertainty. It illustrates
the model of Expected Utility, its properties and the Allais paradox as the main violation of the
model. It describes the Subjective Expected Utility model of decision under uncertainty, and the
Ellsberg paradox as an example of the Knight’s approach to uncertainty.

Introduction
In real economic life, many decisions are taken under risk and uncertainty, for example
investment decisions, decisions about consumption through time, buying and selling insurance,
investment in new industries and countries, choosing new technologies, stock market purchases
and sales.
The literature on decision making under risk and uncertainty can be divided in: (1) the
literature concerning decision making under risk, which includes: (a) the Expected Utility model
(EU) and its axiomatizations (Bernoulli 1954; von Neumann and Morgenstern 1947); (b) the
criticisms to the EU model (Allais 1953) and its alternatives (Kahneman and Tversky 1979;
Quiggin 1982, 1993; Loomes and Sugden 1982); (2) the literature concerning decision making
under uncertainty, which can be divided in two different approaches: (a) the Bayesian approach
(De Finetti 1937; Ramsey 1931; Savage 1954), according to which people assign subjective
probabilities to uncertain events and these probabilities follow the rules of mathematical
probability theory and (b) the approach by Knight (1921) and Keynes (1921) according to which
people cannot assign subjective probabilities to uncertain events that follow the mathematical
rules of probability theory (Ellsberg 1961; Schmeidler 1989; Tversky and Kahneman 1992).
In a situation of risk one does not know with certainty the outcomes of one’s choice, and the
uncertainty related to a decision can be represented by an objective probability distribution over
the events or states of the world which relate actions to outcomes, as for example with a gamble
based on the roll of a die or a roulette wheel.
In a situation of uncertainty, the uncertainty relating a decision cannot be represented by an
objective probability distribution. In this case the individual is either considered able to attach a
subjective esteem of the probability to each event (Savage 1954) - in which case decision under
uncertainty reduces into decision under risk – or probabilities are not known and cannot be
assigned (Knight 1921).

Choice under risk

Let us define now a prospect or lottery like the combination of all the possible outcomes with
the probabilities of the events under which these outcomes occur.
Consider for example having a ticket for the following lottery L = ($100, ½;$0, ½) - where
the outcomes are monetary prizes - in which a coin is tossed: in the event the coin falls Head
(which occurs with probability ½), a prize of $100 is won; in the event the coin falls Tail (which
also occurs with probability ½), nothing is won.
Thus, an action in a risky situation corresponds to playing a lottery or prospect, which
associates each outcome to the probability of its occurrence. As a consequence, in such a
situation choice can be viewed as a choice of the preferred lottery or prospect. But how can the
individual value the different prospects? One possibility is that the individual calculates the
expected value (EV) of the different prospects, and chooses the prospect with the highest of
these values. The concept of expected value was first developed by seventeenth century
mathematicians, for example Pascal. The expected value of a lottery is the average of the
1
monetary prizes associated with the different outcomes, weighted by their respective
probabilities. The expected value of the above lottery L would be (½ * $100 + ½ * $0) = $50.
However, maximizing the expected monetary value does not always appear to be a satisfactory
criterion to choose among different risky situations. Let us consider the two lotteries L1 = ($60,
½;$0, ½) and L2 = ($40, ½;$20, ½). Despite both have the same expected value of $30, some
people might prefer L1 , where there always is a probability of gaining a positive outcome, to L 2,
where there is a 50% chance of winning nothing. As a further example consider the following
lotteries: L3 = ($100, ½;$-1, ½) and L4 = ($100000, ½;$-1000, ½). At a first sight, many people
would be willing to participate to the first lottery, but would they accept so easily to play the
second? It does not seem so: L4 has a 50% probability of winning an outcome 1000 times higher
than L3, but also a 50% probability of losing an outcome 1000 times higher. However, if one
calculates the expected value of the lotteries, it turns out that the EV of L 3 ($49) is much lower
than the EV of L4 ($49500). Clearly, expected value is not always a sufficient criterion to make
a lottery attractive. An alternative is needed.

History
The first important argument on the subject was that of the mathematician Daniel Bernoulli
(1738), who (independently from Gabriel Cramer) developed a new hypothesis based on the
solution of a problem posed by his cousin Nicholas Bernoulli, the ‘St Petersburg Paradox’.
Suppose you have to toss a coin till ‘Head’ comes out; the first time this happens, you stop
tossing the coin and prizes are determined in the following way: if ‘Head’ comes out at the first
toss, you earn $2; if ‘Head’ comes out at the second toss, you earn $2 2; if ‘Head’ comes out at
the third toss, you earn $23 and so on. The probability that at every toss ‘Head’ comes out is
equal to ½, and tosses are independent from each other: then, the probability that ‘Head’ comes
out at the first toss is ½, the probability that ‘Head’ comes out at the second toss is (½) 2 (that is,
the probability that ‘Tail’ comes out at the first toss times the probability that ‘Head’ comes out
at the second toss, that is, ½ * ½), the probability that ‘Head’ comes out at the third toss is (½) 3,
and so on. Thus, the expected value of this lottery is infinite: 2 (½) + 4 (½)2 + 8 (1/8) +…..+2n
(½)n + …= 1 +1 +1 +1 +1+… = ∞, as the coin is thrown if necessary an infinite number of times
and the expected prize from each toss is equal to 1. An individual who looks at the highest
expected value would be willing to pay an infinite amount of money to play this lottery. But this
is very unrealistic: nobody would be willing to pay more than a modest amount for it. In fact,
this is a very risky lottery: it gives the opportunity to gain an increasing big prize with a
decreasing small probability - people would not be willing to undertake this risk.
Bernoulli argued that this and similar choices could be explained by assuming that
individuals do not choose the lottery with the highest expected value, but that with the highest
expected utility, where the utility is represented by a function like the square root of wealth,
where utility increases with wealth, but at a decreasing rate. Bernoulli introduced both the
concepts of expected utility maximization and of decreasing marginal utility of wealth.
However, Bernoulli’s idea of expected utility maximization had little effect on the theory of
decision making under risk till the work ‘Theory of Games and Economic Behaviour’ by von
Neumann and Morgenstern was published in 1947. After that, it had soon to become the
normative and positive theory of decision making under risk.

The Expected Utility model

The expected utility model relies on the hypothesis that the individual possesses – or acts as
if he possesses a von Neumann-Morgenstern utility function over a set of outcomes, and when
he faces alternative lotteries over these outcomes he chooses that lottery which maximizes the
expected value of this utility. In particular, von Neumann and Morgenstern start by assuming
specific conditions on the preference relations between lotteries; these are necessary and
sufficient to show the existence of a utility function, that assigns a numerical value to the
2
‘satisfaction’ of the different outcomes of the lotteries. The individual thus chooses the lottery
for which the expected utility is the highest, where the expected utility of a lottery L = (x1, p1;
…; xn, pn) is given by the sum of the products of the probability and utility over all possible
outcomes, that is,
i = 1,…, n.
Given pi the probability of the outcome xi and U(xi) its utility, the expected utility of a lottery
L = (x1, p1; x2, p2) will be equal to EU(L) = p1U(x1) + p2U(x2). The expected utility of a lottery is
then the expected value of the utilities of the possible outcomes. As an example consider the
two lotteries L1 and L2 illustrated above. Assume that the utility function is of the form U(x) =
√x, where x is the monetary value of the outcome. The expected utility of L 1 is then equal to
EU(L1) = ½ * √60 = $3.87 and EU(L2) = ½ * √40 + ½ * √20 = $3.16 + $2.23 = $5.39.
We can obtain a very important information by observing the expected utilities of these
lotteries. If we compare them with the expected value (equal to $30 for both), we notice that the
expected utilities are not only different from the expected value, but are different from each
other. In particular, the expected utility of the ‘riskier’ lottery L1 (the one which gives a 50%
chance of winning nothing) is lower than that of L2.
The von Neumann-Morgenstern utility function gives us information about the individual’s
attitudes towards risk.

Attitudes towards risk

The possible attitudes to risk of an individual can be represented graphically as follows.
Consider Figure 1, with prizes in monetary value (wealth) on the horizontal axis and their utility
on the vertical axis. The straight line represents the von Neumann-Morgenstern utility function
of the individual and tells us the utility he or she assigns to a given sum of money. In this case in
which the shape of the utility function is a straight line, the individual is neutral to risk: when
facing different lotteries he or she chooses on the basis of their expected monetary value.
Consider the two lotteries L1= x for certain and L2= (x1, ½; x2, ½), where outcome x is also equal
to the expected value of L2. L2 is riskier than L1 , as its prizes have a higher variability. The risk-
neutral individual will not take into account this risk and will value the lotteries only on the
basis of their expected value: he or she will be indifferent between them. In Figure 1, point a
represents the utility of the lowest outcome x1, U(x1), and point b the utility of the highest
outcome x2, U(x2); point c is, therefore, the expected utility of lottery L2, EU(L2) = (½ * U(x1) +
½ * U(x2), and is exactly midway between a and b. The expected utility of the sure outcome x,
U(x) is given by point c. Then, an individual with such a utility function will be indifferent
between having x for certain or a lottery with expected value equal to x: they have the same
expected utility. The fact that L2 is riskier than L1 does not influence choice. For this individual
thus the expected utility of a lottery is equal to the utility of a sure outcome, that in turn is equal
to the expected value of the lottery
Consider now Figure 2 which represents an attitude of aversion to risk: the utility function
represented here is a curve, where the utility increases with wealth, but at a decreasing rate – an
individual with this utility function will not be indifferent between a sure outcome and a lottery.
Suppose the individual is facing the same alternatives as before. In the graph, point a represents
U(x1) and point b U(x2); the expected utility of lottery L2 will be the weighted average of these
two utilities and will be given by point c, in the middle of the segment unifying a and b, ½ U(x1)
+ ½ U(x2). However, we see now that the expected utility of the sure prize x is given by point d,
which is above c: U(x) > ½ U(x1) + ½ U(x2). For this individual thus the expected utility of a
lottery is lower than the utility of a sure outcome, that is equal to the expected value of the
lottery. The individual who is averse to risk will choose the sure alternative to a lottery which

3
gives an expected prize equal to the sure one – he or she prefers the expected value of the lottery
to the lottery itself. The individual is averse to the risk connected to the lottery.
We should note that this risk aversion is implicit in the shape of the utility function, which is
concave. As an illustration of this, consider lotteries ($60, ½;$0, ½) and ($40, ½;$20, ½) above.
In calculating their expected utilities, we have assumed a form for the utility function equal to
U(x) = √x, which has a shape like in Figure 2. This is a utility function implying risk aversion,
as shown by the fact that when calculating the expected utilities of the above lotteries, we find
that the riskier lottery has a lower expected utility than the ‘safer’ lottery.
The case of the utility function of an individual with a proneness to risk is given in Figure 3.
It can be seen here that for such an individual the expected utility of a lottery is higher than the
utility of a sure outcome, that is equal to the expected value of the lottery. The individual who is
risk loving will choose the lottery which gives an expected prize equal to the sure outcome
instead of the sure alternative – he or she prefers the lottery to the expected value of the lottery
itself. This individual loves the risk connected to the lottery.
The individual preferences towards risk can also be expressed using the concepts of (1)
certainty equivalent (CE) and (2) risk premium .
(1) As an example, consider the lottery L = ($100, ½;$0, ½). The expected value of this
lottery is 50. Consider asking this individual what is the amount of money which makes her
indifferent to the lottery. This is defined as the certainty equivalent of the lottery. Therefore, the
utility of the CE of the lottery is defined as equal to its expected utility - U(CE) = EU(L).
Suppose the individual is indifferent between the lottery and $40 for certain, that is, his CE
for the lottery is equal to 40, less than the expected value of the lottery. It follows that the
subject would accept any sum > 40 instead of the lottery and the lottery to each sum < 40. In
this case we say that the subject is averse to risk - U(CE) < EV(L). This is represented in Figure
4. The reverse inequality defines risk proneness, and the equality defines risk neutrality.
(2) The risk premium  is defined as the maximum part of the expected value that the subject
is willing to give up to avoid the risk associated to the lottery,  = EV(L) – CE (see Figure 4).
Consider the example of a CE for the above lottery equal to 40; this means that the individual
is willing to give up at most $10 of the expected monetary value of the lottery, that is, the risk
premium is 10. The sign of the risk premium gives the attitude to risk for that lottery. In general,
when for any lottery  is positive, the individual is averse to risk; when  is negative, he is risk
lover; when  = 0 the individual is risk neutral.
The most common analytical measures of risk aversion have been introduced by Arrow
(1964) and Pratt (1964) and extensively used in finance, insurance markets, health, game theory.

The Allais Paradox

The most important implication of the specific form of the expected utility preference
function is linearity in the probabilities (Marschak 1950; Machina 1987;
Raiffa 1968). For this reason, most of the empirical investigation of the expected utility
hypothesis has focused on this property (MacCrimmon and Larsson 1979; Allais and Hagen
1979; Starmer and Sugden 1989; Starmer 2000), revealing widespread systematic violations.
The best-known violation of linearity in the probability is the Allais paradox (Allais 1953).
Consider that an individual is asked to make a pairwise choice between the two following pairs
of lotteries (outcomes in French francs as in Allais 1953, page 527):

4
• A 100% chance of 100 millions)
or
• B 10% chance of $500 millions
89% chance of $1 million
1% chance of 0
• and
• C 11% chance of $1 million
89% chance of 0
• or
D 10% chance of 500 millions
90% chance of 0

If the individual has expected utility preferences, a choice of A over B in the first pair
implies a choice of C over D in the second pair. However empirical findings show that the
modal choice is for A over B in the first couple of lotteries, but D over C in the second couple.
This pattern of choice violates expected utility. In fact, A preferred to B implies U(100 millions)
> .10U(500 millions) + .89 U(1 million) + .01U(0), that is, .11U(100 millions) + .89U(0)
> .10U(500 millions) + .90U(0), while D preferred to C implies .10U(500 millions) + .90U(0)
> .11U(1 million) + .89U(0), which is a contradiction.
The widespread and systematic empirical violations of the linearity property has put under
discussion the descriptive validity of the theory, giving rise to a growing body of literature of
new theoretical models of choice under risk as, for instance, Prospect Theory (Kahneman and
Tversky 1979; Tversky and Kahneman 1992), Rank Dependent Utility Theory (Quiggin 1982,
1993), Regret Theory (Loomes and Sugden 1982).

Choice under uncertainty

The first distinction between choice under risk and choice under uncertainty was made by
Knight (1921) and by Keynes (1921). They define a choice under uncertainty when we deal
with a choice in which the probability of the occurrence of an outcome is not known. Savage
(1954) extended the Theory of Expected Utility to uncertainty and called it Subjected Expected
Utility. According to Savage all individuals can have a subjective esteem of the probability
attached to an event. If this is true, then uncertainty is reduced to risk. Moreover, according to
Savage this individual estimate of probability follows the mathematical rules of probability.
Consider the following lottery: “you will receive $100 if tomorrow at noon the temperature in
Rome is higher than 30 degree Celsius and $0 otherwise”. The probability that we assign to the
event “at noon the temperature in Rome is higher than 30 degree Celsius” and the probability
that we assign to the opposite event “at noon the temperature in Rome is 30 degree Celsius or
less” follows the rules of probability, that is, they sum to one. In addition, our probability
estimate should be inferred from our choice between lotteries.
Ellsberg (1961) pointed out that this is not true and hence uncertainty cannot be reduced to
risk. Consider the following example given by Ellsberg. You have in front of you two urns: Urn
I and Urn II. Both urns are opaque so you cannot see inside. You are told that in Urns I there are
100 balls: 50 are black and 50 are red. Instead, Urn II contains 100 balls that can be either black
or red but you do not know in which proportion.
You are asked to bet on a colour and choose the urn from which to draw the ball
simultaneously. If the ball that you have drawn is of the same colour you have chosen, then you
win $100, otherwise you will get nothing.
According to Ellsberg, most of the people will decide to draw the ball from Urn I, which
contains 50 red and 50 black balls, independently from which colour they have chosen to bet on.
In this case in fact the probabilities are known to be 50/100 for Red and 50/100 for Black.

5
 The proportion of the red and black balls in the second urn is not known and so we do not
know which probability to assign to red and black. Avoiding to choose to draw a ball from Urn
II is called by Ellsberg uncertainty aversion.
According to the theory of Savage we should be indifferent between the two urns since the
probability of getting red can be represented by a second order probability distribution (a
probability distribution over probability), whose expected probability is 50/100 as in the case of
Urn I. However, people prefer to bet on the first urn showing that they prefer to draw a ball
from an urn in which the probabilities of the two colours balls are precisely defined.
Aversion to uncertainty creates a problem to Savage theory also for another reason. Let’s
consider again the two urns. If you always choose to draw a ball from Urn I, whatever colour
you prefer, this implies that you consider the probability of getting red plus the probability of
getting black in Urn I different from the probability of getting red plus the probability of getting
black in Urn II. If this is the case, then, the sum of your probability estimates of the two colour
balls for the two urns is going to be different from 1. In particular, ambiguity aversion implies
that your estimate of the probability of red plus your estimate of the probability of black in Urn
II is less than 1. Your probability estimates thus are not following the mathematical rules of
Savage theory. Intuitively, the difference of the sum of the two probabilities from one is the
room that we leave to the existence of uncertainty, that is to say, the possibility of occurrence of
some unexpected situation. As a consequence it is very difficult to deduce our probability
estimate from our choice as pointed out by Ellsberg. Ellsberg’s work has been confirmed by a
substantial number of empirical and theoretical research contributions (Camerer and Weber
1992; Trautman et al. 2008; Machina and Siniscalchi 2014; Gilboa and Marinacci 2016).

Cross-References
Certainty Equivalent
Radical Uncertainty
Risk Management, Optimal

References
Allais, Maurice (1953), “Fondements d’une théorie positive des choix comportant un risqué
et critique des postulates et axioms de l’école Américaine”, Econometrica, 21(4), pp.503-546.
Allais, Maurice and Hagen, Ole (eds) (1979), Expected Utility Hypothesis and the Allais
Paradox, Dordrecht, D. Reidel.
Arrow, Kenneth (1964), “The Role of Securities in the Optimal Allocation of Risk-Bearing”,
Review of Economic Studies, 31, pp. 91–96.
Bernoulli, Daniel (1954), “Specimen theoriae novae de mensura sortis”. Commentarii
Academiae Scientiarum Imperialis Petropolitanae, 1738. Trans. as “Exposition of a new theory
on the measurement of risk”, Econometrica, 22, pp. 23-26.
Camerer, Colin E. and Weber, Martin (1992), “Recent Developments in Modeling
Preferences: Uncertainty and Ambiguity”, Journal of Risk and Uncertainty, 5 (4). pp. 325-370.
de Finetti, Bruno (1937), “La prévision: ses lois logiques, ses sources subjectives”, Annales
de l'Institut Henri Poincaré, 7(1), pp.1-68.
Ellsberg Daniel (1961), “Risk Ambiguity and the Savage Axioms”, Quarterly Journal of
Economics, 75, pp.643-69.
Gilboa, Itzhak and Marinacci, Massimo (2016), “Ambiguity and the Bayesian Paradigm”, in
Arló-Costa, Horacio, Hendricks, Vincent F. and van Benthem, Johan (eds), Readings in Formal
Epistemology, new York, Springer.
Kahneman, Daniel and Tversky, Amos (1979), “Prospect Theory: an Analysis of Decision
under Risk”, Econometrica, 47, pp.273-291.
Keynes John, M.(1921), A Treatise on Probability. London, McMillan.
Knight, Frank (1921), Risk, Uncertainty and Profit, Boston, Houghton-Mifflin.
6
 Loomes, Graham and Sugden, Robert (1982), “Regret Theory: An Alternative Theory of
Rational Choice under Uncertainty”, Economic Journal, 92, pp. 805-824.
MacCrimmon, Kenneth R. and Larsson, Stig (1979), “Utility Theory: Axioms versus
‘Paradoxes’, in Allais, Maurice and Hagen, Ole (eds), Expected Utility Hypothesis and the
Allais Paradox, Dordrecht, D. Reidel.
Machina, Mark (1987), “Choice under Uncertainty: Problems Solved and Unsolved”,
Journal of Economic Perspectives, 1(1), pp. 121-154.
Machina, Mark and Siniscalchi, Marciano (2014), “Ambiguity and Ambiguity Aversion”, in
Machina, Mark and Viscusi, Kip (eds), The Handbook of the Economics of Risk and
Uncertainty, Oxford, North-Holland.
Marschak, Jakob (1950), “Rational Behavior, Uncertain Prospects, and Measurable Utility”,
Econometrica, 18, pp. 111-141.
von Neumann, John and Morgenstern, Oskar, Theory of Games and Economic Behaviour. 2nd
edition, Princeton, Princeton University Press, 1947.
Pratt, J.W. (1964), “Risk Aversion in the Small and in the Large”, Econometrica, 32, pp.
122-136.
Quiggin, John (1982), “A Theory of Anticipated Utility, Journal of Economic Behaviour and
Organization, 3, pp. 323-343.
Quiggin, John (1993), Generalized Expected Utility Theory, Dordrecht, Kluwer.
Raiffa, Howard (1968), Decision Analysis: Introductory Lectures on Choices under
Uncertainty, Reading MA, Addison-Wesley.
Ramsey, Frank P. (1931), “Truth and Probability” in The Foundations of Mathematics and
other Logical Essays, New York, Harcourt, Brace.
Savage, Leonard (1954), The Foundations of Statistics, New York, Wiley.
Schmeidler, David (1989), “Subjective Probability and Expected Utility without Additivity,
Econometrica, 57(3), pp.571-587.
Starmer, Chris and Sugden, Robert (1989), “Violations of the Independence Axiom in
Common Ratio Problems: An Experimental Test of Some Competing Hypotheses”, Annals of
Operations Research, 19(1), pp. 79-102.
Starmer, Chris (2000), “Developments in Non-expected Utility Theory”, Journal of
Economic Literature, 38, pp. 332-382.
Trautmann, Stefan T., Vieider, Ferdinand M. and Wakker, Peter P. (2008), “Causes of
Ambiguity Aversion: Known versus Unknown Preferences”, Journal of Risk and Uncertainty,
36(3), pp. 225-243.
Tversky, Amos and Kahneman, Daniel (1992), “Advances in Prospect Theory: Cumulative
Representation of Uncertainty”, Journal of Risk and Uncertainty, 5, pp. 297-323.

7
 U

b
U(x2)

c
U(x)=EU(L2)=(½Ux1+½Ux2)
=U(EV)
a
U(x1)

x1 x=(½x1+½x2)=EV x2 Wealth

Figure 1. von Neumann-Morgenstern utility function of a risk neutral individual

b
U(x2)
d
U(x)= U(½x1+½x2)=U(EV)
EU(L2)=(½Ux1+½Ux2)
c
a
U(x1)

x1 x=(½x1+½x2)=EV x2 Wealth

Figure 2. von Neumann-Morgenstern utility function of a risk averse individual

8
 U

b
U(x2)

c
EU(L2)=(½Ux1+½Ux2)

U(x)=U(½x1+½x2)=U(EV)
a d
U(x1)

x1 x=(½x1+½x2)=EV x2 Wealth

Figure 3. von Neumann-Morgenstern utility function of a risk loving individual

b
U(x2)
d
U(x)=U(½x1+½x2)=U(EV)
EU(L2)=(½Ux1+½Ux2)=
U(CE) c
a
U(x1)

x1 CE(L2) x=(½x1+½x2)=EV x2
Wealth

Figure 4. Utility function of a risk averse individual with certainty equivalent and risk
premium measures